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

Chengzhi Shen; Weixiang Shen; Tobias Susetzky; Chen; Chen; Jun Li; Yuyuan Liu; Xuepeng Zhang; Zhenyu Gong; Daniel Rueckert; Jiazhen Pan

Source-linked AI summary

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

Chengzhi Shen, Weixiang Shen, Tobias Susetzky, Chen, Chen, Jun Li, Yuyuan Liu, Xuepeng Zhang, Zhenyu Gong, Daniel Rueckert, Jiazhen Pan

arXiv:2605.13542v1May 13, 2026cs.AIcs.CLcs.LGcs.MA

TL;DR

ICU decision support requires reasoning over dense, evolving trajectories, but existing benchmarks often evaluate imitation of recorded clinician actions. RealICU addresses this gap with hindsight physician labels across four tasks and two datasets, finding persistent recall–safety and anchoring failures despite structured memory.

Problem
Existing ICU benchmarks often use recorded clinician actions as labels, limiting evaluation of clinical correctness under incomplete and evolving information.
Method
RealICU evaluates Patient Status, Acute Problems, Recommended Actions, and Red Flags using hindsight labels while models see only data available at decision time.
Results
Current LLM agents remain largely unreliable on RealICU, exhibiting a recall–safety tradeoff and anchoring bias; ICU-Evo improves long-horizon reasoning but not safety sufficiently.
Takeaways & Limitations
RealICU provides a clinically grounded testbed for evaluating sequential ICU decision support beyond behavioral imitation.
Takeaways & Limitations
RealICU is based on MIMIC-IV, may not transfer across staffing or documentation settings, omits multimodal data, and reports one experiment per LLM configuration.

Abstract

from arXiv · show

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: https://chengzhi-leo.github.io/RealICU-Bench/

1 Introduction

RealICU targets reliable ICU decision support by evaluating continuous bedside reassessment under dense, evolving data rather than imitating recorded clinician actions. It introduces hindsight-grounded labels, four physician-motivated tasks, two datasets, and analyses of persistent LLM failure modes.

Motivation: ICU physicians must integrate rapidly evolving laboratory, vital-sign, medication, nursing, and imaging data under time pressure.Each measurement captures only a partial slice of physiological state.
Benchmark gap: Existing benchmarks rarely test reasoning across changing ICU trajectories and often treat recorded clinician actions as ground-truth labels.Static question answering, summarization, and single-endpoint prediction provide limited evidence about sequential clinical reasoning.
Benchmark: RealICU evaluates Patient Status, Acute Problems, Recommended Actions, and Red Flags in dense 30-minute windows using hindsight physician judgment.Agents see only information available at each decision time, while labels use the full trajectory.
Findings: Current agents exhibit a recall–safety tradeoff and anchoring bias, while ICU-Evo improves long-horizon reasoning but does not make structured memory sufficient for safe decision support.Up to 47.3% of higher-recall recommendations were flagged as potentially harmful.
Datasets: 930 physician-consensus windows from 94 ICU stays form RealICU-Gold, while RealICU-Scale provides 11,862 windows labeled by physician-validated Oracle.The two subsets combine expert supervision with trajectory-level evaluation at scale.

2 Related Work

Prior clinical benchmarks largely test knowledge recall, static prediction, or conversational diagnosis, while memory-augmented agents use increasingly structured ways to manage long histories. These approaches motivate evaluating sequential ICU reasoning under realistic information constraints.

Clinical Benchmarks for LLMs and Agents: Exam-style benchmarks evaluate clinical knowledge as multiple-choice recall under complete information, revealing little about decisions under uncertainty.Conversational benchmarks instead require history gathering, investigation ordering, and multi-turn diagnostic convergence.
Memory-Augmented LLM Agents: Memory-augmented agents range from sequential ReAct traces to multi-scale summaries, iterative memory refinement, and selective retrieval over long histories.The surveyed designs address context accumulation through different memory mechanisms.

3 RealICU Benchmark

RealICU evaluates sequential ICU decision support with partial observations and hindsight labels across four tasks, using physician-consensus and Oracle-scaled datasets. Its framework defines trajectory windows, annotation protocols, model outputs, semantic matching, and safety metrics.

Benchmark design: RealICU compares model predictions from observations up to time t with labels produced by hindsight review of the full patient trajectory.This asymmetry separates real-time model access from retrospective clinical judgment.
Datasets: RealICU-Gold contains 930 physician-consensus windows from 94 ICU stays, while RealICU-Scale contains 11,862 densely labeled windows produced by Oracle.Oracle is validated against physician consensus for large-scale annotation.
Dataset Construction: Evaluation windows span 30 minutes and are sampled with a 2-hour stride while truncating trajectories before outcome-revealing events.The design preserves short-term dynamics while limiting redundancy and leakage.
Task definitions: The benchmark covers Patient Status, Acute Problems, Recommended Actions, and Red Flag Actions, with supporting evidence drawn from recorded history.Status is categorical, while the other tasks use free-text clinical concepts or actions.
Annotation Protocol: At least two senior ICU physicians independently label sampled windows, and windows without agreement are removed from the 930-window Gold set.Physician inter-rater reliability ranges from 0.826 to 0.985 across tasks.
Evaluation Framework: Patient Status uses accuracy and macro-F1, free-text tasks use Hit@5 and Recall@5, and Red Flags provide the Harmful Recommendation Rate safety check.Semantic matching uses PubMedBERT with a calibrated threshold.

4 ICU-Evo: An ICU Agent System with Evolving Memory

ICU-Evo models the partially observable ICU state with structured memory that is updated over time and used for task-specific predictions. Its components organize raw observations, trends, critical events, trajectory summaries, and patient-specific hypotheses, with specialized agents operating at different temporal scales.

ICU-Evo approximates the latent patient state with structured memory updated at each observation window, producing task-specific predictions from current and static patient context.
The memory decomposes into raw observations, vital and laboratory trends, critical events, trajectory summaries, and patient-specific insights.Critical-event memory logs abnormalities, interventions, and turning points, while trajectory memory periodically compresses the ICU stay into a narrative.
Observation Agent normalizes measurements, aligns them to 30-minute windows, and extracts vital-sign trends using Piecewise Aggregate Approximation.
Assessment Agent periodically summarizes recent observations and detects critical events from working and trend memory.
Insight Agent proposes patient-specific hypotheses with supporting and counter-evidence, accepting them when support exceeds rejection scores.
A task-specific prompted LLM predictor operates over the full memory state and static patient context, decoupled from the agent system.

5 Evaluation & Analysis

Evaluation shows that structured memory improves long-horizon clinical reasoning, but current ICU agents remain unreliable and unsafe. The analysis identifies performance gains alongside persistent agent–Oracle gaps, memory-component differences, and two major failure modes.

ICU-Evo with Gemini-3.1-pro reaches 0.459 accuracy on Patient Status and 0.534 Recall@5 on Action Recommendation, while Red Flags HRR@5 remains non-trivial.These results indicate that RealICU remains unsolved for current frontier LLMs and agent systems.
Structured Memory Consistently Improves Clinical Reasoning: With GPT-5.4, ICU-Evo improves over RAG by 26.8 Hit@5 points on Acute Problems and 19.6 on Action Recommendation.Similar margins occur with Gemini-3.1-pro and Qwen3-235B, and the pattern holds on RealICU-Scale.
Structured Memory Consistently Improves Clinical Reasoning: ICU-Evo’s Acute Problems Hit@5 stays near 0.8 for trajectories up to 1,800 hours, while non-memory baselines remain about 20 points lower and noisier.
The Agent-Oracle Gap: Beyond Behavioral Imitation: Oracle reaches F1 0.987 on Patient Status and 0.964 on Red Flags, whereas ICU-Evo with the same Gemini-3.1-pro backbone reaches 0.365 F1 on Patient Status and 0.300 HRR.The gap shows that the tasks are handled substantially better with full-trajectory hindsight than under real-time agent conditions.
Ablation Study: Removing working memory lowers Acute Problems Hit@5 from 0.823 to 0.761, while trajectory memory affects temporal tasks and Patient Status remains stable without it.Insight memory behaves differently: removing it can produce neutral or beneficial results.
Failure Mode Analysis: Oracle disagreements mainly involve stable–improving or stable–deteriorating boundaries and broad-versus-specific Acute Problems labels.The paper characterizes these disagreements as edge cases rather than systematic errors.
Failure Mode Analysis: The recall–safety tradeoff raises harmful recommendations: ICU-Evo’s Action Recommendation gain over RAG coincides with HRR@5 increasing from 0.234 to 0.473.Harmful cases concentrate in hemodynamic and pressor management, volume and diuresis, anticoagulation, and ventilation or sedation.
Failure Mode Analysis: Anchoring bias leads agents to preserve early interpretations despite later evidence; removing insight memory raises Qwen3-235B Action Recommendation Hit@5 from 0.526 to 0.601.

6 Discussion

RealICU exposes persistent failures in long-horizon ICU reasoning and shows that structured memory improves state tracking without resolving safety. Its findings also establish scope and reproducibility boundaries for the benchmark.

RealICU identifies a gap between frontier LLM medical knowledge and reasoning under partial observability across evolving ICU trajectories.
The benchmark reveals a recall–safety tradeoff and anchoring bias that persist across multiple context configurations.
ICU-Evo uses structured, evidence-grounded memory at multiple temporal scales, improving long-horizon reasoning but not preventing unsafe recommendations.
RealICU’s hindsight-based evaluation methodology is intended to support broader assessment of AI systems in high-stakes sequential decision environments.
The benchmark’s demographic and care-pattern distribution may not transfer to ICUs with different staffing or documentation conventions.

A.1 Experiment Setup

The experiments compare ICU-Evo with full-context, local-window, and retrieval-based systems under shared prediction settings, using three LLM backbones across RealICU datasets.

ICU-Evo is evaluated against full-context, local-window, and top-5 retrieval-augmented baselines sharing the same predictor.For ICU-Evo, both memory horizons are set to 12 windows, or 6 hours.
The study evaluates Gemini-3.1-pro, GPT-5.4, and Qwen3-235B on RealICU-Gold and RealICU-Scale.Oracle uses Gemini-3.1-pro to generate hindsight annotations from the full patient trajectory.

A.2 Evaluation Results on RealICU-Scale

On RealICU-Scale, ICU-Evo generally leads on context-heavy tasks, especially Acute Problems, while Red Flag performance remains weak and temporal analyses show coherent outcome-stratified trajectories.

ICU-Evo achieves the strongest performance on Acute Problems and Action Recommendation across all three backbones.Its Acute Problems Hit margin reaches +0.268 over RAG for Gemini-3.1-pro.
Red Flag HRR remains ICU-Evo’s consistent weak point across backbones, indicating persistent premature anchoring to early patient interpretations.
RealICU-Scale results largely recapitulate the pattern observed on RealICU-Gold, while full-context evaluation is omitted for Gemini-3.1-pro and GPT-5.4 because of inference cost.
Oracle-derived Patient Status trajectories separate survived and died cohorts from admission, with the gap widening as stays progress.Survivors trend toward improving status, whereas non-survivors decline sharply in the final 20% of normalized ICU time.

A.4 Per-Disease Performance on RealICU-GOLD.

Per-disease analyses show that structured longitudinal memory benefits most context-heavy tasks across nearly all disease groups, but its action-recommendation advantage is not universal and carries token costs.

Per-disease results: ICU-Evo achieves the best Acute Problems Hit@5 in every disease group, with margins over the strongest baseline ranging from 0.134 to 0.358.The margins span 0.134 for GI & Hepatic and 0.358 for Sepsis & Infection.
Per-disease results: The gains are most pronounced for Cardiovascular and GI & Hepatic cases, where ICU-Evo wins on six of seven metrics.These groups are described as having protracted trajectories that benefit from explicit trend and trajectory memory.
Per-disease results: RAG matches or surpasses ICU-Evo on Recommended Actions for Respiratory and Sepsis & Infection groups, limiting the memory advantage on these tasks.The passage attributes this competitiveness to recurring, protocol-driven interventions and interaction between memory architecture and task structure.
Matcher calibration: At threshold τ ∗= 0.5, the PubMedBERT matcher reaches F1 0.958, perfect precision, and 92% recall on the calibration set.
Token efficiency: ICU-Evo costs more raw tokens than local-window and RAG baselines but is substantially cheaper than full-context prompting.
Token efficiency: After normalizing for timeline coverage, ICU-Evo becomes the most input-efficient mode, with the highest coverage density and lowest input-token cost per covered window.

B.1 Dataset Statistics

RealICU combines a heterogeneous ICU trajectory dataset with hindsight-validated labels and structured-memory agents designed for evolving clinical states. The datasets capture long-tailed temporal coverage and task-specific label distributions across 94 patients.

Cohort composition: The cohort contains 94 ICU patients characterized by disease category, ICU stay duration, age, sex, survival-stratified stay duration, and event density.Cardiovascular Disorders comprise 32.98% of the cohort, followed by Sepsis and Severe Infection at 15.96%.
RealICU-Gold distributions: RealICU-Gold windows cluster within the first 120 ICU hours but have a long right tail beyond 1,200 hours, while Acute Problems concentrate around two concurrent problems.Recommended Actions are heavier-tailed, and Red Flag Actions have a median of one per window with most windows containing zero or one event.
RealICU-Scale distributions: RealICU-Scale windows cluster within the first 336 ICU hours but extend beyond 1,800 hours, with Acute Problems typically numbering two to three per window.Recommended Actions include a small number of windows with ten or more concurrent recommendations, while Red Flag Actions remain sparse.
Gold annotations: RealICU-Gold contains 930 windows annotated by at least two senior physicians, retaining only labels that pass cross-validation agreement.Active Problems, Recommended Actions, and Red Flags are represented as multi-label sets.
Trajectory construction: The underlying trajectories merge multiple MIMIC sources and include demographics, medications, records, vitals, imaging, procedures, diagnoses, laboratory results, reports, and transfers.Information is temporally aligned, duplicates are removed, coding-system entries are expanded to full-text descriptions, and key-event timestamps receive a conservative 24-hour tolerance.
Structured memory: ICU-Evo represents memory as five components and motivates heterogeneous memory because clinical reasoning requires different temporal scales, fidelities, and update policies.The framework contrasts with homogeneous summary, experience, and linked-note memories that do not separately address within-patient dynamics and hypothesis lifecycle management.

D.2 Failure Case: Anchoring Bias

The anchoring case shows ICU-Evo carrying an earlier refractory-hypoxemia interpretation into a sparse later window. That memory produces an oxygen target contraindicated by the patient’s Eisenmenger baseline, while the empty trend layer provides no corrective signal.

Case context: The current 87.5–88.0-hour window contains only oral water and a daily weight, with no vital-sign stream.The patient has Eisenmenger physiology, severe right ventricular dysfunction, chronic right-to-left shunting, and baseline SpO2 of 85–93%.
Anchoring mechanism: ICU-Evo’s stored Insight #2 carries forward the prior refractory-hypoxemia narrative despite the current window’s sparse observations.The memory snapshot includes working memory, trend, critical events, trajectory, and patient-specific insights.
Unsafe recommendation: The propagated prediction prescribes an active oxygen target of 88–92%, which the gold annotation marks as contraindicated under this patient’s Eisenmenger baseline.The patient’s prior refractory hypoxemia was unresponsive to maximal high-flow oxygen and ongoing diuresis.
Missing corrective evidence: An empty trend layer fails to pull the model back from the earlier interpretation in this window.The prompt separately requires reasoning across hemodynamics, respiratory, renal/metabolic, and neurological domains while distinguishing status from care quality.
Evaluation framing: The evaluation framework uses full-trajectory hindsight to judge actions, identify active risks, and assess current patient direction without conflating eventual outcome with care quality.It asks for temporally grounded summaries and critical events rather than unsupported diagnoses or mechanisms.