Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers Jinliang Deng1 ,
Yiming Niu1 ,
Yibo Pan2 ,
Zhiqi Shao3 ,
Qin Luo4 ,
Yongxin Tong1
1 School of Computer Science and Engineering, Beihang University, Beijing, China 2 School of Artificial Intelligence, Beihang University, Beijing, China 3 School of Economics and Business Administration, Chongqing University, Chongqing, China 4 Department of Cardiology, Fuwai Hospital, Beijing, China
[email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
arXiv:2607.24419v1 [cs.AI] 27 Jul 2026
ABSTRACT Deep models have substantially advanced 12-lead ECG classification, yet their refinement still relies heavily on human experts to inspect failures and iteratively revise classifier designs. Recent LLMbased agents have demonstrated the potential for automated model design, but when guided only by aggregate performance metrics, they lack insight into why individual cases fail and how the classifier should be revised. We present RecursiveECG, an evidence-driven LLM-as-Designer framework in which an LLM serves as an offline model designer that refines ECG classifiers based on concrete failures and objective ECG evidence. To ground failure diagnosis in executable evidence, Criteria-to-Measurement Compilation converts curated ECG criteria into validated deterministic functions that produce reproducible, reference-backed measurements for individual ECGs. Building on these measurements, Evidence-Grounded Failure Review analyzes failed and comparator cases by jointly considering raw waveforms, measurements, and model outputs, enabling the LLM to diagnose classifier limitations and formulate targeted revisions. Candidate revisions are executed and re-evaluated under a fixed problem contract, and only evidence-supported updates are retained. The resulting predictor is frozen after refinement and requires no LLM inference during deployment, while an audit trail links each accepted revision to its supporting evidence. Across PTBXL, Georgia, and CPSC2018, RecursiveECG consistently outperforms strong baselines, achieving an average relative improvement of 10.0%. Extensive ablation and transfer studies further validate the effectiveness of its evidence-grounded refinement process.
CCS CONCEPTS • Computing methodologies → Learning paradigms.
KEYWORDS ECG Analysis, Time-Series Classification, Large Language Models, Autonomous Model Refinement, Evidence Grounding Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY © 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00 https://doi.org/10.1145/nnnnnnn.nnnnnnn
ACM Reference Format: Jinliang Deng, Yiming Niu, Yibo Pan, Zhiqi Shao, Qin Luo, and Yongxin Tong. 2026. Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers. In Proceedings of Make sure to enter the correct conference title from your rights confirmation email (Conference acronym ’XX). ACM, New York, NY, USA, 23 pages. https://doi.org/10.1145/ nnnnnnn.nnnnnnn
1
INTRODUCTION
Electrocardiography remains a core examination for cardiovascular diagnosis because a standard 12-lead recording can characterize cardiac activation and repolarization from multiple electrical perspectives while remaining inexpensive and widely accessible [18]. However, ECG diagnosis is challenging because it requires models to capture subtle local waveform morphologies and their complex interactions across time, waveform components, and leads [23]. The development of deep learning models, together with the construction of large-scale benchmark datasets, has substantially advanced ECG classification [16, 19, 23, 37, 41, 44, 51]. However, further progress still heavily relies on human-led model refinement, which requires expertise in both medicine and deep learning and often involves substantial time and effort. To improve upon existing model designs, developers usually need to inspect failure patterns, diagnose model limitations, formulate revision hypotheses, implement corresponding changes, and evaluate them in a closed loop. This process suggests that advancing ECG classification is not merely a matter of training a stronger classifier, but also a failure-driven model design problem: one must understand where the current classifier fails and how its design should be revised. A natural question is whether large language models can reduce this refinement burden. Many existing LLM-based studies on timeseries analysis, however, use LLMs in a direct LLM-as-Classifier manner, where each time-series sample is fed into an LLM for instance-level prediction or judgment [15]. Although flexible, this paradigm requires per-record LLM inference, incurs high computational cost, and is difficult to reproduce reliably due to hallucination and signal-to-text information loss. Inspired by recent agentic discovery frameworks such as AlphaEvolve [25] and LLM-SR [35], we adopt an LLM-as-Designer paradigm for ECG classification, where the LLM serves as an offline model-design controller rather than an online diagnostic classifier as in the LLM-as-Classifier paradigm. Driven by feedback from iterative evaluation, the agent reasons about the limitations of the current classifier, proposes executable revisions, and validates the effectiveness of these revisions under a fixed protocol. The final
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
1) Failed WPW case
Deng et al.
2) Label is not enough Ground Truth
Model Prediction
WPW
MI
Limited guidance from label
3) Find evidence from data Misleading MI-like Evidence
Pseudo-Q/QS
Know it is wrong Don’t know why Should look into the failed case itself
Pseudo-Q
ST-T Change
Missed WPW/CD Evidence
Slurred δ Wave
Wide QRS
Short PR
Figure 1: Motivation example illustrating how ECG model failures can be analyzed using evidence for refinement. output is a frozen ECG classifier that can be deployed without LLM inference, making deployment-time prediction more reproducible and traceable. The key question, however, is what feedback should drive this design process. In general agentic discovery settings, candidate artifacts are often revised and selected based on aggregate validation metrics, such as accuracy or task-level rewards [13, 21, 25]. While such metric-driven feedback is effective for judging whether a generated solution performs better overall, it is insufficient for ECG model refinement. Specifically, aggregate metrics provide limited guidance on why a particular ECG case fails, which signal characteristics or model limitations caused the error, or which diagnostic criteria or modeling mechanisms should be incorporated into the next revision. This limitation is especially critical for ECG classification, where many errors arise from subtle diagnostic criteria. As shown in Fig. 1, a CNN-based classifier mistakes a Wolff– Parkinson–White (WPW) case under the conduction disturbance (CD) superclass as myocardial infarction (MI). This error is clinically plausible because WPW-related pre-excitation can distort the QRS complex and ST–T segment, producing patterns resembling ECG manifestations of MI. Yet the same ECG contains conduction-related evidence supporting the CD label, including a short PR interval, a slurred delta wave, and a widened QRS complex. The label mismatch exposes the error but not why the classifier favored misleading MIlike cues or how to correct this error. A useful repair signal must ground failures in measurable, reference-backed ECG evidence and translate attribution into concrete classifier revisions. Building on this insight, we introduce RecursiveECG, an evidencedriven LLM-as-Designer framework in which the LLM serves as an offline controller for ECG classifier refinement. RecursiveECG first establishes a problem contract that specifies the label space, data split, evaluation protocol, and leakage constraints and holds them fixed throughout the refinement process. It then applies Criteriato-Measurement Compilation to convert curated ECG criteria into validated deterministic measurement functions with explicit computation rules, which are executed on raw ECG signals to extract structured, clinically interpretable features for failure analysis. Subsequently, it recursively proposes and evaluates candidate classifiers. The evaluation results are used to select informative failure cases for Evidence-Grounded Failure Review, which integrates raw waveforms, reference-backed measurements, model outputs, and
1) LLM-as-Classifier
2) LLM-as-Designer
Update Solution
Metric-driven Feedback
Evidence-driven Feedback
Existing
Our Contribution
Figure 2: Comparison of agentic paradigms for ECG analysis.
comparator cases to identify the underlying classifier weakness and formulate a targeted revision. Each revision is then executed and re-evaluated, and only updates supported by case-level evidence are retained. The process produces a frozen ECG classifier together with an audit trail linking each revision to the failure evidence that motivated it. Contributions. In summary, we make the following contributions: • We formulate ECG model development as an evidence-driven LLM-as-Designer, where an LLM acts as an offline controller that uses failure evidence to recursively refine a deployable and traceable classifier. • We introduce Criteria-to-Measurement Compilation, which converts curated ECG criteria into validated deterministic functions that produce reproducible, reference-backed measurements on individual ECGs. • We introduce Evidence-Grounded Failure Review, which combines failed and comparator cases, waveform measurements, and model outputs to attribute errors and derive targeted, evidence-supported revisions. • Across PTB-XL, Georgia, and CPSC2018, RecursiveECG consistently outperforms the strongest baselines, achieving an average relative macro-F1 improvement of 10.0%, while ablation and transfer studies validate its core mechanisms.
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
2
RELATED WORK
ECG diagnostic classification. Public ECG benchmarks have provided the empirical foundation for modern diagnostic classification. PTB-XL offers large-scale 12-lead ECG waveforms with structured multi-label annotations [41], and subsequent studies established strong baselines for diagnostic prediction [1, 5, 6, 11, 31, 37, 53]. Recent studies improve ECG representations through spatio-temporal masked or contrastive modeling [17, 23], parameter-efficient adaptation of pre-trained ECG models [51], and specialized sequence architectures such as xLSTM-ECG [16]. ECG-language and foundationmodel studies further extend ECG analysis to question answering, report generation, and multimodal pretraining [19, 26, 42, 44], while clinically guided localization improves interpretability. Whereas these studies primarily focus on prediction and representation learning, RecursiveECG focuses on auditable revision of an ECG classifier based on deterministic measurements and case-level failure evidence. Agents for time-series reasoning. Existing LLM-based timeseries methods can be categorized by the role assigned to the language model. First, direct-inference methods convert numerical observations into LLM-compatible inputs and then use an LLM for reasoning [8, 15, 43, 46, 52]. This paradigm offers flexibility at inference time, but its performance depends strongly on the chosen input representation and prompting strategy [27]. Second, agentic orchestration methods use LLMs to generate executable detection rules or iteratively invoke statistical tools [9, 33, 34, 39, 48, 49]. These systems coordinate reasoning with tools, but their adaptability is largely bounded by the capabilities of the available tools and predefined workflows. This limits their ability to adapt when the existing toolkit does not adequately capture the domain criteria or failure modes required by the task. RecursiveECG addresses this ECG model-revision setting by using an agent to review predictor failures with knowledge-grounded tools, while ensuring that the deployed pipeline requires no LLM inference. Agents for scientific discovery. Recent work increasingly uses LLM agents to generate, evaluate, and iteratively revise scientific artifacts [2, 7, 21, 36, 47]. MLAgentBench studies whether agents can inspect experimental results and modify machine-learning code to improve model performance, while Scientific Generative Agent combines LLM-generated scientific structures with differentiable simulation and numerical optimization [13, 20]. LLM-SR applies a related closed-loop paradigm to symbolic regression, using an LLM to propose executable equations and refining them through data-driven evaluation and evolutionary search [35]. AlphaEvolve generalizes this approach to algorithm discovery by combining LLM-based program mutation, automated evaluators, and evolutionary selection [25, 32]. Beyond candidate optimization, POPPER emphasizes hypothesis falsification by translating scientific hypotheses into measurable implications and testing them through sequential experiments [12], whereas DiscoveryWorld evaluates whether agents can complete scientific cycles involving hypothesis formation, experimentation, and explanatory inference [14]. These systems typically guide revision using aggregate objectives and hypothesis-level test results. RecursiveECG instead uses case-level waveform evidence, clinical measurements, comparator cases, and predictor outputs to guide revisions of an ECG analysis pipeline.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
3 PROBLEM FORMULATION 3.1 ECG Classification Given a collection of ECG recordings, the goal of ECG classification is to learn a classifier that maps raw ECG signals to diagnostic labels. Formally, the 𝑖-th ECG recording is represented as 𝑋𝑖 ∈ R𝐶 ×𝑇 , where 𝐶 denotes the number of ECG leads and𝑇 denotes the number of temporal samples. Because multiple diagnostic conditions may coexist in a single recording, its label is represented as a multi-hot vector y𝑖 ∈ Y = {0, 1}𝐿 , where 𝐿 denotes the number of diagnostic labels. An ECG classifier parameterized by 𝜃 is defined as 𝑓𝜃 : X → [0, 1] 𝐿 . Given a recording 𝑋𝑖 , the classifier outputs label-wise probabilities b p𝑖 = 𝑓𝜃 (𝑋𝑖 ), where 𝑝b𝑖,𝑙 indicates the predicted probability that the 𝑙-th diagnostic condition is present. The corresponding binary predictions are obtained using a fixed decision rule: b y𝑖 = Decide (b p𝑖 ; 𝝂) , where 𝝂 denotes the predefined label-wise decision thresholds. The dataset is divided into pairwise disjoint training, validation, and held-out test sets: D = Dtr ∪¤ Dval ∪¤ Dte . The training set is used to optimize model parameters, whereas the validation set provides feedback during classifier development and refinement. The held-out test set remains inaccessible throughout refinement and is used only once for final evaluation.
3.2
ECG Classifier Refinement
Conventional ECG classification focuses on learning model parameters from training data. In contrast, we consider the problem of automatically refining an existing ECG classifier through executable modifications to its design and implementation. Given a trained classifier 𝑓𝜃 , a revision operation 𝑟 ∈ R produces an updated classifier 𝑓𝜃 ′ = 𝑟 (𝑓𝜃 ), where R denotes the space of admissible executable revisions. The goal of classifier refinement is to identify a revision that addresses limitations of the current classifier while improving predictive performance: 𝑟 ∗ = arg max 𝑀 (𝑟 (𝑓𝜃 ), Dval ) , 𝑟 ∈R
where 𝑀 (·) denotes the predefined evaluation metric. While validation performance provides a necessary criterion for comparing candidate revisions, it does not explain why the current classifier fails or what should be modified. Therefore, unlike conventional search that explores predefined model spaces primarily guided by validation metrics, RecursiveECG generates revision hypotheses from case-level failure evidence. Validation performance is used to verify and select evidence-supported revisions, while the held-out test set remains inaccessible throughout refinement and is used exclusively for final evaluation.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY 1) Overview 1. Inputs Evaluation Protocol
2) Criteria-to-Measurement Compilation 1 Criteria Formalization
2
Function Compilation
3 Representative Subsets
Criteria for WPW
2. Diagnostic Classifier Proposal 3. Criteria-to-Measurement Compilation
Deng et al.
PR Interval
� Short PR interval � Delta wave � Wide QRS complex � Secondary STTC � Cross-lead evidence � Pre-excitation pattern � ...
Reference Corpus
Function Validation
Measurement Functions
Delta Wave
Label WPW
QRS Duration Lead Pattern
Expected
Output
PR interval
90 ms
Delta wave
None
QRS duration
130 ms
Lead pattern
2nd STTC
…
Match
Revise if mismatch
4. Evidence-Grounded Failure Review
3) Evidence-Grounded Failure Review 1 Iter.1 Iter.2
Case Selection
5. Diagnostic Classifier Revision 6. Output
Revision Log
2 Evidence Construction
3
Contrastive Diagnosis
Iter.3
Final Classifier
Ground Truth WPW
Model Prediction MI
Extractor
Check
Evidence
PR interval 90ms
Confirm short PR
Delta wave Slurred upstroke
Slurred upstroke
QRS duration 130ms
Wide complex
Correctly Classified Cases
Compare the evidence Misclassified Cases
Failure Patterns and Limitations � The model applies an incorrect threshold when evaluating the PR interval. � The model has limited sensitivity to delta waves. � The model is overly sensitive to QRS duration. � …
Figure 3: Overview of RecursiveECG. It compiles ECG criteria into validated measurement functions, selects influential failure cases, constructs evidence, diagnoses classifier limitations through contrastive review, and performs classifier revision.
4 METHOD 4.1 Overview of RecursiveECG RecursiveECG is an agentic framework for automatically refining existing ECG classifiers. Rather than using an LLM for samplelevel ECG prediction, RecursiveECG adopts an LLM-as-Designer paradigm, in which an offline LLM agent analyzes classifier failures, directly modifies the classifier implementation, and validates candidate revisions under a fixed evaluation protocol. Given an initial ECG classifier, RecursiveECG performs refinement through three stages. First, Criteria-to-Measurement Compilation transforms expert-defined ECG diagnostic criteria into functionally validated executable measurement functions, which provide clinically grounded signal measurements. Second, EvidenceGrounded Failure Review selects informative failures, constructs case-level evidence, and contrastively analyzes them against correctly classified reference cases to identify limitations of the current classifier. Finally, Diagnostic Classifier Revision translates the identified limitations into executable modifications and retains only revisions that improve validation performance. The accepted classifier then enters the next refinement iteration, while the held-out test set remains inaccessible until final evaluation.
4.2
Criteria-to-Measurement Compilation
Aggregate performance metrics provide limited insight into the ECG characteristics associated with prediction failures. To construct clinically grounded evidence, RecursiveECG introduces Criteria-toMeasurement Compilation, which converts expert-defined ECG diagnostic criteria into functionally validated executable measurement functions. The resulting measurement inventory and validation contract are summarized in Tab. 5 and Sec. A.5.
𝑄
Clinical Criterion Formalization. Let C = {𝑐𝑞 }𝑞=1 denote the collection of curated ECG diagnostic criteria. Each criterion is formalized as a structured specification 𝑐 = (𝑑, 𝑚, 𝜏), where 𝑑 denotes the diagnostic concept, 𝑚 denotes the target ECG measurement, and 𝜏 denotes the condition used to interpret the measurement. For example, the criterion “a prolonged PR interval indicates first-degree atrioventricular block” can be formalized as 𝑐 = (first-degree AV block, PR interval, PR > 200 ms) . This formalization specifies which clinically meaningful property should be extracted and how its value should be interpreted, providing a computable specification for subsequent function compilation. Measurement Function Compilation and Validation. Given a formalized criterion 𝑐, the LLM agent generates an executable measurement function under the compilation instruction 𝜌 cmp : 𝑔𝑐(0) = A𝜙 𝑐; 𝜌 cmp . The key compilation prompt excerpt corresponding to 𝜌 cmp is reported in Sec. A.2. 𝑔𝑐 : X → R𝑘𝑐 , where z𝑖,𝑐 = 𝑔𝑐 (𝑋𝑖 ) denotes the structured measurements extracted from ECG recording 𝑋𝑖 , and 𝑘𝑐 is the number of values produced for criterion 𝑐. Although the function is generated by an LLM, its execution is deterministic: the same ECG input produces the same measurements. The compiled function therefore operationalizes the clinical criterion as an explicit and reproducible computational procedure rather than an implicit latent representation. RecursiveECG validates each function on a representative validation subset V𝑐 covering relevant
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
labels and criterion conditions. The validation process examines execution correctness and the reliability of the extracted ECG measurements. At iteration 𝑡, the resulting feedback is represented as 𝜂𝑐(𝑡 ) . When a function fails the predefined tests, the feedback is returned to the agent for correction: 𝑔𝑐(𝑡 +1) = A𝜙 𝑔𝑐(𝑡 ) , 𝑐, 𝜂𝑐(𝑡 ) ; 𝜌 cmp . This process continues until the generated function satisfies the validation requirements. The functionally validated measurements form a reusable measurement library G = {𝑔𝑐 | 𝑐 ∈ C}, which provides the clinically grounded measurements used in subsequent failure review.
4.3
Evidence-Grounded Failure Review
RecursiveECG introduces Evidence-Grounded Failure Review, which transforms failures into actionable evidence through case selection, evidence construction, and batch-level contrastive diagnosis. Selection and retrieval details are provided in Sec. A.4. Failure-driven Case Selection. Although evaluation may reveal diverse errors, not all failure cases provide equally useful feedback for model refinement. The agent should therefore prioritize cases that are most likely to expose limitations of the current classifier. Let the validation failure set of classifier 𝑓𝜃 be − Dval,𝜃 = {(𝑋𝑖 , y𝑖 ) ∈ Dval | b y𝑖 ≠ y𝑖 } . − , the agent conRather than uniformly reviewing all cases in Dval,𝜃 structs a targeted failure batch: − B𝜃− ⊆ Dval,𝜃 .
The selection process prioritizes diagnostic labels and error modes that exert substantial influence on validation performance, thereby defining a focused set of failures for further review. Failure Evidence Construction. For each recording in the selected failure batch, RecursiveECG constructs an evidence representation that connects classifier behavior with clinically interpretable ECG characteristics. For an ECG recording 𝑋𝑖 , the clinically grounded measurements are: 𝑍𝑖 = z𝑖,𝑐 = 𝑔𝑐 (𝑋𝑖 ) 𝑐 ∈ C . The corresponding case-level evidence is defined as: p𝑖 , b y𝑖 , 𝑍𝑖 ) , 𝐸𝑖 = (𝑋𝑖 , y𝑖 , b which contains the raw ECG waveform, ground-truth labels, labelwise prediction probabilities, and criterion-grounded measurements. For a batch B, its evidence representation is: 𝐸 (B) = {𝐸𝑖 | (𝑋𝑖 , y𝑖 ) ∈ B} . In particular, 𝐸 (B𝜃− ) provides the signal-level and decision-level context required to diagnose recurring failures in the selected batch. Contrastive Failure Diagnosis. A characteristic observed in the failure batch is not necessarily specific to prediction errors, as it may also occur in correctly classified cases. To identify systematic limitations of the current classifier, RecursiveECG contrasts the selected failure batch with correctly classified reference cases exhibiting comparable clinical evidence profiles.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
+ Let Dval,𝜃 = (𝑋 𝑗 , y 𝑗 ) ∈ Dval b y 𝑗 = y 𝑗 denote the set of validation samples correctly classified by the current classifier. Given the selected failure batch B𝜃− , a reference retrieval procedure constructs + B𝜃+ = R B𝜃− , Dval,𝜃 , where R (·) retrieves correctly classified cases whose clinical evidence profiles are comparable to those of the failure batch. Given the evidence representations of the failure and reference batches, the agent performs batch-level contrastive diagnosis: ℓ = A𝜙 𝐸 (B𝜃− ), 𝐸 (B𝜃+ ); 𝜌 diag , where ℓ denotes the identified failure pattern together with the corresponding limitation of the current classifier. The key failurereview prompt excerpt corresponding to 𝜌 diag is reported in Sec. A.2. By comparing failed and successful cases, the agent identifies signal characteristics or diagnostic patterns that are insufficiently captured by the current classifier. The diagnosed limitation ℓ is then passed to the subsequent classifier revision stage.
4.4
Diagnostic Classifier Revision
Given the diagnosed limitation ℓ, RecursiveECG uses the LLM agent to revise the ECG classifier implementation and evaluate the resulting candidate model under a fixed protocol. Code-level Revision Generation. Let 𝑃𝜃 denote the executable implementation of the current classifier 𝑓𝜃 , including its model architecture and associated training procedure. Conditioned on the diagnosed limitation ℓ and the revision instruction 𝜌 rev , the LLM agent generates a code-level revision: 𝑟 = A𝜙 (𝑃𝜃 , ℓ; 𝜌 rev ) . Here, 𝑟 denotes an executable modification to the current implementation rather than a natural-language recommendation. The key prompt excerpt corresponding to 𝜌 rev is reported in Sec. A.2. Applying the generated revision produces an updated program: 𝑃𝜃 ′ = Apply (𝑃𝜃 , 𝑟 ) . The revised program is then instantiated and trained on the fixed training set to obtain the candidate predictor 𝑓𝜃 ′ . Revision Execution and Selection. LLM-generated code may fail to execute or reduce classifier performance. Each candidate revision is therefore executed and evaluated under the same training and validation protocol as the current classifier. A revision is admissible only when its program executes successfully and the resulting classifier satisfies the predefined constraints. Among admissible revisions, a candidate is retained when it improves validation performance: 𝑀 (𝑓𝜃 ′ , Dval ) > 𝑀 (𝑓𝜃 , Dval ) . An accepted revision replaces the current implementation and initiates the next iteration. Refinement terminates when no admissible improvement is found or when the refinement budget is exhausted. Because each generated revision is conditioned on the evidencederived limitation ℓ, RecursiveECG records the provenance of every accepted update in a refinement history: H ← H ∪ {(ℓ, 𝑟, 𝑀before, 𝑀after )} . This history links each accepted code modification to its motivating model limitation and observed validation outcome.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Deng et al.
Table 1: Classification results on PTB-XL, Georgia, and CPSC2018. Values are reported as mean ± standard deviation over five runs. The best overall results are in bold, and the second-best results are underlined. The final column reports the relative improvement of the best-performing RecursiveECG variant over the strongest non-RecursiveECG baseline. Dataset
Deep Models
Metric
Pretrained Models
PatchTST
TimesNet
TS2Vec
xLSTM-ECG
UniTS
MERL
Agentic Models Argos
Impr (%)
RecursiveECG
RecursiveECG
(Qwen3.5-27B)
(DeepSeek-V4-Pro)
PTB-XL
AUC Macro F1 Micro F1 Hamming Acc.
0.8271±0.0020 0.5899±0.0057 0.6424±0.0026 0.7941±0.0056
0.8710±0.0020 0.6557±0.0051 0.6975±0.0054 0.8279±0.0044
0.8683±0.0015 0.6566±0.0022 0.6903±0.0017 0.8229±0.0017
0.9130±0.0019 0.7250±0.0036 0.7614±0.0062 0.8705±0.0058
0.8465±0.0030 0.6202±0.0033 0.6783±0.0059 0.8160±0.0082
0.8964±0.0007 0.6912±0.0019 0.7394±0.0045 0.8561±0.0096
0.8511±0.0056 0.9166±0.0187 0.9373±0.0066 0.6055±0.1196 0.7338±0.0415 0.7653±0.0088 0.6281±0.0071 0.7821±0.0326 0.7930±0.0039 0.7498±0.0047 0.9118±0.0149 0.8929±0.0177
+2.66% +5.56% +4.15% +4.74%
Georgia
AUC Macro F1 Micro F1 Hamming Acc.
0.8459±0.0054 0.5574±0.0087 0.5720±0.0061 0.8544±0.0078
0.8522±0.0047 0.5579±0.0091 0.6325±0.0073 0.8750±0.0060
0.9163±0.0026 0.6216±0.0064 0.6961±0.0041 0.8979±0.0032
0.9233±0.0034 0.6827±0.0068 0.7469±0.0057 0.9178±0.0039
0.7149±0.0029 0.5511±0.0075 0.7567±0.0052 0.9199±0.0048
0.9693±0.0018 0.8130±0.0046 0.8426±0.0038 0.9486±0.0027
0.8849±0.0086 0.5911±0.0064 0.6483±0.0085 0.8699±0.0094
0.9698±0.0124 0.8021±0.0368 0.8662±0.0297 0.9587±0.0116
0.9814±0.0052 0.8826±0.0069 0.9170±0.0043 0.9735±0.0021
+1.25% +8.56% +8.83% +2.62%
CPSC2018
AUC Macro F1 Micro F1 Hamming Acc.
0.7840±0.0049 0.4375±0.0102 0.4953±0.0066 0.8835±0.0042
0.7285±0.0068 0.4089±0.0115 0.4683±0.0094 0.8802±0.0057
0.8390±0.0031 0.4935±0.0076 0.5988±0.0059 0.9051±0.0028
0.9016±0.0042 0.5775±0.0096 0.6030±0.0074 0.8691±0.0069
0.8668±0.0037 0.5470±0.0089 0.5915±0.0071 0.8848±0.0065
0.8047±0.0029 0.6931±0.0063 0.7459±0.0051 0.9363±0.0034
0.8708±0.0067 0.5332±0.0065 0.6007±0.0086 0.9054±0.0068
0.9437±0.0169 0.6988±0.0442 0.7204±0.0395 0.9199±0.0173
0.9746±0.0036 0.8027±0.0078 0.8352±0.0056 0.9613±0.0030
+8.10% +15.81% +11.97% +2.67%
RecursiveECG
Macro F1
0.75
Bubble diameter = parameters (M)
xLSTM-ECG
MERL
0.70
PTB-XL
TimesNet
0.65
0.9223
UniTS PatchTST Computational cost (GFLOPs per sample) 100
101
0.90
0.9652 0.9663
0.92
0.94
Macro AUC
0.96
0.98
Figure 5: Mechanism ablation study across three datasets.
EXPERIMENTS
In the following section, we conduct a series of experiments to comprehensively evaluate RecursiveECG and demonstrate its effectiveness. We have made our implementation publicly available at https://github.com/neumyor/RecursiveECG.
5.1
0.9746 0.9728 0.9735
CPSC2018
Figure 4: The effectiveness–efficiency tradeoff on PTB-XL.
5
0.9814 0.9788 0.9765 0.9808
Georgia
TS2Vec
10 1
0.9235 0.9273 0.9060
Better performance with lower cost
0.60
RecursiveECG w/o EGFR (metric-driven) w/o CMC w/o Measurement Exec. Direct LLM Design
0.9373 0.9149
Experimental Setup
Datasets. We use three public 12-lead ECG multi-label diagnostic datasets. PTB-XL [41] provides a large-scale diagnostic benchmark with official splits and hierarchical superclass/subclass annotations. Georgia, drawn from the PhysioNet/Computing in Cardiology Challenge 2020 corpus [28], evaluates the method on a scored multi-label setting used by prior ECG baselines [16]. CPSC2018 provides an additional multi-label arrhythmia benchmark, for which we use a fixed train/validation/test split. Detailed label definitions and dataset-specific preprocessing choices are provided in Sec. A.1. Baselines and Protocols. We compare RecursiveECG with both general time-series baselines and ECG-specific baselines, covering diverse paradigms such as deep learning models, pretrained models, and agentic solutions. Following the order used in the results tables, the deep learning baselines are PatchTST [24], TimesNet [45], TS2Vec [50], and xLSTM-ECG [16]; the pretrained or agentic baselines are UniTS [4], MERL [19], and Argos [9]. For a fair comparison, all methods use the same data splits and label mapping, with key hyperparameters independently tuned on the validation set of each
dataset and model selection performed based on validation performance. The default implementations of RecursiveECG and Argos use DeepSeek-V4-Pro [3] as the LLM agent backbone. For these agentic methods, each run re-executes the entire model design process, allowing different design trajectories, rather than fixing a single architecture and only varying training randomness. Results are reported as mean ± standard deviation over five runs; implementation details are provided in Sec. A.1 and Tab. 4. Evaluation Metrics. Since all three tasks are imbalanced multilabel classification problems, we use Macro F1, Macro AUC, Micro F1, and Hamming accuracy. For deployment efficiency, we evaluate FLOPs, GPU memory usage, and per-sample inference latency.
5.2
Main Results
We evaluate the classifiers discovered by RecursiveECG from two perspectives: predictive effectiveness and computational efficiency. Overall Performance Improvement. Table 1 summarizes the performance comparison on PTB-XL, Georgia, and CPSC2018 under two backbone LLMs. RecursiveECG consistently achieves the best performance across all datasets and evaluation metrics compared with other non-RecursiveECG baselines, outperforming representative deep models, pretrained models, and agentic approaches. For Macro F1, RecursiveECG improves by 5.56%, 8.56%, and 15.81% over
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers (Initial) Iteration 1: Unified 12-Lead ResNet
Global Average Pooling
Res Block
Res Block
Limb Conv1d Leads
Res Block
Global Average Pooling
Chest Leads
…
Linear
Conv1d
Global Average Pooling
12Lead ECG
Linear
(Final) Iteration 5: Lead-wise Dual-Branch Attention
…
Chest Leads
Cross-Lead Attn
Temporal Attn
Extract CLS Token
12Lead Limb ECG Leads
Cross-Lead Attn
Temporal Attn
Extract CLS Token
concat
Res Block
Conv1d
(Intermediate) Iteration 3: Lead-wise Dual-Branch CNN
concat
12Lead ECG
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Linear
Figure 6: Model design evolution on PTB-XL. Absolute F1 gain
Gap +0.0458
Absolute AUC gain
0.8368
Georgia +9.1%
AF
+14.6%
SB
+1.2%
AF
94.8% retained
+1.4%
SB
0.8826
Gap +0.1016 0.7011
CPSC2018
0.8027
87.3% retained
+24.1%
STE
+7.6%
STE
0.0 +32.6%
PAC
+46.7%
ISCA
0
10
20
30
40
50
0.6
0.8
1.0
PTB-XL-evolved architecture
Additional gain from target-specific evolution
+7.6%
NST_
60
0.4
Macro F1
+5.3%
ISCA
+55.1%
NST_
0.2
+20.2%
PAC
0
5
10
15
20
Figure 8: Dataset-adaptive architecture evolution.
25
Absolute improvement over the strongest baseline (percentage points) PTB-XL
CPSC2018
Georgia
the capability of LLMs for initial model design. However, its gap with RecursiveECG shows that LLM knowledge alone is insufficient without evidence-grounded diagnosis and iterative refinement. Removing EGFR causes notable degradation on PTB-XL, indicating that metric-driven refinement cannot provide sufficient guidance for identifying model limitations and generating targeted revisions. Disabling measurement execution causes a measurable the strongest baselines on PTB-XL, Georgia, and CPSC2018, respecbut relatively small drop on CPSC2018, while its effect varies across tively. Moreover, RecursiveECG exhibits well-controlled variance datasets. CMC further provides complementary benefits by translatacross independent runs, demonstrating that it can reliably evolve ing clinical criteria into structured measurements. Their effects are high-performing architectures. Using Qwen3.5-27B [29] as an alcomplementary rather than interchangeable within the iterative ternative LLM backbone also achieves competitive performance, evolution loop. indicating that the proposed evidence-grounded refinement paraThese results show that effective ECG classifier evolution dedigm remains effective across different backbone LLMs. pends not only on LLM-based design or metric optimization, but Effectiveness–Efficiency Trade-off. We evaluate whether the classifiers discovered by RecursiveECG achieve a favorable effectiveness– on a structured, evidence-grounded process of failure diagnosis, executable validation, and iterative refinement across datasets. efficiency trade-off. Figure 4 compares representative classifiers in Figure 7: Representative gains on fine-grained labels. Values indicate the absolute gains of RecursiveECG over the strongest baseline for Macro F1 and AUC in percentage.
terms of predictive performance and computational cost. RecursiveECG achieves superior predictive effectiveness with 10.6% fewer parameters than xLSTM-ECG and 90.5% fewer FLOPs than UniTS, while reducing peak memory by 65.8% relative to MERL. These advantages indicate that its performance gains do not arise from increased model complexity.
5.3
Ablation Study
To determine whether the gains of RecursiveECG arise from LLMbased model design or from its refinement mechanisms, we conduct ablation studies to assess the roles of all components in RecursiveECG. We evaluate variants without Evidence-Grounded Failure Review (w/o EGFR), Criteria-to-Measurement Compilation (w/o CMC), or measurement execution (w/o Measurement Exec.), together with a Direct LLM Design variant that removes evidencegrounded failure review and iterative refinement. Detailed definitions are provided in Sec. C.1. As shown in Fig. 5, the complete framework provides the best performance, although the contribution of each mechanism varies across datasets. Interestingly, Direct LLM Design remains competitive and even outperforms several existing baselines, demonstrating
5.4
Model Design Evolution
To understand whether RecursiveECG discovers meaningful model designs, we analyze the evolution trajectory of the ECG classifier on PTB-XL. As shown in Fig. 6, RecursiveECG progressively evolves a generic multi-lead architecture into an ECG-specific design by introducing lead-aware representation learning and dependency modeling across leads and temporal patterns. Starting from a unified 12-lead ResNet, RecursiveECG first separates chest and limb leads to capture heterogeneous lead information, and then incorporates cross-lead and temporal attention to enhance feature interaction. This evolution trajectory demonstrates that RecursiveECG performs targeted architecture refinement rather than arbitrary model modifications, progressively discovering ECG-specific inductive biases through evidence-driven evolution.
5.5
Fine-Grained Label Behavior
While aggregate metrics demonstrate the effectiveness of RecursiveECG, they do not show how gains are distributed across diagnostic categories. We therefore conduct a fine-grained label-level analysis. Fig. 7 highlights representative categories with notable
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Deng et al.
Evidence-Grounded Failure Review
WPW Example Ground Truth
1
Enhance delta wave detection
WPW
better capture slurred upstrokes and early pre-excitation patterns
Model Prediction
MI
2 reference-backed measurements
Compiled Measurement Functions def extract_delta_wave(signal): # detect and quantify delta wave
Delta wave measurement
def compute_pr_interval(signal): # compute PR interval
PR interval measurement
def compute_qrs_duration(signal): # compute QRS duration from onset to offset
QRS duration measurement
Delta wave
QRS duration
Reduce MI/ST-T confusion MI?
Axis deviation
3 Slurred upstroke
Duration:
130ms
Left axis:
-35°
Penalize false MI predictions when deltawave evidence is present.
Strengthen integration Combine PR interval, QRS duration, axis deviation，and morphology
Figure 9: Evidence-Grounded Failure Review of a WPW pseudo-infarction case. improvements, while complete label-wise and negative-label results are provided in Sec. B.3 and Sec. B.4. The gains are most pronounced in diagnoses with explicit, clinically interpretable ECG patterns. On PTB-XL, RecursiveECG improves substantially on NST_ and ISCA, which depend on local waveform morphology and ST-T characteristics. It also achieves notable gains on PAC in CPSC2018 and on rhythm-based diagnoses such as SB and AF in Georgia. These results suggest that diagnoses with well-defined ECG characteristics are particularly suitable for evidence-grounded model evolution, as their failure modes can be described using clinical measurements and case-level evidence.
5.6
Dataset-Adaptive Architecture Evolution
An important capability of an autonomous model designer is to adapt its design strategy according to the characteristics of different datasets. To investigate this capability, we perform dataset-adaptive architecture evolution experiments. Specifically, we first evolve an ECG classifier on PTB-XL and apply the discovered architecture to other datasets, where model parameters are retrained using the corresponding target dataset. We then further evolve the architecture with RecursiveECG on each target dataset and evaluate the additional improvements obtained from dataset-specific evolution. As shown in Fig. 8, the architecture evolved on PTB-XL already provides a strong architectural prior for other datasets, achieving Macro F1 scores of 0.8368 and 0.7011 on Georgia and CPSC2018, respectively; the complete transfer results are reported in Tab. 10. After target-specific evolution, RecursiveECG further improves Macro F1 by 0.0458 and 0.1016 on Georgia and CPSC2018, reaching 0.8826 and 0.8027. These results demonstrate that RecursiveECG discovers reusable ECG modeling patterns while adapting designs to dataset-specific characteristics, rather than relying on a fixed architecture.
5.7
Case Study
We use a WPW pseudo-infarction case from PTB-XL to show how evidence-grounded review converts a failed prediction into an identified model limitation. The case analysis is shown in Fig. 9, with the full audit trace in Tab. 21. The initial unified 12-lead ResNet jointly encoded all leads but did not explicitly distinguish their spatial and diagnostic roles. In this case, WPW-related pre-excitation produced infarction-like initial
deflections, causing the model to overemphasize local morphology and incorrectly predict myocardial infarction. The review identified a short PR interval, slurred QRS onset, and QRS widening, all consistent with WPW. Moreover, the apparent infarction pattern conflicted with the overall electrical axis and ventricular activation pattern. The failure therefore revealed a model limitation: insufficient use of complementary cross-lead evidence to reconcile local morphology with global activation patterns.
6
CONCLUSION
Metric-driven feedback can evaluate whether a model improves, but provides limited guidance for understanding failures and refining model designs. In this work, we introduce RecursiveECG, an evidence-driven LLM-as-Designer paradigm that transforms prediction failures into actionable evidence and executable revisions for recursive ECG classifier refinement. Experiments on three ECG benchmarks demonstrate that RecursiveECG achieves consistent performance improvements, efficient deployment, and datasetadaptive evolution by discovering ECG-specific inductive biases. Future work will explore extending this paradigm toward a general-purpose medical time-series model refinement agent, capable of leveraging task-specific failures and domain evidence to autonomously improve predictive models across diverse physiological signals and healthcare applications.
7
LIMITATIONS AND ETHICAL CONSIDERATIONS
Our experiments use publicly released, de-identified ECG benchmarks under their respective licenses. RecursiveECG may inherit errors or biases from the underlying LLM and curated clinical criteria. The system has only been evaluated retrospectively and is not intended for autonomous diagnosis or clinical deployment without external clinician oversight.
8
GENERATIVE AI USAGE
DeepSeek-V4-Pro served as the default backbone LLM for the offline RecursiveECG workflow. Qwen3.5-27B was used in the LLMbackbone sensitivity experiments, and ChatGPT was used to assist with language polishing and graphical icon editing in the manuscript. The authors verified all outputs and accept responsibility.
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
REFERENCES [1] Zachi I. Attia, Suraj Kapa, Francisco Lopez-Jimenez, Paul M. McKie, Dorothy J. Ladewig, Gaurav Satam, Patricia A. Pellikka, Maurice Enriquez-Sarano, Peter A. Noseworthy, Thomas M. Munger, Samuel J. Asirvatham, Christopher G. Scott, Rickey E. Carter, and Paul A. Friedman. 2019. Screening for Cardiac Contractile Dysfunction Using an Artificial Intelligence-Enabled Electrocardiogram. Nature Medicine 25, 1 (2019), 70–74. https://doi.org/10.1038/s41591-018-0240-2 [2] Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D. White, and Philippe Schwaller. 2024. Augmenting Large Language Models with Chemistry Tools. Nature Machine Intelligence 6 (2024), 525–535. https://doi.org/10.1038/ s42256-024-00832-8 [3] DeepSeek-AI. 2026. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence. [4] Shanghua Gao, Teddy Koker, Owen Queen, Thomas Hartvigsen, Theodoros Tsiligkaridis, and Marinka Zitnik. 2024. UniTS: A Unified Multi-Task Time Series Model. In Advances in Neural Information Processing Systems, Vol. 37. [5] Tomer Golany, Gal Lavee, Shai Tejman Yarden, and Kira Radinsky. 2020. Improving ECG Classification Using Generative Adversarial Networks. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 13280–13285. https://doi.org/10.1609/aaai.v34i08.7037 [6] Tomer Golany and Kira Radinsky. 2019. PGANs: Personalized Generative Adversarial Networks for ECG Synthesis to Improve Patient-Specific Deep ECG Classification. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33. 557–564. https://doi.org/10.1609/aaai.v33i01.3301557 [7] Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Petar Sirkovic, Artiom Myaskovsky, Grzegorz Glowaty, Felix Weissenberger, Alessio Orlandi, Dan Popovici, et al. 2026. Accelerating scientific discovery with Co-Scientist. Nature 655, 8122 (2026), 487–496. https://doi.org/10.1038/s41586-026-10644-y [8] Nate Gruver, Marc Finzi, Shikai Qiu, and Andrew G Wilson. 2023. Large language models are zero-shot time series forecasters. Advances in neural information processing systems 36 (2023), 19622–19635. [9] Yile Gu, Yifan Xiong, Jonathan Mace, Yuting Jiang, Yigong Hu, Baris Kasikci, and Peng Cheng. 2025. Argos: Agentic Time-Series Anomaly Detection with Autonomous Rule Generation via Large Language Models. arXiv preprint arXiv:2501.14170 (2025). [10] E. William Hancock, Barbara J. Deal, David M. Mirvis, Peter Okin, Paul Kligfield, Leonard S. Gettes, James J. Bailey, Rory Childers, Anton Gorgels, Mark Josephson, Jan A. Kors, Peter Macfarlane, Jay W. Mason, Olle Pahlm, Pentti M. Rautaharju, Borys Surawicz, Gerard van Herpen, Galen S. Wagner, and Hein Wellens. 2009. AHA/ACCF/HRS Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part V: Electrocardiogram Changes Associated With Cardiac Chamber Hypertrophy: A Scientific Statement From the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society: Endorsed by the International Society for Computerized Electrocardiology. Journal of the American College of Cardiology 53, 11 (2009), 992–1002. https://doi.org/10.1016/j.jacc.2008.12.015 [11] Awni Y. Hannun, Pranav Rajpurkar, Masoumeh Haghpanahi, Geoffrey H. Tison, Codie Bourn, Mintu P. Turakhia, and Andrew Y. Ng. 2019. Cardiologist-Level Arrhythmia Detection and Classification in Ambulatory Electrocardiograms Using a Deep Neural Network. Nature Medicine 25, 1 (2019), 65–69. https: //doi.org/10.1038/s41591-018-0268-3 [12] Kexin Huang, Ying Jin, Ryan Li, Michael Y. Li, Emmanuel Candes, and Jure Leskovec. 2025. Automated Hypothesis Validation with Agentic Sequential Falsifications. In Proceedings of the 42nd International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 267). PMLR, 25372– 25437. https://proceedings.mlr.press/v267/huang25n.html [13] Qian Huang, Jian Vora, Percy Liang, and Jure Leskovec. 2024. MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 20271–20309. https: //proceedings.mlr.press/v235/huang24y.html [14] Peter A. Jansen, Marc-Alexandre Côté, Tushar Khot, Erin Bransom, Bhavana Dalvi Mishra, Bodhisattwa Prasad Majumder, Oyvind Tafjord, and Peter Clark. 2024. DiscoveryWorld: A Virtual Environment for Developing and Evaluating Automated Scientific Discovery Agents. In Advances in Neural Information Processing Systems, Vol. 37. Curran Associates, Inc., 10088–10116. https://doi.org/10.52202/ 079017-0324 [15] Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y. Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. 2024. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. In The Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=Unb5CVPtae [16] Lei Kang, Xuanshuo Fu, Javier Vazquez-Corral, Ernest Valveny, and Dimosthenis Karatzas. 2025. xLSTM-ECG: Multi-label ECG Classification via Feature Fusion with xLSTM. arXiv preprint arXiv:2504.16101 (2025). [17] Dani Kiyasseh, Tingting Zhu, and David A. Clifton. 2021. CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients. In Proceedings of
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 139). PMLR, 5606–5615. https://proceedings.mlr.press/ v139/kiyasseh21a.html [18] Paul Kligfield, Leonard S. Gettes, James J. Bailey, Rory Childers, Barbara J. Deal, E. William Hancock, Gerard van Herpen, Jan A. Kors, Peter Macfarlane, David M. Mirvis, Olle Pahlm, Pentti Rautaharju, Galen S. Wagner, Mark Josephson, Jay W. Mason, Peter Okin, Borys Surawicz, Hein Wellens, American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology, American College of Cardiology Foundation, and Heart Rhythm Society. 2007. Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part I: The Electrocardiogram and Its Technology: A Scientific Statement from the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society Endorsed by the International Society for Computerized Electrocardiology. Journal of the American College of Cardiology 49, 10 (March 2007), 1109–1127. https://doi.org/10.1016/j.jacc.2007.01.024 [19] Che Liu, Zhongwei Wan, Cheng Ouyang, Anand Shah, Wenjia Bai, and Rossella Arcucci. 2024. Zero-Shot ECG Classification with Multimodal Learning and Testtime Clinical Knowledge Enhancement. In International Conference on Machine Learning. PMLR, 31949–31963. [20] Pingchuan Ma, Tsun-Hsuan Wang, Minghao Guo, Zhiqing Sun, Joshua B. Tenenbaum, Daniela Rus, Chuang Gan, and Wojciech Matusik. 2024. LLM and Simulation as Bilevel Optimizers: A New Paradigm to Advance Physical Scientific Discovery. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 33940–33962. https://proceedings.mlr.press/v235/ma24m.html [21] Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2024. Eureka: Human-Level Reward Design via Coding Large Language Models. In The Twelfth International Conference on Learning Representations. https://openreview.net/ forum?id=IEduRUO55F [22] Jay W. Mason, E. William Hancock, Leonard S. Gettes, James J. Bailey, Rory Childers, Barbara J. Deal, Mark Josephson, Paul Kligfield, Jan A. Kors, Peter Macfarlane, Olle Pahlm, David M. Mirvis, Peter Okin, Pentti M. Rautaharju, Borys Surawicz, Gerard van Herpen, Galen S. Wagner, and Hein Wellens. 2007. Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part II: Electrocardiography Diagnostic Statement List: A Scientific Statement From the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society: Endorsed by the International Society for Computerized Electrocardiology. Journal of the American College of Cardiology 49, 10 (2007), 1128–1135. https://doi.org/10.1016/j.jacc.2007.01.025 [23] Yeongyeon Na, Minje Park, Yunwon Tae, and Sunghoon Joo. 2024. Guiding Masked Representation Learning to Capture Spatio-Temporal Relationship of Electrocardiogram. In The Twelfth International Conference on Learning Representations. [24] Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In The Eleventh International Conference on Learning Representations. [25] Alexander Novikov, Ngân Vũ, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. 2025. AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery. arXiv preprint arXiv:2506.13131 (2025). https://doi.org/10.48550/arXiv.2506.13131 [26] Jungwoo Oh, Gyubok Lee, Seongsu Bae, Joon-myoung Kwon, and Edward Choi. 2023. ECG-QA: A Comprehensive Question Answering Dataset Combined With Electrocardiogram. In Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, Vol. 36. 66277–66288. [27] Junwoo Park, Kyudan Jung, Dohyun Lee, Hyuck Lee, Daehoon Gwak, ChaeHun Park, Jaegul Choo, and Jaewoong Cho. 2025. Delving into Large Language Models for Effective Time-Series Anomaly Detection. In Advances in Neural Information Processing Systems, Vol. 38. https://proceedings.neurips.cc/paper_files/paper/ 2025/hash/950a34c7b107111570077078e1b5b8ef-Abstract-Conference.html [28] Erick A. Perez Alday, Annie Gu, Amit J. Shah, Chad Robichaux, An-Kwok Ian Wong, Chengyu Liu, Feifei Liu, Ali Bahrami Rad, Andoni Elola, Salman Seyedi, Qiao Li, Ashish Sharma, Gari D. Clifford, and Matthew A. Reyna. 2020. Classification of 12-lead ecgs: the physionet/computing in cardiology challenge 2020. 41, 12 (2020), 124003. [29] Qwen Team. 2026. Qwen3.5: Towards Native Multimodal Agents. https: //qwen.ai/blog?id=qwen3.5 [30] Pentti M. Rautaharju, Borys Surawicz, Leonard S. Gettes, James J. Bailey, Rory Childers, Barbara J. Deal, Anton Gorgels, E. William Hancock, Mark Josephson, Paul Kligfield, Jan A. Kors, Peter Macfarlane, Jay W. Mason, David M. Mirvis, Peter Okin, Olle Pahlm, Gerard van Herpen, Galen S. Wagner, and Hein Wellens. 2009. AHA/ACCF/HRS Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part IV: The ST Segment, T and U Waves, and the QT Interval: A Scientific Statement From the American Heart Association
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society: Endorsed by the International Society for Computerized Electrocardiology. Journal of the American College of Cardiology 53, 11 (2009), 982–991. https://doi.org/10.1016/j.jacc.2008.12.014 [31] Antônio H. Ribeiro, Manoel Horta Ribeiro, Gabriela M. M. Paixão, Derick M. Oliveira, Paulo R. Gomes, Jéssica A. Canazart, Milton P. S. Ferreira, Carl R. Andersson, Peter W. Macfarlane, Wagner Meira Jr., Thomas B. Schön, and Antonio Luiz P. Ribeiro. 2020. Automatic Diagnosis of the 12-Lead ECG Using a Deep Neural Network. Nature Communications 11 (2020), 1760. https://doi.org/10.1038/s41467020-15432-4 [32] Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. 2024. Mathematical Discoveries from Program Search with Large Language Models. Nature 625 (2024), 468–475. https://doi.org/10.1038/s41586-023-06924-6 [33] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. Advances in neural information processing systems 36 (2023), 68539–68551. [34] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Reinforcement Learning. In Advances in Neural Information Processing Systems, Vol. 36. [35] Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K. Reddy. 2025. LLM-SR: Scientific Equation Discovery via Programming with Large Language Models. In The Thirteenth International Conference on Learning Representations. https://proceedings.iclr.cc/paper_files/paper/2025/ hash/28df8e730c054c5331855fd4d5403ba9-Abstract-Conference.html [36] Henry W. Sprueill, Carl Edwards, Khushbu Agarwal, Mariefel V. Olarte, Udishnu Sanyal, Conrad Johnston, Hongbin Liu, Heng Ji, and Sutanay Choudhury. 2024. CHEMREASONER: Heuristic Search over a Large Language Model’s Knowledge Space Using Quantum-Chemical Feedback. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 46351–46374. [37] Nils Strodthoff, Patrick Wagner, Tobias Schaeffter, and Wojciech Samek. 2021. Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL. IEEE Journal of Biomedical and Health Informatics 25, 5 (2021), 1519–1528. https: //doi.org/10.1109/JBHI.2020.3022989 [38] Borys Surawicz, Rory Childers, Barbara J. Deal, Leonard S. Gettes, James J. Bailey, Anton Gorgels, E. William Hancock, Mark Josephson, Paul Kligfield, Jan A. Kors, Peter Macfarlane, Jay W. Mason, David M. Mirvis, Peter Okin, Olle Pahlm, Pentti M. Rautaharju, Gerard van Herpen, Galen S. Wagner, and Hein Wellens. 2009. AHA/ACCF/HRS Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part III: Intraventricular Conduction Disturbances: A Scientific Statement From the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society: Endorsed by the International Society for Computerized Electrocardiology. Journal of the American College of Cardiology 53, 11 (2009), 976–981. https://doi.org/10.1016/j.jacc.2008.12.013 [39] Xiaoyu Tao, Yuchong Wu, Mingyue Cheng, Ze Guo, and Tian Gao. 2026. AnomaMind: Agentic Time Series Anomaly Detection with Tool-Augmented Reasoning. arXiv preprint arXiv:2602.13807 (2026). https://doi.org/10.48550/arXiv.2602.13807 [40] Galen S. Wagner, Peter Macfarlane, Hein Wellens, Mark Josephson, Anton Gorgels, David M. Mirvis, Olle Pahlm, Borys Surawicz, Paul Kligfield, Rory Childers, Leonard S. Gettes, James J. Bailey, Barbara J. Deal, E. William Hancock, Jan A. Kors, Jay W. Mason, Peter Okin, Pentti M. Rautaharju, and Gerard van Herpen. 2009. AHA/ACCF/HRS Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part VI: Acute Ischemia/Infarction: A
Deng et al.
Scientific Statement From the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society: Endorsed by the International Society for Computerized Electrocardiology. Journal of the American College of Cardiology 53, 11 (2009), 1003–1011. https: //doi.org/10.1016/j.jacc.2008.12.016 [41] Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffter. 2020. PTB-XL, a large publicly available electrocardiography dataset. Scientific data 7, 1 (2020), 154. [42] Zhongwei Wan, Che Liu, Xin Wang, Chaofan Tao, Hui Shen, Jing Xiong, Rossella Arcucci, Huaxiu Yao, and Mi Zhang. 2025. MEIT: Multimodal electrocardiogram instruction tuning on large language models for report generation. In Findings of the association for computational linguistics: ACL 2025. Association for Computational Linguistics, Vienna, Austria, 14510–14527. https: //doi.org/10.18653/v1/2025.findings-acl.749 [43] Chengsen Wang, Qi Qi, Jingyu Wang, Haifeng Sun, Zirui Zhuang, Jinming Wu, Lei Zhang, and Jianxin Liao. 2025. ChatTime: A Unified Multimodal Time Series Foundation Model Bridging Numerical and Textual Data. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 12694–12702. https://doi.org/ 10.1609/aaai.v39i12.33384 [44] Fuying Wang, Jiacheng Xu, and Lequan Yu. 2025. From Token to Rhythm: A Multi-Scale Approach for ECG-Language Pretraining. In International Conference on Machine Learning. PMLR, 65059–65074. [45] Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2023. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. In The Eleventh International Conference on Learning Representations. [46] Xiongxiao Xu, Haoran Wang, Yueqing Liang, Philip S. Yu, Yue Zhao, and Kai Shu. 2026. Can Multimodal LLMs Perform Time Series Anomaly Detection?. In Proceedings of the ACM Web Conference 2026. 5392–5403. https://doi.org/10.1145/ 3774904.3792376 [47] Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. 2024. Large Language Models as Optimizers. In The Twelfth International Conference on Learning Representations. [48] Tiankai Yang, Junjun Liu, Michael Siu, Jiahang Wang, Zhuangzhuang Qian, Chanjuan Song, Cheng Cheng, Xiyang Hu, and Yue Zhao. 2025. AD-AGENT: A Multi-agent Framework for End-to-end Anomaly Detection. In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics. The Asian Federation of Natural Language Processing and The Association for Computational Linguistics, Mumbai, India, 191–205. https://doi.org/10.18653/v1/2025.findings-ijcnlp.11 [49] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In The Eleventh International Conference on Learning Representations. https://openreview.net/forum?id=WE_vluYUL-X [50] Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, Yunhai Tong, and Bixiong Xu. 2022. TS2Vec: Towards Universal Representation of Time Series. Proceedings of the AAAI Conference on Artificial Intelligence 36, 8 (2022), 8980–8987. https://doi.org/10.1609/aaai.v36i8.20881 [51] Rushuang Zhou, Yuanting Zhang, and Yining Dong. 2025. H-Tuning: Toward LowCost and Efficient ECG-based Cardiovascular Disease Detection with Pre-Trained Models. In International Conference on Machine Learning. PMLR, 79548–79569. [52] Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al. 2023. One fits all: Power general time series analysis by pretrained lm. Advances in neural information processing systems 36 (2023), 43322–43355. [53] Hongling Zhu, Cheng Cheng, Hang Yin, Xingyi Li, Ping Zuo, Jia Ding, Fan Lin, Jingyi Wang, Beitong Zhou, Yonge Li, Shouxing Hu, Yulong Xiong, Binran Wang, Guohua Wan, Xiaoyun Yang, and Ye Yuan. 2020. Automatic Multilabel Electrocardiogram Diagnosis of Heart Rhythm or Conduction Abnormalities with Deep Learning: A Cohort Study. The Lancet Digital Health 2, 7 (2020), e348–e357. https://doi.org/10.1016/S2589-7500(20)30107-2
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
A
EXPERIMENTAL PROTOCOL AND EVIDENCE INTERFACES
RecursiveECG denotes the complete workflow with candidate search, Criteria-to-Measurement Compilation, Evidence-Grounded Failure Review, independent subagent execution, regression checks, governed candidate promotion, and a frozen final predictor.
A.1
Evaluation Protocol and Data Splits
All experiments evaluate multi-label diagnosis from 12-lead ECGs. Methods use the same label mapping, lead order, preprocessing, validation-based model selection, and validation-based threshold selection within each dataset. Test sets are used only for final reporting. Unless otherwise stated, RecursiveECG uses DeepSeek-V4-Pro, candidate count 𝑘 = 3, maximum iteration count 5. This subsection expands the compact experimental setup in Sec. 5.1: Tab. 2 gives the evaluation scale, and Tab. 3 specifies the full label sets and preprocessing choices that define the fixed problem contracts. The three datasets cover different label organizations. PTB-XL provides a superclass/subclass diagnostic hierarchy, Georgia uses a seven-label scored setting, and CPSC2018 is treated as a nine-label arrhythmia task under a fixed 60/20/20 split. This makes Macro F1 the primary metric because the label supports are highly imbalanced and several clinically meaningful labels are rare. All reported standard deviation is computed with a denominator of 𝑁 (i.e., ddof=0). Tab. 4 summarizes the default configuration used to instantiate the workflow in Fig. 3. These settings are held fixed for the main comparison unless a later appendix table explicitly studies a sensitivity setting, such as the candidate budget or LLM backbone. Tab. 5 lists the measurement interface produced by one representative CMC run. The table is not a new diagnostic rule set; it is an inventory of the deterministic evidence channels available to EGFR. The WPW example in Fig. 9 and Tab. 21 shows how several of these functions are used in an individual failure review. Tab. 5 is intentionally reported at the function-family level rather than as source code. The validation procedure, artifact requirements, and deterministic execution contract for these functions are detailed next in Sec. A.3 and Sec. A.5.
A.2
Prompt Templates and Control Instructions
Tab. 6 reports the key prompt excerpts used by the agent calls in Sec. 4.2, Sec. 4.3, and Sec. 4.4. We report only the operationally important parts of each prompt, because the full runtime prompt also includes workspace paths, current iteration state, available tools, and task-specific attachments.
A.3
Reference Corpus and Evidence Binding
The reference corpus K is a fixed development-time artifact rather than a source that is searched during deployment-time inference. For each dataset, it contains the task contract, dataset label definitions, and ECG interpretation references relevant to the target label space. For the ECG criteria used to construct the reference-backed measurement functions and support Evidence-Grounded Failure Review, we use standard ECG interpretation material covering ECG acquisition and intervals, diagnostic statement terminology, intraventricular conduction, ST-segment and repolarization patterns,
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
chamber- and axis-related changes, and acute ischemia/infarction morphology [10, 18, 22, 30, 38, 40]. These references support the construction of auditable measurement functions; they are not used to claim clinical guideline validation of the final predictor. Reference selection follows three constraints. First, the corpus is frozen before iterative solving starts, and the test split is never used to select or modify reference criteria. Second, a candidate measurement must be linked to the current problem contract, label set, or a concrete need identified during Evidence-Grounded Failure Review; generic ECG knowledge that cannot affect the task is not compiled into the measurement interface. Third, every compiled measurement function must carry an evidence record containing a reference path, section or page pointer, and a concise description of the operationalized rule. If the reference material or waveform quality does not support a reliable judgment for an input condition, the measurement function remains available but emits an unsupported or indeterminate status rather than a speculative normal/abnormal judgment. This evidence-binding mechanism is designed to prevent knowledge drift across iterations. During candidate generation and Evidence-Grounded Failure Review, the agent may inspect the emitted measurement records and their linked evidence, but it does not repeatedly reinterpret unconstrained clinical text. The same reference-backed computation is therefore reused across candidate testing, failure attribution, case replay, and final audit reporting.
A.4
Failure Selection and Comparator Retrieval
The failure-review stage is executed only on the validation split and is never allowed to inspect test labels. We make the selection and retrieval contract explicit here because aggregate validation loss alone is not a sufficient description of the cases used to formulate a revision. For a record 𝑖 and label 𝑘, let 𝑦𝑖𝑘 be the ground-truth label, 𝑝𝑖𝑘 the predicted probability, and 𝑦b𝑖𝑘 the thresholded decision. The per-label binary cross-entropy contribution is ℓ𝑖𝑘 = −𝑦𝑖𝑘 log(𝑝𝑖𝑘 + 𝜖) − (1 − 𝑦𝑖𝑘 ) log(1 − 𝑝𝑖𝑘 + 𝜖), and the per-record error score is 𝐿
𝑢𝑖 =
1 ∑︁ ℓ𝑖𝑘 . 𝐿 𝑘=1
We quantify the influence of label 𝑘 on the current failure pool by Í 𝑦𝑖𝑘 ≠ 𝑦𝑖𝑘 ]ℓ𝑖𝑘 𝑖 ∈ D 1[b 𝐼𝑘 = Í𝐿 Í val . 𝑦𝑖 𝑗 ≠ 𝑦𝑖 𝑗 ]ℓ𝑖 𝑗 + 𝜖 𝑗=1 𝑖 ∈ Dval 1[b Thus, failure influence is not a subjective label importance score: it is the fraction of error-weighted validation loss attributable to a label. The reported error-mode taxonomy is the Cartesian product of the affected label set and the decision error type: false negative, false positive, or mixed error. We additionally mark a case as persistent when it recurs for the incumbent across consecutive iterations, and as borderline when the decision margin is within the pre-specified review band around the label threshold. The selection procedure has two stages. First, it forms a failure − pool Dval,𝜃 and assigns each case its 𝑢𝑖 , affected-label set, error type, persistence flag, and measurement-quality status. Second, it samples within error-mode strata, prioritizing the largest 𝑢𝑖 cases
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Deng et al.
Table 2: Dataset and evaluation scope. Dataset
Train
Validation
Test
Evaluated labels
PTB-XL Georgia CPSC2018
17,441 4,946 4,126
2,193 581 1,375
2,203 599 1,376
5 diagnostic superclasses; 23 diagnostic subclasses for fine-grained analysis NSR, AF, IAVB, LBBB, RBBB, SB, STach Normal/NSR, AF, IAVB, LBBB, RBBB, PAC, PVC, STD, STE
Table 3: Dataset-specific labels and preprocessing used in the experiments. Dataset
Complete evaluated label set
Split protocol
Preprocessing
PTB-XL
Main task: NORM, MI, STTC, CD, HYP. Fine-grained analysis: NORM, IMI, AMI, LMI, PMI, STTC, NST_, ISC_, ISCA, ISCI, LAFB/LPFB, IRBBB, _AVB, IVCD, CRBBB, CLBBB, WPW, ILBBB, LVH, LAO/LAE, RVH, RAO/RAE, SEHYP.
Official folds: 1–8 for training, 9 Official 100 Hz 12-lead waveforms are used in for validation, and 10 for testing. canonical lead order. Signals are loaded as float32, non-finite values are set to zero, and each lead is normalized with training-set statistics before model fitting and evaluation. Records keep the official 10 s duration.
Georgia
NSR, AF, IAVB, LBBB, RBBB, SB, STach.
Record groups 3–11 for training, Header metadata and SNOMED codes are group 2 for validation, and group mapped to the seven scored labels used by the 1 for testing. baseline protocol. Signals are converted to physical units, reordered to canonical 12-lead order, resampled to 100 Hz, padded or cropped to 10 s, cleaned for missing or non-finite values, and normalized using training-set statistics.
CPSC2018
Normal/NSR, AF, IAVB, LBBB, RBBB, PAC, PVC, STD, STE.
Fixed record-level 60/20/20 train/validation/test split with primary-label stratification.
SNOMED codes are mapped to the nine target labels. Signals are converted to physical units, reordered to canonical 12-lead order, resampled to 250 Hz, padded or cropped to 10 s, cleaned for missing or non-finite values, and normalized with statistics computed from the training split.
Table 4: Default RecursiveECG configuration used for the main experiments. Component
Setting
Offline controller Candidate budget Iteration budget Data access Promotion objective
DeepSeek-V4-Pro for development-time planning, evidence review, and candidate proposal; no LLM calls at test time. 𝑘 = 3 executable candidate revisions per iteration. Maximum of five refinement iterations, unless explicitly changed in a sensitivity study. Train for fitting, validation for model/threshold selection and promotion, and test only for final reporting. Promote only when validation score improves over the incumbent and execution, reproducibility, and regression checks pass. Label-wise validation thresholds; metrics reported as mean ± standard deviation over five runs. Problem contract, measurement interface, selected and comparator cases, candidate logs, promotion decisions, and final outputs.
Thresholds and runs Audit artifacts
and retaining at least one case for each targeted failure mode whenever that mode is present. If a stratum contains fewer cases than its quota, the unused quota is assigned to the next highest influence stratum. Correctly classified cases are sampled as controls and are never mixed into the failure pool. The concrete audit traces use a small, fixed review budget. PTBXL reviews use five bad cases and two correctly classified controls per iteration in the final case-review trace; CPSC2018 uses stratified samples of 12–15 bad cases when the validation failure pool is large; and the Georgia audit uses 15 stratified cases from the 86-case error pool in its available case-review trace. These counts describe review
artifacts, not the number of cases used to train or evaluate the classifier. For comparator retrieval, the candidate pool is + Dval,𝜃 = {(𝑋 𝑗 , y 𝑗 ) : b y 𝑗 = y 𝑗 },
restricted to the same dataset and split and excluding the failure record itself. A comparator is ranked by the following deterministic evidence key, in descending lexicographic order: (1) label-profile agreement, measured by Jaccard overlap of the ground-truth label sets and agreement on the relevant errormode labels;
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Table 5: Measurement-function inventory from a representative RecursiveECG run. Each function emits a numeric value or structured status together with its source evidence pointer and applicability flag. Measurement function
Emitted evidence and review use
Signal quality and lead availability
Missing leads, non-finite values, flat segments, and usable waveform length; used to mark unreliable evidence. Detected beat times, representative beat, and confidence; used by rhythm and interval measurements. Mean heart rate and RR statistics; used for bradycardia, tachycardia, and rhythm plausibility. RR variability and irregularity status; used for AF, premature-beat patterns, and comparator selection. PR duration and short/long PR status; used for IAVB and pre-excitation evidence. QRS width and wide-QRS status; used for bundle-branch block and ventricular-conduction evidence. QT/QTc estimates with unsupported flags when landmarks are unreliable; used for repolarization review. Limb-lead polarity and axis-quadrant summary; used for conduction and hypertrophy evidence. Lead-wise ST elevation/depression summaries; used for STTC, STD, STE, ischemia, and infarction-like failures. T-wave polarity and repolarization cues; used for ST–T subclasses and comparator evidence. Initial deflection and Q-wave-like morphology by lead group; used for MI, pseudo-infarction, and WPW review. Precordial R/S progression and transition pattern; used for MI, conduction, and hypertrophy-related failures. V1–V2 and lateral-lead terminal-force patterns; used for LBBB/RBBB failure review. PR prolongation and conduction-delay status; used for IAVB and AV-block-related errors. Short PR, slurred QRS onset, widened QRS, and axis consistency; used for WPW and pseudo-infarction cases. Limb/precordial voltage summaries and cross-lead high-voltage pattern; used for LVH/RVH and HYP errors. Early-beat and compensatory-pause cues when detectable; used for PAC, PVC, and rhythm-label failures. Agreement or conflict among local morphology, rhythm, and axis evidence; used by EGFR for comparator construction and revision hypotheses.
R-peak and beat localization Heart-rate summary Rhythm regularity PR interval QRS duration QT and corrected QT Frontal QRS axis ST-segment deviation T-wave morphology Pathological Q-wave screen R-wave progression Bundle-branch morphology AV-conduction screen Pre-excitation screen Voltage hypertrophy screen Premature-beat evidence Cross-lead consistency summary
(2) measurement-profile agreement, using exact agreement for categorical statuses and robustly normalized distance for numeric measurements that are available for both cases; (3) waveform-profile agreement, using lead-wise normalized morphology summaries derived from the same 12-lead input, including rhythm, QRS, ST–T, axis, and voltage summaries; and (4) deterministic record-ID order as a tie-breaker. Only correctly classified cases are eligible. We retain at most two comparators per failure stratum to prevent a large class from dominating the diagnosis. A missing measurement is not treated as a match: it is excluded from the numeric distance and retained as an explicit indeterminate status. If no eligible comparator satisfies the label and evidence filters, the case is retained with the status no-comparator; EGFR may use its raw waveform and measurement records, but it must not infer a contrastive mechanism from an absent reference case. This fallback is reported as an evidence limitation rather than silently replacing retrieval with unconstrained LLM judgment. The resulting provenance is summarized in Tab. 7. It makes the split, incumbent, failure-pool size, reviewed-case budget, and stratification rule explicit for each available case-review trace, so that the subsequent evidence-grounded analysis can be interpreted as a development-time audit rather than as an additional evaluation protocol. The evaluation-record counts follow the official dataset split sizes reported in the experimental setup. The failure-pool and reviewedcase counts are the corresponding case-review trace statistics after
the trace-specific filtering and sampling procedure. The Georgia row is a validation-set case-review trace: it analyzes the 581-record validation split used by the development protocol. It therefore belongs to the development-time refinement evidence and is distinct from the held-out test split, which remains reserved for final reporting. The row is included to document the provenance and limits of the validation case-review artifact.
A.5
Measurement Function Construction and Validation
The Criteria-to-Measurement Compilation stage builds the deterministic measurement interface as a set of validated executable artifacts under a fixed input-output contract. The implementation requires an evidence map, measurement plan, Python measurement extractor, manifest, reference-rule file, README, test cases, and evaluation report before the interface can be used downstream. The backend validator then checks that these artifacts are complete, mutually consistent, deterministic, executable, and grounded in the cited reference evidence. The validator checks are listed in Tab. 8. They define the minimum artifact, evidence-binding, determinism, execution, and control-case requirements that must be satisfied before emitted measurements are used by Evidence-Grounded Failure Review. The resulting dataset-level status counts are reported separately in Tab. 9.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Deng et al.
Table 6: Key prompt excerpts used by the main RecursiveECG agent calls. Symbol / prompt
Role
Key prompt excerpt
p_cmp
Criteria-to-Measurement Compilation
Generate a deterministic and auditable reference feature extractor from the current task definition and fixed ECG references. Read the problem contract, data specification, and listed references first. Do not add thresholds, diagnostic rules, or feature definitions from model common sense. Every feature, computation method, threshold, and judgment must be traceable to reference text; unsupported or unobservable conditions must return indeterminate. Write an evidence map and measurement plan before implementing the extractor, then validate the Python API and CLI on at least one real workspace ECG sample.
p_diag
Evidence-Grounded Failure Review
Use validation feedback, selected failed cases, correctly classified comparator cases, model outputs, raw ECG waveforms, and reference-backed measurements to diagnose why the current predictor fails. Compare failed cases with clinically similar successful cases. Base each important judgment on data observations, tool outputs, or reference-backed evidence. If the bad case cannot be explained by numerical or clinical evidence, explicitly state that the evidence is insufficient rather than forcing a domain explanation. Output a concise, evidence-supported classifier limitation for the next revision.
p_rev
Diagnostic Classifier Revision
Given the current predictor implementation and the diagnosed limitation, generate executable code-level revisions rather than natural-language suggestions. Preserve the fixed label space, data split, decision rule, evaluation metric, leakage constraints, and validation protocol. Execute and evaluate each candidate under the same training and validation protocol. Reject candidates that fail to run, violate the contract, leak test information, or degrade validation performance. Retain a revision only when executable validation evidence supports promotion.
Main orchestrator
Node routing and workflow control
Use the backend-provided workspace progress as the source of truth for routing. Enter the recommended node only when the workflow should continue; do not perform substantive node work in the main session. Query the knowledge graph, reference QA, or extractor-source QA only when the corresponding tool is explicitly available. Do not start a second node while another node is active.
Node execution and Reproducible node execution finish control
Complete only the current node’s responsibility and cite background knowledge, data observations, or tool outputs for important judgments. Use only node-native tools and injected Harness tools. Avoid broad process-killing commands. At completion, call the finish-control tool with success status, a short summary, key artifact paths, and, for iterative solving, an explicit continue-or-exit decision.
Table 7: Failure-selection audit traces from the available case-review workspaces. The table reports the split and incumbent used by each trace; it is not a cross-dataset performance comparison. Dataset / trace
Evaluation records
Failure pool
Bad cases reviewed
Controls
PTB-XL / final review trace
2,193 validation
862
5
2
Georgia / validation case-review trace
581 validation
86
15
CPSC2018 / final review trace
1,375 validation
300
15
The validation reports provide three distinct kinds of evidence. First, the extractor is executable and deterministic: each dataset report is generated from fixed JSON cases and the extractor contract forbids network access, randomness, hidden mutable state, and model-dependent computation. Second, the status counts expose uncertainty instead of converting failed landmark detection into a normal or abnormal decision. Third, the control cases reveal
Sampling rule
Stratified by HYP false negatives, persistent failures, borderline cases, and multi-label confusion. 6 Stratified across AF false negatives, NSR/IAVB false positives, multi-label errors, and diverse edge cases; extractor unavailable in this trace. 5 Stratified by error type and severity, with coverage of PAC, RBBB, PVC, STD, and STE failure modes.
dataset- and sampling-specific failure modes. In particular, the PTBXL report attributes three control abnormalities to 100 Hz timeresolution limitations, the Georgia report identifies P-wave and axis weaknesses, and the CPSC2018 report records indeterminate QRS/PR measurements together with non-judgmental descriptive features. These results support the claim that the functions are executable, traceable, and useful for model-development evidence; they do not establish expert-level clinical measurement accuracy.
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Table 8: Validation checks for deterministic measurement functions. Check
Requirement
Schema completeness
The manifest must declare the entrypoint, Python API, input and output schemas, and non-empty measurement definitions. Each manifest measurement, evidence-map entry, and reference rule must contain reference evidence with a path and section or page pointer. Measurement names must match across the manifest, measurement plan, evidence map, and reference-rule file. The measurement extractor cannot use network calls, random number generators, system time, external processes, or hidden mutable state. The importable Python API and command-line wrapper must return the same JSON object on the same input. Repeated execution on the same JSON input must produce identical output. At least one test case must be extracted from a real workspace sample following the data specification. The evaluation report must include control or reference cases and record abnormal burden, warnings, unsupported conditions, and indeterminate measurements.
Evidence binding Rule consistency Deterministic source Execution consistency Repeatability Real-sample behavior Control behavior
Table 9: Quantitative validation report for the reference-backed measurement extractors. Counts are aggregated over emitted feature records; descriptive denotes a non-judgmental signal summary and is therefore excluded from the normal/abnormal/indeterminate denominator. Dataset
Cases
Control cases
Normal
Abnormal
Indeterminate
PTB-XL
2
1
16
16
0
Georgia
7
1
27
13
CPSC2018
28
2
144
196
The construction workflow is intentionally conservative. The agent first writes an evidence map and measurement plan before writing executable code. The evidence map binds each proposed measurement to its source reference, while the measurement plan records its unit, computation procedure, judgment rules, control expectation, uncertainty conditions, and expected failure modes. Only after these planning artifacts exist does the agent implement the measurement extractor. The extractor receives all waveform data through explicit JSON input and is not allowed to read raw data files, reference files, network resources, or model state during execution. Validation uses both artifact checks and real-sample execution checks. For each test case, the backend executes the measurement extractor twice and compares the complete JSON outputs. When command-line execution is enabled, it also compares the Python module API against the command-line interface. A real sample is required so that validation is not limited to synthetic smoke tests. The evaluation report summarizes measurement-status counts for real and control cases, including warnings, unsupported states, and indeterminate outputs. If a control case produces an unexpectedly high abnormal burden, the corresponding measurement function must be debugged, weakened to an advisory cue, or marked with an explicit limitation before Evidence-Grounded Failure Review can use its output as case-level evidence. Thus, the measurement interface is validated as a reproducible evidence interface rather than as an independent clinical diagnostic system.
Observed limitation
3/16 abnormal judgments on the NORM control; QTc, T-wave discordance, and pathological-Q detection are sensitive to 100 Hz resolution. 9 The NSR control has 1 abnormal and 3 indeterminate judgments out of 7; P-wave and area-based axis heuristics are unreliable. 24 The report additionally contains 84 descriptive summaries; all judgments are deterministic heuristics rather than clinical diagnoses.
A.6
Reference-Backed Measurement Interface Example
Algorithm 1 gives a compact description of the deterministic reference-backed measurement interface used during EvidenceGrounded Failure Review. The example omits low-level signalprocessing details but shows the auditable output contract: each measurement function emits a measured value, unit, judgment, operational rule, and reference evidence pointer. The interface is deterministic and receives all waveform data as explicit JSON input; it does not call an LLM, train a model, or read external state during either development-time replay or deployment-time inference.
B ADDITIONAL DIAGNOSTIC RESULTS B.1 Aggregate Diagnostic Performance The primary aggregate result is that RecursiveECG improves macro F1 over the strongest baseline on all three datasets. The absolute gain is 0.0403 on PTB-XL, 0.0696 on Georgia, and 0.1096 on CPSC2018. The aggregate results support a focused but important claim. RecursiveECG does not improve only one benchmark or one label organization. It improves the primary metric under the PTBXL diagnostic hierarchy, the Georgia scored-label setting, and the CPSC2018 arrhythmia label set. The largest absolute gain occurs
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Algorithm 1: Reference-backed deterministic measurement emission Input: ECG case 𝑋 with 12-lead waveform and sampling rate; fixed measurement-rule registry R Output: Measurement report 𝐸 (𝑋 ) with measured values, judgments, and evidence pointers 1 Initialize 𝐸 (𝑋 ) ← ∅; 2 foreach registered measurement criterion 𝑟 ∈ R do 3 Select the required leads, fiducial points, and measurement window specified by 𝑟 ; 4 Compute the numeric measurement 𝑣𝑟 using deterministic signal-processing routines; 5 if 𝑣𝑟 is unavailable or the required waveform support is invalid then 6 Set judgment 𝑞𝑟 ← indeterminate; 7 else 8 Compare 𝑣𝑟 with the threshold or predicate encoded in 𝑟 and set 𝑞𝑟 ← normal or abnormal; Append the measurement record {𝑛𝑎𝑚𝑒 : 𝑟 .𝑛𝑎𝑚𝑒, 𝑣𝑎𝑙𝑢𝑒 : 𝑣𝑟 , 𝑢𝑛𝑖𝑡 : 𝑟 .𝑢𝑛𝑖𝑡, 𝑗𝑢𝑑𝑔𝑚𝑒𝑛𝑡 : 𝑞𝑟 , 𝑟𝑢𝑙𝑒 : 𝑟 .𝑡𝑒𝑥𝑡, 𝑒𝑣𝑖𝑑𝑒𝑛𝑐𝑒 : 𝑟 .𝑟𝑒 𝑓 𝑒𝑟𝑒𝑛𝑐𝑒𝑠} to 𝐸 (𝑋 );
9
10
return 𝐸 (𝑋 );
on CPSC2018. However, this dataset also contains the strongest label-level F1 exception, RBBB, which is examined below. Thus, the aggregate improvement should be interpreted together with the fine-grained positive and negative results rather than as uniform per-label dominance.
B.2
Dataset-Adaptive Architecture Evolution
The architecture-transfer experiment asks whether the architecture and governed toolchain structure discovered on PTB-XL can be reused on new ECG datasets. This setting transfers architecture and toolchain design rather than trained weights or zero-shot predictions. The target output layer is replaced and retrained, but the agent workflow is not rerun on the target dataset. The transferred architecture remains competitive, but it does not match the corresponding target-specific rerun under the transfer protocol. This result indicates that RecursiveECG discovers reusable ECG model structures, but that architecture reuse alone does not recover the complete benefit observed under target-specific refinement. In particular, target-dataset Evidence-Grounded Failure Review (EGFR), the execution and use of reference-backed measurements produced through Criteria-to-Measurement Compilation (CMC), and governed toolchain revision expose label-set-specific and distribution-specific failure patterns that the PTB-XL-derived architecture alone cannot repair. The transfer result therefore separates reusable architecture discovery from target-specific evidence-grounded refinement under this protocol.
B.3
Fine-Grained Label Behavior
Fine-grained analysis shows that the aggregate gains are not uniform, and that F1 and AUC provide complementary views of
Deng et al.
model behavior. In terms of F1, RecursiveECG improves many morphology-sensitive and rhythm-sensitive labels, particularly the PTB-XL ST–T and conduction subclasses and the CPSC2018 PAC, STE, STD, NSR, and AF labels. The principal F1 exceptions are PTB-XL RVH and SEHYP and CPSC2018 RBBB. The complete per-label F1 results are reported in Tab. 11, Tab. 12, and Tab. 13; the corresponding AUC results are reported in Tab. 14, Tab. 15, and Tab. 16. These tables provide the numerical breakdown behind the aggregate and exception patterns described above, with deltas computed against the strongest baseline for each label. It is worth noting that, since PTB-XL provides an explicit 23-subclass classification setting, the PTB-XL results here are obtained by replacing the classification head with a 23-class classification head and recomputing the metrics. The AUC results show a broader improvement in ranking quality. RecursiveECG improves over the strongest baseline on 17 of the 23 PTB-XL subclasses, ties on ISC_ and CLBBB, improves on 6 of the 7 Georgia labels, and improves on 8 of the 9 CPSC2018 labels. The remaining AUC regressions are PTB-XL CRBBB, ILBBB, LAFB/LPFB, and SEHYP, together with Georgia RBBB and CPSC2018 RBBB. The label-level evidence clarifies the aggregate result. RecursiveECG is strongest on labels for which failures analyzed through Evidence-Grounded Failure Review can be grounded in reproducible morphology, rhythm, or cross-lead measurements. EGFR does not infer the cause of an error from the label mismatch alone; it jointly considers raw waveforms, reference-backed measurements, model outputs, and comparator cases. The complementary metrics also reveal different failure modes. PTB-XL RVH trails the strongest baseline in F1 but exceeds it in AUC, suggesting that its remaining weakness may be related to the operating point, threshold selection, or calibration rather than a complete loss of discriminative ranking. Conversely, the AUC regressions on PTB-XL CRBBB, ILBBB, LAFB/LPFB, and SEHYP, Georgia RBBB, and CPSC2018 RBBB show that RecursiveECG is not uniformly dominant in ranking quality. Several of these labels have very low support, so their estimates should be interpreted cautiously. We report all labels and both metrics to make these exceptions explicit rather than excluding unfavorable cases post hoc.
B.4
Failure Modes and Negative-Label Analysis
The negative label-level deltas identify where the current workflow still leaves label-specific errors, rather than where the overall method fails. Across the three datasets, the principal F1 exceptions are CPSC2018 RBBB and the low-support PTB-XL hypertrophy subclasses RVH and SEHYP. These labels are harder than the highgain cases because their evidence is either sparsely represented, heterogeneous across cases, or less directly captured by the retained measurement and revision path. CPSC2018 RBBB. RBBB is the clearest negative case in the labellevel comparison. RecursiveECG reaches an F1 of 0.5556, whereas MERL reaches 0.9084 on the same evaluation split. This exception does not overturn the aggregate CPSC2018 gain, where RecursiveECG improves macro F1 from 0.6931 to 0.8027 and improves eight of the nine evaluated labels, but it shows that the learned decision rule is not uniformly reliable across all labels.
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Table 10: Architecture transfer from PTB-XL to target datasets under the transfer-control protocol. Architecture
Target
PTB-XL-derived Target-specific RecursiveECG PTB-XL-derived Target-specific RecursiveECG
Georgia Georgia CPSC2018 CPSC2018
F1 Δ NORM
MI
AUC Δ
Both improve
At least one improves
STTC
Target workflow rerun
Macro F1
Gap
No Yes No Yes
0.8368 0.8826 0.7011 0.8027
0.0458 0.0000 0.1016 0.0000
Negative delta
CD
HYP
F1 improvement
1.0 0.8
Δ F1
0.6 0.4 0.2 0.0 −0.2
Δ AUC
0.150
AUC improvement
0.100 0.050 0.000
HY P SE
RV H
RA E
LV H
RA O/
E
PW
LA LA O/
B PF /L FB LA
W
CD IV
BB B IR
BB B IL
BB B CR
BB B CL
_A VB
ST TC
T_ NS
C_ IS
CI IS
I
CA IS
PM
I LM
I IM
I AM
NO
RM
−0.025
PTB-XL diagnostic subclass
Figure 10: Fine-grained F1 and AUC changes of RecursiveECG relative to the strongest baseline on PTB-XL. Table 11: PTB-XL diagnostic subclass F1 results reported as mean ± standard deviation over five runs. Delta is computed from mean values. Superclass Subclass NORM MI MI MI MI STTC STTC STTC STTC STTC CD CD CD CD CD CD CD CD HYP HYP HYP HYP HYP
NORM AMI IMI LMI PMI ISCA ISCI ISC_ NST_ STTC _AVB CLBBB CRBBB ILBBB IRBBB IVCD LAFB/LPFB WPW LAO/LAE LVH RAO/RAE RVH SEHYP
Support
PatchTST
TimesNet
TS2Vec
xLSTM
UniTS
MERL
Argos
RecursiveECG
Best baseline
Delta
957 307 330 20 2 92 40 127 77 226 83 54 55 7 112 79 181 8 43 213 10 12 3
0.7104±0.0049 0.5907±0.0078 0.5025±0.0036 0.0625±0.0182 0.0000±0.0000 0.1151±0.0125 0.0882±0.0110 0.3049±0.0098 0.0727±0.0118 0.2615±0.0111 0.0796±0.0114 0.7172±0.0059 0.6541±0.0121 0.0000±0.0000 0.5201±0.0125 0.1387±0.0115 0.6511±0.0026 0.1333±0.0233 0.0155±0.0161 0.3461±0.0109 0.0000±0.0000 0.1875±0.0175 0.0000±0.0000
0.7106±0.0076 0.5385±0.0120 0.4214±0.0086 0.0000±0.0000 0.0000±0.0000 0.1505±0.0100 0.1667±0.0100 0.4843±0.0055 0.0449±0.0077 0.3978±0.0056 0.1606±0.0153 0.5854±0.0136 0.7321±0.0058 0.0000±0.0000 0.3486±0.0138 0.1101±0.0145 0.5646±0.0067 0.0000±0.0000 0.0000±0.0000 0.4935±0.0061 0.0000±0.0000 0.2295±0.0177 0.0000±0.0000
0.7592±0.0052 0.7568±0.0063 0.6762±0.0087 0.1395±0.0140 0.0000±0.0000 0.3704±0.0045 0.4190±0.0176 0.6806±0.0117 0.2660±0.0151 0.4794±0.0116 0.4662±0.0145 0.8947±0.0141 0.8525±0.0140 0.1250±0.0169 0.5687±0.0115 0.2420±0.0110 0.7420±0.0031 0.2000±0.0250 0.2029±0.0117 0.6748±0.0034 0.2381±0.0194 0.2500±0.0168 0.5714±0.0000
0.8301±0.0057 0.6182±0.0047 0.6529±0.0116 0.0800±0.0107 0.0000±0.0000 0.2930±0.0115 0.2407±0.0107 0.5579±0.0100 0.1707±0.0156 0.4625±0.0105 0.2364±0.0148 0.8205±0.0119 0.7737±0.0080 0.0000±0.0000 0.5918±0.0044 0.1299±0.0169 0.6530±0.0091 0.0000±0.0000 0.0213±0.0112 0.5806±0.0067 0.0000±0.0000 0.1702±0.0130 0.0000±0.0000
0.8037±0.0068 0.6783±0.0065 0.6078±0.0059 0.1875±0.0097 0.0000±0.0000 0.3065±0.0111 0.3377±0.0102 0.5570±0.0033 0.2929±0.0109 0.4510±0.0037 0.4701±0.0127 0.8430±0.0137 0.8293±0.0103 0.3750±0.0203 0.5950±0.0120 0.1769±0.0064 0.6536±0.0047 0.4000±0.0179 0.1644±0.0120 0.4799±0.0112 0.2105±0.0219 0.1818±0.0076 0.0000±0.0000
0.7931±0.0104 0.6967±0.0052 0.6478±0.0095 0.1026±0.0089 0.0000±0.0000 0.3581±0.0097 0.3433±0.0169 0.6139±0.0074 0.1935±0.0114 0.4325±0.0105 0.1648±0.0108 0.8785±0.0147 0.7597±0.0133 0.2857±0.0229 0.5077±0.0136 0.1808±0.0072 0.7019±0.0038 0.2000±0.0177 0.0959±0.0119 0.6497±0.0064 0.2222±0.0264 0.3226±0.0171 0.0000±0.0000
0.6818±0.0002 0.4286±0.0026 0.4035±0.0027 0.0000±0.0000 0.0000±0.0000 0.1821±0.0097 0.1483±0.0381 0.4316±0.0019 0.0746±0.0192 0.3399±0.0067 0.1397±0.0066 0.6869±0.0052 0.7369±0.0080 0.0000±0.0000 0.3380±0.0087 0.1933±0.0134 0.5206±0.0013 0.0000±0.0000 0.0085±0.0148 0.4904±0.0024 0.1465±0.0063 0.3531±0.0091 0.5000±0.0000
0.8757±0.0048 0.8502±0.0051 0.7606±0.0079 0.2500±0.0208 0.9200±0.0980 0.8370±0.0059 0.7500±0.0164 0.9449±0.0058 0.8442±0.0058 0.7920±0.0065 0.6265±0.0100 0.9630±0.0137 0.9091±0.0112 0.8571±0.0199 0.8036±0.0128 0.5190±0.0081 0.9282±0.0041 0.7500±0.0149 0.2791±0.0076 0.7559±0.0028 0.3000±0.0165 0.2500±0.0178 0.3333±0.0242
xLSTM TS2Vec TS2Vec UniTS None TS2Vec TS2Vec TS2Vec UniTS TS2Vec UniTS TS2Vec TS2Vec UniTS UniTS TS2Vec TS2Vec UniTS TS2Vec TS2Vec TS2Vec Argos TS2Vec
0.0456 0.0934 0.0844 0.0625 0.9200 0.4666 0.3310 0.2643 0.5513 0.3126 0.1564 0.0683 0.0566 0.4821 0.2086 0.2770 0.1862 0.3500 0.0762 0.0811 0.0619 -0.1031 -0.2381
Reliable bundle-branch-block characterization depends on QRS duration, terminal forces, and lead-specific morphology in V1–V2 and lateral leads. Although RecursiveECG improves several arrhythmia and ST-segment labels, the RBBB drop appears in both F1 and AUC. This suggests a label-specific ranking and calibration problem, not only an unfavorable threshold. The ablation results show the same local weakness. For RBBB, each ablated variant obtains higher F1 than the full system, even
though the full system gives the best aggregate CPSC2018 result. This pattern suggests a trade-off between aggregate optimization and some label-wise decision boundaries. RBBB should therefore be treated as a target for more specialized conduction-block measurement functions, stronger label-wise regression checks, and separate failure-selection quotas within EGFR.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Deng et al.
Table 12: Georgia label F1 results reported as mean ± standard deviation over five runs. Delta is computed from mean values. Label
Support
PatchTST
TimesNet
TS2Vec
xLSTM
UniTS
MERL
Argos
RecursiveECG
Best baseline
Delta
NSR AF IAVB LBBB RBBB SB STach
192 70 61 18 41 146 137
0.6069±0.0055 0.2150±0.0164 0.2426±0.0109 0.7701±0.0143 0.7490±0.0123 0.6079±0.0049 0.7105±0.0090
0.6129±0.0083 0.2687±0.0121 0.3751±0.0110 0.5092±0.0124 0.6173±0.0109 0.7354±0.0057 0.7866±0.0096
0.7789±0.0021 0.3864±0.0137 0.3110±0.0098 0.6639±0.0184 0.6607±0.0101 0.7511±0.0079 0.7991±0.0114
0.8413±0.0062 0.5306±0.0092 0.3504±0.0142 0.5947±0.0181 0.7738±0.0113 0.8203±0.0096 0.8676±0.0097
0.6720±0.0036 0.2295±0.0075 0.4047±0.0101 0.6236±0.0139 0.5660±0.0184 0.6467±0.0067 0.7152±0.0097
0.8751±0.0065 0.6854±0.0101 0.7968±0.0151 0.7833±0.0120 0.7981±0.0126 0.8187±0.0090 0.9333±0.0033
0.6662±0.0060 0.2622±0.0114 0.3090±0.0253 0.6784±0.0103 0.7751±0.0097 0.6557±0.0000 0.7912±0.0050
0.9514±0.0108 0.7761±0.0096 0.8387±0.0134 0.8571±0.0153 0.8395±0.0192 0.9667±0.0070 0.9485±0.0053
MERL MERL MERL MERL MERL xLSTM MERL
0.0763 0.0907 0.0419 0.0738 0.0414 0.1464 0.0152
Table 13: CPSC2018 label F1 results reported as mean ± standard deviation over five runs. Delta is computed from mean values. Label
Support
PatchTST
TimesNet
TS2Vec
xLSTM
UniTS
MERL
Argos
RecursiveECG
Best baseline
Delta
NSR AF IAVB LBBB RBBB PAC PVC STD STE
184 237 144 372 45 121 142 43 180
0.3419±0.0112 0.4841±0.0104 0.2308±0.0078 0.7500±0.0071 0.8152±0.0164 0.1951±0.0097 0.5415±0.0104 0.3608±0.0112 0.2182±0.0106
0.4109±0.0021 0.4171±0.0078 0.2115±0.0089 0.7333±0.0044 0.7658±0.0140 0.1860±0.0109 0.4120±0.0035 0.3508±0.0176 0.1923±0.0030
0.5596±0.0080 0.5923±0.0032 0.2624±0.0028 0.7387±0.0068 0.8254±0.0144 0.0448±0.0083 0.6731±0.0072 0.5030±0.0188 0.2424±0.0086
0.5662±0.0082 0.7869±0.0093 0.4075±0.0059 0.7636±0.0075 0.8667±0.0132 0.2633±0.0051 0.7101±0.0059 0.6114±0.0159 0.2222±0.0108
0.4963±0.0032 0.6715±0.0081 0.4304±0.0112 0.8431±0.0047 0.8184±0.0126 0.1709±0.0109 0.6190±0.0101 0.5265±0.0108 0.3469±0.0069
0.6779±0.0028 0.8718±0.0066 0.8291±0.0060 0.8723±0.0100 0.9084±0.0127 0.3333±0.0067 0.5414±0.0090 0.6907±0.0132 0.5128±0.0075
0.5431±0.0084 0.5636±0.0147 0.3232±0.0385 0.8629±0.0050 0.7964±0.0111 0.1635±0.0071 0.6815±0.0099 0.5601±0.0115 0.3042±0.0211
0.9112±0.0046 0.9494±0.0021 0.8529±0.0068 0.9667±0.0106 0.5556±0.0206 0.6593±0.0093 0.8021±0.0066 0.7737±0.0067 0.7534±0.0039
MERL MERL MERL MERL MERL MERL xLSTM MERL MERL
0.2333 0.0776 0.0238 0.0944 -0.3528 0.3260 0.0920 0.0830 0.2406
Table 14: PTB-XL diagnostic subclass AUC results reported as mean ± standard deviation over five runs. Delta is computed from mean values. Superclass
Subclass
NORM MI MI MI MI STTC STTC STTC STTC STTC CD CD CD CD CD CD CD CD HYP HYP HYP HYP HYP
NORM AMI IMI LMI PMI ISCA ISCI ISC_ NST_ STTC _AVB CLBBB CRBBB ILBBB IRBBB IVCD LAFB/LPFB WPW LAO/LAE LVH RAO/RAE RVH SEHYP
Support
PatchTST
TimesNet
TS2Vec
xLSTM
UniTS
MERL
Argos
957 307 330 20 2 92 40 127 77 226 83 54 55 7 112 79 181 8 43 213 10 12 3
0.8632±0.0045 0.8865±0.0043 0.8567±0.0015 0.7278±0.0022 0.6872±0.0037 0.7714±0.0023 0.7484±0.0055 0.8709±0.0010 0.6682±0.0017 0.7335±0.0052 0.6786±0.0051 0.9952±0.0037 0.9952±0.0027 0.8198±0.0021 0.9273±0.0057 0.6496±0.0032 0.9513±0.0026 0.7547±0.0056 0.6961±0.0055 0.7981±0.0054 0.7090±0.0042 0.9264±0.0015 0.8031±0.0048
0.8865±0.0036 0.8630±0.0033 0.8354±0.0052 0.7192±0.0046 0.6596±0.0051 0.8387±0.0029 0.7789±0.0039 0.9196±0.0029 0.7828±0.0033 0.8141±0.0047 0.7115±0.0043 0.9884±0.0012 0.9914±0.0030 0.8695±0.0015 0.8483±0.0051 0.6495±0.0044 0.9432±0.0048 0.5929±0.0056 0.7273±0.0026 0.8881±0.0033 0.6974±0.0027 0.8711±0.0051 0.7017±0.0058
0.9055±0.0048 0.8617±0.0018 0.8452±0.0050 0.7677±0.0034 0.7783±0.0021 0.8269±0.0014 0.8499±0.0016 0.8639±0.0010 0.7473±0.0069 0.7963±0.0050 0.8655±0.0048 0.8971±0.0041 0.8974±0.0058 0.8268±0.0023 0.8520±0.0051 0.6281±0.0061 0.8640±0.0042 0.7537±0.0044 0.7059±0.0045 0.8473±0.0031 0.8700±0.0016 0.8516±0.0049 0.8890±0.0054
0.9328±0.0034 0.9495±0.0044 0.9188±0.0022 0.8834±0.0054 0.6541±0.0063 0.8905±0.0026 0.8772±0.0040 0.9587±0.0050 0.7861±0.0023 0.8821±0.0053 0.7790±0.0047 0.9964±0.0022 0.9972±0.0033 0.8869±0.0046 0.9515±0.0039 0.6487±0.0054 0.9519±0.0017 0.7831±0.0039 0.7258±0.0043 0.9254±0.0024 0.7632±0.0016 0.8602±0.0031 0.8083±0.0033
0.9225±0.0017 0.9431±0.0047 0.9074±0.0043 0.8812±0.0014 0.9137±0.0034 0.9012±0.0031 0.8971±0.0019 0.9459±0.0012 0.8361±0.0033 0.8672±0.0028 0.9495±0.0023 0.9983±0.0025 0.9970±0.0010 0.9110±0.0045 0.9588±0.0042 0.7018±0.0021 0.9665±0.0015 0.9181±0.0033 0.8092±0.0031 0.8778±0.0039 0.8980±0.0029 0.9689±0.0013 0.9340±0.0043
0.9308±0.0019 0.9392±0.0021 0.9245±0.0053 0.8809±0.0040 0.8579±0.0029 0.9052±0.0018 0.8820±0.0012 0.9465±0.0033 0.8297±0.0052 0.8709±0.0020 0.7694±0.0059 0.9983±0.0025 0.9965±0.0017 0.8567±0.0047 0.9484±0.0018 0.7401±0.0042 0.9455±0.0037 0.8386±0.0038 0.7653±0.0054 0.9417±0.0037 0.9018±0.0037 0.9195±0.0053 0.9238±0.0050
0.8777±0.0110 0.8659±0.0086 0.8272±0.0067 0.7975±0.0214 0.7789±0.1101 0.7966±0.0048 0.8050±0.0151 0.9135±0.0119 0.7009±0.0055 0.8236±0.0006 0.6962±0.0065 0.9970±0.0002 0.9933±0.0004 0.9117±0.0333 0.8683±0.0041 0.6627±0.0043 0.9609±0.0231 0.7831±0.0641 0.7043±0.0123 0.9318±0.0009 0.8818±0.0426 0.9143±0.0084 0.9961±0.0013
RecursiveECG Best baseline
0.9540±0.0017 0.9591±0.0015 0.9337±0.0023 0.9191±0.0017 0.9263±0.0026 0.9581±0.0012 0.9205±0.0012 0.9587±0.0047 0.9124±0.0026 0.9454±0.0049 0.9515±0.0012 0.9983±0.0019 0.9718±0.0041 0.8933±0.0015 0.9704±0.0032 0.8828±0.0046 0.9516±0.0040 0.9672±0.0015 0.8247±0.0021 0.9419±0.0028 0.9381±0.0055 0.9885±0.0031 0.9613±0.0057
Delta
xLSTM 0.0212 xLSTM 0.0096 MERL 0.0092 xLSTM 0.0357 UniTS 0.0126 MERL 0.0529 UniTS 0.0234 xLSTM 0.0000 UniTS 0.0763 xLSTM 0.0633 UniTS 0.0020 UniTS/MERL 0.0000 xLSTM -0.0254 Argos -0.0184 UniTS 0.0116 MERL 0.1427 UniTS -0.0149 UniTS 0.0491 UniTS 0.0155 MERL 0.0002 MERL 0.0363 UniTS 0.0196 Argos -0.0348
Table 15: Georgia label AUC results reported as mean ± standard deviation over five runs. Delta is computed from mean values. Label
Support
PatchTST
TimesNet
TS2Vec
xLSTM
UniTS
MERL
Argos
RecursiveECG
Best baseline
Delta
NSR AF IAVB LBBB RBBB SB STach
192 70 61 18 41 146 137
0.7716±0.0038 0.6760±0.0057 0.7235±0.0021 0.9920±0.0025 0.9447±0.0018 0.8946±0.0038 0.9189±0.0025
0.8007±0.0022 0.6908±0.0062 0.7928±0.0069 0.8912±0.0029 0.9199±0.0040 0.9194±0.0027 0.9503±0.0038
0.9247±0.0030 0.8646±0.0050 0.7647±0.0019 0.9890±0.0021 0.9625±0.0028 0.9547±0.0013 0.9536±0.0045
0.9702±0.0026 0.8671±0.0050 0.7717±0.0061 0.9341±0.0035 0.9643±0.0051 0.9745±0.0027 0.9814±0.0010
0.7277±0.0058 0.5445±0.0016 0.4937±0.0051 0.8382±0.0046 0.8414±0.0027 0.7131±0.0039 0.8460±0.0021
0.9756±0.0010 0.9331±0.0026 0.9773±0.0024 0.9637±0.0040 0.9607±0.0047 0.9843±0.0012 0.9902±0.0046
0.8427±0.0012 0.7840±0.0077 0.7765±0.0060 0.9746±0.0016 0.9440±0.0032 0.9164±0.0016 0.9558±0.0010
0.9934±0.0012 0.9455±0.0011 0.9775±0.0035 0.9988±0.0014 0.9592±0.0013 0.9981±0.0017 0.9976±0.0028
MERL MERL MERL PatchTST xLSTM MERL MERL
0.0178 0.0124 0.0002 0.0068 -0.0051 0.0138 0.0074
PTB-XL RVH and SEHYP. The PTB-XL negative labels differ from CPSC2018 RBBB because their results are dominated by support constraints. RVH has 12 test cases and SEHYP has 3 test cases in the diagnostic-subclass evaluation. At this scale, changing the prediction of a single case can cause a large F1 movement. These rows are therefore useful as audit signals but should not be interpreted as stable estimates of diagnostic superiority.
The negative deltas do not imply that the framework cannot model hypertrophy. The related LVH label has a support of 213 and improves from the strongest-baseline F1 of 0.6748 to 0.7559. The case-level evidence suggests a more specific limitation. Hypertrophy subtyping requires the joint interpretation of amplitude criteria, axis deviation, chamber-specific patterns, and secondary repolarization changes. The current compiled measurement set captures voltage and cross-lead evidence sufficiently to improve LVH,
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Table 16: CPSC2018 label AUC results reported as mean ± standard deviation over five runs. Delta is computed from mean values. Label
Support
PatchTST
TimesNet
TS2Vec
xLSTM
UniTS
MERL
Argos
RecursiveECG
Best baseline
Delta
NSR AF IAVB LBBB RBBB PAC PVC STD STE
184 237 144 372 45 121 142 43 180
0.7533±0.0053 0.7869±0.0065 0.6987±0.0049 0.9818±0.0042 0.9401±0.0046 0.5698±0.0026 0.8617±0.0052 0.7191±0.0042 0.7445±0.0042
0.7617±0.0050 0.6933±0.0056 0.6304±0.0042 0.9190±0.0038 0.9045±0.0053 0.6143±0.0046 0.7092±0.0043 0.6771±0.0019 0.6469±0.0026
0.8693±0.0026 0.8722±0.0019 0.7641±0.0013 0.9740±0.0045 0.9622±0.0024 0.6511±0.0034 0.8869±0.0043 0.8293±0.0032 0.7420±0.0029
0.9368±0.0014 0.9673±0.0025 0.8924±0.0041 0.9790±0.0028 0.9846±0.0023 0.7477±0.0023 0.9025±0.0028 0.8928±0.0018 0.8114±0.0037
0.8762±0.0046 0.9067±0.0035 0.8054±0.0022 0.9772±0.0022 0.9612±0.0025 0.6848±0.0065 0.8726±0.0018 0.8534±0.0033 0.8638±0.0041
0.8190±0.0010 0.7953±0.0041 0.7467±0.0057 0.9615±0.0032 0.9050±0.0042 0.6358±0.0054 0.8068±0.0032 0.7342±0.0035 0.8384±0.0038
0.8755±0.0021 0.8476±0.0020 0.7964±0.0011 0.9857±0.0020 0.9495±0.0006 0.7374±0.0073 0.8981±0.0029 0.8651±0.0013 0.8819±0.0076
0.9880±0.0024 0.9944±0.0009 0.9825±0.0034 0.9995±0.0010 0.9721±0.0022 0.9502±0.0039 0.9611±0.0045 0.9665±0.0030 0.9575±0.0016
xLSTM xLSTM xLSTM Argos xLSTM xLSTM xLSTM xLSTM Argos
0.0512 0.0271 0.0901 0.0138 -0.0125 0.2025 0.0586 0.0737 0.0756
but the rare right-ventricular and severe-hypertrophy subclasses provide too few informative failures for stable subtype-specific revision. This is primarily a data-efficiency boundary of the present evaluation rather than direct evidence that the self-iteration mechanism fails. Effect of imbalance. Failure-case selection within EGFR could in principle overemphasize frequent error patterns. RecursiveECG mitigates this risk by selecting false positives, false negatives, lowmargin predictions, calibration errors, and label-specific failures rather than simply reviewing the most numerous aggregate errors. The label-level results are consistent with this design. Large gains occur on several minority- or moderate-support labels, including PTB-XL NST_, ISCA, and WPW and CPSC2018 PAC and STE. The remaining negative labels show that imbalance is not fully resolved, but the observed failure pattern is localized rather than systemic. These observations refine the claim of RecursiveECG. The method is most effective when failed cases expose reproducible morphology, rhythm, or cross-lead evidence that can be compiled into measurements and replayed during review. When a label has extremely low support or depends on unstable, baseline-sensitive morphology, the workflow records the evidence boundary rather than converting an inconclusive case into a speculative rule. Additional label-specific measurement functions or failure-selection quotas may nevertheless be required. This behavior is consistent with the intended role of RecursiveECG as an auditable modeldevelopment workflow rather than a claim of uniform per-label dominance.
C
MECHANISM ABLATIONS AND ITERATIVE REFINEMENT PROCESS C.1 Mechanism Ablations The ablation study reports the full mechanism-sensitivity evaluation used in the main text. RecursiveECG is the complete workflow used for the main results and serves as the reference configuration. We evaluate four controlled workflow variants that remove one design choice at a time. The comparisons hold the candidate budget, iteration budget, data splits, evaluation protocol, and final predictor interface fixed. The two core evidence-grounding mechanisms are Criteria-to-Measurement Compilation (CMC) and Evidence-Grounded Failure Review (EGFR). • w/o EGFR removes informative failure-case selection and the structured Evidence-Grounded Failure Review procedure. The agent retains access to the validated measurement functions produced by CMC, but it cannot jointly inspect selected
failures, comparator cases, raw waveforms, reference-backed measurements, and model behavior to formulate revision hypotheses. Candidate revisions must therefore be selected primarily from aggregate validation metrics and candidate-level evaluation summaries. This variant isolates the contribution of failure-evidence-grounded revision. • w/o CMC removes Criteria-to-Measurement Compilation, including the compilation and validation of curated ECG criteria as deterministic measurement functions. Domain knowledge remains accessible through a direct referencereading interface, but EGFR receives no reproducible, reference-backed measurement records emitted from individual ECGs. It can still inspect raw waveforms, model outputs, and comparator cases. This variant isolates the contribution of operationalizing textual clinical criteria as validated deterministic measurements. • w/o Measurement Execution retains the measurement specifications, validated function implementations, computation rules, and evidence back-pointers produced by CMC, but prevents these functions from being executed on individual review cases. EGFR therefore has access to the static measurement definitions but not to their emitted case-level measurement records. It must reason from raw waveforms, model outputs, comparator cases, and static rule descriptions. This variant distinguishes the availability of compiled measurement knowledge from the use of executed measurement evidence during failure review. • Direct LLM Design replaces the governed workflow with a single coding-agent session. It removes the explicit problemcontract gate, CMC, independent candidate exploration, structured EGFR, promotion ledger, and adaptive stopping procedure. Candidate changes are not organized through the same separation between evidence collection, revision formulation, execution, re-evaluation, and promotion. This configuration serves as a simplified workflow baseline and isolates the contribution of the governed multi-stage organization. Together, these five configurations provide the controlled comparison used in the main text. The w/o EGFR and w/o CMC variants isolate the two core evidence-grounding mechanisms. The w/o Measurement Execution variant separates compiled-measurement availability from case-level measurement usage. The Direct LLM Design variant evaluates the contribution of the governed workflow organization. Backbone sensitivity is reported separately in the capability-gradient table below, while RecursiveECG anchors the
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
mechanism comparison. The rightmost column of Tab. 17 reports the Macro F1 difference between RecursiveECG and each variant. Positive values indicate degradation after ablation, whereas negative values indicate that the ablated run exceeds the reference run on that dataset. The dominant ablation effect differs across datasets. On PTB-XL, removing EGFR reduces Macro F1 by 0.0613, while replacing the governed workflow with Direct LLM reduces it by 0.0606. This indicates that both structured failure review and workflow organization contribute substantially in this setting. On Georgia, the largest degradation is produced by Direct LLM Design (0.2524), followed by the CMC ablation (0.0423), whereas disabling per-case measurement execution is effectively neutral in this run. CPSC2018 shows lower aggregate sensitivity than Georgia but still degrades under the workflow variants: Direct LLM Design reduces Macro F1 by 0.0467, removing EGFR by 0.0411, removing CMC by 0.0226, and disabling per-case measurement execution by 0.0099. These results support an aggregate mechanism claim rather than requiring every workflow mechanism to improve every dataset or label. Backbone sensitivity is summarized separately in Tab. 18. The aggregate mechanism effects are not attributable to a single label. PTB-XL is most consistently sensitive to removing EGFR, while Georgia shows the clearest dependence on the governed workflow. CPSC2018 has a more heterogeneous profile, with some labels neutral or improved under individual ablations. The intended conclusion is robustness at the aggregate and mechanism levels, rather than uniform per-label dominance by every component.
C.2
Search Width, Backbone, and Iterative Refinement Process
Search-width sensitivity indicates that 𝑘 = 3 provides the best accuracy–cost operating point among the evaluated PTB-XL settings. Increasing the search width to 𝑘 = 5 increases the number of candidates that pass the execution and regression checks but does not improve the final Macro F1, whereas the single-candidate setting 𝑘 = 1 produces a weaker final predictor. The pass/fail counts in Tab. 19 summarize candidate-level checks and should not be read as the number of promoted revisions in the main iterative chain. The resulting trajectory across the five refinement rounds is reported in Tab. 20. It separates the evidence or update source used at each round from the final performance change, making clear how the frozen predictor was selected rather than presenting only the final endpoint. The trajectory shows that the final predictor was not selected from a single isolated trial. The largest improvement occurs after the regression-aware revision in Round 4. This is consistent with the method definition: EGFR formulates targeted revision hypotheses from failure evidence, but a hypothesis does not become an update solely because it is supported by an individual case. It must first be implemented, executed under the fixed problem contract, reevaluated on the prescribed validation protocol, and accepted by
Deng et al.
the regression and promotion gates before it enters the frozen toolchain.
D CASE AUDITS AND DEPLOYMENT D.1 WPW Pseudo-Infarction Failure Audit The audit trace in Tab. 21 records the case-specific evidence retained from Evidence-Grounded Failure Review (EGFR). It is intentionally presented as a development-time model-design trace rather than as a clinical diagnostic explanation. The claim is that RecursiveECG uses reproducible measurements produced through Criteria-toMeasurement Compilation (CMC) to formulate a testable revision hypothesis, execute it under the fixed problem contract, and retain it only after replay and regression evaluation; the trace supports the specific revision on this reviewed case, not a general guarantee about all WPW cases or all future revisions.
D.2
Evidence-Grounded Failure Review Example
For completeness, Tab. 22 gives a second, audited PTB-XL failurereview trace. Unlike the WPW example above, this trace focuses on a cross-lead hypertrophy error and shows the initial model scores, measured waveform evidence, candidate revision, replay outcome, and promotion decision in one development-time record. This case illustrates the intended role of Evidence-Grounded Failure Review. The review begins with a concrete model failure and jointly examines the raw ECG, reference-backed measurements produced by CMC, model outputs, and the relevant cross-lead evidence. It then formulates a targeted revision hypothesis and admits that hypothesis to executable candidate testing. The revision is promoted only after case replay and regression evaluation support its retention. The LLM is therefore used as an offline model-design controller, not as an online diagnostic judge.
D.3
Deployment Efficiency
Efficiency is reported for the frozen predictor produced by the governed refinement process, rather than for the offline agent search. The PTB-XL efficiency evaluation shows that RecursiveECG is not the fastest model in raw inference latency, but it uses few parameters and FLOPs while achieving higher macro F1 than the strongest compared baselines. The deployed predictor has higher latency than TS2Vec but lower latency than most of the compared neural baselines. It also has low FLOPs and peak memory usage, and it requires no LLM query at inference time. Under the default offline development setting, the candidate budget is 𝑘 = 3 and the workflow runs for five iterations. This produces 15 candidate proposals, of which five pass the candidatelevel checks and 10 fail the checks or are rejected. The average agent time is 44.9 minutes per iteration. These quantities characterize one-time model development cost and should not be interpreted as per-record diagnostic cost.
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Table 17: Overall mechanism-ablation results reported as mean ± standard deviation over five runs. Delta is RecursiveECG minus the variant Macro F1 mean; positive values indicate degradation. Backbone sensitivity is reported separately in the capability-gradient table. Dataset
Variant
PTB-XL PTB-XL PTB-XL PTB-XL PTB-XL Georgia Georgia Georgia Georgia Georgia CPSC2018 CPSC2018 CPSC2018 CPSC2018 CPSC2018
RecursiveECG w/o EGFR w/o CMC w/o Measurement Execution Direct LLM Design RecursiveECG w/o EGFR w/o CMC w/o Measurement Execution Direct LLM Design RecursiveECG w/o EGFR w/o CMC w/o Measurement Execution Direct LLM Design
Macro AUC
Macro F1
Micro F1
Hamming Acc.
Delta Macro F1
0.9373±0.0066 0.9149±0.0048 0.9235±0.0037 0.9273±0.0042 0.9060±0.0055 0.9814±0.0052 0.9788±0.0016 0.9765±0.0024 0.9808±0.0015 0.9223±0.0049 0.9746±0.0036 0.9728±0.0020 0.9735±0.0018 0.9652±0.0031 0.9663±0.0027
0.7653±0.0088 0.7040±0.0093 0.7354±0.0071 0.7430±0.0066 0.7047±0.0108 0.8826±0.0069 0.8735±0.0057 0.8403±0.0082 0.8829±0.0064 0.6302±0.0116 0.8027±0.0078 0.7616±0.0089 0.7801±0.0073 0.7928±0.0061 0.7560±0.0096
0.7930±0.0039 0.7546±0.0061 0.7682±0.0054 0.7761±0.0047 0.7392±0.0070 0.9170±0.0043 0.9220±0.0038 0.8951±0.0049 0.9148±0.0035 0.7918±0.0067 0.8352±0.0056 0.7987±0.0063 0.8105±0.0058 0.8329±0.0045 0.8182±0.0069
0.8929±0.0177 0.8783±0.0084 0.8789±0.0068 0.8892±0.0059 0.8505±0.0091 0.9735±0.0021 0.9757±0.0025 0.9675±0.0031 0.9731±0.0022 0.9376±0.0055 0.9613±0.0030 0.9444±0.0046 0.9517±0.0039 0.9609±0.0032 0.9588±0.0036
0.0000 0.0613 0.0299 0.0223 0.0606 0.0000 0.0091 0.0423 -0.0003 0.2524 0.0000 0.0411 0.0226 0.0099 0.0467
Table 18: LLM-backbone sensitivity across three datasets. All metrics are reported as mean ± standard deviation over five runs. The candidate budget and maximum iteration count are fixed at 𝑘 = 3 and 5, respectively. Dataset
Backbone
PTB-XL Georgia CPSC2018 PTB-XL Georgia CPSC2018
Capability tier
𝑘
Max iter.
Macro AUC
Macro F1
Micro F1
Hamming Acc.
DeepSeek-V4-Pro Advanced reasoning DeepSeek-V4-Pro Advanced reasoning DeepSeek-V4-Pro Advanced reasoning
3 3 3
5 5 5
0.9373±0.0066 0.9814±0.0052 0.9746±0.0036
0.7653±0.0088 0.8826±0.0069 0.8027±0.0078
0.7930±0.0039 0.9170±0.0043 0.8352±0.0056
0.8929±0.0177 0.9735±0.0021 0.9613±0.0030
Qwen3.5-27B Qwen3.5-27B Qwen3.5-27B
3 3 3
5 5 5
0.9166±0.0187 0.9698±0.0124 0.9437±0.0169
0.7338±0.0415 0.8021±0.0368 0.6988±0.0442
0.7821±0.0326 0.8662±0.0297 0.7204±0.0395
0.9118±0.0149 0.9587±0.0116 0.9199±0.0173
Lightweight reasoning Lightweight reasoning Lightweight reasoning
Table 19: PTB-XL candidate-count sensitivity. 𝑘
Max iter.
AUC
Macro F1
Micro F1
Hamming Acc.
Passed / failed checks
1 3 5
5 5 5
0.9231 0.9373 0.9325
0.7265 0.7653 0.7584
0.7638 0.7930 0.7869
0.8822 0.8929 0.8927
3/2 5 / 10 12 / 13
Note greedy single-candidate update default setting wider search without a final Macro-F1 gain
Table 20: PTB-XL iterative refinement trajectory. Round
Main evidence or update source
0 1 2 3 4 5
initial model and default threshold informative validation failures and initial EGFR reference-backed ECG measurements and label-wise error analysis minority-label and false-negative failure review regression-aware candidate evaluation final frozen RecursiveECG
Macro AUC
Macro F1
0.9259 0.9286 0.9286 0.9304 0.9373 0.9373
0.7183 0.7270 0.7287 0.7345 0.7653 0.7653
Micro F1 Hamming Acc. Gain over Round 0 0.7641 0.7723 0.7723 0.7700 0.7930 0.7930
0.8811 0.8871 0.8871 0.8865 0.8929 0.8929
0.0000 0.0087 0.0104 0.0162 0.0470 0.0470
Deng et al.
0.78
0.940
0.76
0.935
0.74
0.930
0.72 0.70
Macro F1 Macro AUC
0
1
2 3 Iteration round
4
5
Macro AUC
Macro F1
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
0.925 0.920
Figure 11: PTB-XL iterative refinement trajectory. Table 21: Complete retained audit trace for WPW case 461. The reference-backed measurements are computed from the supplied 12-lead ECG array around the reviewed QRS complex. Audit item
Evidence
Sample Ground-truth condition
Case 461, represented as a 12-lead ECG array with shape 12 × 1000, sampled at 100 Hz and stored in mV. WPW / ventricular pre-excitation. The case is reviewed because WPW can produce pseudo-infarction morphology that resembles myocardial infarction in individual leads. The evaluation stage selected the record as an informative MI-like error. Its local morphology contains Q-like initial negative deflections, whereas the ground-truth condition is ventricular pre-excitation. The reviewed record contains WPW-compatible pre-excitation evidence, including a short PR interval, a slurred QRS onset consistent with a delta wave, and a widened QRS complex. At the same time, several leads show local pseudo-infarction morphology, which provides plausible evidence for the initial MI-like prediction. The replayed measurements match the evidence shown in Fig. 9: the PR interval is below the clinical short-PR threshold of 120 ms, the QRS duration is 130 ms, and the frontal-plane axis is approximately −35◦ , pointing toward the left-superior/aVL sector. The QRS onset is slurred, supporting a delta-wave interpretation rather than a purely infarction-driven explanation. The current predictor overweights lead-local pseudo-infarction morphology and does not adequately evaluate whether the activation pattern is consistent across limb and precordial leads. The review links the apparent infarction cues to WPW physiology: activation through an accessory pathway changes the initial ventricular activation vector and can generate Q-wave-like pseudo-infarction patterns without requiring an MI explanation. Introduce a multi-lead mechanism that replaces purely lead-local morphology reasoning with cross-lead attention, allowing the predictor to evaluate axis-level and morphology-level consistency across limb and precordial leads. The proposed cross-lead revision is implemented as an executable candidate and evaluated under the fixed data split, metric, threshold-selection, and leakage constraints. Its behavior is then replayed on the reviewed case and checked for validation-set regressions. After replay, the revised predictor suppresses the MI interpretation and recovers the WPW label. The revision is retained only after the executable candidate passes case replay and regression checks. The resulting trace therefore connects a concrete failure, raw waveform evidence, reference-backed measurements, model behavior, and a promoted model-design update.
Failure selected for Evidence-Grounded Failure Review Raw waveform evidence
Reference-backed measurements
Identified pipeline weakness Evidence-grounded interpretation
Revision hypothesis
Execution and re-evaluation
Replay result Promotion decision
Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Table 22: Audited PTB-XL Evidence-Grounded Failure Review example. Item
Evidence
Sample Ground truth Initial error Initial model outputs Raw ECG evidence
PTB-XL validation sample 1103. HYP= 1, with all other diagnostic superclasses equal to 0. NORM false positive and HYP false negative. NORM score = 0.859 and HYP score = 0.151. High R-wave amplitude in leads II, aVF, and V4–V6; a deep S wave in aVR; and an LVH-like cross-lead voltage pattern. Lead II maximum = 10.96, aVF maximum = 10.57, aVR minimum = −9.14, V4 maximum = 10.34, V5 maximum = 8.49, and V6 maximum = 7.81. The independent-channel CNN fails to combine the distributed high-voltage evidence across leads and therefore misses the cross-lead pattern associated with HYP. Introduce multi-head cross-lead attention so that the predictor can jointly evaluate voltage relationships across limb and precordial leads. The proposed revision is implemented and evaluated as an executable candidate in Iteration 004 under the fixed problem contract. The revised candidate corrects HYP to 1 and suppresses NORM to 0. The HYP score increases from 0.151 to 0.755, while the NORM score decreases from 0.859 to 0.357. The candidate passes the replay and regression checks and is therefore retained in the final Round 5.
Reference-backed measurements Identified pipeline weakness Revision hypothesis Candidate execution Replay result Score change Promotion decision
Table 23: Frozen-predictor deployment efficiency on PTB-XL. Method
Type
Params
FLOPs/sample
ms/sample
PatchTST TimesNet TS2Vec xLSTM-ECG
patch Transformer frequency-temporal temporal representation ECG-specific xLSTM
1.368M 2.505M 0.640M 0.545M
1.478G 15.620G 1.274G 2.860G
0.461 0.573 0.081 1.548
2,169 1,746 12,376 646
1,309.7 170.0 188.0 4,890.5
1.066M 5.382M 0.487M
0.903G 1.155G 0.086G
0.236 0.194 0.116
4,244 5,149 8,601
181.8 56.1 19.2
UniTS unified time-series MERL medical ECG representation RecursiveECG final model offline-agent-designed frozen predictor
Throughput Peak mem. MB