arXiv CSopen access

PEACE: Cross-modal Enhanced Pediatric-Adult ECG Alignment for Robust Pediatric Diagnosis

arXiv:2605.00647v1 [cs.LG] 1 May 2026

PEACE: Cross-modal Enhanced Pediatric-Adult ECG Alignment for Robust Pediatric Diagnosis

Xinran Liu∗ School of Instrument Science and Engineering Southeast University Nanjing 210096, China

Yuwen Li∗ School of Instrument Science and Engineering Southeast University Nanjing 210096, China

Hongxiang Gao School of Instrument Science and Engineering Southeast University Nanjing 210096, China

Heyang Xu School of Instrument Science and Engineering Southeast University Nanjing 210096, China Zongmin Wang† Zhengzhou University Zhengzhou 450001, China [email protected]

Jianqing Li Nanjing Medical University Nanjing 211166, China

Chengyu Liu† School of Instrument Science and Engineering Southeast University Nanjing 210096, China [email protected]

∗

Equal contribution. † Corresponding authors.

Abstract Automated pediatric electrocardiogram (ECG) diagnosis remains challenging because models trained predominantly on adult data suffer from substantial crosspopulation mismatch, while pediatric labels are often scarce. We present PEACE (Pediatric–Adult ECG Alignment via Cross-modal Enhancement), a structured cross-modal alignment framework for adult-to-pediatric ECG transfer. PEACE integrates tri-axial clinical semantic decomposition, label-query feature extraction, and curriculum-gated optimization to align transferable adult ECG representations with pediatric diagnostic targets. Since ZZU-pECG provides no paired clinical reports, we generate label-conditioned semantic descriptors using Gemini with concise clinical prompts and use them only as auxiliary training supervision; inference remains ECG-only. On ZZU-pECG, PEACE achieves 59.39%, 79.03%, and 90.89% AUC under zero-shot, 50-shot, and full fine-tuning settings, respectively, and reaches 96.65% AUC on the shared PTB-XL label space. These results suggest that structured clinical semantic supervision can improve low-resource adult-to-pediatric ECG transfer, while prospective clinical validation and more explicit age-aware modeling remain necessary before real-world deployment. Preprint.

1

Introduction

Cross-population transfer is challenging when the source domain is large and semantically annotated, while the target domain is distribution-shifted and label-scarce. We study this setting in pediatric ECG classification, where adult ECG datasets are abundant but pediatric labels and paired reports are limited. Distribution mismatch induced by developmental electrophysiology causes adult-trained models to degrade on pediatric signals Ribeiro et al. [2020], Leone et al. [2024], Chen et al. [2024], and limited pediatric supervision further constrains adaptation Zhu et al. [2022]. State-of-the-art models for arrhythmia detection in adults Chen et al. [2023], Leone et al. [2024], Yang et al. [2025] have demonstrated strong performance, yet their efficacy diminishes in pediatric scenarios with out-of-distribution morphology Leone et al. [2024]. From an ML perspective, we focus on transfer under target-domain scarcity: given abundant labeled adult ECGs and limited labeled pediatric ECGs, we seek an adaptation strategy that remains stable across zero-shot, few-shot, and full-supervision regimes. While recent efforts have explored cross-racial generalization Tian et al. [2024], adult-to-pediatric transfer under severe distribution shift remains less studied Liang et al. [2021]. To address this cross-population distribution shift, we propose PEACE (PediatricAdult ECG Alignment via Cross-modal Enhancement), a structured multimodal framework that uses clinical semantics as transfer anchors. By introducing structured semantic supervision and label-guided alignment, PEACE learns representations that are more robust to pediatric distribution shift. Category-wise gain overview. PEACE achieves substantial area under the receiver operating characteristic curve (AUC) gains across multiple pediatric categories, including an 11.1% absolute improvement on TAB_ and 98.5% on LAFB (1.6% above the single-modal ResNet1d baseline Strodthoff et al. [2020]). The full category-wise bar-chart comparison is provided in Appendix 3. The technical core of PEACE comprises two components: Label-Specific Bidirectional Contrastive Learning (LSBC) and Curriculum-Adaptive Fusion (CAF). LSBC constructs disease-aware positive pairs across three clinical dimensions, namely rhythm, morphology, and ST-T repolarization, enabling fine-grained alignment between waveform features and diagnostic semantics. CAF employs a multistage curriculum to dynamically calibrate modality weights. This process stabilizes cross-population transfer and mitigates representation disruption during early-stage optimization. We validate PEACE on ZZU-pECG Tan et al. [2025], a high-fidelity pediatric database comprising ECG records from 11,643 children. We additionally evaluate on PTB-XL Wagner et al. [2020] as an external set over shared/evaluable labels. The distinct morphological variances between populations are shown in Appendix A, Figure 4, where the same label IRBBB exhibits markedly different Lead II waveforms. PEACE outperforms evaluated baselines on these benchmarks. The primary contributions are summarized as follows: • Tri-axial semantic decomposition for transfer: We decompose diagnostic semantics into rhythm/morphology/ST-T repolarization axes based on clinical reporting practice, reducing semantic entanglement and mitigating attenuation of label-specific cues. In this paper, attenuation of label-specific cues means that subtle pathology-relevant information is lost when heterogeneous diagnostic cues are fused into a single undifferentiated semantic stream. • Label-query feature extraction and structured alignment: We design an LQN module that performs label-guided cross-attention on ECG/report features and combine it with label-specific bidirectional alignment, improving over global pooling-based matching. • Curriculum-gated optimization for stable transfer: We introduce CAF to gate alignment intensity across training stages, preventing early-stage disruption of pretrained discriminative features. Without CAF, zero-shot AUC drops to 47.10%, which is below the random baseline. • Comprehensive cross-regime validation: On ZZU-pECG (11,643 children), PEACE improves AUC over evaluated baselines under zero-shot, few-shot, and full fine-tuning; external PTB-XL results provide supportive transfer evidence on the shared label space.

2

Figure 1: Pediatric–Adult ECG Alignment via Cross-modal Enhancement (PEACE).

2

Related Work

2.1

ECG Representation Learning

ECG analysis evolved from foundational CNN and RNN architectures Kiranyaz et al. [2015], Yildirim [2018], Chen et al. [2020] to attention-based Transformer paradigms Vaswani et al. [2017], Akan et al. [2023]. These methods capture temporal dependencies, but direct transfer from adult-heavy corpora to pediatric cohorts remains difficult under developmental distribution shift. 2.2

Multimodal ECG-Text Alignment

Early integration relied on coarse strategies such as feature concatenation or late fusion Lalam et al. [2023]. These often fail to capture nuanced interactions between waveforms and clinical text. Contemporary frameworks Tian et al. [2024], Li et al. [2025], Wang et al. [2022] use attention-based global alignment, but pediatric transfer requires finer local-to-semantic correspondence because morphology deviates from adult-centered patterns Gliner et al. [2025]. 2.3

Label Semantics and Supervised Contrastive Learning

Contrastive learning is widely used in medical representation learning, including supervised diagnostic hierarchies Li and Gao [2022], Lu et al. [2023], multi-task objectives Hou et al. [2023], and bidirectional alignment Xiong et al. [2025]. Standard formulations are often label-agnostic in pair construction, which can underutilize label semantics when target-domain labels are scarce. Our LSBC objective is SupCon-style in spirit but differs in three transfer-oriented choices: (i) alignment is performed on label-conditioned queried features instead of global pooled representations, (ii) semantic supervision is decomposed into three clinical axes instead of a single fused text stream, and (iii) alignment intensity is dynamically gated by CAF to reduce early-stage degradation of pretrained features. We do not claim LSBC as a new generic contrastive primitive; the contribution is its instantiation within a structured transfer framework that combines LQN, tri-axial semantic descriptors, and curriculum-gated optimization.

3

2.4

Cross-Population Transfer under Target Scarcity

Recent advances explore spatial-temporal memory Na et al. [2024], self-supervised pretraining Liu et al. [2024], structured knowledge injection Tang et al. [2026], medical prompting Wang et al. [2025], and language supervision Zhou et al. [2025]. Curriculum-based denoising also improves multimodal optimization Xu et al. [2025]. Our setting differs in that the target domain is pediatric, strongly shifted, and label-scarce; this motivates label-conditioned alignment plus curriculum gating rather than relying on global alignment alone. 2.5

Domain Adaptation Perspective

Physiological discrepancies across heart rate, morphology, and thresholds create substantial crosspopulation distribution shift between children and adults Leone et al. [2024], Dickinson [2005]. Developmental shifts, such as neonatal ventricular dominance, require age-aware interpretation criteria. Models trained predominantly on adult data therefore show notable mismatch on pediatric cohorts Gow et al. [2023], Han et al. [2023]. Existing adaptation techniques target adult scenarios via unsupervised alignment He et al. [2023], patient reweighting Fan et al. [2025], or distribution matching Xiong et al. [2025]. Standard diagnostic guidelines Surawicz et al. [2009] assume relatively static physiology, which conventional strategies struggle to reconcile with pediatric maturation. These methods are ill equipped for pediatric transfer, where pathological manifestations are governed by maturational mechanisms and symptomatic overlaps with conditions like Brugada syndrome Brugada and Brugada [1992]. PEACE performs supervised cross-population transfer. LSBC uses label-specific feature matching to mitigate pediatric shift, and CAF modulates alignment strength to improve stability during adult-topediatric adaptation Xu et al. [2025], Liu et al. [2025].

3

Methodology

3.1

Overview

The PEACE framework is engineered to bridge morphological drift, which represents the fundamental physiological variance in ECG waveforms between adult and pediatric populations. As illustrated in Figure 1, adult ECGs serve as the source of physiological knowledge, while diagnostic labels provide dual-purpose supervision for multi-label classification and the derivation of label-driven semantic prompts. PEACE mimics the clinical diagnostic process by isolating pathological markers through the LQN, aligning them via LSBC, and stabilizing the transfer via CAF. Notation. We denote batch size by B, the number of time tokens by T , and the number of diagnostic labels by C. For mathematical simplicity, we set the latent dimensions of all encoders to be equal, i.e., decg = drep = dlbl = d. The ECG encoder output is denoted by X ecg , the semantic descriptor encoder output by X rep , and label text embeddings by Z lbl . We keep the symbol “rep” for continuity with earlier drafts; this encoder processes label-conditioned semantic descriptors, not patient-specific reports. All modalities are projected into a shared latent space of dimension dshare , where dshare ≪ d and dshare ∈ N+ . We use [·| · |·] for concatenation and sim(·, ·; τ ) for cosine similarity with temperature τ . 3.2

Multi-Modal Encoders

ECG Encoder. We employ the resnet1d_wang implementation from the torch_ecg library, based on a 1D ResNet architecture Strodthoff et al. [2020], to project raw 12-lead ECG signals into tokenlevel representations X ecg ∈ RB×T ×d . Each token represents a localized physiological segment and captures both micro-morphologies such as QRS spikes and broader macro-dynamics. Global average pooling yields the classification embedding T 1X X̄ ecg = X ecg [:, t, :] ∈ RB×d . (1) T t=1 while the full token sequence X ecg ∈ RB×T ×d is preserved for LQN cross-attention with label queries. Semantic Descriptor Encoder. To reflect clinical interpretation workflows and align with 4

the diagnostic taxonomy defined in Section 4.1, we decompose diagnostic prompts into three semantic axes: rhythm, morphology, and ST-T repolarization. This design follows structured clinical semantic factorization: rhythm disturbances in the temporal domain and structural morphological changes in the spatial domain are modeled separately from repolarization-related changes. This structured semantic decomposition prevents attenuation of label-specific cues, where subtle pediatric markers such as J-point or ST-segment patterns are obscured by dominant adult patterns. We encode these dimensions via BioClinicalBERT Alsentzer et al. [2019] and fuse them via linear projection: X rep = W fus [X rhythm ∥ X morph ∥ X isch ] ∈ RB×d , (2) d×3d where W fus ∈ R is a learnable projection matrix, and ∥ denotes feature concatenation. Label Embeddings and Shared Projection. Each diagnostic label is encoded via BioClinicalBERT to produce Z lbl ∈ RC×d . Features of all modalities are projected into a shared latent space Rdshare via task-specific MLP heads: X̃ ecg = MLPecg (X ecg ), X̃ rep = MLPrep (X rep ), and Z̃ lbl = MLPlbl (Z lbl ). Note that for the LQN input, the MLP is applied point-wise to the token sequence X ecg . The inclusion of ST-T repolarization as a dedicated axis follows standard ECG interpretation practice: repolarization changes such as ST-segment and T-wave abnormalities are clinically important and complementary to rhythm and structural morphology. In pediatric settings, this axis is particularly useful for separating repolarization patterns from developmental T-wave variations. Gemini-Generated Semantic Descriptors. Since ZZU-pECG does not contain paired free-text clinical reports, we use Gemini to generate short label-conditioned diagnostic descriptors along the three semantic axes under clinical reporting constraints. These descriptors provide structured semantic supervision during training and are not used as sample-specific notes. Inference remains strictly ECG-only. Additional controls (no-text and single-text variants) are reported in Table 2. 3.3

Label-Query Network (LQN): Task-Steered Feature Probing

Standard representations often suffer from feature entanglement in multi-label scenarios, where markers for different pathologies such as tachycardia and hypertrophy overlap in a single vector. The LQN resolves this by treating diagnostic labels as label-conditioned queries that extract label-specific features from modality embeddings via cross-attention. For each diagnostic label c ∈ {1, 2, . . . , C} and sample i ∈ {1, 2, . . . , B}, we leverage the projected label embedding Z̃ lbl [c, :] as a query the modality  to interrogate  keys and values: z̄ ecg ∈ Rdshare , i,c = LN Attn Z̃ lbl [c, :], X̃ ecg [i, :]    (3) dshare z̄ rep = LN Attn Z̃ [c, :], X̃ [i, :] ∈ R . lbl rep i,c where Attn(·, ·) denotes the Multi-Head Cross-Attention operator with layer normalization LN(·). rep This selective attention ensures z̄ ecg i,c and z̄ i,c isolate patterns specific to label c, filtering out coexisting pathological noise and providing a stable foundation for cross-modal alignment. Each token corresponds to a short temporal window after encoder downsampling, enabling sub-beat pattern selection rather than only global averaging. 3.4

Label-Specific Bidirectional Contrastive Learning (LSBC)

LSBC enforces category-aware consistency within the feature manifolds defined by the LQN outputs. For each diagnostic label c, we define the positive sample set Pc (samples annotated with label c) and negative sample set Nc (samples without label c) in the current batch. The directional InfoNCE-style contrastive loss for label c (from modality m 1 to m2 ) is: 1 X m1 →m2 Lc =− |Pc | i∈Pc
P (4) m1 m2 j∈Pc exp sim(z̄ i,c , z̄ j,c )/τ  . log P m1 m2 k∈Pc ∪Nc exp sim(z̄ i,c , z̄ k,c )/τ The total LSBC loss is averaged over active labels in the batch CB and both cross-modal directions: 1 X 1 ecg→rep LLSBC = (L + Lrep→ecg ). (5) c |CB | 2 c c∈CB

5

3.5

Curriculum-Adaptive Fusion (CAF): Gated Optimization

During transfer, early-stage cross-modal alignment can introduce strong contrastive gradients before unimodal encoders have stabilized, disrupting discriminative features acquired during pretraining. CAF acts as a convergence-aware regulator to delay aggressive alignment until training is stable. The dynamic weight of the LSBC loss is jointly governed by normalized training progress t ∈ [0, 1] and a stability gate: wLSBC,t = βt · I (∆t < ε) , (6) where I(·) is the indicator function and ε is a pre-defined stability threshold. The stability gate monitors the exponential moving average (EMA) of the multi-label classification loss to quantify training convergence: EMAt = (1 − γ)EMAt−1 + γLce,t , (7) ∆t = |EMAt − EMAt−K | . where γ is the EMA decay rate, Lce,t is the instantaneous multi-label classification loss at step t, and K is the sliding window size for convergence monitoring. The alignment intensity is further modulated by a three-stage curriculum weight βt (normalized to  [0, 1]): t t ∈ [0, 0.3), 0.1 − 0.1 · 0.3 , t−0.3 (8) βt = 0.3 · 0.4 , t ∈ [0.3, 0.7),  0.3 + 0.7 · t−0.7 , t ∈ [0.7, 1]. 0.3 We additionally tested multiple curriculum breakpoints in the 50-shot regime and observed robust performance: [0.2, 0.6]→0.782 AUC, [0.3, 0.7]→0.790 AUC, and [0.4, 0.8]→0.788 AUC. 3.6

Final Training Objective

The entire PEACE framework is jointly optimized by fusing multi-label classification loss and curriculum-weighted LSBC contrastive loss. The total training objective is: L = (Lce,ecg + Lce,rep ) + λmax · wLSBC,t · LLSBC , (9) where Lce,ecg and Lce,rep denote the multi-label Binary Cross-Entropy loss for ECG and semantic descriptor encoders (respectively) on the original diagnostic label task, and λmax is a hyper-parameter that controls the maximum alignment strength of the LSBC loss. For the semantic branch, descriptors are constructed from training labels and therefore serve as privileged auxiliary supervision during training; this branch is discarded at inference. We report No-text and single-text controls to quantify gains from structured semantic content beyond ECG-only training.

4

Experiments

4.1

Datasets

The PEACE model is pretrained on MIMIC-IV Gow et al. [2023], a comprehensive adult ECG dataset comprising over 800,000 records. Evaluation is conducted on two target domains: ZZU-pECG Tan et al. [2025] (pediatric, ages 0-14) and PTB-XL Wagner et al. [2020]. ZZU-pECG contains 11,643 children in the released database. After filtering and label mapping, we retained 7,593 valid ECG records for experiments, corresponding to 9,198 label instances due to multi-label annotation. As shown in Figure 2, the pediatric domain exhibits a significant domain shift, characterized by morphological divergence and extreme label imbalance. Full names for all diagnostic labels are detailed in the note of Table 3. To enable structured clinical analysis, we categorize the 12 diagnostic labels into three evaluation axes based on their electrophysiological properties Mason et al. [2007], matching the tri-axial semantic structure used in our semantic descriptor encoder: • Rhythm (1 label): This category includes the NORM label. • Morphology (10 labels): This axis aggregates structural abnormalities and morphological variants, including conduction blocks (e.g., CRBBB, IRBBB, LAFB), hypertrophy (e.g., LVH, RVH, LAO/LAE, RAO/RAE), and other morphological variants (e.g., TAB_, LVOLT, LQTS). 6

Figure 2: Sample distribution across diagnostic labels, illustrating the relative proportion of each category within the adult and pediatric datasets. Table 1: Comprehensive Performance Comparison and Clinical Dimension Fingerprinting (AUC %). "FT" means Fine-Tuning. PTB-XL full-FT scores are computed on shared/evaluable labels only. Axiswise scores are descriptive subgroup averages rather than balanced clinical-dimension benchmarks, because the three axes contain different numbers of labels. Best performance is highlighted in bold with a gray background . Pediatric (ZZU-pECG)

Model

Zero-shot 50-shot

Full FT

Axis-wise Proficiency (ZZU Full FT) Cross-Pop. (PTB-XL) Rhythm Morphology

ST-T

Full FT

ST-MEM Na et al. [2024] 50.91±1.46 66.08±2.21 81.55±0.53

80.41

81.45

83.68

89.22±0.98

MERL Liu et al. [2024]

58.58±0.00 70.65±1.91 80.97±0.92

82.09

81.06

78.85

92.49±0.36

KED Tian et al. [2024]

57.77±0.00 50.66±1.00 80.54±3.50

80.90

81.13

74.21

89.10±9.36

PEACE (Ours)

59.39±0.00 79.03±1.00 90.89±0.96

91.68

91.15

87.49

96.65±0.24

• ST-T repolarization (1 label): Represented by the STTC label. Specifically, TAB_ constitutes 34.2% of ZZU-pECG but only 0.8% of PTB-XL, while STTC exhibits inverse prevalence patterns. Detailed sample distributions are provided in Appendix A.2. 4.2

Experimental Settings

Architecture and Optimization. PEACE is instantiated using a ResNet1D backbone Strodthoff et al. [2020] for waveform encoding and BioClinicalBERT for clinical text processing. To balance knowledge retention with domain adaptation, we employ a graduated freezing strategy where the first 10 transformer blocks of BioClinicalBERT remain fixed. The model is optimized using AdamW with a cosine annealing learning rate scheduler. To address the long-tailed distribution of clinical labels, we incorporate class-balanced loss weighting. Transfer Learning Regimes. We evaluate PEACE under three clinical deployment scenarios. Zeroshot Transfer: the pretrained model remains frozen, and classification thresholds are optimized on the validation set. Few-shot Adaptation: following standard few-shot protocols, the model is fine-tuned on 50 samples per class (50-shot) for 20 epochs; this configuration was determined based on sample efficiency saturation in Appendix D. Full Supervised Fine-tuning: the model is optimized on the complete target dataset while maintaining foundational text encoder freezing. Baseline Models. To evaluate the efficacy of PEACE, we compare it against a diverse set of strong baselines representing various paradigms in ECG representation learning: • ST-MEM Na et al. [2024]: Uses a masked auto-encoding framework to capture morphological dependencies, benchmarking generative self-supervised priors in pediatric transfer. • MERL Liu et al. [2024]: Employs multi-resolution feature extraction and global alignment, representing the standard for multimodal ECG-text synchronization. • KED Tian et al. [2024]: Integrates medical knowledge via textual-signal alignment to evaluate knowledge distillation stability in data-scarce pediatric scenarios.

7

Table 2: Ablation and semantic-supervision controls on ZZU-pECG (AUC %). “–” indicates that the setting was not evaluated in that regime. Best performance is highlighted in bold. No-text removes the semantic branch within the PEACE training framework and should not be interpreted as a separately optimized single-modal ECG baseline. Configuration

Zero-shot

50-shot

Full FT

No-text w/o LSBC w/o CAF PEACE(single text) PEACE(full)

46.01±0.00 56.23±0.00 47.10±0.00 49.20±3.00 59.39±0.00

66.37±1.80 77.74±0.33 77.69±0.81 74.30±1.50 79.03±1.00

66.86±3.89 89.63±0.86 90.45±0.35 88.70±1.10 90.89±0.96

By selecting these baselines, we cover the spectrum of generative, contrastive, and knowledge-guided architectures, ensuring a comprehensive validation of our proposed LSBC and CAF modules. Evaluation Metrics. We adopt AUC as the primary metric. AUC provides a threshold-independent measure of discriminative capacity that is robust against the class imbalance inherent in medical diagnosis. Comprehensive hyperparameter configurations and hardware protocols are detailed in Appendix C. 4.3

Results and Analyses

We consolidate the cross-domain evaluation and clinical dimension analysis in Table 1. Pediatric Domain Adaptation on ZZU-pECG. As presented in Table 1, PEACE shows strong ability to reduce adult-to-pediatric distribution shift in pediatric ECG classification. In the data-constrained 50-shot regime, PEACE achieves 79.03% AUC, yielding an absolute gain of 8.38% over the strongest multimodal baseline, MERL. While the generative ST-MEM model and the contrastive MERL model exhibit monotonic scaling with increasing samples, KED’s performance degrades to 50.66% in the sparse 50-shot setting. This pattern is consistent with optimization instability or negative transfer in standard knowledge distillation when fine-tuning data is insufficient to bridge the pediatric-specific morphological gap. In contrast, PEACE scales consistently from 59.39% to 79.03% to 90.89%, suggesting that LSBC helps fine-tuning refine the learned representation space instead of disrupting it. Clinical Dimension Proficiency. Fine-grained evaluation across the three diagnostic axes introduced in Section 4.1 provides deeper insight into the framework’s specialized proficiency. As presented in Table 1, PEACE demonstrates superior performance on the Morphology diagnostic axis with 91.15% AUC. This result is particularly significant because pediatric ECGs are characterized by maturational conduction variants, including high-prevalence TAB_ patterns that account for 34.2% in ZZU-pECG as discussed in Section 4.1, and these variants typically trigger false positives in adult-trained systems. Furthermore, the model achieves 91.68% in Rhythm and 87.49% in ST-T repolarization. Consistent performance across all evaluation axes supports the efficacy of our tri-axial semantic alignment. By decoupling these clinical primitives, PEACE isolates specific pathological markers, such as repolarization changes or conduction delays, without interference from dominant, age-related physiological artifacts. The match between the tri-axial semantic descriptor encoder design and final diagnostic outcomes suggests that the framework captures robust cross-population electrophysiological features. External Validation on PTB-XL. PTB-XL is used as an auxiliary external validation set on the shared label space. Since several labels are absent or extremely rare in PTB-XL, as listed in Table 7, this result should be interpreted as supportive evidence of external transferability rather than comprehensive cross-population validation. Under this setting, PEACE achieves 96.65% AUC, exceeding MERL by 4.16% and ST-MEM by 7.43%. 4.4

Ablation Studies

Table 2 summarizes both module ablations and semantic-supervision controls. The No-text condition removes the semantic descriptor branch from within the PEACE training framework. This setting is not equivalent to an independently optimized ECG-only baseline; rather, it isolates the contribution of structured semantic supervision within our framework. The substantial gap under 8

full fine-tuning (66.86% vs. 90.89%) reflects that the ECG encoder in PEACE is co-trained with the semantic alignment objective, and removing this supervision mid-framework degrades performance significantly. This result further motivates the design of structured semantic supervision as an integral component rather than an optional add-on. The single-text variant also lags behind the tri-axial design, supporting the benefit of decomposing clinical semantics into rhythm, morphology, and ST-T repolarization-related descriptions. Removing LSBC reduces performance across all regimes, indicating that label-specific bidirectional alignment remains important for robust transfer from adult to pediatric ECGs. Finally, w/o CAF remains competitive under full fine-tuning but degrades in zero-shot and few-shot regimes. This pattern indicates that CAF mainly benefits low-label transfer regimes; under full fine-tuning, the performance gap narrows substantially. Baseline Stability under Few-shot Adaptation. We further examined whether KED’s 50-shot degradation was caused by insufficient tuning. We evaluated multiple settings: frozen no-update baseline (0.578 AUC, close to zero-shot), 20-epoch fine-tuning (0.507), and 100-shot fine-tuning (0.494). The degradation persisted and did not recover with more data, which is consistent with negative transfer under severe cross-population shift. 4.5

Interpretability Analyses

To qualitatively examine that PEACE’s diagnostic decisions are grounded in physiologically meaningful patterns, we employ Grad-CAM++ Chattopadhay et al. [2018] to visualize the model’s attention across 12-lead ECG waveforms. Figure 6 illustrates the attention heatmap for a representative case of IRBBB. Across multiple consecutive cardiac cycles, PEACE exhibits high temporal consistency, with saliency peaks precisely aligned with the QRS complexes. Model inference appears to be driven by ventricular depolarization morphology rather than stochastic baseline fluctuations or environmental noise. Critically, the most intense and structured activations are concentrated in the right precordial leads V1 and V2, which are the primary diagnostic windows for IRBBB. The highlighted regions coincide with the terminal R′ wave and the characteristic rSR′ morphology, a hallmark electrophysiological indicator of right bundle branch conduction delay. In contrast, the attention mapped to limb leads and lateral precordial leads is notably more diffused and lower in magnitude. This lead-specific selectivity shows that the LQN and LSBC objective guide the model toward diagnostically relevant leads for each pathology. By aligning ECG features with fine-grained clinical semantics, PEACE’s attention patterns remain consistent with standard cardiological diagnostic criteria and focus on relevant wave segments. For readability, the representative IRBBB heatmap and all category-wise visualizations are placed in Appendix F. On the AUC–F1 divergence. Under full fine-tuning, ST-MEM achieves higher macro-F1 at 0.8885 than PEACE at 0.5288, while AUC rankings are reversed, with PEACE at 90.89% and ST-MEM at 81.55%. This divergence reflects different operating points under severe class imbalance: ST-MEM’s high F1 is accompanied by low specificity at 0.4544 and low overall accuracy at 0.3585, indicating a recall-heavy decision boundary, whereas PEACE attains higher specificity at 0.9315 and accuracy at 0.9175. We therefore retain AUC as the primary threshold-independent metric, while reporting F1 explicitly for transparency.

5

Discussion

5.1

Cross-Population Findings and Practical Scope

The multi-regime evaluation on the ZZU pECG dataset, detailed in Table 3, shows consistent performance gains for PEACE across data-availability settings. The framework achieves a macro average AUC of 90.89% by combining label-specific alignment with structured clinical semantic factorization. Without pediatric training updates, PEACE reaches a zero-shot AUC of 59.39%, outperforming KED and MERL. As supervision increases to the 50-shot regime, AUC rises to 79.03%, consistent with CAF helping stabilize adult-to-pediatric transfer under limited target supervision. CAF mainly benefits low-label transfer regimes; under full fine-tuning, the performance gap with w/o CAF narrows substantially. The weakest zero-shot categories, NORM and LVH, are also those most affected by pediatric developmental variation in heart rate and voltage criteria. This indicates that PEACE should not 9

be interpreted as a reliable zero-shot pediatric diagnostic model. Its main value is providing a transferable initialization that becomes effective after limited pediatric supervision. A primary challenge in pediatric interpretation is differentiating benign developmental variants from pathological conduction defects. IRBBB serves as a critical case, frequently manifesting as a normal developmental variant in children due to delayed cardiac maturation. While zero-shot performance on IRBBB starts at 83.82%, it scales to 95.52% under full fine-tuning, outperforming MERL by 14.52%. This performance gap across regimes indicates that label-specific alignment may help separate age-related morphological nuances, whereas generic feature fusion tends to conflate pediatric physiological maturation with adult pathological patterns. Similarly, pediatric ventricular hypertrophy often causes false positives in adult-centric models due to natural right ventricular dominance. PEACE improves LVH detection as it transitions from zero-shot to full fine-tuning, ultimately exceeding KED by 11.85%. This progression indicates that per-label feature alignment may help recalibrate voltage thresholds to age-appropriate clinical norms. The expressive capacity of the framework is further evidenced by its sensitivity to rare life-threatening arrhythmias within the long-tailed pediatric distribution. For LQTS, PEACE improves from a 51.61% zero-shot AUC to 89.55% under full fine-tuning, a 19.18% lead over KED. These results suggest that LSBC helps prevent rare classes from collapsing into dominant clusters by preserving a labelconditioned feature space throughout the learning curriculum. Crucially, this high sensitivity for rare conditions is achieved without compromising performance on the majority NORM class, which reaches 91.68% AUC. Balanced performance across zero-shot, 50-shot, and full fine-tuning regimes suggests potential decision-support relevance, pending clinical validation, in retrospective pediatric ECG classification under data scarcity. The adult-to-pediatric transfer gap is not driven by a single age mechanism but by interacting developmental processes. Neonatal right-ventricular dominance, age-dependent heart-rate decrease, and maturation-related interval changes jointly alter morphology and rhythm statistics over different timescales. Therefore, pediatric-to-adult discrepancy is heterogeneous both across and within age groups. PEACE is designed to bridge such cross-population differences through structured semantic supervision and gated alignment, rather than assuming one uniform causal factor. 5.2

Limitations

First, diagnostic descriptors are generated by Gemini conditioned on labels. Although our controls show that label identity alone does not explain gains, we cannot fully exclude label-correlated stylistic bias in synthetic text. The generated descriptors are label-level and age-agnostic; some thresholds, such as the normal sinus rhythm range, reflect adult-oriented ECG conventions and may be suboptimal for pediatric age bands. Second, PEACE currently performs static label alignment without explicit developmental stage modeling, which may limit robustness in rapidly changing neonatal physiology. Third, despite strong quantitative performance, PEACE has not yet been prospectively validated in real clinical workflows with expert-in-the-loop assessment. Fourth, age-aware continuous transfer modeling remains future work. 5.3

Interpretability and Unresolved Challenges

As visualized via Grad-CAM++ (Figure. 6), PEACE exhibits localized attention on clinically relevant waveform segments, specifically the QRS complex morphology in the right precordial leads (V1-V3) for IRBBB. This spatial specificity is consistent with standard cardiological diagnostic criteria (e.g., prolonged QRS duration in V1-V3 for IRBBB) rather than spurious signal artifacts. This level of interpretability likely comes from CAF, which anchors waveform features to clinical semantics and reinforces structured clinical semantic factorization between physiological and pathological patterns. Despite strong performance, PEACE exhibits performance lags in TAB_ and RAO/RAE, which reveals the intrinsic challenges of neonatal ECG interpretation. The hemodynamic transition from fetal to postnatal circulation induces transient physiological right-heart dominance in neonates, producing ECG signatures that are confounded with structural pathology. These two categories lack clear orthogonal boundaries in neonatal data, which limits the effectiveness of purely data-driven alignment. This bottleneck suggests that future work must incorporate developmental stage-aware clinical priors to model the dynamic cardiac maturation process, rather than relying on static label alignment. Furthermore, current saliency maps only provide spatial interpretability for single-label 10

Table 3: Per-class performance comparison of four frameworks on ZZU-pECG across three evaluation regimes. "FT" means Fine-Tuning. Best performance in each regime is highlighted in bold with a gray background . Zero-shot

Label

50-shot

Full FT

PEACE

KED

MERL

ST-MEM

PEACE

KED

MERL

ST-MEM

PEACE

KED

MERL

ST-MEM

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_

67.86 83.82 69.60 66.65 51.61 35.93 60.53 46.16 59.67 55.51 58.45 56.90

47.14 56.89 46.40 61.02 53.48 48.08 66.65 52.90 75.46 73.03 52.43 59.91

87.42 58.25 66.01 44.31 64.71 37.97 64.04 48.83 56.80 73.55 51.77 49.28

48.86 49.18 56.68 50.79 51.34 50.32 53.60 47.32 46.64 49.32 55.67 51.22

97.34 85.42 88.75 71.03 72.72 81.30 73.69 80.01 82.51 84.68 74.59 56.27

55.89 60.18 56.76 51.28 55.27 37.95 41.97 53.51 43.35 47.49 49.62 54.66

94.68 65.61 76.11 63.58 60.87 71.70 56.48 73.08 76.75 82.68 68.13 58.18

89.59 60.67 82.04 50.00 67.61 59.29 56.64 70.79 56.59 69.88 74.08 55.77

98.20 95.52 98.50 88.05 89.55 86.92 85.32 91.68 88.24 97.17 87.49 84.05

94.46 87.12 87.14 78.54 70.37 75.07 74.73 80.90 84.13 91.44 74.21 68.35

98.21 81.00 89.44 69.24 77.82 74.48 73.61 82.09 84.89 91.73 78.85 70.23

97.08 82.38 94.36 67.97 81.80 71.89 75.93 80.42 82.00 90.79 83.68 70.33

Avg.

59.39

57.77

58.58

50.91

79.03

50.66

70.65

66.08

90.89

80.54

80.97

81.55

Note: All results are presented as AUC %. Best results in each evaluation regime are highlighted with gray backgrounds. CRBBB: complete right bundle branch block; IRBBB: incomplete RBBB; LAFB: left anterior fascicular block; LAO/LAE: left atrial enlargement; LQTS: long QT syndrome; LVH: left ventricular hypertrophy; LVOLT: low QRS voltage; NORM: normal ECG; RAO/RAE: right atrial enlargement; RVH: right ventricular hypertrophy; STTC: ST/T changes; TAB_: T-wave abnormality.

diagnoses and lack causal explanations for multi-label comorbidities. Future research will explore structural constraints (e.g., temporal t-step relationships between cardiac cycles) to emulate the longitudinal reasoning process required for clinical pediatric ECG interpretation.

6

Conclusion

In this work, we presented PEACE, a structured multimodal framework for adult-to-pediatric ECG transfer grounded in structured clinical semantic factorization. By integrating tri-axial semantic decomposition, label-query alignment, and curriculum-adaptive fusion, our approach bridges crosspopulation mismatch and achieves 90.89% AUC on ZZU-pECG. Our findings indicate that structured semantic supervision can improve transfer robustness under severe pediatric label scarcity while preserving clinically meaningful attention behavior. Future work will focus on developmental-stageaware modeling, prospective clinical validation, and tighter integration between synthetic and real report supervision.

Impact Statement PEACE is a retrospective computational framework for studying data-efficient pediatric ECG classification. It is not intended for autonomous diagnosis or direct clinical deployment. False positives may increase unnecessary follow-up examinations and caregiver burden, while false negatives may delay intervention in high-risk cases; therefore, PEACE should be used only as decision support with human-in-the-loop review, confidence calibration, and conservative escalation protocols. Dataset bias is another key concern: ZZU-pECG is a Chinese pediatric cohort, and model behavior may shift across regions, acquisition systems, and demographic subgroups. Before real-world use, subgroupstratified evaluation (including age bands and site-specific protocols) is required to assess fairness and avoid systematic underperformance in underrepresented populations. We also emphasize strict data governance and auditability, including transparent reporting of failure cases and post-deployment monitoring.

References Taymaz Akan, Sait Alp, and Mohammad Alfrad Nobel Bhuiyan. Ecgformer: Leveraging transformer for ecg heartbeat arrhythmia classification. In 2023 International Conference on Computational Science and Computational Intelligence (CSCI), pages 1412–1417. IEEE, 2023.

11

Emily Alsentzer, John Murphy, William Boag, Wei-Hung Weng, Di Jindi, Tristan Naumann, and Matthew McDermott. Publicly available clinical bert embeddings. In Proceedings of the 2nd clinical natural language processing workshop, pages 72–78, 2019. Pedro Brugada and Josep Brugada. Right bundle branch block, persistent st segment elevation and sudden cardiac death: a distinct clinical and electrocardiographic syndrome: a multicenter report. Journal of the American College of Cardiology, 20(6):1391–1396, 1992. Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Gradcam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 839–847. IEEE, 2018. Aiyun Chen, Fei Wang, Wenhan Liu, Sheng Chang, Hao Wang, Jin He, and Qijun Huang. Multiinformation fusion neural networks for arrhythmia automatic detection. Computer methods and programs in biomedicine, 193:105479, 2020. Jintai Chen, Shuai Huang, Ying Zhang, Qing Chang, Yixiao Zhang, Dantong Li, Jia Qiu, Lianting Hu, Xiaoting Peng, Yunmei Du, et al. Congenital heart disease detection by pediatric electrocardiogram based deep learning integrated with human concepts. Nature Communications, 15(1):976, 2024. Tsai-Min Chen, Yuan-Hong Tsai, Huan-Hsin Tseng, Kai-Chun Liu, Jhih-Yu Chen, Chih-Han Huang, Guo-Yuan Li, Chun-Yen Shen, and Yu Tsao. Srecg: Ecg signal super-resolution framework for portable/wearable devices in cardiac arrhythmias classification. IEEE Transactions on Consumer Electronics, 69(3):250–260, 2023. David F Dickinson. The normal ecg in childhood and adolescence. Heart, 91(12):1626–1630, 2005. Wei Fan, Yujuan Si, Meiqi Sun, Lin Zhou, Weiyi Yang, Adi Alhudhaif, and Fayadh Alenezi. Classspecific weighted broad learning system-based domain adaptation for patient-specific ecg classification. Expert Systems with Applications, 273:126824, 2025. Vadim Gliner, Idan Levy, Kenta Tsutsui, Moshe Rav Acha, Jorge Schliamser, Assaf Schuster, and Yael Yaniv. Clinically meaningful interpretability of an ai model for ecg classification. NPJ Digital Medicine, 8(1):109, 2025. Brian Gow, Tom Pollard, Larry A Nathanson, Alistair Johnson, Benjamin Moody, Chrystinne Fernandes, Nathaniel Greenbaum, Jonathan W Waks, Parastou Eslami, Tanner Carbonati, et al. Mimic-iv-ecg: Diagnostic electrocardiogram matched subset. Type: dataset, 6:13–14, 2023. Chuang Han, Wenge Que, Zhizhong Wang, Songwei Wang, Yanting Li, and Li Shi. A review on intelligent auxiliary diagnosis methods based on electrocardiograms for myocardial infarction. Sheng wu yi xue Gong Cheng xue za zhi= Journal of Biomedical Engineering= Shengwu Yixue Gongchengxue Zazhi, 40(5):1019–1026, 2023. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. Ziyang He, Yufei Chen, Shuaiying Yuan, Jianhui Zhao, Zhiyong Yuan, Kemal Polat, Adi Alhudhaif, Fayadh Alenezi, and Arwa Hamid. A novel unsupervised domain adaptation framework based on graph convolutional network and multi-level feature alignment for inter-subject ecg classification. Expert Systems with Applications, 221:119711, 2023. Zhenyu Hou, Yukuo Cen, Ziding Liu, Dongxue Wu, Baoyan Wang, Xuanhe Li, Lei Hong, and Jie Tang. Mtdiag: an effective multi-task framework for automatic diagnosis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 14241–14248, 2023. Serkan Kiranyaz, Turker Ince, and Moncef Gabbouj. Real-time patient-specific ecg classification by 1-d convolutional neural networks. IEEE transactions on biomedical engineering, 63(3):664–675, 2015. Sravan Kumar Lalam, Hari Krishna Kunderu, Shayan Ghosh, Harish Kumar, Samir Awasthi, Ashim Prasad, Francisco Lopez-Jimenez, Zachi I Attia, Samuel Asirvatham, Paul Friedman, et al. Ecg representation learning with multi-modal ehr data. Transactions on Machine Learning Research, 2023. 12

David M Leone, Donnchadh O’Sullivan, and Katia Bravo-Jaimes. Artificial intelligence in pediatric electrocardiography: A comprehensive review. Children, 12(1):25, 2024. Jun Li, Aaron D Aguirre, Valdery Moura Junior, Jiarui Jin, Che Liu, Lanhai Zhong, Chenxi Sun, Gari Clifford, M Brandon Westover, and Shenda Hong. An electrocardiogram foundation model built on over 10 million recordings. NEJM AI, 2(7):AIoa2401033, 2025. Rui Li and Jing Gao. Multi-modal contrastive learning for healthcare data analytics. In 2022 IEEE 10th International Conference on Healthcare Informatics (ICHI), pages 120–127. IEEE, 2022. Jian Liang, Dapeng Hu, and Jiashi Feng. Domain adaptation with auxiliary target domain-oriented classifier. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16632–16642, 2021. Che Liu, Zhongwei Wan, Cheng Ouyang, Anand Shah, Wenjia Bai, and Rossella Arcucci. Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement. In Forty-first International Conference on Machine Learning, 2024. Che Liu, Cheng Ouyang, Zhongwei Wan, Haozhe Wang, Wenjia Bai, and Rossella Arcucci. Knowledge-enhanced multimodal ECG representation learning with arbitrary-lead inputs. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 7298–7316, 2025. Chang Lu, Chandan Reddy, Ping Wang, and Yue Ning. Towards semi-structured automatic icd coding via tree-based contrastive learning. In Advances in Neural Information Processing Systems, volume 36, pages 68300–68315. Curran Associates, Inc., 2023. Jay W Mason, E William Hancock, and Leonard S Gettes. 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):1128–1135, 2007. Yeongyeon Na, Minje Park, Yunwon Tae, and Sunghoon Joo. Guiding masked representation learning to capture spatio-temporal relationship of electrocardiogram. In International Conference on Learning Representations, 2024. Antônio H Ribeiro, Manoel Horta Ribeiro, Gabriela MM Paixão, Derick M Oliveira, Paulo R Gomes, Jéssica A Canazart, Milton PS Ferreira, Carl R Andersson, Peter W Macfarlane, Wagner Meira Jr, et al. Automatic diagnosis of the 12-lead ecg using a deep neural network. Nature communications, 11(1):1760, 2020. Nils Strodthoff, Patrick Wagner, Tobias Schaeffter, and Wojciech Samek. Deep learning for ecg analysis: Benchmarks and insights from ptb-xl. IEEE journal of biomedical and health informatics, 25(5):1519–1528, 2020. Borys Surawicz, Rory Childers, Barbara J Deal, and Leonard S Gettes. 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):976–981, 2009. Jian Tan, Haoyi Fan, Jiawei Luo, Yanjie Zhou, Ning Wang, Xizheng Wang, Guizhi Liu, Chengyu Liu, and Zongmin Wang. A pediatric ecg database with disease diagnosis covering 11643 children. Scientific Data, 12(1):867, 2025. Jialu Tang, Hung Manh Pham, Ignace De Lathauwer, Henk S Schipper, Yuan Lu, Dong Ma, and Aaqib Saeed. Interpretable multimodal zero shot ecg diagnosis via structured clinical knowledge alignment. npj Cardiovascular Health, 3(1):1, 2026. Yuanyuan Tian, Zhiyuan Li, Yanrui Jin, Mengxiao Wang, Xiaoyang Wei, Liqun Zhao, Yunqing Liu, Jinlei Liu, and Chengliang Liu. Foundation model of ecg diagnosis: Diagnostics and explanations of any form and rhythm on ecg. Cell Reports Medicine, 5(12), 2024. 13

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffter. Ptb-xl, a large publicly available electrocardiography dataset. Scientific data, 7(1):1–15, 2020. Ning Wang, Haiyan Wang, Jian Tan, Panpan Feng, Shihua Li, Zongmin Wang, and Bing Zhou. Ecg-text multi-modal learning for zero-shot detection via time-frequency alignment and medical prompt learning. Expert Systems with Applications, page 131064, 2025. Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing, volume 2022, page 3876, 2022. Jie Xiong, Yu Li, and Xi Niu. Big: A bidirectional group-wise contrastive learning method for multi-label text classification. Expert Systems with Applications, page 127977, 2025. Mingying Xu, Kui Peng, Jie Liu, Qing Zhang, Linqi Song, and Yinqiao Li. Multimodal named entity recognition based on topic prompt and multi-curriculum denoising. Information Fusion, page 103405, 2025. Kai Yang, Massimo Hong, Jiahuan Zhang, Yizhen Luo, Suyuan Zhao, Ou Zhang, Xiaomao Yu, Jiawen Zhou, Liuqing Yang, Ping Zhang, et al. Ecg-lm: Understanding electrocardiogram with a large language model. Health Data Science, 5:0221, 2025. Özal Yildirim. A novel wavelet sequence based on deep bidirectional lstm network model for ecg signal classification. Computers in biology and medicine, 96:189–202, 2018. Xue Zhou, Tianhui Li, Hiromasa Hayama, Keijiro Nakamura, Shing-Hong Liu, Wenxi Chen, and Xin Zhu. Diagnosis of cardiac conditions from 12-lead electrocardiogram through natural language supervision. npj Digital Medicine, 8(1):697, 2025. Yanna Zhu, Pengfei Guo, Zhiyong Zou, Xiuhong Li, Muqing Cao, Jun Ma, and Jin Jing. Status of cardiovascular health in chinese children and adolescents: a cross-sectional study in china. JACC: Asia, 2(1):87–100, 2022.

14

A

Dataset and Label Mapping

A.1

Supplementary Figures

Figure 3: Category-wise performance evaluation of PEACE on ZZU-pECG.

Figure 4: Visual comparison of IRBBB morphologies across adult and pediatric populations. The Lead II waveform segments highlight the significant rhythm and interval discrepancies between age groups. Figure 4 illustrates 5-second Lead II segments of IRBBB from three benchmark datasets, showcasing the pediatric-adult domain shift even within the same pathological category: • Top (MIMIC-IV, Adult): Displays IRBBB in an adult context with standard resting rhythm and characteristic late ventricular activation. • Middle (ZZU-pECG, Pediatric): Exhibits physiological tachycardia typical of pediatric patients, characterized by markedly shorter R-R intervals compared to adult counterparts. • Bottom (PTB-XL, Adult): Shows adult-typical IRBBB morphology, reinforcing the distinction in heart rate and interval parameters between adult and developmental stages. These physiological discrepancies, particularly the accelerated rhythm in pediatric signals, validate the challenge of direct knowledge transfer and justify our domain-adaptive cross-modal approach.

15

A.2

Dataset Preprocessing Protocols

Note on Sample Counts: The total sample counts reported in Table 7 may exceed the unique record counts in each dataset due to the multi-label nature of the diagnosis, where a single record can be associated with multiple pathological categories. Table 4: MIMIC-IV Label Mapping Original Label

Mapping Target

Full Name

normal ECG normal ECG except for rate left anterior fascicular block incomplete right bundle branch block

NORM NORM LAFB IRBBB

right bundle branch block

CRBBB

generalized low QRS voltages low QRS voltages in limb leads low QRS voltages in precordial leads low voltage in the extremity leads low voltage in the precordial leads left atrial enlargement right atrial abnormality

LVOLT LVOLT LVOLT LVOLT LVOLT LAO/LAE RAO/RAE

left ventricular hypertrophy right ventricular hypertrophy non-specific ST segment junctional depression non-specific lateral ST segment changes T wave abnormality non-specific anterior T wave changes non-specific anterolateral T wave changes non-specific anteroseptal T wave changes non-specific extensive T wave changes non-specific inferior T wave changes non-specific inferior and lateral T wave changes non-specific inferior/lateral T wave changes non-specific lateral T wave changes non-specific septal T wave changes prolonged QT interval

LVH RVH STTC STTC TAB_ TAB_ TAB_ TAB_ TAB_ TAB_ TAB_ TAB_ TAB_ TAB_ LQTS

normal ECG normal ECG left anterior fascicular block incomplete right bundle branch block complete right bundle branch block low QRS voltage low QRS voltage low QRS voltage low QRS voltage low QRS voltage left atrial enlargement/overload right atrial enlargement/overload left ventricular hypertrophy right ventricular hypertrophy ST/T changes ST/T changes T wave abnormality T wave abnormality T wave abnormality T wave abnormality T wave abnormality T wave abnormality T wave abnormality T wave abnormality T wave abnormality T wave abnormality long QT syndrome

A.2.1

MIMIC-IV Preprocessing

Original MIMIC-IV labels were provided as unstructured clinical diagnostic descriptions. We applied a four-stage pipeline to standardize the records: 1. Temporal Unification: ECG records (3–10s) were resampled to 500Hz and uniformly adjusted to 10 seconds (5,000 samples) via truncation or zero-padding. 2. Ontology Mapping: We mapped raw clinical strings to 12 target categories (e.g., mapping "left ventricular hypertrophy" to LVH) to ensure cross-dataset label consistency. 3. Signal Quality Filtering: After merging machine_measurements and label records via subject_id, we excluded records with missing annotations or excessive noise, resulting in 483,051 valid samples. 4. Splitting: Using a fixed seed (42), we partitioned the data into training, validation, and test sets (8:1:1 ratio). A.2.2

ZZU-pECG Preprocessing

Labels in ZZU-pECG follow the AHA code format. Our workflow addresses pediatric-specific signal characteristics:

16

1. Label Conversion: We mapped 14 AHA codes (e.g., I101, L147) to our 12 core disease labels (e.g., LAFB, TAB_). Specifically, codes A1/A2 were assigned to NORM, while L145/L146 were mapped to STTC. 2. Signal Refinement: Signals were standardized to 10s at 500Hz. We applied a filtering suite consisting of high-pass (0.5Hz), low-pass (100Hz), and 50Hz notch filters to remove baseline wander and power-line interference. 3. Amplitude Calibration: To mitigate gain variances across acquisition hardware, we scaled pediatric signal amplitudes relative to the peak-to-peak statistics of the MIMIC-IV dataset, ensuring a consistent input distribution. 4. Stratified Splitting: From 12,334 original records, 7,593 valid multi-label samples were selected. We employed multi-label stratified sampling to maintain class balance across subsets (8:1:1 ratio). Table 5: ZZU-pECG Label Mapping

A.2.3

Original Label

Mapping Target

Full Name

A1 A2 I101 I105 I106 J125 K140 K141 K142 K143 L145 L146 L147 L148

NORM NORM LAFB IRBBB CRBBB LVOLT LAO/LAE RAO/RAE LVH RVH STTC STTC TAB_ LQTS

normal ECG normal ECG left anterior fascicular block incomplete right bundle branch block complete right bundle branch block low QRS voltage left atrial enlargement/overload right atrial enlargement/overload left ventricular hypertrophy right ventricular hypertrophy ST/T changes ST/T changes T wave abnormality long QT syndrome

PTB-XL Preprocessing

PTB-XL labels utilize the SCP-ECG standard. Preprocessing focused on diagnostic ontology alignment: 1. Ontology Refinement: We mapped SCP codes to our 12-label system based on the subdiagnostic hierarchy. Notably, T-wave abnormalities (e.g., NDT) were subsumed under the STTC category according to PTB-XL’s diagnostic logic, resulting in a zero count for the standalone TAB_ label in this dataset. 2. Standardization: 10-second signals (500Hz) were processed using a StandardScaler fitted on training set statistics to normalize the signal mean and variance. 3. Data Splitting: We selected 16,363 valid samples containing target labels and split them into training, validation, and test sets at an 8:1:1 ratio. After preprocessing, the number of samples of each category used in the final experiment for the three datasets is shown in the table below.

17

Table 6: PTB-XL Label Mapping Original Label

Mapping Target

Full Name

NDT DIG LNGQT ANEUR EL NORM LAFB IRBBB CRBBB LAO/LAE RAO/RAE LVH RVH LVOLT LQTS TAB_

STTC STTC STTC STTC STTC NORM LAFB IRBBB CRBBB LAO/LAE RAO/RAE LVH RVH LVOLT LQTS TAB_

ST changes ST changes ST changes ST changes ST changes normal ECG left anterior fascicular block incomplete right bundle branch block complete right bundle branch block left atrial enlargement/overload right atrial enlargement/overload left ventricular hypertrophy right ventricular hypertrophy low QRS voltage long QT syndrome T wave abnormalities

Table 7: Distribution of Abnormality Labels across MIMIC-IV, ZZU-pECG, and PTB-XL. Label

MIMIC-IV (Counts)

ZZU-pECG (Counts)

PTB-XL (Counts)

NORM TAB_ LVOLT LVH CRBBB LAFB LQTS LAO/LAE STTC IRBBB RVH RAO/RAE

153,895 102,610 72,634 69,067 56,537 47,514 36,799 15,834 14,197 7,340 6,468 3,827

2,380 2,544 277 348 290 323 427 184 1,195 336 693 201

9,514 0 0 2,132 541 1,623 0 426 2,239 1,118 126 99

Total

586,722

9,198

17,818

18

B

Label-derived Clinical Semantic Descriptors

To obtain concise, label-specific diagnostic rationales, we employed Gemini 3 to generate short expert-like ECG interpretations. For each diagnostic label, we prompted the model with the following template: I want you to play the role of a professional Electrocardiologist, and I need you to teach me how to diagnose <LABEL> from 12-lead ECG, such as what leads or what features to focus on, etc. Your answer must be less than 50 words. The full collection of generated diagnostic reports is as follows: These descriptors are label-level and age-agnostic. They were used as auxiliary semantic supervision rather than pediatric diagnostic guidelines; age-specific thresholds are not encoded in the current descriptor set. The third field was originally generated under an “ischemia” prompt but is interpreted in this paper as ST-T repolarization-related information, because these descriptors primarily describe ST-segment and T-wave patterns rather than clinically adjudicated myocardial ischemia. • CRBBB: rhythm: Regular sinus rhythm showing stable atrioventricular conduction; morphology: QRS ≥ 120 ms with characteristic rsR’ or rSR’ in right precordial leads V1-V2 and slurred terminal S wave in I and V6; ST-T repolarization: Secondary T-wave inversion in V1-V3 without territorial ST-segment elevation or reciprocal changes. • IRBBB: rhythm: Normal sinus rhythm with consistent R-R intervals; morphology: QRS duration 100-119 ms exhibiting rSr’ or rsR’ in V1-V2 and narrow terminal S in I and V6; ST-T repolarization: Absence of acute ST-T changes; no pathologic Q waves or ST-segment deviation in contiguous leads. • LAFB: rhythm: Sinus rhythm with normal heart rate; morphology: Marked left axis deviation −45 to −90 degrees, qR pattern in lateral leads I and aVL, and rS pattern in inferior leads II, III, aVF; ST-T repolarization: Negative for acute ST-segment deviation or localized T-wave inversion. • LAO/LAE: rhythm: Sinus rhythm at a regular rate; morphology: P mitrale with notched P wave in lead II or terminal negative P component in V1 ≥ 1 mm deep and ≥ 40 ms duration; ST-T repolarization: ST-segments are isoelectric; no pathologic Q waves suggesting prior injury patterns. • LQTS: rhythm: Sinus rhythm with prolonged ventricular repolarization; morphology: Prolonged QTc interval > 470 ms measured in lead II or V5-V6 with normal QRS duration; ST-T repolarization: Absence of acute ST-T abnormalities; T-waves may be broad but lack a specific territorial pattern. • LVH: rhythm: Regular sinus rhythm; morphology: Increased QRS voltage (S V1 + R V5-V6 > 35 mm) with left axis deviation; ST-T repolarization: Asymmetric downsloping ST-segment depression and T-wave inversion in lateral leads I, aVL, V5-V6 representing a left ventricular strain pattern. • LVOLT: rhythm: Sinus rhythm with attenuated signal amplitude; morphology: QRS voltage < 5 mm in all limb leads and < 10 mm in all precordial leads; ST-T repolarization: No evidence of territorial ST-elevation or depression; T-waves are concordant but low in amplitude. • NORM: rhythm: Normal sinus rhythm 60-100 bpm with consistent P-P intervals; morphology: Normal P wave, PR interval, QRS duration, and QRS axis; ST-T repolarization: No diagnostic ST-segment elevation, depression, or T-wave inversion; no pathologic Q waves in any lead. • RAO/RAE: rhythm: Sinus rhythm with prominent atrial signals; morphology: Tall peaked P wave > 2.5 mm in inferior leads II, III, aVF or initial positive P in V1 > 1.5 mm; ST-T repolarization: Negative for acute myocardial injury or primary repolarization abnormalities. • RVH: rhythm: Sinus rhythm with rightward QRS vector; morphology: Right axis deviation with dominant R wave in V1 (R/S > 1) and deep S wave in lateral leads V5-V6; ST-T repolarization: ST-segment depression and T-wave inversion in right precordial leads V1-V3 consistent with right ventricular strain. • STTC: rhythm: Stable sinus rhythm; morphology: Normal QRS morphology and axis; ST-T repolarization: Nonspecific ST-T abnormalities including minor ST-segment flattening or T-wave inversion without a specific lead-territory pattern or reciprocal ST-elevation. • TAB_: rhythm: Sinus rhythm with normal intervals; morphology: QRS duration and Pwave morphology are unremarkable; ST-T repolarization: Generalized T-wave flattening

19

or inversion in contiguous leads without significant ST-segment deviation, excluding acute focal injury patterns.

20

C

Implementation and Reproducibility Details

C.1

Compute Resources

All experiments in this work are conducted on a single NVIDIA RTX 4090 GPU (24 GB VRAM) with 48 GB system RAM. For our proposed model training, one training epoch takes approximately 20 minutes under the above hardware configuration. The total GPU computation time for full model training and ablation experiments is summarized based on the per-epoch runtime. All computational resources belong to our institutional laboratory server and are used strictly for academic research only. C.2

Asset Licenses and Data Access

MIMIC-IV-ECG is used under the PhysioNet credentialed access and corresponding data usage license agreement. PTB-XL and ZZU-pECG are utilized in accordance with their public dataset release terms for non-commercial academic research. The torch_ecg toolkit and BioClinicalBERT pre-trained model are employed following their official open-source licenses and research usage policies.

21

D

Additional Experimental Analyses

To determine the optimal few-shot configuration for our pediatric transfer learning task, we conducted a systematic sample efficiency analysis on the ZZU-pECG dataset. We evaluated the model’s performance (Area Under the Receiver Operating Characteristic curve, AUC) across a logarithmic scale of samples per class (N ∈ {5, 10, 20, 50, 100}). D.1

Performance Saturation Analysis

As summarized in Table 8, we identified three distinct learning phases: 1. Cold Start (N < 10): Performance gains are marginal (< 1.3%), as the model struggles to align sparse pediatric samples with the pretrained adult clinical priors. 2. Rapid Adaptation (10 ≤ N ≤ 50): A significant performance surge is observed, with AUC improving by approximately 13% (from 66.13% to 79.03%). At this stage, 50 samples appear sufficient to capture core pediatric-specific morphological features. 3. Saturation Phase (N > 50): Beyond 50-shot, performance gains hit a bottleneck (< 1%), suggesting that further increasing the annotation burden provides diminishing returns for cross-modal transfer. Based on this empirical evidence, we established the 50-shot protocol as our primary few-shot adaptation benchmark, as it achieves 99.2% of the 100-shot performance while reducing the required clinical annotation effort by half. Table 8: Impact of Sample Size on PEACE Performance (AUC %) on ZZU-pECG. Configuration AUC (Mean±Std) ∆ Gain Relative Imp. 5-shot 10-shot 20-shot 50-shot 100-shot

D.2

64.85±4.48 66.13±3.33 73.10±1.64 79.03±1.00 79.68±1.18

– +1.28 +6.97 +5.93 +0.65

– +2.0% +10.5% +8.1% +0.8%

General Training Configuration

For all experimental phases, we employ the AdamW optimizer (β1 = 0.9, β2 = 0.999) with an initial learning rate of 4 × 10−4 and a weight-decay coefficient of 1.2 × 10−3 . The learning rate follows a 5-epoch linear warm-up He et al. [2016] (starting from 1 × 10−5 ), followed by a cosine annealing schedule that decays to a minimum of 1 × 10−6 over the remaining 95 epochs. To ensure optimization stability, gradient norms are clipped at 5.0. We utilize a batch size of 1,024 and compute the Binary Cross-Entropy loss using per-class weights proportional to the inverse frequency of positive labels, clipped to the interval [1.0, 5.0] to prevent extreme gradient updates. D.3

Pre-training on MIMIC-IV

PEACE is pre-trained on the MIMIC-IV dataset for 100 epochs. The text encoder, BioClinicalBERT, has its foundational nine transformer layers frozen to preserve broad clinical semantics. Gradient updates are restricted to the top three layers, the pooler, and the cross-modal projection head. Conversely, the ResNet1D ECG encoder Strodthoff et al. [2020] and the LQN fusion module are trained from scratch to learn the initial alignment between electrical signals and clinical descriptions. D.4

Transfer Learning Regimes

We evaluate PEACE under three progressively intensive adaptation regimes to simulate realistic clinical deployment scenarios: 1. Zero-shot Evaluation Protocol. In this regime, all model parameters remain frozen at the state obtained from MIMIC-IV pre-training. This protocol strictly isolates the cross-population 22

generalization capacity of the learned representations. We optimize class-specific classification thresholds on the target validation set by identifying the values that maximize the macro F1-score. These thresholds are subsequently applied to the held-out test set to report final performance metrics. 2. Few-shot Adaptation Protocol (50-shot). The model is initialized with pre-trained weights. The ECG encoder and LQN fusion module are unfrozen, while the text encoder maintains its foundational 10 transformer layers in a frozen state to prevent catastrophic forgetting of medical priors. We randomly sample 50 labeled instances per diagnostic class (or the maximum available for rare classes). Fine-tuning proceeds for 20 epochs at a reduced learning rate of 2.5 × 10−5 . We monitor the validation macro-averaged F1-score after each epoch to update thresholds and save the optimal checkpoint. 3. Full Supervised Fine-tuning Protocol. This regime utilizes the entire labeled training portion of the target dataset. The optimization schedule and the graduated freezing strategy (first 10 transformer layers of BioClinicalBERT remain fixed) are consistent with the few-shot protocol to ensure stability during cross-domain adaptation. Training is conducted for 20 epochs at a learning rate of 1×10−4 . We store the checkpoint attaining the highest macro-average validation AUC and evaluate its performance on the final test set.

23

E

Full Performance Tables

This section provides a detailed performance breakdown of PEACE and baseline models across 12 pediatric ECG categories. We report five key metrics: AUC, Accuracy (ACC), Sensitivity, Specificity, and F1-Score. The results across Zero-shot, Few-shot, and Full fine-tuning scenarios demonstrate the robustness of PEACE in diverse clinical data regimes. E.1

Zero-shot Performance

Zero-shot transfer performance without any pediatric-specific fine-tuning is presented in Table 9. E.2

Few-shot Performance

Few-shot (50-shot per class) adaptation results, where only limited labeled pediatric samples are available, are detailed in Table 10. E.3

Full fine-tuning Performance

Full supervised fine-tuning results using the complete ZZU-pECG training set are shown in Table 11.

24

Table 9: Performance Comparison of Models on ZZU-pECG Dataset (Zero-shot, Mean ± Std) Indicator

Label

PEACE

KED

MERL

ST-MEM

AUC

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.6786±0.0000 0.8382±0.0000 0.6960±0.0000 0.6665±0.0000 0.5161±0.0000 0.3593±0.0000 0.6053±0.0000 0.4616±0.0000 0.5967±0.0000 0.5551±0.0000 0.5845±0.0000 0.5690±0.0000 0.5939±0.0000

0.4714±0.0000 0.5689±0.0000 0.4640±0.0000 0.6102±0.0000 0.5348±0.0000 0.4808±0.0000 0.6654±0.0000 0.5290±0.0000 0.7546±0.0000 0.7303±0.0000 0.5243±0.0000 0.5991±0.0000 0.5777±0.0000

0.8742±0.0000 0.5825±0.0000 0.6601±0.0000 0.4431±0.0000 0.6471±0.0000 0.3797±0.0000 0.6404±0.0000 0.4883±0.0000 0.5680±0.0000 0.7355±0.0000 0.5177±0.0000 0.4928±0.0000 0.5858±0.0000

0.4886±0.0674 0.4918±0.0531 0.5668±0.0493 0.5079±0.0147 0.5134±0.0687 0.5032±0.0883 0.5360±0.0600 0.4732±0.0473 0.4664±0.0883 0.4932±0.0631 0.5567±0.0356 0.5122±0.0244 0.5091±0.0146

ACC

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.9358±0.0000 0.9067±0.0000 0.9404±0.0000 0.6988±0.0000 0.4052±0.0000 0.2508±0.0000 0.9037±0.0000 0.3440±0.0000 0.8654±0.0000 0.2982±0.0000 0.6514±0.0000 0.4939±0.0000 0.6412±0.0000

0.5902±0.0000 0.2003±0.0000 0.8593±0.0000 0.7554±0.0000 0.1468±0.0000 0.9358±0.0000 0.8502±0.0000 0.3685±0.0000 0.7385±0.0000 0.2768±0.0000 0.1651±0.0000 0.3639±0.0000 0.5209±0.0000

0.9618±0.0000 0.9281±0.0000 0.9572±0.0000 0.9740±0.0000 0.9419±0.0000 0.9664±0.0000 0.9602±0.0000 0.6774±0.0000 0.9480±0.0000 0.9174±0.0000 0.8532±0.0000 0.6529±0.0000 0.8949±0.0000

0.2969±0.1855 0.4058±0.1842 0.4480±0.2686 0.6128±0.2295 0.5153±0.2585 0.6404±0.3571 0.7606±0.2099 0.3391±0.0294 0.7214±0.2669 0.4141±0.1659 0.3823±0.0897 0.3575±0.0077 0.4912±0.0839

Sensitivity

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.1304±0.0000 0.3667±0.0000 0.0741±0.0000 0.5000±0.0000 0.6389±0.0000 0.5882±0.0000 0.0800±0.0000 0.7915±0.0000 0.0000±0.0000 0.7885±0.0000 0.3474±0.0000 0.7797±0.0000 0.4238±0.0000

0.2174±0.0000 0.9000±0.0000 0.1111±0.0000 0.3333±0.0000 0.9444±0.0000 0.0588±0.0000 0.3600±0.0000 0.9621±0.0000 0.6000±0.0000 0.9615±0.0000 0.9789±0.0000 0.9471±0.0000 0.6146±0.0000

0.2174±0.0000 0.0667±0.0000 0.0000±0.0000 0.0000±0.0000 0.0000±0.0000 0.0000±0.0000 0.0000±0.0000 0.0000±0.0000 0.0667±0.0000 0.0769±0.0000 0.0000±0.0000 0.0000±0.0000 0.0356±0.0000

0.7217±0.2391 0.6467±0.2464 0.6593±0.3382 0.4833±0.3472 0.5111±0.3194 0.3765±0.3948 0.2400±0.2918 0.9393±0.1029 0.2533±0.2473 0.6231±0.2599 0.7811±0.1089 0.9709±0.0256 0.6005±0.2434

Specificity

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.9651±0.0000 0.9327±0.0000 0.9777±0.0000 0.7025±0.0000 0.3916±0.0000 0.2418±0.0000 0.9364±0.0000 0.1309±0.0000 0.8858±0.0000 0.2558±0.0000 0.7030±0.0000 0.3419±0.0000 0.6221±0.0000

0.6038±0.0000 0.1667±0.0000 0.8915±0.0000 0.7632±0.0000 0.1003±0.0000 0.9592±0.0000 0.8696±0.0000 0.0858±0.0000 0.7418±0.0000 0.2176±0.0000 0.0268±0.0000 0.0539±0.0000 0.4567±0.0000

0.9889±0.0000 0.9696±0.0000 0.9984±0.0000 0.9922±0.0000 0.9968±0.0000 0.9922±0.0000 0.9984±0.0000 1.0000±0.0000 0.9687±0.0000 0.9900±0.0000 0.9982±0.0000 1.0000±0.0000 0.9911±0.0000

0.2815±0.1999 0.3942±0.2043 0.4389±0.2946 0.6153±0.2401 0.5155±0.2918 0.6474±0.3771 0.7812±0.2298 0.0533±0.0920 0.7324±0.2790 0.3960±0.2012 0.3145±0.1219 0.0314±0.0201 0.4335±0.2127

F1-Score

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.1250±0.0000 0.2651±0.0000 0.0930±0.0000 0.0574±0.0000 0.1057±0.0000 0.0392±0.0000 0.0597±0.0000 0.4377±0.0000 0.0000±0.0000 0.1516±0.0000 0.2245±0.0000 0.5168±0.0000 0.1730±0.0000

0.0360±0.0000 0.0936±0.0000 0.0612±0.0000 0.0476±0.0000 0.1086±0.0000 0.0455±0.0000 0.1552±0.0000 0.4957±0.0000 0.0952±0.0000 0.1745±0.0000 0.2541±0.0000 0.5083±0.0000 0.1730±0.0000

0.2857±0.0000 0.0784±0.0000 0.0000±0.0000 0.0000±0.0000 0.0000±0.0000 0.0000±0.0000 0.0000±0.0000 0.0000±0.0000 0.0556±0.0000 0.1290±0.0000 0.0000±0.0000 0.0000±0.0000 0.0457±0.0000

0.0666±0.0116 0.0888±0.0126 0.0855±0.0101 0.0339±0.0186 0.0928±0.0376 0.0328±0.0269 0.0384±0.0373 0.4770±0.0186 0.0337±0.0169 0.1350±0.0387 0.2689±0.0121 0.5119±0.0066 0.1554±0.0036

25

Table 10: Performance Comparison of Models on ZZU-pECG Dataset (Few-shot, Mean ± Std) Indicator

Label

PEACE

KED

MERL

ST-MEM

AUC

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.9734±0.0059 0.8542±0.0209 0.8875±0.0184 0.7103±0.0197 0.7272±0.0399 0.8130±0.0358 0.7369±0.0358 0.8001±0.0183 0.8251±0.0327 0.8468±0.0567 0.7459±0.0283 0.5627±0.0160 0.7903±0.0100

0.5589±0.0417 0.6018±0.0235 0.5676±0.0265 0.5128±0.0353 0.5527±0.0308 0.3795±0.0096 0.4197±0.0198 0.5351±0.0388 0.4335±0.0181 0.4749±0.0140 0.4962±0.0305 0.5466±0.0249 0.5066±0.0103

0.9468±0.0195 0.6561±0.0318 0.7611±0.0607 0.6358±0.0542 0.6087±0.0374 0.7170±0.0606 0.5648±0.0446 0.7308±0.0164 0.7675±0.0565 0.8268±0.0379 0.6813±0.0110 0.5818±0.0109 0.7065±0.0191

0.8959±0.0300 0.6067±0.0520 0.8204±0.0467 0.5000±0.1243 0.6761±0.0272 0.5929±0.0603 0.5664±0.0619 0.7079±0.0165 0.5659±0.0529 0.6988±0.0382 0.7408±0.0197 0.5577±0.0214 0.6608±0.0221

ACC

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.9633±0.0037 0.8297±0.0384 0.8639±0.0417 0.9394±0.0161 0.8560±0.0397 0.9557±0.0141 0.8174±0.0542 0.7174±0.0122 0.9431±0.0320 0.8067±0.0656 0.7514±0.0486 0.4113±0.0105 0.8213±0.0081

0.7113±0.1615 0.6731±0.1664 0.9214±0.0503 0.9758±0.0049 0.6208±0.0632 0.9443±0.0272 0.6052±0.2921 0.3434±0.0272 0.2933±0.2578 0.4853±0.3222 0.1841±0.0463 0.4012±0.0195 0.5966±0.0570

0.9664±0.0090 0.8682±0.0422 0.9171±0.0113 0.9535±0.0313 0.8462±0.0584 0.9578±0.0115 0.8813±0.0486 0.6865±0.0221 0.9569±0.0105 0.9000±0.0064 0.7765±0.0407 0.4780±0.0370 0.8490±0.0154

0.9508±0.0141 0.8920±0.0057 0.9245±0.0216 0.9630±0.0064 0.8703±0.0257 0.9468±0.0165 0.9223±0.0227 0.6642±0.0400 0.8969±0.0740 0.8566±0.0444 0.7596±0.0543 0.5364±0.0961 0.8486±0.0205

Sensitivity

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.5130±0.1413 0.6933±0.0827 0.6741±0.1032 0.1333±0.0850 0.3833±0.1585 0.2353±0.1441 0.5040±0.0999 0.7763±0.0583 0.2533±0.1485 0.6808±0.0591 0.5389±0.0966 0.9348±0.0227 0.5267±0.1000

0.3391±0.2241 0.4067±0.2342 0.0519±0.1037 0.0000±0.0000 0.4667±0.0643 0.0000±0.0000 0.3520±0.2987 0.9791±0.0417 0.7333±0.3293 0.5346±0.3847 0.9053±0.0818 0.9295±0.0263 0.4748±0.1491

0.5913±0.1364 0.2333±0.1229 0.3111±0.0763 0.1167±0.0408 0.2278±0.1308 0.1529±0.0880 0.1520±0.1170 0.6057±0.0746 0.0933±0.0680 0.3500±0.1231 0.3284±0.0928 0.7595±0.0880 0.3268±0.0966

0.5043±0.1120 0.1600±0.0827 0.3778±0.0544 0.0667±0.0333 0.3278±0.0539 0.0706±0.0686 0.0800±0.0566 0.5962±0.1333 0.0800±0.0777 0.2769±0.1036 0.5368±0.1265 0.5542±0.2312 0.3026±0.0945

Specificity

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.9797±0.0060 0.8362±0.0438 0.8721±0.0467 0.9545±0.0179 0.8835±0.0490 0.9749±0.0180 0.8299±0.0600 0.6894±0.0364 0.9593±0.0357 0.8176±0.0720 0.7875±0.0675 0.1330±0.0269 0.8098±0.0400

0.7249±0.1749 0.6859±0.1855 0.9589±0.0569 0.9941±0.0050 0.6298±0.0703 0.9695±0.0279 0.6153±0.3156 0.0406±0.0600 0.2829±0.2711 0.4811±0.3832 0.0615±0.0667 0.1204±0.0436 0.5471±0.1384

0.9800±0.0122 0.8987±0.0493 0.9432±0.0125 0.9692±0.0324 0.8822±0.0687 0.9793±0.0136 0.9103±0.0545 0.7251±0.0547 0.9772±0.0120 0.9475±0.0132 0.8526±0.0624 0.3283±0.1027 0.8661±0.0407

0.9670±0.0150 0.9272±0.0067 0.9480±0.0205 0.9798±0.0069 0.9019±0.0283 0.9702±0.0186 0.9558±0.0252 0.6966±0.1220 0.9161±0.0776 0.9066±0.0567 0.7975±0.0823 0.5269±0.2686 0.8745±0.0607

F1-Score

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.4890±0.0687 0.2770±0.0289 0.3053±0.0683 0.0644±0.0381 0.2230±0.0512 0.1996±0.0730 0.1809±0.0308 0.6387±0.0187 0.1760±0.0528 0.3737±0.0699 0.3879±0.0444 0.5243±0.0036 0.3200±0.0457

0.0753±0.0118 0.0952±0.0224 0.0214±0.0427 0.0000±0.0000 0.1202±0.0085 0.0000±0.0000 0.0578±0.0110 0.4904±0.0012 0.0449±0.0080 0.1291±0.0232 0.2434±0.0092 0.5187±0.0020 0.1497±0.0117

0.5539±0.0559 0.1338±0.0435 0.2361±0.0463 0.0992±0.0313 0.1344±0.0317 0.1442±0.0765 0.0814±0.0347 0.5534±0.0292 0.0807±0.0495 0.3479±0.0833 0.2956±0.0277 0.5015±0.0125 0.2635±0.0435

0.4265±0.1022 0.1167±0.0572 0.3057±0.0887 0.0599±0.0313 0.2222±0.0445 0.0495±0.0459 0.0652±0.0452 0.5291±0.0294 0.0229±0.0187 0.2319±0.0295 0.3928±0.0365 0.4410±0.0469 0.2386±0.0123

26

Table 11: Performance Comparison of Models on ZZU-pECG Dataset (Full fine-tuning, Mean ± Std) Indicator

Label

PEACE

KED

MERL

ST-MEM

AUC

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.9820±0.0127 0.9552±0.0118 0.9850±0.0056 0.8805±0.0398 0.8955±0.0146 0.8692±0.0331 0.8532±0.0260 0.9168±0.0038 0.8824±0.0242 0.9717±0.0044 0.8749±0.0113 0.8405±0.0022 0.9089±0.0096

0.9446±0.0115 0.8712±0.0198 0.8714±0.0219 0.7854±0.0236 0.7037±0.0844 0.7507±0.0253 0.7473±0.0263 0.8090±0.0790 0.8413±0.0137 0.9144±0.0148 0.7421±0.1405 0.6835±0.1075 0.8054±0.0350

0.9821±0.0038 0.8100±0.0157 0.8944±0.0312 0.6924±0.0639 0.7782±0.0296 0.7448±0.0189 0.7361±0.0123 0.8209±0.0159 0.8489±0.0221 0.9173±0.0094 0.7885±0.0195 0.7023±0.0117 0.8097±0.0092

0.9708±0.0090 0.8238±0.0209 0.9436±0.0121 0.6797±0.0339 0.8180±0.0203 0.7189±0.0335 0.7593±0.0133 0.8042±0.0159 0.8200±0.0164 0.9079±0.0121 0.8368±0.0099 0.7033±0.0128 0.8155±0.0053

ACC

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.9697±0.0180 0.9232±0.0187 0.9688±0.0123 0.9740±0.0116 0.9177±0.0237 0.9737±0.0056 0.9220±0.0314 0.8343±0.0126 0.9419±0.0239 0.9401±0.0124 0.8670±0.0124 0.7777±0.0053 0.9175±0.0078

0.7954±0.1715 0.8266±0.1028 0.8572±0.0601 0.6550±0.3023 0.2917±0.1470 0.9070±0.0468 0.5887±0.1818 0.3382±0.0113 0.9517±0.0218 0.8474±0.0354 0.2128±0.0514 0.3694±0.0305 0.6368±0.0287

0.9734±0.0023 0.8590±0.0499 0.8884±0.0291 0.9642±0.0153 0.8795±0.0390 0.9486±0.0138 0.9263±0.0207 0.7450±0.0140 0.9514±0.0107 0.8945±0.0179 0.7997±0.0402 0.6034±0.0489 0.8694±0.0044

0.6145±0.0819 0.3056±0.0578 0.4884±0.0609 0.0471±0.0941 0.3323±0.0645 0.0877±0.0128 0.1269±0.0231 0.6396±0.0213 0.1238±0.0420 0.4782±0.0463 0.4884±0.0303 0.5692±0.0024 0.3585±0.0155

Sensitivity

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.7565±0.1014 0.7867±0.0748 0.8074±0.0277 0.3833±0.1716 0.6389±0.0703 0.3176±0.1211 0.4880±0.0960 0.8483±0.0393 0.3733±0.1236 0.8731±0.0432 0.5895±0.0359 0.7304±0.0445 0.6328±0.0791

0.8522±0.1471 0.6600±0.2380 0.5037±0.3268 0.6333±0.3232 0.9000±0.1345 0.2000±0.1561 0.7360±0.2111 1.0000±0.0000 0.2133±0.1655 0.7923±0.1436 0.9684±0.0437 0.9859±0.0156 0.7038±0.1587

0.5043±0.0806 0.4667±0.1445 0.6296±0.0811 0.0500±0.0667 0.3333±0.1165 0.3294±0.1091 0.1520±0.0588 0.7791±0.0824 0.1600±0.0904 0.5962±0.1053 0.5137±0.1223 0.7198±0.0976 0.4362±0.0963

0.9676±0.0117 0.9135±0.0142 0.9492±0.0137 0.9688±0.0145 0.8765±0.0438 0.9514±0.0169 0.9055±0.0328 0.7122±0.0222 0.9453±0.0318 0.8899±0.0315 0.8312±0.0303 0.6193±0.0210 0.8775±0.0072

Specificity

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.9775±0.0202 0.9298±0.0227 0.9758±0.0130 0.9850±0.0129 0.9340±0.0289 0.9912±0.0079 0.9393±0.0342 0.8275±0.0344 0.9552±0.0272 0.9458±0.0148 0.9141±0.0192 0.8028±0.0305 0.9315±0.0222

0.7933±0.1819 0.8346±0.1178 0.8724±0.0751 0.6555±0.3131 0.2563±0.1618 0.9259±0.0521 0.5828±0.1972 0.0230±0.0167 0.9690±0.0259 0.8522±0.0504 0.0844±0.0658 0.0417±0.0497 0.5743±0.1090

0.9905±0.0014 0.8779±0.0588 0.8995±0.0319 0.9813±0.0154 0.9113±0.0465 0.9651±0.0168 0.9571±0.0233 0.7287±0.0503 0.9700±0.0122 0.9203±0.0248 0.8483±0.0675 0.5415±0.1246 0.8826±0.0395

0.7217±0.1546 0.4267±0.1405 0.5704±0.0602 0.0333±0.0667 0.5167±0.0451 0.0941±0.0471 0.1920±0.1085 0.7905±0.0275 0.1867±0.1485 0.6423±0.1890 0.5537±0.0909 0.7251±0.0458 0.4544±0.0937

F1-Score

CRBBB IRBBB LAFB LAO/LAE LQTS LVH LVOLT NORM RAO/RAE RVH STTC TAB_ Overall Average

0.6643±0.1331 0.4922±0.0497 0.6905±0.0736 0.3510±0.1081 0.4691±0.0452 0.3785±0.0853 0.3391±0.0815 0.7677±0.0094 0.2343±0.0262 0.7012±0.0429 0.5633±0.0178 0.6948±0.0093 0.5288±0.0568

0.3257±0.1509 0.2816±0.0611 0.1977±0.0732 0.0892±0.0500 0.1244±0.0148 0.0841±0.0449 0.1405±0.0507 0.4937±0.0043 0.1389±0.0864 0.4561±0.0238 0.2636±0.0084 0.5207±0.0118 0.2597±0.0484

0.5679±0.0636 0.2353±0.0235 0.3256±0.0561 0.0595±0.0970 0.2328±0.0390 0.2481±0.0318 0.1352±0.0303 0.6620±0.0251 0.1243±0.0547 0.4734±0.0453 0.4254±0.0159 0.5571±0.0121 0.3372±0.0412

0.9765±0.0151 0.9369±0.0201 0.9656±0.0152 0.9863±0.0144 0.8974±0.0487 0.9743±0.0186 0.9339±0.0384 0.6749±0.0322 0.9631±0.0359 0.9113±0.0486 0.8784±0.0482 0.5630±0.0565 0.8885±0.0326

27

F

Qualitative Heatmap Visualizations

This appendix section presents Grad-CAM++ attention heatmap examples of the PEACE model across all 12 diagnostic categories. All samples are sourced from the ZZU-pECG pediatric dataset. These visualizations provide qualitative evidence that the model’s focus aligns with standard cardiological diagnostic criteria, such as the QRS complex in precordial leads for bundle branch blocks and ST-T segments for repolarization abnormalities.

Figure 5: Grad-CAM++ attention visualization for a CRBBB sample from ZZU-pECG.

Figure 6: Grad-CAM++ attention visualization for an IRBBB sample from ZZU-pECG.

28

Figure 7: Grad-CAM++ attention visualization for an LAFB sample from ZZU-pECG.

Figure 8: Grad-CAM++ attention visualization for an LAOLAE sample from ZZU-pECG.

29

Figure 9: Grad-CAM++ attention visualization for an LQTS sample from ZZU-pECG.

Figure 10: Grad-CAM++ attention visualization for an LVH sample from ZZU-pECG.

30

Figure 11: Grad-CAM++ attention visualization for an LVOLT sample from ZZU-pECG.

Figure 12: Grad-CAM++ attention visualization for a NORM sample from ZZU-pECG.

31

Figure 13: Grad-CAM++ attention visualization for an RAORAE sample from ZZU-pECG.

Figure 14: Grad-CAM++ attention visualization for an RVH sample from ZZU-pECG.

32

Figure 15: Grad-CAM++ attention visualization for an STTC sample from ZZU-pECG.

Figure 16: Grad-CAM++ attention visualization for a TAB_ sample from ZZU-pECG.

33