Conceptio › Archive › arXiv CS

arXiv CSopen access

RealICU: Do LLM Agents Understand Long-Context ICU Data? A Benchmark Beyond Behavior Imitation

2026 · arxiv_cs

arXiv CS · Papers · License: Open Access · 2026

Open Source ↗Direct PDF ↓

knowledge-representation reasoning

artificial intelligence, reasoning, knowledge representation

RealICU: Do LLM Agents Understand Long-Context

ICU Data? A Benchmark Beyond Behavior Imitation

arXiv:2605.13542v1 [cs.AI] 13 May 2026

Chengzhi Shen1,2,10 Weixiang Shen1,2,3 Tobias Susetzky1,2 Chen (Cherise) Chen4 Jun Li1 Yuyuan Liu5 Xuepeng Zhang6 Zhenyu Gong7,† Daniel Rueckert1,2,8,9,10,† Jiazhen Pan1,2,9,† 1

Technical University of Munich (TUM) 2 TUM University Hospital 3 LMU Munich University of Sheffield 5 University of Oxford 6 Zhongshan Hospital Fudan University 7 Sun Yat-sen University Cancer Center 8 Imperial College London 9 Munich Center for Machine Learning (MCML) 10 relAI – Konrad Zuse School of Excellence in Reliable AI † Corresponding Authors

Abstract Intensive care units (ICU) generate long, dense and evolving streams of clinical information, where physicians must repeatedly reassess patient states under time pressure, underscoring a clear need for reliable AI decision support. Existing ICU benchmarks typically treat historical clinician actions as ground truth. However, these actions are made under incomplete information and limited temporal context of the underlying patient state, and may therefore be suboptimal, making it difficult to assess the true reasoning capabilities of AI systems. We introduce RealICU, a hindsight-annotated benchmark for evaluating large language models (LLMs) under realistic ICU conditions, where labels are created after senior physicians review the full patient trajectory. We formulate four physician-motivated tasks: assess Patient Status, Acute Problems, Recommended Actions, and Red Flag actions that risk unsafe outcomes. We partition each trajectory with 30-min windows and release two datasets: RealICU-Gold with 930-window annotations from 94 MIMIC-IV patients, and RealICU-Scale with 11,862 windows extended by Oracle, a physician-validated LLM hindsight labeler. Existing LLMs including memory-augmented ones performed poorly on RealICU, exposing two failure modes: a recall-safety tradeoff for clinical recommendations, and an anchoring bias to early interpretations of the patient. We further introduce ICU-Evo to study structured-memory agents that improves long-horizon reasoning but does not fully eliminate safety failures. Together, RealICU provides a clinically grounded testbed for measuring and improving AI sequential decision-support in high-stakes care. Project page: chengzhi-leo.github.io/RealICU-Bench

Introduction

The Intensive Care Unit (ICU) is one of the most information-dense environments in the hospital. Within hours, a single patient can generate large volumes of laboratory results, vital signs, medications, nursing observations, and imaging reports [20, 24]. Physicians must integrate this evolving stream under time pressure, where each measurement captures only a partial slice of the patient’s physiological state, and decisions made in one moment may shape outcomes hours or days later [23, 27]. This underscores a clear need for AI decision support system in real-time monitoring and decision-making in the ICU, which usually acts as a clinical co-pilot. In consultations with over 30 board-certified clinicians, including five senior ICU physicians who later served as annotators, four Preprint.

Long-Context ICU Data Stream

ICU Co-Pilot with Clinical Decision Support

Patient Status

Acute Problems Major Concerns

IMPROVING

STABLE DETERIORATING

Overall Status: Deteriorating Hemodynamics

↓ Deteriorating

Oxygenation

↓ Deteriorating

Inflammation

↓ Deteriorating

Organ Function

- Stable

INFECTION

Temp. 38.6°C

SEPTIC SHOCK

Pulmonary

LACTIC ACIDOSIS

Lactate 3.8

Recommended Actions

Red Flag Actions

Recommended Treatments

Actions to Avoid

NOREPINEPHRINE

AGGRESSIVE DIURESIS

0.15 μg/kg/min

Hypovolemic, will worsen shock

VASOPRESSIN

ETOMIDATE FOR INTUBATION

0.03 U/min IV

Adrenal suppression

LACTATED RINGER’S

BETA-BLOCKERS

500 mL

Compensatory tachycardia

Figure 1: ICU decisions are made under massive data volume and time pressure. An ICU AI co-pilot integrates data streams into a decision-support panel that assesses Patient Status, identifies Acute Problems, proposes Recommended Actions, and warns against unsafe Red Flag actions.

capabilities emerged as core requirements for a useful ICU co-pilot: assess Patient Status, identify Acute Problems, propose Recommended Actions, and warn against Red Flag actions that may cause unsafe outcomes. Figure 1 illustrates the use case of an AI co-pilot in ICU decision support. Benchmark gap. Despite rapid progress in Large Language Models (LLMs) and agentic systems, few benchmarks evaluate these four capabilities in real-world ICU settings. Most clinical benchmarks reduce clinical reasoning to static question answering, diagnosis, or summarization [18, 31, 13, 14, 3], or to single-endpoint prediction (e.g., mortality [38], shock [7, 37], or acute kidney injury [19, 5]). Such benchmarks aggregate clinical care into isolated predictions, offering little signal on whether a model can reason across a changing patient trajectory. More importantly, benchmarks built on electronic health record (EHR) databases such as MIMIC-IV [16], HiRID [11], and eICU-CRD [25] treat recorded clinician actions as ground-truth labels. But this assumption is fragile. A recorded action reflects what clinicians believed best given incomplete information at the bedside, whereas the optimal action often becomes clear only after reviewing the trajectory using hindsight. Evaluating AI models against such labels therefore rewards behavioral imitation rather than clinical correctness. Proposed benchmark. To address this gap, we introduce RealICU, a hindsight-grounded benchmark built from MIMIC-IV [16] for evaluating LLM-based clinical decision support in the ICU. RealICU evaluates four physician-motivated tasks over dense 30-minute windows across the ICU trajectory: Patient Status, Acute Problems, Recommended Actions, and Red Flags. At each window, the agent observes only information available up to that time, while labels are produced by hindsight physician judgment over the full trajectory. This design scores agents on clinical correctness rather than on recorded behavior. RealICU contains two subsets. RealICU-Gold provides 930 physician-labeled windows from 94 ICU stays, and RealICU-Scale extends evaluation to 11,862 windows using Oracle, a physician-validated LLM-based hindsight evaluator calibrated against expert consensus. Failure mode identification and mitigation. Using RealICU, we benchmark frontier LLM-based ICU agents across diverse context configurations including memory. Current agents show poor reliability over long ICU contexts, with two failure modes: (i) Recall-safety tradeoff, where higher recommendation recall comes with up to 47.3% of these recommendations flagged as potentially harmful; (ii) Anchoring bias, where agents preserve early interpretations of the patient despite later contradictory evidence. To mitigate these, we introduce ICU-Evo, a structured-memory agent framework that maintains recent observations, temporal trends, critical events, trajectory summaries, and patient-specific insights. ICU-Evo is backbone-agnostic and improves clinical reasoning, but its safety failures show that structured memory alone is insufficient for reliable ICU co-pilots. 2

Our key contributions are as follows: • We formulate ICU co-pilot evaluation around four physician-motivated tasks: Patient Status, Acute Problems, Recommended Actions, and Red Flags. Unlike static clinical QA or outcome prediction benchmarks, these tasks evaluate whether an AI system can support continuous bedside reassessment across an evolving ICU trajectory. • We release RealICU, a hindsight-annotated benchmark for clinical correctness rather than behavioral imitation. Agents observe only data available at decision time, while labels are produced by hindsight physician judgment over the full trajectory. RealICU-Gold provides 930 physician-consensus windows from 94 ICU stays, and RealICU-Scale extends this to 11,862 windows using Oracle, a physician-validated LLM-based hindsight evaluator. • We identify gaps in current LLM ICU agents and study structured memory as a mitigation. Across frontier LLMs and multiple context strategies, RealICU remains largely unsolved. We identify a recall–safety tradeoff and anchoring bias as major failure modes, and introduce ICUEvo, a structured-memory agent that improves long-horizon reasoning but shows that memory alone is insufficient for safe ICU decision support.

Related Work

Clinical Benchmarks for LLMs and Agents. Exam-style benchmarks such as MedQA [13], PubMedQA [14], and MedXpertQA [39] evaluate clinical knowledge as multiple-choice recall under complete information, a format well-addressed by state-of-the-art models that reveals little about decisions under uncertainty. Conversational benchmarks such as AI Hospital [6], AgentClinic [28], and VivaBench [3] require agents to gather history, order investigations, and converge on a diagnosis over multiple turns, exposing failure modes such as premature diagnostic closure. MedAgentBench [12] moves closer to real EHR environments but retains a task-completion framing rather than evaluating overall patient management. None of these benchmarks evaluates sequential decision-making over long ICU trajectories or distinguishes behavioral imitation from clinical correctness. RealICU addresses both by grounding evaluation in hindsight physician judgment over the full ICU trajectory, providing dense and trajectory-level signal of clinical correctness. Memory-Augmented LLM Agents. Recent LLM agent architectures have explored a range of memory designs. ReAct [35] appends all reason-action results sequentially but saturates quickly as context accumulates. AgentFold [36] addresses this by summarizing completed sub-tasks at multiple temporal scales. Evo-Memory [32] unifies reasoning, action, and memory refinement in a test-time loop. Retrieval-based systems such as RAG [1, 4] and A-MEM [33] enable selective access over long histories. However, these systems treat clinical context equally, making no distinction between static patient background [21], time-sensitive physiological trends [17], and high-level trajectory [29, 26], which play fundamentally different roles in clinical reasoning. ICU-Evo organizes clinical context into heterogeneous memory types aligned with these distinctions, enabling systematic study of how structured memory design shapes ICU decision-making.

RealICU Benchmark

RealICU evaluates LLM agents on sequential clinical decision-making across ICU trajectories, mirroring standard medical quality review: model outputs are assessed against hindsight physician labels produced with full knowledge of patient trajectory rather than against logged clinician actions. RealICU consists of two datasets. RealICU-Gold contains 930 sparsely sampled windows from 94 ICU stays labeled by physician consensus. To scale beyond manual annotation, we introduce Oracle, an LLM-based hindsight evaluator validated against RealICU-Gold, yielding RealICU-Scale with 11,862 densely labeled windows. Both datasets are released test-only to prevent leakage. Detailed statistics are in Figure 8, Figure 9, and Figure 10. Each window Wt = (Xt ; St , Pt , At , Rt ) contains clinical observations up to time t, annotated for four tasks: Patient Status St , Acute Problems Pt , Recommended Actions At , and Red Flag Actions Rt . The model predicts (Ŝt , P̂t , Ât ) from Xt ; Rt serves as a safety check against Ât . This 3

Data Pipeline

Patient Context · Age: 45 ·Gender: Female · Pre-ICU history: ruptured saccular aneurysm, left cerebellar parenchymal hemorrhage, rising lactate (1.3 → 3.6) · Allergies: penicillins, sulfa

MIMIC-IV Cohort

Cohort Sampling 94 ICU stays, balanced sampling ICU Trajectory Window Discretization 30-min eval windows, 2h stride Raw Dataset 11,862 windows

Window Selection action density

Physicians Anno. five ICU experts

Oracle Anno hindsight eval

RealICU-Gold 930 windows

RealICU-Scale 11,862 windows

truncate at discharge ... ... ... ...

Enter ICU

22:34 22:41 22:41 22:55 23:00 23:00 23:00 23:02 23:04 23:05 23:06 ...

hour 4

hour 8

admit Neurosurgery, hx HTN phenylephrine 14.55 mg, D5W 60 ml weight 70 kg intubated, aspiration precautions, HOB >30 EVD placed: ventricular, level 15cm ICP 11, CPP 67, drainage bloody R femoral angio site, DSD, C/D/I GCS 3 (E1 V1 M1), pupils 3mm non-reactive corneas absent, gag/cough intact RLE pale & cool, dorsal pulses doppler only ART 126/60, MAP 85, HR 87, SpO2 100% ... ...

Labels

hour 12

07:40 08:00 08:12 08:12 08:15 09:01 09:02 09:03 09:04 09:05 09:06 ...

sent to OR (aneurysm clipping) EVD output 10 ml since last check post-op care plan initiated restraints initiated, HOB >30 weight 70 kg, vent settings unchanged ABG: pH 7.44, pCO2 35, pO2 221 acidosis resolved (was pH 7.25) Hb 9.4, Hct 28% — mild anemia Na 134, glucose 128, lactate 1.4 K 3.2, iCa 0.93 — repletion needed lactate trending down (3.6 → 1.4) ... ...

Labels

Patient Status

Acute Problems

Patient Status

Acute Problems

stable

Aneurysm rupture EVD obstruction

improving

Electrolyte derangements Hydrocephalus / ICP

Action Recomm.

Red Flags

Action Recomm.

Red Flags

Phenylephrine EVD patency check

✗ Anticoagulants ✗ Aggressive BP lowering

ABG monitoring Replete K, iCa

✗ Hypotonic IV fluids ✗ Anticoagulants

Quality Assessment Phys.-Phys. IRR, Oracle validation

Label Construction status · problems · actions · red flags

hour 96

Figure 2: Left: Data pipeline for RealICU-Gold and RealICU-Scale. Right: Data samples for a patient ICU trajectory. For each evaluation window, RealICU provides raw observation data and clinical labels, including patient status, acute problems, action recommendation, and red flag action.

asymmetry between partial observation and hindsight annotation mirrors the gap between real-time decision-making and hindsigth review. Figure 2 illustrates the data construction pipeline and samples. 3.1

Dataset Construction

Cohort. We sample 94 ICU stays from the MIMIC-IV [16] cohort, each from a distinct patient and balanced by ICU outcome. Stays shorter than 4 hours are discarded. To capture both early stabilization and long trajectories, we balance stays by duration above and below 96 hours. Windowing. We define 30-minute windows as our evaluation unit and sample them along each ICU trajectory with a 2-hour stride, preserving short-term dynamics while limiting redundancy across adjacent windows. At inference time, the trajectory visible to the model is truncated prior to outcome-revealing events such as ICU discharge or the discharge summary. 3.2

Tasks

We identify four crucial ICU reasoning tasks below after consulting more than 30 clinicians, including five senior ICU physicians who later served as annotators. Together they cover the key capabilities of a useful ICU co-pilot. For all four tasks, each prediction is accompanied by supporting evidence E ⊆ Xt drawn from the raw events in the recorded history. Patient Status. A classification of whether the patient is improving, stable, or deteriorating relative to recent context: St = (st , Et ), where st ∈ {improving, stable, deteriorating} and Et ⊆ Xt . Acute Problems. A free-text set of acute problems or emerging risks that require active management: Pt = {(pi , Ei )}ki=1 , where Ei ⊆ Xt . Action Recommendation. A free-text set of actions likely to benefit the patient within one hour, such as stabilizing physiology or preventing deterioration: At = {(aj , Ej )}m j=1 , where Ej ⊆ Xt . Red Flags. A free-text set of high-risk actions that should be avoided because they may be harmful under the patient’s current physiology or trajectory: Rt = {(rl , El )}nl=1 , where El ⊆ Xt . 4

3.3

Annotation Protocol

RealICU-Gold with physician consensus. We begin from sampling approximately 10 windows per ICU stay by action density ρt = |Etaction |/|Et |, i.e. the fraction of action events inside each window. We draw 80% of windows from the ρt ≥ 0.5 regime, where interventions are frequent, and 20% from ρt < 0.5 as a control set. Each window is independently labeled by at least two of five senior ICU physicians. Inter-rater reliability (IRR) among physicians ranges from 0.826 to 0.985 across the four tasks (Table 1), confirming both strong label reproducibility and that the task definitions are sufficiently precise for consistent clinical judgment. Windows without physician agreement are dropped, yielding 930 validated windows in RealICU-Gold. RealICU-Scale with Oracle scaling. Despite high qual- Table 1: RealICU-Gold label quality and ity, manual annotation covers only a sparse sample of each Oracle validation. ICU stay. We therefore introduce Oracle, an LLM evaluator operating under the same hindsight conditions as Task Phys. IRR Oracle F1 the physicians, and apply it to densely label every window across the cohort, yielding 11,862 annotated windows Patient Status 0.985 0.987 in RealICU-Scale. We validate Oracle by measuring its 0.980 0.987 F1 score against physician consensus on RealICU-Gold. Acute Problems Oracle achieves more than 0.895 F1 score across all four Action Recom. 0.826 0.895 tasks (Table 1), supporting its use as a reliable hindsight Red Flags 0.916 0.964 annotator at scale. While Oracle is backbone-agnostic, we instantiate it with Gemini-3.1-pro [8] in this work. Detailed Oracle prompt is in Appendix E. Label construction. Labels for Patient Status, Acute Problems, and Red Flags are taken directly from annotations. For Action Recommendation, we restrict the annotation space to critical clinical interventions, discarding routine monitoring. Annotators review each action as best-practice, acceptable, or potentially-harmful, and may add free-text actions that should have been taken but were not observed. At is constructed as the union of best-practice and acceptable actions together with these free-text additions. Red Flags are annotated independently as a separate label, not derived from potentially-harmful actions. 3.4

Evaluation Framework (k)

(k)

A model under test M maps observations Xt to predictions (Ŝt , P̂t , Ât ), where P̂t and Ât are top-k ranked lists, with access only to events up to time t. In this paper we focus on LLM agents, but M can be any model. Models are evaluated against RealICU-Gold and RealICU-Scale, providing sparse gold-standard supervision and trajectory-level evaluation at scale respectively. Algorithm 1 summarizes the complete evaluation framework. Semantic matching. To score free-text tasks (Acute Problems, Recommended Actions, Red Flag Actions), we adopt PubMedBERT [9] and define a binary match, where τ is calibrated against 100 expert-annotated pairs, achieving 0.96 F1 at τ = 0.5 (Appendix A.5): match(xpred , xref ) = 1[cos(epred , eref ) ≥ τ ] .

(1)

Metrics. Patient Status is evaluated with accuracy and macro-F1 to avoid dominance by the majority class (stable). Acute Problems and Recommended Actions are set-matching tasks evaluated with Hit@k and Recall@k at k=5. Red Flag Actions serves as a safety check via the Harmful (k) Recommendation Rate (HRR). Let S be the set of ICU stays, Ws the windows in stay s, Ât the top-k recommendations, and Rt the red-flag set at window t; HRR averages the fraction of recommended actions that are flagged across stays: 1 X HRR(M) = |S| s∈S

(k)

t∈Ws

Ât

t∈Ws

∩ Rt

(k) Ât

(2)

Algorithm 1 RealICU Evaluation Framework. Require: model M; label source R ∈ {Gold, Scale}; ICU stay set S; per-stay window sets {Ws }s∈S 1: for each ICU stay s ∈ S do 2: h ← 0, n ← 0 ▷ red-flag hits / total recommendations 3: for each window t ∈ Ws in chronological order do (k) (k) 4: (Ŝt , P̂t , Ât ) ← M(Xt ) ▷ model sees events up to t 5: (St , Pt , At , Rt ) ← R(t) ▷ pre-labeled by hindsight annotator 6: evaluate (Ŝt , St ) for Patient Status accuracy and F1 (k) 7: evaluate (P̂t , Pt ) for Acute Problems Hit@k / Recall@k ▷ semantic matching (k) 8: evaluate (Ât , At ) for Action Recommendation Hit@k / Recall@k ▷ semantic matching (k) (k) 9: h += |Ât ∩ Rt |; n += |Ât | ▷ safe recommendation check 10: end for 11: aggregate per-window scores 12: end for 13: return scores across S for each task

ICU-Evo: An ICU Agent System with Evolving Memory

ICU decision-making is sequential, where the underlying patient state is only partially observable with clinical measurements and can only be updated via new observations. We model this as a partially observable Markov process [2, 30] and approximate the latent patient state with a structured memory Mt . We introduce ICU-Evo as an instance of the memory-augmented agent frameworks to study how structured memory design shapes clinical decision-making. 4.1

Memory as a Structured Belief State

Given the context Xt and static patient context c (e.g. demographics, allergies, pre-ICU history), ICU-Evo maintains a structured memory state Mt updated at each window by incorporating the new (k) measurements xt = Xt − Xt−1 , and produces task-specific predictions yt via (k)

Mt = U(Mt−1 , xt ),

= f (k) (Mt , c).

The memory decomposes into five components following clinical reasoning: Mt = Mtwork , Mttrend , Mtevent , Mttraj , Mtinsight .

(3)

(4)

Working memory Mtwork holds the most recent raw observations at detailed resolution. Trend memory Mttrend captures signal trends of vital and lab values. Critical-event memory Mtevent is a persistent, append-only log of clinically critical events that change the patient story, such as abnormal physiology, interventions, and turning points. Trajectory memory Mttraj provides a compressed narrative of the stay at periodic intervals. Insight memory Mtinsight maintains patient-specific hypotheses constructed as deviations from population-level expectation. Every memory component carries evidence from raw observations, so any clinical decision is explainable and verifiable against the patient record. In Table 12, we summarize the memory components with corresponding agent sources. 4.2

ICU-Evo Agent Pipeline

ICU-Evo realizes the memory update operator U through three specialized agents operating at different temporal scales over the shared memory. ICU-Evo belongs to a broader family of memoryaugmented agent systems. We discuss it alongside recent agent systems in Appendix C. Detailed prompts are reported in Appendix E. Observation Agent (Aobs ). A rule-based agent that turns raw measurements into structured signals at every window. It normalizes units, aligns observations to the 30-minute window grid, and extracts trend signals from vitals using Piecewise Aggregate Approximation [10]: work trend Mtwork , Mttrend = Aobs (Mt−1 , Mt−1 , xt ). (5) 6

Table 2: Evaluation results on RealICU-GOLD. Within each backbone, bold marks the best system per column and underline the second best. Patient Status

Acute Problems

Action Recom.

Red Flags

Backbone

System

Acc.↑

F1↑

Hit@5↑

R@5↑

Hit@5↑

R@5↑

HRR@5↓

Gemini-3.1-pro [8]

Full-context Local-window RAG ICU-Evo

0.298 0.315 0.402 0.459

0.258 0.239 0.348 0.365

0.486 0.459 0.596 0.823

0.308 0.258 0.342 0.526

0.259 0.395 0.496 0.676

0.152 0.260 0.313 0.534

0.137 0.151 0.216 0.300

GPT-5.4 [22]

Full-context Local-window RAG ICU-Evo

0.294 0.233 0.288 0.312

0.233 0.184 0.256 0.264

0.510 0.500 0.599 0.867

0.348 0.293 0.349 0.570

0.404 0.380 0.480 0.676

0.300 0.281 0.398 0.534

0.298 0.165 0.234 0.473

Qwen3-235B [34]

Full-context Local-window RAG ICU-Evo

0.225 0.152 0.315 0.253

0.188 0.154 0.271 0.197

0.384 0.213 0.379 0.600

0.226 0.126 0.211 0.362

0.329 0.352 0.453 0.526

0.222 0.242 0.324 0.357

0.117 0.080 0.095 0.117

Assessment Agent (Aassess ). For every ka cumulative windows, an LLM transforms recent observations into a trajectory summary and detects critical events. It consumes the working and trend memory accumulated over the past ka windows, producing a trajectory summary zt appended to Mttraj and critical events et appended to Mtevent : work trend zt , et = Aassess Mt−k , Mt−k . (6) a :t a :t Insight Agent (Ainsight ). Every ki windows, an LLM proposes hypotheses about what is driving the patient’s clinical course and gathers supporting evidence es and counter-evidence er from Mtevent . A hypothesis ht is accepted if s(ht ) > r(ht ) and rejected otherwise. The Insight Agent actively reasons about patient-specific patterns, such as unusual drug responses or persistent abnormalities, promoting individualized care beyond averaged guidelines: insight event Mtinsight = Ainsight Mt−1 , Mt−k . (7) i :t Predictor (f (k) ). The predictor is a task-specific prompted LLM over the full memory state and static patient context, decoupled from the agent system (Equation 3).

Evaluation & Analysis

RealICU Remains Unsolved for Current LLM Systems

RealICU remains unsolved for current frontier LLMs and agent systems. Across all evaluation setups in Table 2, ICU-Evo with Gemini-3.1-pro [8] reaches only 0.459 accuracy on Patient Status and 0.534 Recall@5 on Action Recommendation. More concerning, Red Flags HRR@5 stays non-trivial across all configurations, indicating current LLM systems still recommend potentially harmful actions in high-stake ICU setting. Together, these gaps establish RealICU as a clinically grounded safety check for future AI decision-support systems. 5.2

Structured Memory Consistently Improves Clinical Reasoning

Structured memory improves performance across all four tasks. With GPT-5.4 [22], ICU-Evo improves over RAG by 26.8 Hit@5 points on Acute Problems and 19.6 on Action Recommendation, 7

1.0

Local window RAG ICU-Evo

0.8

Patient Status

1.0 0.8

Hit@5

0.4

0.2

0.2 0

250

500

750

1000

1250

Hours Since Admission

1500

0.0

1750

Action Recommendation

1.0

Local window RAG ICU-Evo

0.8

250

500

750

1000

1250

1500

1750

1250

1500

1750

Hours Since Admission

Local window RAG ICU-Evo

0.8

Red Flags

0.6

HRR

Hit@5

0.6 0.4

0.4

0.2 0.0

Acute Problems

0.6

Accuracy

0.6

0.0

Local window RAG ICU-Evo

0.2 0

250

500

750

1000

1250

Hours Since Admission

1500

0.0

1750

250

500

750

1000

Hours Since Admission

Figure 3: Temporal performance on RealICU-Scale (Gemini-3.1-pro [8]). ICU-Evo demonstrates its advantage on Patient Status and Acute Problems even up to 1,800-hour trajectory.

with similar margins on Gemini-3.1-pro [8] and Qwen3-235B [34] (Table 2). The pattern holds on the densely labeled RealICU-Scale (Table 4, Figure 3). ICU-Evo’s Hit@5 on Acute Problems stays near 0.8 even for stays up to 1,800 hours, while non-memory baselines remain about 20 points lower and visibly noisier. Future ICU decision-support agents will benefit from memory that actively tracks the patient’s evolving state and scales to long stays. 5.3

The Agent-Oracle Gap: Beyond Behavioral Imitation

We observe a large performance gap between Agent and Oracle on RealICU-Gold. The bottleneck of current ICU agents is not medical knowledge in the LLM backbone but how an agent integrates evidence over time. With Gemini-3.1-pro [8], Oracle reaches F1 0.987 on Patient Status and 0.964 on Red Flags identification (Table 1), while ICU-Evo on the same backbone reaches only 0.365 F1 on Patient Status, with a concerningly high rate of harmful recommendations with 0.300 HRR (Table 2). The four clinical tasks are therefore well handled given the full trajectory but break down under real-time conditions. This gap also indicates the value of hindsight evaluation, since scoring agents against recorded clinician actions can only measure how closely the agents imitate human behavior under limited information. Progress on ICU decision support therefore depends on both stronger real-time reasoning architectures and the broader adoption of hindsight evaluation. 5.4

Ablation Study

We ablate each component of ICU-Evo’s memory in a leave-one-out setup (Table 3). Working memory is crucial for local clinical reasoning, and removing it degrades every task with notable drops on Acute Problems Hit@5 (Gemini-3.1-pro [8], 0.823 to 0.761). Trajectory memory matters for temporal understanding tasks. Without it, Acute Problems and Action Recommendation both drop, while Patient Status stays stable since it leans on local observations. In contrast, insight memory causes fluctuations, and removing it sometimes leads to neutral or beneficial results. This suggests that current LLMs default to medical-generalist priors and is not yet capable to identify reliable personalized clinical patterns across long stays. We examine the failure modes in the next section. 8

Table 3: Memory ablation on RealICU-GOLD. For each row, we remove one component from ICU-Evo’s memory. Within each backbone, bold marks the best result and underline the second best. Patient Status Acute Problems Backbone

Memory Variant

Action Recom.

Red Flags

Acc.↑

F1↑

Hit@5↑ R@5↑ Hit@5↑ R@5↑

HRR↓

ICU-Evo − working memory − trend Gemini-3.1-pro [8] − critical events − trajectory − insight

0.459 0.383 0.451 0.445 0.443 0.462

0.365 0.294 0.352 0.351 0.362 0.356

0.823 0.761 0.811 0.819 0.789 0.823

0.526 0.461 0.527 0.527 0.500 0.534

0.676 0.507 0.521 0.528 0.506 0.555

0.534 0.308 0.330 0.328 0.304 0.357

0.300 0.087 0.097 0.099 0.090 0.088

ICU-Evo − working memory − trend − critical events − trajectory − insight

0.253 0.159 0.248 0.236 0.270 0.250

0.197 0.127 0.188 0.187 0.249 0.202

0.600 0.421 0.552 0.546 0.486 0.587

0.362 0.233 0.333 0.320 0.290 0.348

0.526 0.447 0.559 0.595 0.557 0.601

0.357 0.307 0.393 0.420 0.393 0.420

0.117 0.117 0.122 0.128 0.138 0.141

Qwen3-235B [34]

5.5

Failure Mode Analysis

Oracle failure modes. Oracle reaches around 90% F1 across all tasks (Table 1). The remaining disagreements concentrate on two patterns: (i) boundary mis-calibration on Patient Status, where failures fall on stable–improving or stable–deteriorating borders; (ii) granularity mismatch on Acute Problems, where Oracle reaches for broad descriptors (e.g. hemodynamic instability) while physicians name specific complications (e.g. ventilator-associated pneumonia). These are edge cases rather than systematic errors, supporting Oracle’s reliability as a large-scale annotator. Agent failure modes. The most consequential failure is the recall–safety tradeoff, where higher recommendation recall increases the incidence of harmful clinical suggestions. In Table 2, ICU-Evo (GPT-5.4 [22]) gains more than 20 Action Recommendation Hit@5 points over RAG, but its HRR@5 doubles from 0.234 to 0.473. We use an LLM-based classifier to group these 394 cases, and the majority concentrate in four high-stakes families: hemodynamic and pressor management (n=135), volume and diuresis (n=64), anti-coagulation (n=54), and ventilation/sedation (n=53). We find that current LLM agents tend to recognize part of a syndrome and propose the full treatment bundle before contraindications are ruled out. The second failure is anchoring bias, where agents over-commit to early interpretations and ignore later evidence. Removing insight memory improves Action Recommendation Hit@5 from 0.526 to 0.601 on Qwen3-235B [34] (Table 3), indicating that generated insights actively mislead the agent. The agent maintains around 6 hypotheses per patient, 80% containing anticipatory exceptions (e.g. below-average tolerance, paradoxical response). These priors push the agent toward rescue bundles even when current evidence is weak. Case studies are in Appendix D.

Discussion

RealICU reveals a substantial gap between the medical knowledge of frontier LLMs and their ability to reason under partial observability across an evolving ICU trajectory. We identify two recurring failure modes that persist across multiple context configurations. (i) the recall–safety tradeoff, where gains in Recommended Actions coverage are accompanied by a higher rate of unsafety. (ii) anchoring bias, where agents commit to an early read of the patient and fail to update as new evidence accumulates. ICU-Evo uses structured, evidence-grounded memory at multiple temporal scales to track the evolving patient state, but multi-scale memory alone does not prevent unsafe recommendations. Reliable ICU co-pilots will require advances in long-context clinical reasoning together with better safety mechanisms. 9

Beyond the ICU, RealICU offers a methodology for evaluating AI systems where recorded human actions are imperfect and the right action is visible only in hindsight. We hope this framing supports broader work on evaluating AI systems in high-stakes sequential decision environments. Limitations. RealICU is built on the MIMIC-IV [16] cohort, and its demographic and care-pattern distribution may not transfer to ICUs with different staffing or documentation conventions. Extending to multi-center and international data is an important direction. Due to compute constraints, we run a single experiment per LLM configuration and omit variance over long ICU trajectories. We also focus on text-based data, leaving multi-modal data such as imaging and signals to future work.

Acknowledgement

This paper is supported by the DAAD programme Konrad Zuse Schools of Excellence in Artificial Intelligence, sponsored by the Federal Ministry of Research, Technology and Space. This work is partially funded by the European Research Council (ERC) project Deep4MI (884622).

References [1] Muhammad Arslan, Hussam Ghanem, Saba Munawar, and Christophe Cruz. A survey on rag with llms. Procedia computer science, 246:3781–3790, 2024. [2] Anthony Rocco Cassandra. Exact and approximate algorithms for partially observable Markov decision processes. Brown University, 1998. [3] Christopher Chiu, Silviu Pitis, and Mihaela van der Schaar. Simulating viva voce examinations to evaluate clinical reasoning in large language models. arXiv preprint arXiv:2510.10278, 2025. [4] Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. The power of noise: Redefining retrieval for rag systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 719–729, 2024. [5] Junzi Dong, Ting Feng, Binod Thapa-Chhetry, Byung Gu Cho, Tunu Shum, David P Inwald, Christopher JL Newth, and Vinay U Vaidya. Machine learning model for early prediction of acute kidney injury (aki) in pediatric critical care. Critical Care, 25(1):288, 2021. [6] Zhihao Fan, Lai Wei, Jialong Tang, Wei Chen, Wang Siyuan, Zhongyu Wei, and Fei Huang. Ai hospital: Benchmarking large language models in a multi-agent medical interaction simulator. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10183–10213, 2025. [7] Shameek Ghosh, Jinyan Li, Longbing Cao, and Kotagiri Ramamohanarao. Septic shock prediction for icu patients via coupled hmm walking on sequential contrast patterns. Journal of biomedical informatics, 66:19–31, 2017. [8] Google DeepMind. Gemini 3.1 pro model card. Technical report, Google DeepMind, 2026. URL https://deepmind.google/models/model-cards/gemini-3-1-pro/. Accessed: May 6, 2026. [9] Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1–23, 2021. [10] Chonghui Guo, Hailin Li, and Donghua Pan. An improved piecewise aggregate approximation based on statistical features for time series mining. In International conference on knowledge science, engineering and management, pages 234–244. Springer, 2010. [11] Stephanie L Hyland, Martin Faltys, Matthias Hüser, Xinrui Lyu, Thomas Gumbsch, Cristóbal Esteban, Christian Bock, Max Horn, Michael Moor, Bastian Rieck, et al. Early prediction of circulatory failure in the intensive care unit using machine learning. Nature medicine, 26(3): 364–373, 2020. [12] Yixing Jiang, Kameron C Black, Gloria Geng, Danny Park, James Zou, Andrew Y Ng, and Jonathan H Chen. Medagentbench: a virtual ehr environment to benchmark medical llm agents. Nejm Ai, 2(9):AIdbp2500144, 2025. [13] Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences, 11(14):6421, 2021. [14] Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 2567–2577, 2019. [15] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 11

[16] Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, et al. Mimic-iv, a freely accessible electronic health record dataset. Scientific data, 10(1):1, 2023. [17] H Lehman Li-wei, Ryan P Adams, Louis Mayaud, George B Moody, Atul Malhotra, Roger G Mark, and Shamim Nemati. A physiological time series dynamics-based approach to patient monitoring and outcome prediction. IEEE journal of biomedical and health informatics, 19(3): 1068–1076, 2014. [18] Mingyu Derek Ma, Chenchen Ye, Yu Yan, Xiaoxuan Wang, Peipei Ping, Timothy S Chang, and Wei Wang. Clibench: A multifaceted and multigranular evaluation of large language models for clinical decision making. arXiv preprint arXiv:2406.09923, 2024. [19] Rakesh Malhotra, Kianoush B Kashani, Etienne Macedo, Jihoon Kim, Josee Bouchard, Susan Wynn, Guangxi Li, Lucila Ohno-Machado, and Ravindra Mehta. A risk prediction score for acute kidney injury in the intensive care unit. Nephrology Dialysis Transplantation, 32(5): 814–822, 2017. [20] Orit Manor-Shulman, Joseph Beyene, Helena Frndova, and Christopher S Parshuram. Quantifying the volume of documented clinical information in critical illness. Journal of critical care, 23(2):245–250, 2008. [21] Paola P Mattey-Mora, Connor A Begle, Candice K Owusu, Chen Chen, and Maria A Parker. Hospitalised versus outpatient covid-19 patients’ background characteristics and comorbidities: a systematic review and meta-analysis. Reviews in Medical Virology, 32(3):e2306, 2022. [22] OpenAI. Gpt-5.4 technical report and model card. Technical report, OpenAI, March 2026. URL https://openai.com/index/introducing-gpt-5-4/. Accessed: May 6, 2026. [23] Nicolas Paul, Elena Ribet Buse, Anna-Christina Knauthe, Monika Nothacker, Björn Weiss, and Claudia D Spies. Effect of icu care bundles on long-term patient-relevant outcomes: a scoping review. BMJ open, 13(2):e070962, 2023. [24] Brian W Pickering, Vitaly Herasevich, Adil Ahmed, and Ognjen Gajic. Novel representation of clinical information in the icu. Applied Clinical Informatics, 1(02):116–131, 2010. [25] Tom J Pollard, Alistair EW Johnson, Jesse D Raffa, Leo A Celi, Roger G Mark, and Omar Badawi. The eicu collaborative research database, a freely available multi-center database for critical care research. Scientific data, 5(1):180178, 2018. [26] Elizabeth Reed and Jessica Corner. Defining the illness trajectory of metastatic breast cancer. BMJ supportive & palliative care, 5(4):358–365, 2015. [27] Regis Goulart Rosa, Giovanni Esteves Ferreira, Thiago Wendt Viola, Caroline Cabral Robinson, Renata Kochhann, Paula Pinheiro Berto, Livia Biason, Paulo Ricardo Cardoso, Maicon Falavigna, and Cassiano Teixeira. Effects of post-icu follow-up on subject outcomes: a systematic review and meta-analysis. Journal of critical care, 52:115–125, 2019. [28] Samuel Schmidgall, Rojin Ziaei, Carl Harris, Eduardo Reis, Jeffrey Jopling, and Michael Moor. Agentclinic: a multimodal agent benchmark to evaluate ai in simulated clinical environments. arXiv preprint arXiv:2405.07960, 2024. [29] Helena Sousa, Susana Almeida, Joao Bessa, and M Graca Pereira. The developmental trajectory of cancer-related cognitive impairment in breast cancer patients: a systematic review of longitudinal neuroimaging studies. Neuropsychology review, 30(3):287–309, 2020. [30] Matthijs TJ Spaan. Partially observable markov decision processes. In Reinforcement learning: State-of-the-art, pages 387–414. Springer, 2012. [31] Robin Van De Water, Hendrik Schmidt, Paul Elbers, Patrick Thoral, Bert Arnrich, and Patrick Rockenschaub. Yet another icu benchmark: A flexible multi-center framework for clinical ml. arXiv preprint arXiv:2306.05109, 2023. 12

[32] Tianxin Wei, Noveen Sachdeva, Benjamin Coleman, Zhankui He, Yuanchen Bei, Xuying Ning, Mengting Ai, Yunzhe Li, Jingrui He, Ed H Chi, et al. Evo-memory: Benchmarking llm agent test-time learning with self-evolving memory. arXiv preprint arXiv:2511.20857, 2025. [33] Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-mem: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110, 2025. [34] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. [35] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In The eleventh international conference on learning representations, 2022. [36] Rui Ye, Zhongwang Zhang, Kuan Li, Huifeng Yin, Zhengwei Tao, Yida Zhao, Liangcai Su, Liwen Zhang, Zile Qiao, Xinyu Wang, et al. Agentfold: Long-horizon web agents with proactive context management. arXiv preprint arXiv:2510.24699, 2025. [37] Christopher R Yee, Niven R Narain, Viatcheslav R Akmaev, and Vijetha Vemulapalli. A datadriven approach to predicting septic shock in the intensive care unit. Biomedical informatics insights, 11:1178222619885147, 2019. [38] Zirun Zhao, Anne Chen, Wei Hou, James M Graham, Haifang Li, Paul S Richman, Henry C Thode, Adam J Singer, and Tim Q Duong. Prediction model and risk scores of icu admission and mortality in covid-19. PloS one, 15(7):e0236618, 2020. [39] Yuxin Zuo, Shang Qu, Yifei Li, Zhangren Chen, Xuekai Zhu, Ermo Hua, Kaiyan Zhang, Ning Ding, and Bowen Zhou. Medxpertqa: Benchmarking expert-level medical reasoning and understanding. arXiv preprint arXiv:2501.18362, 2025.

Appendix Appendix Contents A. Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 A.1. Experiment Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 A.2. Evaluation Results on RealICU-Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 A.3. Averaged Patient Trajectory on RealICU-Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 A.4. Per-Disease Performance on RealICU-GOLD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 A.5. Semantic Matcher Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 A.6. Token Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 B. Dataset Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 B.1. Dataset Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .23 B.2. RealICU-Gold Cross Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 B.3. Dataset Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 C. Memory-Augmented Agents for Clinical Decision Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 C.1. Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 C.2. Instantiations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 C.3. ICU-Evo as Heterogeneous Clinical Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 C.4. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 D. Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 D.1. Failure Case: Recall Safety Tradeoff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 D.2. Failure Case: Anchoring Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 D.3. Memory Snapshot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 E. Prompts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 E.1. Oracle Prompt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 E.2. Agent Prompt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

Performance Analysis

A.1

Experiment Setup

We evaluate ICU-Evo on RealICU-Gold and RealICU-Scale against three baselines sharing the same predictor: (i) full-context, all prior observations up to the window; (ii) local-window, the current window only; (iii) RAG, top-5 windows retrieved via PubMedBERT [9] embeddings. We set ka and ki to 12 windows (6 hours) for ICU-Evo. For Action Recommendation, we strip the current window’s recorded actions before prediction to prevent label leakage. We evaluate every window in RealICU-Gold and every fourth window along the trajectory in RealICU-Scale. We use two closed-source LLMs (Gemini-3.1-pro [8], GPT-5.4 [22]) and one open-source LLM (Qwen3-235B-A22B [34]) as backbones for evaluation. Evaluation results on RealICU-Gold and RealICU-Scale are reported in Table 2 and Table 4 respectively. Full-context Gemini and GPT runs on RealICU-Scale are omitted due to compute budget on stays beyond hundreds of hours. Oracle uses Gemini-3.1-pro [8] to generated hindsight annotations with access to the full patient trajectory. A.2

Evaluation Results on RealICU-Scale

Table 4 reports full evaluation results on RealICU-Scale across all three backbones and four systems. Full-context evaluation is omitted for Gemini-3.1-pro [8] and GPT-5.4 [22] due to prohibitive inference cost over multi-day ICU trajectories. Qwen3-235B [34] is included as a reference openweight upper bound. The results on RealICU-Scale largely recapitulate the pattern observed on RealICU-GOLD (Table 2). ICU-Evo achieves the strongest performance on Acute Problems and Action Recommendation across all three backbones, with particularly large margins on Acute Problems Hit (up to +0.268 over RAG for Gemini-3.1-pro [8]). The Red Flag HRR remains the consistent weak point of ICU-Evo regardless of backbone, suggesting the same premature anchoring failure mode (see Sec. 5.5), where current agent systems over-commit to early interpretation of the patient instead of updating hypothesis with new observations. Qwen3-235B [34] achieves lower overall performance compared to Gemini3.1-pro [8] and GPT-5.4 [22], suggesting that weaker instruction-following reduces the benefit of structured memory on tasks requiring precise categorical judgment. We further illustrate the temporal performance of LLM agents with full-context, local-window, retrieval-augmentation, memory-augmentation configurations in Figure 4 and Figure 5. Table 4: Evaluation results on the RealICU-Scale. Within each backbone, bold marks the best system per column and underline the second best. Patient Status

Acute Problems

Action Recom.

Red Flags

Backbone

System

Acc.↑

F1↑

Hit@5↑

R@5↑

Hit@5↑

R@5↑

HRR↓

Gemini-3.1-pro [8]

Full-context Local-window RAG ICU-Evo

– 0.405 0.442 0.519

– 0.264 0.312 0.348

– 0.487 0.568 0.827

– 0.265 0.315 0.518

– 0.447 0.466 0.514

– 0.307 0.331 0.3301

– 0.066 0.073 0.087

GPT-5.4 [22]

Full-context Local-window RAG ICU-Evo

– 0.415 0.411 0.438

– 0.265 0.269 0.327

– 0.475 0.584 0.852

– 0.266 0.321 0.562

– 0.451 0.509 0.575

– 0.308 0.435 0.368

– 0.073 0.096 0.090

Qwen3-235B [34]

Full-context Local-window RAG ICU-Evo

0.201 0.175 0.367 0.304

0.116 0.159 0.282 0.177

0.401 0.254 0.379 0.649

0.232 0.142 0.207 0.375

0.455 0.440 0.446 0.515

0.299 0.295 0.342 0.327

0.215 0.207 0.225 0.292

1.0

Local window RAG ICU-Evo

0.8

Patient Status

1.0 0.8

Hit@5

0.4

0.2

0.2 0

250

500

750

1000

1250

Hours Since Admission

1500

0.0

1750

Action Recommendation

1.0

Local window RAG ICU-Evo

0.8

250

500

750

1000

1250

1500

1750

1250

1500

1750

Hours Since Admission

Local window RAG ICU-Evo

0.8

Red Flags

0.6

HRR

Hit@5

0.6 0.4

0.4

0.2 0.0

Acute Problems

0.6

Accuracy

0.6

0.0

Local window RAG ICU-Evo

0.2 0

250

500

750

1000

1250

Hours Since Admission

1500

0.0

1750

250

500

750

1000

Hours Since Admission

Figure 4: Temporal performance over the full ICU stay on RealICU-Scale (GPT-5.4 [22]).

Full history Local window RAG ICU-Evo

0.8

Accuracy

0.6

Patient Status

1.0 0.8 0.6

0.4

0.2 0

200

400

600

800

Hours Since Admission

1000

0.0

1200

Action Recommendation

1.0

250

500

750

Full history Local window RAG ICU-Evo

0.8 0.6

1000

1250

1500

1750

1250

1500

1750

Hours Since Admission

Red Flags

HRR

Hit@5

0.6

1.0

Full history Local window RAG ICU-Evo

0.8

0.4

0.2 0.0

Acute Problems

0.4

0.2 0.0

Full history Local window RAG ICU-Evo

Hit@5

1.0

0.2 0

250

500

750

1000

1250

Hours Since Admission

1500

0.0

1750

250

500

750

1000

Hours Since Admission

Figure 5: Temporal performance over the full ICU stay on RealICU-Scale (Qwen3-235B [34]).

A.3

Averaged Patient Trajectory on RealICU-Scale

We visualize patient trajectories on RealICU-Scale using the Patient Status label in Figure 6. We map each window-level label to an ordinal score (deteriorating = -1, stable = 0, improving = 1) and normalize time within each ICU stay to the interval [0,1]. After binning each trajectory into 20 normalized time bins and averaging repeated observations within patient-bin pairs, we plotted all individual trajectories as low-opacity curves and overlaid outcome-stratified cohort means with 95% confidence bands. This highlights both patient-level heterogeneity and the average temporal separation between survivors and non-survivors. The survived and died cohorts are already separated at admission, with survivors hovering near stable and non-survivors sitting consistently below it, and the gap widens over the course of the stay as the survivor mean drifts toward improving while the non-survivor mean declines sharply in the final 20% of normalized ICU time. Both cohorts show substantial patient-level heterogeneity in the thin lines, which is expected given the diversity of admission diagnoses, but the cohort means recover the clinically intuitive ordering that survivors trend upward and non-survivors trend downward. This pattern indicates that the window-level labels produced by Oracle aggregate into a coherent patient-level signal, and supports the use of RealICU-Scale for trajectory-level analyses despite its labels being generated rather than physician-annotated. Survived mean (n=47) Died mean (n=47)

Patient status score

Improving

Stable

Deteriorating 0%

20%

40%

60%

Normalized ICU progression

80%

100%

Figure 6: Averaged patient status trajectories from Oracle on RealICU-Scale. Window-level Patient Status labels are mapped to an ordinal score (deteriorating = −1, stable = 0, improving = +1) and aggregated into normalized duration time. Thin lines show individual trajectories, and thick lines and shaded regions show cohort means with 95% confidence bands. The survived and died cohorts separate from admission onward and diverge further over the stay.

A.4

Per-Disease Performance on RealICU-GOLD.

Tables 5 report a breakdown across the six disease groups exceeding 8% prevalence in RealICU-Gold using Gemini-3.1-pro [8] backbone. The decomposition tests whether the gains of ICU-Evo are driven by a subset of phenotypes or hold across the case mix. The dominance of ICU-Evo on context-heavy tasks (Patient Status, Acute Problems, Recommended Actions) is consistent across nearly all disease groups. ICU-Evo achieves the best Hit@5 on Acute Problems for every group, with margins over the strongest baseline ranging from 0.134 (GI & Hepatic) to 0.358 (Sepsis & Infection), so the benefit of structured longitudinal memory is not phenotypespecific. The pattern is most pronounced on Cardiovascular and GI & Hepatic cases, where ICU-Evo wins on six of seven metrics, suggesting that diseases with protracted trajectories benefit most from explicit trend and trajectory memory. Respiratory and Sepsis & Infection expose the limits of the current memory design on Recommended Actions. RAG matches or surpasses ICU-Evo on Hit@5 and R@5 for these two groups, while ICU-Evo retains its lead on upstream tasks. Respiratory and septic management is dominated by recurring, protocol-driven interventions such as ventilator adjustments and antimicrobial escalation, for which lexical retrieval over recent context is competitive with longitudinal memory. This aligns with the backbone-dependent memory tolerance reported in the Qwen ablation and reinforces that memory architecture and task structure interact. We report more detailed per-disease results with GPT-5.4 [22] backbone in Table 6, and with Qwen235B [34] backbone in Table 7. Table 5: Per-disease performance on RealICU-GOLD (Gemini-3.1-pro [8] backbone). Disease groups are omitted with less than 8% proportion. Within each group, bold marks the best system per column and underline the second best. Patient Status

Acute Problems

Action Recom.

Red Flags

Disease Group

System

Acc.↑

F1↑

Hit@5↑

R@5↑

Hit@5↑

R@5↑

HRR@5↓

Cardiovascular

Full-context Local-window RAG ICU-Evo

0.337 0.313 0.409 0.472

0.269 0.246 0.370 0.374

0.584 0.450 0.593 0.828

0.367 0.245 0.337 0.526

0.299 0.400 0.442 0.595

0.177 0.273 0.284 0.375

0.066 0.059 0.066 0.089

Sepsis & Infection

Full-context Local-window RAG ICU-Evo

0.147 0.327 0.353 0.453

0.166 0.268 0.283 0.335

0.287 0.458 0.574 0.816

0.154 0.258 0.312 0.491

0.156 0.378 0.535 0.514

0.075 0.226 0.329 0.296

0.065 0.067 0.089 0.098

Injury & Poisoning

Full-context Local-window RAG ICU-Evo

0.303 0.303 0.474 0.490

0.199 0.209 0.379 0.399

0.469 0.512 0.636 0.817

0.338 0.306 0.369 0.534

0.225 0.390 0.526 0.451

0.145 0.257 0.361 0.299

0.057 0.090 0.078 0.095

Respiratory

Full-context Local-window RAG ICU-Evo

0.340 0.270 0.290 0.320

0.248 0.218 0.260 0.278

0.621 0.456 0.594 0.851

0.390 0.255 0.340 0.559

0.379 0.425 0.505 0.551

0.207 0.241 0.284 0.323

0.112 0.056 0.084 0.120

GI & Hepatic

Full-context Local-window RAG ICU-Evo

0.438 0.350 0.412 0.500

0.396 0.311 0.391 0.437

0.527 0.510 0.645 0.779

0.338 0.298 0.408 0.512

0.347 0.427 0.587 0.592

0.226 0.317 0.360 0.382

0.115 0.122 0.073 0.098

All Diseases (Table 2)

Full-context Local-window RAG ICU-Evo

0.298 0.315 0.402 0.459

0.258 0.239 0.348 0.365

0.486 0.459 0.596 0.823

0.308 0.258 0.342 0.526

0.259 0.395 0.496 0.676

0.152 0.260 0.313 0.534

0.137 0.151 0.216 0.300

Table 6: Per-disease performance on RealICU-GOLD (GPT-5.4 [22] backbone). Disease groups are omitted with less than 8% proportion. Within each group, bold marks the best system per column and underline the second best. Patient Status

Acute Problems

Action Recom.

Red Flags

Disease Group

System

Acc.↑

F1↑

Hit@5↑

R@5↑

Hit@5↑

R@5↑

HRR@5↓

Cardiovascular

Full-context Local-window RAG ICU-Evo

0.350 0.262 0.300 0.275

0.250 0.190 0.243 0.246

0.589 0.484 0.578 0.853

0.396 0.286 0.313 0.558

0.518 0.327 0.487 0.705

0.387 0.297 0.456 0.562

0.157 0.114 0.138 0.136

Sepsis & Infection

Full-context Local-window RAG ICU-Evo

0.153 0.153 0.227 0.347

0.164 0.213 0.292 0.266

0.314 0.528 0.609 0.864

0.191 0.293 0.329 0.539

0.137 0.368 0.430 0.700

0.087 0.233 0.327 0.558

0.111 0.114 0.095 0.158

Injury & Poisoning

Full-context Local-window RAG ICU-Evo

0.249 0.319 0.328 0.328

0.157 0.150 0.182 0.267

0.510 0.479 0.545 0.922

0.373 0.299 0.364 0.648

0.404 0.482 0.545 0.618

0.306 0.318 0.431 0.455

0.161 0.138 0.157 0.104

Respiratory

Full-context Local-window RAG ICU-Evo

0.340 0.322 0.378 0.270

0.254 0.189 0.290 0.237

0.653 0.533 0.678 0.923

0.438 0.297 0.424 0.631

0.543 0.386 0.319 0.767

0.477 0.211 0.247 0.623

0.176 0.137 0.113 0.182

GI & Hepatic

Full-context Local-window RAG ICU-Evo

0.425 0.100 0.214 0.363

0.349 0.165 0.256 0.347

0.575 0.561 0.682 0.827

0.441 0.373 0.512 0.570

0.474 0.331 0.512 0.657

0.323 0.279 0.412 0.490

0.306 0.098 0.148 0.157

All Diseases (Table 2)

Full-context Local-window RAG ICU-Evo

0.294 0.233 0.288 0.312

0.233 0.184 0.256 0.264

0.510 0.500 0.599 0.867

0.348 0.293 0.349 0.570

0.404 0.380 0.480 0.676

0.300 0.281 0.398 0.534

0.298 0.165 0.234 0.473

Table 7: Per-disease performance on RealICU-GOLD (Qwen3-235B [34] backbone). Disease groups are omitted with less than 8% proportion. Within each group, bold marks the best system per column and underline the second best. Patient Status

Acute Problems

Action Recom.

Red Flags

Disease Group

System

Acc.↑

F1↑

Hit@5↑

R@5↑

Hit@5↑

R@5↑

HRR@5↓

Cardiovascular

Full-context Local-window RAG ICU-Evo

0.218 0.156 0.307 0.268

0.206 0.164 0.278 0.219

0.455 0.188 0.350 0.552

0.249 0.109 0.189 0.316

0.390 0.351 0.451 0.552

0.270 0.246 0.332 0.363

0.129 0.087 0.090 0.134

Sepsis & Infection

Full-context Local-window RAG ICU-Evo

0.107 0.147 0.333 0.140

0.122 0.186 0.282 0.130

0.233 0.224 0.348 0.575

0.137 0.113 0.176 0.328

0.156 0.322 0.443 0.530

0.108 0.198 0.277 0.353

0.083 0.098 0.093 0.098

Injury & Poisoning

Full-context Local-window RAG ICU-Evo

0.269 0.197 0.362 0.305

0.153 0.165 0.252 0.220

0.432 0.301 0.516 0.669

0.289 0.221 0.328 0.457

0.291 0.373 0.481 0.452

0.232 0.280 0.357 0.321

0.093 0.065 0.102 0.087

Respiratory

Full-context Local-window RAG ICU-Evo

0.280 0.140 0.340 0.380

0.145 0.117 0.311 0.205

0.444 0.253 0.451 0.658

0.266 0.147 0.272 0.404

0.419 0.354 0.463 0.675

0.236 0.222 0.331 0.457

0.116 0.052 0.083 0.162

GI & Hepatic

Full-context Local-window RAG ICU-Evo

0.375 0.175 0.338 0.262

0.315 0.150 0.336 0.258

0.459 0.214 0.329 0.646

0.295 0.138 0.199 0.413

0.382 0.334 0.459 0.508

0.270 0.273 0.369 0.394

0.188 0.123 0.099 0.132

All Diseases (Table 2)

Full-context Local-window RAG ICU-Evo

0.225 0.152 0.315 0.253

0.188 0.154 0.271 0.197

0.384 0.213 0.379 0.600

0.226 0.126 0.211 0.362

0.329 0.352 0.453 0.526

0.222 0.242 0.324 0.357

0.117 0.080 0.095 0.117

A.5

Semantic Matcher Calibration

We adopt PubMedBERT [9] (NeuML/pubmedbert-base-embeddings) to generate embeddings for semantic match for Acute Problems, Action Recommendation, and Red Flags tasks. Calibration set. We sampled 100 action-string pairs from held-out ICU windows and asked a board-certified intensivist to label each pair as a binary classification for semantic match or non-match. The set is balanced by construction with 50 matched pairs and 50 non-matched pairs. Threshold sweep. Table 8 reports precision, recall, F1, and accuracy at seven candidate thresholds, and Figure 7 visualises the trade-off. PubMedBERT cosine similarity separates the two classes almost perfectly (AUROC = 0.996). Precision reaches 1.00 for all τ ≥ 0.5, while recall decays monotonically as τ increases. We select τ ∗ = 0.5, which maximises F1 (0.958) and eliminates false positives while retaining 92% of true matches. This operating point is used in all reported evaluations. Table 8: Semantic matcher performance on the 100-pair calibration set across candidate thresholds. The selected threshold (τ ∗ = 0.5) maximises F1 and achieves perfect precision. τ

Accuracy

Precision

Recall

0.3 0.4 0.5∗ 0.6 0.7 0.8 0.9

0.86 0.95 0.96 0.82 0.66 0.53 0.50

0.78 0.91 1.00 1.00 1.00 1.00 0.00

1.00 1.00 0.92 0.64 0.32 0.06 0.00

0.88 0.95 0.96 0.78 0.48 0.12 0.00

Calibration Curve: Precision / Recall vs Threshold

1.0 0.8

Score

0.6 0.4 0.2 0.0 0.3

Precision Recall 0.4

0.5

0.6 Threshold

0.7

0.8

0.9

Figure 7: Evaluating PubMedBERT [9] matcher on the calibration set under different thresholds. The selected τ ∗ = 0.5 (dashed line) achieves the best overall performance.

A.6

Token Efficiency

We assess token efficiency from two complementary perspectives, namely the per-prediction cost and the longitudinal coverage delivered per input token. A direct comparison of raw token counts suggests that ICU-Evo is more expensive than the local-window and RAG baselines. This view, however, omits a central design objective of ICU-Evo, which is to surface broad trajectory context at every prediction step. We therefore report a coverage-normalized metric alongside the raw cost. Per-prediction cost. We report the average input and total tokens per prediction on the Qwen run, with RAG projected to match the call volume of the local-window baseline. ICU-Evo is not the cheapest configuration in raw tokens, yet it is substantially cheaper than full-context prompting while remaining more expensive than the local-window and RAG baselines, as shown in Table 9. Coverage-normalized efficiency. To account for the trajectory context that each mode actually surfaces, we define the covered windows per prediction as 1 for the local-window baseline, 1 + k for RAG with k retrieved windows, and window_index + 1 for ICU-Evo and the full-context baseline, reflecting the current window together with all accumulated prior context. We then report the covered windows per million input tokens and the input tokens consumed per covered window on the Patient Status task. Once normalized by timeline coverage, ICU-Evo becomes the most input-efficient mode, achieving the highest coverage density and the lowest input-token cost per covered window, as shown in Table 10. Taken together, these two views indicate that, although ICU-Evo consumes more tokens per prediction than the local-window baseline, it delivers substantially denser longitudinal context per input token, which reflects better token utilization for timeline-aware reasoning in the ICU. Table 9: Per-prediction token cost on RealICU-Scale with Qwen3-235B [34]. Mode Full-context Local-window RAG (projected) ICU-Evo

Predictions

Input tokens

Avg. input / pred.

Avg. total / pred.

11,065 11,862 11,862 11,862

405,300,948 25,382,499 83,797,771 254,272,587

36,629.10 2,139.82 7,064.39 21,435.90

36,763.84 2,319.49 7,318.39 21,971.55

Table 10: Coverage-normalized input efficiency on Patient Status. Mode Full-context Local-window RAG ICU-Evo

Covered windows

Windows / 1M input tok.

Input tok. / window

5,112,589 11,862 6,130 6,304,410

12,614.30 467.33 541.32 24,793.90

79.28 2,139.82 1,847.34 40.33

Dataset Details

B.1

Dataset Statistics

Cohort Statistics Figure 8 summarizes the demographic and clinical composition of the selected 94-patient cohort across six dimensions: disease category, ICU stay duration, age, sex, stay duration stratified by survival outcome, and mean event density per window stratified by outcome. To categorize each ICU stay with disease types, we extract the diagnosis code closest in time to ICU admission. ICD-9 and ICD-10 codes were then mapped to broad disease categories using a rule-based grouping based on ICD chapters, with sepsis-related codes grouped into a dedicated Sepsis and Severe Infection category. Specifically, the largest disease group was Cardiovascular Disorders (32.98%), followed by Sepsis and Severe Infection (15.96%), Injury/Poisoning (13.83%), Respiratory Disorders (10.64%), and Gastrointestinal/Hepatic Disorders (8.51%). The remaining categories were less common: Neurological Disorders (4.26%); Clinical Signs/Symptoms, Congenital Disorders, Infectious Diseases, and Oncology (2.13% each); and Endocrine/Metabolic Disorders, Hematologic Disorders, Musculoskeletal Disorders, Psychiatric Disorders, and Renal/Genitourinary Disorders (1.06% each). In the pie chart, disease categories below 5% were merged into Others 25

Disease Category Distribution

Patient age distribution

ICU stay duration distribution median = 208 h 72 h bins

Respiratory Disorders Injury/Poisoning 10.6%

13.8%

16.0%

33.0%

ICU stays

8.5%

Sepsis & Severe Infection

Cardiovascular Disorders

4 2 0

216

Patient sex distribution

432

648

864

1080

1296

ICU stay duration (hours)

1512

1750

10 Male

750

Age (years)

Mean event density per window by outcome

500 250 0

1000

Mean events per window

ICU stay duration (hours)

Patients

1250

1500

Female

1728 1872

ICU duration grouped by outcome 55

Others

18.1%

Count

Gastrointestinal/Hepatic Disorders

median = 69.3

20 Survived

Died

Survived

Died

Figure 8: Cohort Demographics and Clinical Characteristics of the 94-Patient RealICU Cohort RealICU-Gold Label Statistics Figure 9 summarizes the distributional properties of R EAL ICUG OLD. The coverage histogram exhibits a long-tail distribution, with windows concentrated within the first 120 hours after ICU admission and a long right tail extending past 1,200 hours, yielding a median position of 74.8 hours. The Patient Status distribution is dominated by Stable windows (63.0%), followed by Deteriorating (22.4%) and Improving (14.6%). For the set-valued tasks, Acute Problems is tightly concentrated around two concurrent problems per window, whereas Recommended Actions exhibits a heavier-tailed distribution with a small number of windows reaching twelve or more concurrent recommendations, reflecting the variable cognitive load of ICU management. Red Flag Actions remain rare by design, with a median of one per window and most windows containing zero or one event. RealICU-Scale Label Statistics Figure 10 summarizes the distributional properties of RealICUScale. The coverage histogram exhibits a long-tail distribution, with windows concentrated within the first 336 hours after ICU admission and a long right tail extending past 1,800 hours, yielding a median position of 207.8 hours. The Patient Status distribution is dominated by Stable windows (68.8%), followed by Deteriorating (23.1%) and Improving (8.2%). For the set-valued tasks, Acute Problems is tightly concentrated around two to three concurrent problems per window, whereas Recommended Actions exhibits a heavier-tailed distribution with a small number of windows reaching ten or more concurrent recommendations, reflecting the variable cognitive load of ICU management. Red Flag Actions has a median of one per window and most windows containing zero or one event. 23

Patient status label distribution

Coverage across ICU timeline median = 74.8

300

24 h bins

Improving

Deteriorating

250

14.6%

22.4%

Windows

200 150 100

63.0%

50 0

120

240

360

480

600

720

840

960

Hours since ICU admission

Active problems per window

1080

1200 1248

Stable

Red flags per window

Recommended actions per window median = 2

500

median = 2

300

median = 1

400

250

400

300

Windows

200

300

150

200

100

Number of active problems

Number of recommended actions

12 13 14

Number of red flags

Figure 9: RealICU-Gold statistics and label distribution for Patient Status, Active Problems, Recommended Action, and Red Flags.

1000

Patient status label distribution

Coverage across ICU timeline 24 h bins

800

Oracle windows

Improving

median = 207.8

Deteriorating

8.2%

23.1% 600 400 200

168

336

504

672

840

Active problems per window median = 2

4000

6000

3500

5000

3000

Recommended actions per window median = 2

Windows

2000

1000

500 0

Number of active problems

median = 1

3000

1500

2000

Red flags per window

4000

2000

3000

6000 5000

2500

4000

Stable

1008 1176 1344 1512 1680 1848

Hours since ICU admission

Windows

68.8%

Number of recommended actions

Number of red flags

Figure 10: RealICU-Scale statistics and label distribution for Patient Status, Active Problems, Recommended Action, and Red Flags.

B.2

RealICU-Gold Cross Validation

RealICU-Gold contains 930 windows in total. For each window, we invite at least two out of five senior physicians for annotation. And we run a cross-validation check after annotation to maintain the golden-standard labels. In Table 11, we report the detailed number of each labels before and after cross-validation. Only labels with agreements are kept into RealICU-Gold. Note that Active Problems, Action Recomm., and Red Flags are stored as sets with multiple labels per window. Table 11: Label-wise statistics of R EAL ICU-G OLD after cross-validation filtering. Task Patient Status Active Problems Action Recomm. Red Flags

B.3

N labels raw

N labels kept

Keep rate

930 2,170 2,328 1,220

921 2,066 2,198 1,058

99.0% 95.2% 94.4% 86.7%

Dataset Pre-processing

To obtain our underlying base dataset of trajectories that cover ICU stays as well as their preceding patient journey, we merge MIMIC-IV [16], MIMIC-ED [16], MIMIC-Note [16], MIMIC-IVECHO [16], MIMIC-IV-ECG and MIMIC-CXR [15]. By this, we include not only patient meta data such as demographics, insurance, etc., but a diverse holistic timeline of medication, online medical records, vital measurements, X-ray, electro- and echocardiograms, procedures, diagnosis, lab results, text reports, and transfers. We also include triaging data, subject to availability. From MIMIC-Note, we use the entire contents of the discharge summaries and the findings sections from radiology reports. In total, our resulting base dataset comprises 73,181 ICU stays from 50,920 patients. We arrange all charted information and measurements along a time axis together with patient age and time delta to the beginning of the specific ICU stay and sort them temporally ascending. Full duplicates are eliminated. Encoded categorical information from established ontologies and coding systems, e.g. for diagnosis (ICD) or medication (GSN), are resolved to their full-text descriptions. Text data is cleaned according to a permissive policy, only adjusting e.g. consecutive whitespace characters and unambiguous processing artifacts. Numerical data is also represented textually together with the respective unit of measurement and description. While we directly include all textually representable information and numeric measurements, we limit the integration of imaging and waveforms to their metadata, leaving the utilization of the X-ray, ECG, and ECHO contents to future work. We ensure that patient data is not leaked across our dataset splits. Further, we account for inaccurate charting and limitations of raw data collection by conservatively establishing an adversarial tolerance of 24h for key events such as discharge. In case of multiple records for the same event with different precision (e.g. death), usually originating from different tables in the raw dataset, we default to the most fine-grain timestamp.

Memory-Augmented Agents for Clinical Decision Support

We position ICU-Evo as an instance of the broader class of memory-augmented language agents. In the following, we discuss a generic formulation of the class, several specific instantiations from recent work, and the design choices that motivate ICU-Evo. C.1

Formulation

A memory-augmented agent processes a stream of inputs {x1 , x2 , . . . , xT } while maintaining an evolving memory state. At step t, the update and decision rules take the generic form Mt = U(Mt−1 , xt ),

(k)

= f (k) (Mt ),

(8)

where Mt is the memory state, U is an update operator that integrates the latest input into memory, and f (k) is a task-specific decision function realized as a prompted call of the underlying language model. Different memory systems differ primarily in the structure of Mt and in the choice of U, and the structural choices that define a memory system reduce to three questions. What types of content does Mt contain, at what temporal scale is each type maintained, and under what update policy does each type evolve? C.2

Instantiations

We describe three instantiations of Eq. 8 in which Mt takes increasingly heterogeneous forms. Compressive Stream Memory. AgentFold [36] sets Mt as an ordered sequence of summary blocks together with a high-fidelity record of the latest interaction. The update operator U is a learned folding policy that, at each step, either condenses the latest interaction into a fine-grained block or consolidates a contiguous span of prior blocks into a single coarse-grained block. This instantiation supports streaming inputs and adaptive scale, while committing all memory content to a single representational type (textual summary) under a single update rule (replacement by summarization). Cross-Task Experience Memory. Evo-Memory [32] sets Mt as an unordered set of prior task experiences, each encoded as a structured tuple (xi , ŷi , fi ), where fi is a feedback signal. The update operator U is append-with-pruning, and a separate refine action lets the agent reorganize or discard memory entries during decision-making. This instantiation targets cross-task transfer rather than within-task dynamics, and treats each task as the atomic unit of memory. Linked Note Memory. A-Mem [33] sets Mt as a collection of atomic notes, where each note is a tuple of raw content, timestamp, LLM-generated keywords, tags, and contextual description, a dense embedding, and a set of links to other notes. The update operator U is realized in two LLM-driven steps. On arrival of a new note, top-k retrieval over the embedding space surfaces candidate neighbors, and an LLM decides which neighbors deserve a semantic link. The same neighbors are then re-examined, and the LLM may rewrite the contextual description, keywords, or tags of any neighbor in light of the new note. This instantiation supports streaming inputs and introduces evolution of prior entries, while committing all memory content to a single note schema under a single LLM-driven update rule. Table 12: ICU-Evo’s memory components and the corresponding agent update operator. Component work

M M trend M event M traj M insight

Definition

Updated by

Recent raw observations at full resolution. Piecewise-constant segmentations of vitals and labs. Append-only log of critical events. Compressed episode-level narrative of the stay. Patient-specific hypotheses with supporting and counter-evidence.

Observation Agent Observation Agent Assessment Agent Assessment Agent Insight Agent

C.3

ICU-Evo as Heterogeneous Clinical Memory

ICU-Evo sets Mt as a tuple of five components, Mt = Mtwork , Mttrend , Mtevent , Mttraj , Mtinsight ,

(9)

defined in Table 12. Algorithm 2 formalizes the full inference loop and the pipeline of three agents that realize the update operator U at different temporal cadences over the shared memory state. At every window t, the Observation Agent ingests the new measurements xt and updates Mtwork by per-window overwrite and Mttrend by piecewise aggregation. Every ka windows, the Assessment Agent compresses the recent working and trend memory into a trajectory summary zt , appended to Mttraj as a multi-scale rollup, and detects newly emerging critical events Ẽt , appended to Mtevent under severity gating. Every ki windows, the Insight Agent proposes patient-specific hypotheses, gathers supporting and counter-evidence from Mtevent , and the Orchestrator commits the accepted hypotheses to Mtinsight via lifecycle transitions. The Predictor then queries the consolidated memory (k) state to emit task-specific predictions yt , decoupled from the memory update cycle. The five components form principled correspondences with prior designs, recombined under a common formulation. Mtwork and Mttraj mirrors the multi-scale summaries of AgentFold [36], the lifecycle-managed update of Mtinsight mirrors both the rewriting of prior notes in A-Mem [33] and the refine action of Evo-Memory [32]. Algorithm 2 ICU-Evo Memory-Augmented Agent System. Require: LLM backbone F; ICU stay s; window sequence {xt }Tt=1 ; static context c; agent periods ka , ki 1: Initialize memory M0 ▷ work, trend, event, traj, insight 2: for each window t = 1, . . . , T do work trend 3: Mtwork , Mttrend ← Observe Mt−1 , Mt−1 , xt ▷ Observation Agent; every window 4: if t mod ka = 0 then ▷ Assessment Agent fires every ka windows trend work , Mt−k 5: zt , Ẽt ← F Mt−k a :t a :t traj traj event 6: Mt ← Mt−1 ∪ {zt }; Mtevent ← Mt−1 ∪ Ẽt 7: end if 8: if t mod ki = 0 then ▷ Insight Agent fires every ki windows insight event 9: ∆H ← F Mt−1 , Mt−k ▷ propose/update hypotheses :t i 10: for each hypothesis h ∈ ∆H do 11: state(h) ← accept if s(h) > r(h) else reject 12: end for insight 13: Mtinsight ← Orchestrator Mt−1 , ∆H 14: end if 15: for each task k do ▷ Predictor decoupled from memory update (k) 16: yt ← F (k) Mt ; c 17: end for 18: end for (k) 19: return predictions {yt } for evaluation against RealICU labels

C.4

Discussion

The instantiations above demonstrate the flexibility of Eq. 8, yet alternative combinations remain possible. The heterogeneous decomposition we adopt reflects that clinical reasoning under partial observability proceeds along multiple simultaneous modes. A homogeneous memory forces a single answer to three independent questions: at what temporal scale to retain content, at what fidelity, and under what update policy. AgentFold [36] couples scale and fidelity under a uniform textual summary type, fitting neither append-only event logs nor lifecycle-managed hypotheses. A-Mem [33] couples all three under a uniform note schema and LLM-driven evolution rule, providing no mechanism for the distinct update policies that event detection and hypothesis lifecycle management each require. Evo-Memory [32] treats each task as the unit of experience, fitting cross-task transfer but leaving within-patient dynamics unaddressed. The heterogeneous memory structure of ICU-Evo in Eq. 9 offers a frameworks to align previous designs into real clinical reasoning over evolving patient states.

Case Study

D.1

Failure Case: Recall Safety Tradeoff

Patient background: Age 55. Female, entering ICU with severe subarachnoid hemorrhage. Course complicated by severe intracranial hypertension (ICP >40 mmHg) managed with mannitol, propofol, and right frontal EVD placement. Current 191.5–192.0 hours after ICU admission (day-8). Current-window events 11:16 Impaired Tissue Perfusion NCP – Therapeutic temperature management 11:16 Impaired Tissue Perfusion NCP – Pressors 11:17 Infection NCP – Monitor and assess respiratory secretions 11:17 Impaired Tissue Perfusion NCP – Monitor neuro assessment 11:17 Impaired Tissue Perfusion NCP – Monitor hemodynamic status q1h . . . (63 further entries)

Memory snapshot Working memory: 68 events from the current window. Trend memory: None at current window. Critical events: 08:16 VITALS Richmond-RAS Scale=−4.00; 11:00 VITALS Intra Cranial Pressure 21.0 mmHg; . . . Trajectory: This patient exhibits a highly coupled neuro-respiratory deterioration, demonstrating acute and severe hypoxemia simultaneously with sudden intracranial pressure spikes . . . Insight #6: This patient exhibits extreme ICP volatility, demonstrating sudden, severe ICP spikes that break through maximal medical management, including deep sedation, neuromuscular blockade, and targeted temperature management. Expect a higher-than-average risk of sudden, refractory intracranial crises and secondary cerebral hypoperfusion despite seemingly stable baseline ICP, requiring aggressive, anticipatory hyperosmolar or surgical interventions.

Gold annotation Patient Status. Deteriorating. Active Problems. Refractory intracranial hypertension and impending herniation; worsening hypernatremia and hyperosmolality (Na 155 mEq/L, osm 325 mOsm/kg); ventilator-associated pneumonia. Recommended Actions. Therapeutic temperature management; monitor and assess respiratory secretions. Red Flags. Routine hyperosmolar therapy (mannitol or hypertonic saline) — further routine administration is contraindicated and dangerous under current sodium and osmolality; emergency bolusing only as a last resort for acute herniation.

ICU-Evo predictions (recommended actions) 1. Urgent neurosurgery and neurocritical care consultation. [unmatch] 2. Administer hyperosmolar therapy. Boluses of mannitol or 3% hypertonic saline for sustained ICP >20– [red flag] 22 mmHg. 3. Maintain CPP >70 mmHg. Continuous norepinephrine to meet MAP goals. [match] 4. Initiate goals-of-care discussion. [unmatch] 5. Strict glycemic and electrolyte monitoring (q1–2h K, glucose). [unmatch]

Figure 11: Recall-safety tradeoff case study. ICU-Evo’s stored insight #6 prescribes “aggressive, anticipatory hyperosmolar interventions,” which propagates to prediction 2 — flagged as contraindicated by the gold annotation under current Na/osm. The trend layer carries no sodium signal forward, so nothing counterweighs the stored frame.

D.2

Failure Case: Anchoring Bias

Patient background: Adult patient with Eisenmenger’s physiology, severe right ventricular dysfunction, and chronic right-to-left shunting with polycythemia. Baseline SpO2 85–93%. Prior course marked by refractory hypoxemia (SpO2 nadir 77% on FiO2 100%) unresponsive to maximal high-flow oxygen and ongoing diuresis. Current 87.5–88.0 hours after ICU admission (day-4). Current-window events 18:48 BODY_INPUT Oral/Gastric Ingredient = 300 ml 18:48 BODY_INPUT PO Intake = 300 ml 18:48 BODY_INPUT Water = 300 ml 18:48 VITALS Weight = 51.8 kg (4 events total; no vitals stream this window)

Memory snapshot Working memory: 4 events from the current window. Trend memory: No vital signal this window (HR/RR/SpO2 /BP/MAP/Temp all empty). Critical events: 02-03 13:00 VITALS SpO2 77% on FiO2 100%; 02-04 06:00 VITALS O2 Flow 35 L/min; 02-04 22:00 BODY_OUTPUT Void 700 ml; . . . Trajectory: This patient demonstrates a coupled refractory-hypoxemia and preload-dependence pattern, where standard oxygen escalation fails to correct shunt-driven desaturation and successful volume offloading precipitates hemodynamic collapse . . . Insight #2: This patient demonstrates refractory hypoxemia that is unresponsive to maximal high-flow oxygen therapy and ongoing diuresis. Expect below-average improvement in oxygenation indices from standard volume offloading, likely due to underlying pulmonary hypertension and fixed V/Q mismatch or shunting.

Gold annotation Patient Status. Stable. Active Problems. Hypotension and reduced cardiac output (preload-sensitive); volume overload risk under chronic right-heart strain. Recommended Actions. BODY_INPUT Water (oral intake); VITALS Weight (daily tracking). Red Flags. Titrating supplemental oxygen to SpO2 > 95% — baseline SpO2 is 85–93% from chronic right-to-left shunting and polycythemia; oxygen should be administered only if SpO2 drops below 85% or the patient becomes symptomatic.

ICU-Evo predictions (recommended actions) 1. Hold or reduce diuretic therapy. Suspend or decrease loop diuretics and reassess volume status before [unmatch] further diuresis. 2. Titrate norepinephrine to maintain MAP > 65 mmHg, weaning cautiously if hemodynamics remain stable. [unmatch] 3. Monitor serum potassium via basic metabolic panel or venous blood gas. [unmatch] 4. Maintain targeted oxygenation. Continue 4 L/min nasal cannula to target SpO2 88–92%, avoiding over- [red flag] oxygenation.

Figure 12: Premature-anchoring case study. The window contains four events — oral water and a daily weight — and the gold status is stable. ICU-Evo’s stored insight #2 carries forward the prior cardiopulmonary story of refractory hypoxemia, propagating to prediction 4, which prescribes an active oxygen target (88–92%) that the gold annotation marks as contraindicated under this patient’s Eisenmenger baseline. The trend layer is empty for the current window, so nothing pulls the model back to the simpler window-grounded interpretation.

D.3

Memory Snapshot

ICU-Evo memory snapshot — window 120, 87.5–88.0 h Working Memory recent windows, raw events previous windows: 17 events (windows 118–119, omitted) current window (window 120, 87.5–88.0 h, 4 events): 2152-02-06 18:48 BODY_INPUT Oral/Gastric Ingredient, ml = 300.00 2152-02-06 18:48 BODY_INPUT PO Intake, ml = 300.00 2152-02-06 18:48 BODY_INPUT Water, ml = 300.00 2152-02-06 18:48 VITALS Weight = 51.80 Trend Memory vital-sign aggregates, two scopes current window none global (windows 0–120, 0.0–88.0 h, 3422 raw events): heart_rate_bpm: mean = 68.97, min = 58.00, max = 85.00, count = 97 resp_rate_per_min: mean = 13.16, min = 8.00, max = 30.00, count = 96 spo2_percent: mean = 89.87, min = 77.00, max = 100.00, count = 97 map_mmhg: mean = 75.00, min = 47.00, max = 99.00, count = 88 . . . (sbp, dbp, temperature omitted) Critical Events Memory salient events that change patient story previous episodes: 38 events (episodes 1–9, hours 0.0–79.5, omitted) current episode (episode 10, hours 80.0–87.5): (no critical events extracted) Trajectory Memory episode-level summaries episode 1 (hours 0.0–7.5): The patient was admitted to the MICU for management of acute decompensated heart failure, acute kidney injury, and hypercapnic respiratory failure. Respiratory support was initiated with high-flow nasal cannula at 35 L/min and 65% FiO2, . . . ... episode 10 (hours 80.0–87.5): The patient began the block with stable hemodynamics (MAP 70 mmHg) and borderline oxygenation (SpO2 88%) on 4 L/min nasal cannula. Throughout the period, mean arterial pressures were maintained between 70 and 80 mmHg, demonstrating sustained hemodynamic stability. Respiratory status remained stable, with oxygen saturations ranging from 88% to 94% on unchanged nasal cannula support . . . Insight Memory personalized hypotheses with supporting and counter evidence insight #1: This patient exhibits a paradoxical and rapid escalation in serum potassium despite ongoing loop diuretic therapy. Expect an above-average risk of severe hyperkalemia and resistance to standard potassium-wasting effects of furosemide. supporting: 03:20 LAB_TEST Potassium = 7.30 mEq/L counter: 03:20 LAB_TEST Creatinine = 1.90 mg/dL insight #2: This patient demonstrates refractory hypoxemia that is unresponsive to maximal high-flow oxygen therapy and ongoing diuresis. Expect below-average improvement in oxygenation indices from standard volume offloading, likely due to underlying pulmonary hypertension and fixed V/Q mismatch or shunting. supporting: 13:00 VITALS O2 saturation pulseoxymetry, =77.00 % 13:00 VITALS Inspired O2 Fraction =1 00.00 counter: 06:00 VITALS Inspired O2 Fraction =60.00 06:00 VITALS O2 Flow = 35.00 L/min . . . ...

Figure 13: ICU-Evo memory snapshot at 87.5–88.0 hours after admission. The five layers of memory together constitute the full state available to the prediction modules at this window, including working memory, trend, critical events, trajectory, and patient-specific insights. Red highlights mark the thread most relevant to the case study in Figure 12.

Prompts

E.1

Oracle Prompt

Oracle Prompt You are Oracle, a clinical AI evaluator with hindsight access to a patient’s full ICU trajectory. Your task is to evaluate a specific local observation window {window_time} across two parts.

=== PART 1: PATIENT ASSESSMENT === [1A. CURRENT STATUS] Assess the patient’s clinical direction at this window by reasoning across four domains: Hemodynamics, Respiratory, Renal/Metabolic, Neurology Synthesize your domain reasoning into a single overall status label: - improving: indicators trending toward recovery relative to the provided context - stable: no meaningful change in either direction - deteriorating: indicators trending toward worsening relative to the provided context - insufficient_data: available information is not sufficient to make a reliable judgment (e.g., sparse events, conflicting signals) Important nuances: - Do NOT conflate outcome with care quality. A patient may be deteriorating despite excellent care. - A patient may be labeled stable or improving even if they eventually die, if the trajectory at this window genuinely reflects that direction.

[1B. ACTIVE PROBLEMS] Using the full trajectory and current window, identify active clinical problems or emerging risks this patient faces going forward from this window. - Only include risks that are real and imminent or already developing --- not distant or hypothetical - Each risk must be tied to specific trajectory evidence - An empty list is expected and acceptable when no urgent risks are present

=== PART 2: ACTION REVIEW === [2A. ACTION EVALUATION] Evaluate each clinical action taken during this window. For each action, integrate two perspectives into a single judgment: - Guideline alignment: does this action follow established ICU guidelines (e.g., Surviving Sepsis Campaign, ARDSNet, PADIS, AHA/ACC where relevant)? - Contextual appropriateness: given this patient’s specific condition, trajectory, comorbidities, and the eventual outcome known to you, was this action appropriate? Assign one overall label: - best_practice: action is both guideline-aligned and well-suited to this patient’s specific situation - acceptable: action is reasonable given the context, even if not optimal or if guidelines are ambiguous - potentially_harmful: action poses real risk of harm to this patient, whether due to guideline violation, patient-specific contraindication, or both - insufficient_data: not enough context to evaluate this action reliably Use your hindsight knowledge to inform the judgment, but be fair: judge the action against what the context reveals, not against impossible foresight. If an action was reasonable at the time but later context revealed a missed diagnosis, note this nuance explicitly in the rationale. The action can be identified by its code, which includes and are not limited to: DRUG_START, DRUG_STOP, DRUG_PRESCRIPTION, BODY_INPUT, TRANSFER, LAB_TEST, DIAGNOSIS.

[2B. RED FLAGS] Using the full trajectory and current window, identify any actions that should be strictly avoided for this specific patient going forward. - Only flag actions that a reasonable clinician might consider but would be harmful for this specific patient --- do not list generic contraindications unless directly applicable here - Each flag must be justified by patient-level evidence (comorbidities, trajectory events, organ function, known sensitivities) - An empty list is expected and acceptable when no red flags are present

=== PATIENT ICU CONTEXT WINDOW ===

{patient_icu_trajectory}

Now, evaluate the CURRENT OBSERVATION WINDOW according to the instructions above.

E.2

Agent Prompt

Assessment Agent Prompt You are an ICU assessment agent. You compress {k} consecutive 30-minute windows ({duration} hours total) into one episode summary and identify the block’s critical events.

=== INPUT === [PATIENT METADATA] (read-only context) {patient_metadata} [PRIOR EPISODE SUMMARY] (read-only context) Used only to distinguish new from continuing findings and to calibrate baseline. Do not cite from it. {prior_episode_summary_text} [EPISODE TIME RANGE] {episode_start_time} to {episode_end_time} [WINDOWED ICU DATA] You receive: - Raw events for each window, formatted as ‘[<event_id>] <time> <event_name> <payload>‘ - One grouped ‘selected vital trends‘ section summarizing tracked vitals across the whole block - In the trend section, only windows with at least one value for that vital are shown, plus an overall summary across the block {episode_input}

=== TASK === Produce (1) an episode summary and (2) a list of critical events.

[TASK 1: EPISODE SUMMARY] (3--6 sentences) Narrate the block’s clinical trajectory: where the patient started the block, what meaningfully happened, where they ended, and what remains unresolved. Focus on direction of travel and inflection points, not a window-by-window recap. - Use specific numeric values at inflection points; avoid adjective-only descriptions ("rising", "unstable") without numbers. - Include sustained abnormal states only when their persistence is the point (e.g., tachycardia held across the block despite intervention). - When the prior summary establishes a condition or intervention, frame current findings as continuation, escalation, or resolution, not new onset. - Stay descriptive and temporal. Do not assign diagnoses, syndromes, or mechanisms that are not explicitly in the input. - Refer to events by clinical name in prose. Event IDs appear only in ‘supporting_event_ids‘.

[TASK 2: CRITICAL EVENTS] A critical event is a high-SNR inflection point in the patient’s ICU trajectory: a moment that materially changes the clinical story. Reading only the critical events, a clinician should be able to reconstruct the shape of the block. Critical events typically fall into one of these categories: - New or worsening organ dysfunction (respiratory, cardiovascular, renal, hepatic, neurological). - Resolution or meaningful improvement of existing organ dysfunction. - Initiation of a major intervention (intubation, vasopressor start, dialysis, transfusion for active bleed, emergency procedure). - Diagnosis scores, such as GCS, RASS, or SOFA scores. - Significant escalation or de-escalation of care reflecting a change in trajectory. An event qualifies only if it is clinically meaningful on its own, changes how

subsequent data should be read, and is corroborated by surrounding trend or events rather than an isolated outlier. Exclude routine readings, scheduled medications without clinical context, or noisy measurements. Err toward under-listing. If no event meets the bar, return an empty list.

Insight Agent You are a clinical insight agent. Your job is to identify how THIS patient deviates from the population-average ICU patient in similar circumstances, specifically in how they respond to illness and interventions, or how their physiology is trending relative to what the current illness and treatment would predict. You are generating patient-specific response profiles and trajectory deviations that would change how future data should be interpreted or how future decisions should be weighted. You will be given: - The patient’s existing hypothesis bank (all hypotheses, active and retired) - Patient metadata with compressed pre-ICU history, for background context only - The latest episode: - Clinical trajectory summary - Critical events, formatted as ‘<event_id> <time> <event_name> <payload>‘ - Vital trend statistics

=== INPUT === [EXISTING HYPOTHESES] {hypothesis_bank} [PATIENT METADATA] {patient_metadata} [LATEST EPISODE SUMMARY] {episode_summary} [LATEST EPISODE CRITICAL EVENTS] {critical_events} [LATEST EPISODE VITAL TRENDS] {vital_trends}

=== TASK === [TASK 1: EVIDENCE FOR EXISTING HYPOTHESES] This is a matching task. Scan the episode for evidence bearing on each active hypothesis. Do not infer beyond what is stated. For each hypothesis where the episode provides relevant signal, report: - ‘supporting_evidence‘: event IDs or ‘vital_trend‘ that reinforce it - ‘counter_evidence‘: event IDs or ‘vital_trend‘ that weaken it Updates must cite at least one piece of evidence. No citation, no update. Skip hypotheses with no relevant signal.

[TASK 2: NEW HYPOTHESIS GENERATION] This task requires reasoning. Before proposing a new hypothesis, mentally construct the population-average trajectory for a patient with this illness receiving these interventions, then compare to what you observe. Only flag a new hypothesis where this patient’s pattern meaningfully departs from that reference. A valid new hypothesis must: - Describe an individualized response profile or trajectory deviation (not a diagnosis, not an event restatement) - Be grounded in evidence - Include an expected deviation from population-average response (e.g., "below-average," "slower than typical") - Be clinically actionable or prognostically relevant Over-generation is worse than under-generation. Generate at most 2 new hypotheses; if more candidates exist, report only the strongest. Do not flag obvious, trivial, or diagnosis-shaped claims. Empty list is acceptable and often correct.

[GROUNDING RULES] - Cite only event IDs that appear in the critical events block, or the token ‘vital_trend‘ for claims grounded in trend statistics. - Do not invent events, values, or clinical facts not present in the provided inputs.

[EXAMPLES OF VALID NEW HYPOTHESES] Example A --- individualized response profile: "This patient shows diminished hemodynamic response to fluid resuscitation. Expect below-average MAP rise to standard fluid boluses." Example B --- trajectory deviation: "This patient’s respiratory recovery is progressing slower than typical for their current ventilator settings and sedation level. Expect below-average improvement in oxygenation indices over the next shift."

E.2.1

Predictor Prompt

Shared Prompt You are a clinical decision support AI. Assess the patient’s overall clinical status for the current ICU window.

=== INPUT === You will receive one of: (A) Raw ICU events in chronological order (B) A structured Memory object with the following layers: - patient_metadata - working_memory : raw events from recent windows --- current local status - trend_memory : per-window vital trend statistics - critical_events_memory : episode-level critical events - trajectory_memory : episode summaries - insights : patient-specific deviations from typical ICU trajectories

Patient Status Predictor === INSTRUCTIONS === Assess the patient’s clinical direction at this window by reasoning across four domains: - Hemodynamics : heart rate, MAP, lactate, perfusion, vasopressor requirements - Respiratory : SpO2, PaO2/FiO2, respiratory rate, ventilator settings - Renal/Metabolic: creatinine, urine output, electrolytes, acid-base status - Neurology : GCS, mental status, RASS sedation score Synthesize your domain reasoning into a single overall status label, weighted by the relative clinical importance of each domain for this specific patient: - improving : indicators trending toward stability or recovery relative to the provided context - stable : no meaningful change in either direction - deteriorating : indicators trending toward worsening or decompensation relative to the provided context - insufficient_data : available information is not sufficient to make a reliable judgment (e.g., sparse events, conflicting signals)

Active Problems Predictor === INSTRUCTIONS === Identify active clinical problems or emerging risks this patient faces going forward from this window. - Only include risks that are real and imminent or already developing --- not distant or hypothetical - Each risk must be tied to specific trajectory evidence - An empty list is expected and acceptable when no urgent risks are present

Action Recommendation Predictor === INSTRUCTIONS === 1. Recommend up to {int(top_k_actions)} distinct actions that are clinically actionable in the next {float(prediction_horizon_hours):g}-hour horizon. 2. Only recommend actions that are clearly justified by the available data. 3. It is totally acceptable to return fewer than {int(top_k_actions)}. If data is insufficient to justify a recommendation with at least low confidence, omit it. 4. Order actions from highest to lowest clinical priority (rank 1 = most urgent). 5. Prioritize interventions with the highest expected impact on short-term stability and outcome. 6. Ground every recommendation strictly in the provided context. Do not infer or invent missing data.

Red Flag Actions Predictor === INSTRUCTIONS === Using the full trajectory and current window, identify any actions that should be strictly avoided for this specific patient going forward. - Only flag actions that a reasonable clinician might consider but would be harmful for this specific patient --- do not list generic contraindications unless directly applicable here - Each flag must be justified by patient-level evidence (comorbidities, trajectory events, organ function, known sensitivities) - An empty list is expected and acceptable when no red flags are present

Record · ID 180712 · SHA-256 7a6b93409d204d08
Retrieved via Conceptio — every document is proof-bundled with source, license, and retrieval metadata.