GLoC-EHR: Evidence-Cited Clinical Reasoning over Global Context and Local EHR Events
GLoC-EHR addresses clinical outcome prediction and reasoning over structured EHRs by equipping a language model with two complementary memories: a compact global trajectory summary and a bounded local set of selected events. The architecture aligns both routes with clinical text through hospital-course generation and masked-concept description recovery, then learns record-grounded, evidence-cited reasoning via GRPO with rewards for answer correctness and lexical evidence matching. Its core contribution is showing that a small multimodal LLM can match strong direct-prediction AUROCs while producing interpretable rationales tied to the patient record.
The paper delivers a carefully engineered multimodal EHR-to-language interface. Direct GLoC-EHR achieves the highest macro AUROC among compared methods on three MIMIC-IV tasks, and the evidence reward measurably suppresses unsupported citations relative to ablations. Nevertheless, the evaluation hinges on a single GRPO run, the lexical matcher leaves roughly 44% of evidence bullets unsupported, and the reasoning branch suffers a non-trivial AUPRC penalty (0.042 macro) for limited transparency gains. The teacher rationales are post-hoc and label-conditioned, raising the risk of spurious reasoning patterns.
The global-local architecture is principled and ablations support its specialization: removing the local route raises concept-description loss by over tenfold while leaving trajectory-summary loss nearly unchanged. Direct GLoC-EHR attains the highest macro AUROC and AUPRC (0.879 and 0.549) among all compared models on MIMIC-IV, far surpassing zero-shot LLMs. The evidence-aware GRPO objective also mechanistically reduces unsupported evidence, lowering the unsupported bullet rate from 0.58 after rationale SFT to 0.44, and cited-slot masking confirms the local route shapes which findings are cited without detectably changing answers.
First, the evidence reward and its evaluation rely entirely on a lexical matcher—IDF-weighted Dice overlap $\geq 0.55$ or simple containment—that cannot verify clinical correctness or catch paraphrases, and still permits a 44% unsupported bullet rate after optimization. Second, supervised fine-tuning uses teacher rationales generated by Qwen3.5-397B-A17B given the gold label; as the authors acknowledge, these are label-conditioned, post-hoc rationales, not independent annotations of clinical causality, which risks teaching the model to rationalize an outcome rather than infer it from the structured record. Third, GLoC-EHR (Reasoning) is a single GRPO run, so reported variance reflects only decoding chains, not training stability, and the reasoning model trails its direct multi-task counterpart by 0.042 macro AUPRC—a gap concentrated on the rarer mortality and long-LOS classes.
The comparison to zero-shot LLMs is fair and aligns with ClinicalBench findings, though those baselines face a much harder zero-shot textual-serialization regime. GLoC-EHR (Reasoning) stays within 0.008 macro AUROC of its direct multi-task counterpart, yet this narrow margin conceals a larger AUPRC deficit and a 44% unsupported evidence rate. Under stricter lexical matching ($\tau = 0.85$), the local route does increase distinct supported findings, but the absolute yield is modest (~2.7 per response). EHR-BERT remains competitive despite lacking an LLM backbone or description embeddings, suggesting the performance lift may stem partly from the rich multimodal interface rather than the evidence-citing mechanism itself.
Appendix A exhaustively documents hyperparameters, cross-attention placements, and reward coefficients, and the authors pledge to release code and model weights through PhysioNet upon acceptance. However, the full pipeline spans five stages with numerous tuned thresholds, and the GRPO run required 55 hours on four NVIDIA H200 GPUs. Because reasoning results derive from a single training run with uncertainty reported only over sampled decoding chains, training stability is unknown. Reproduction also depends on the specific teacher model (Qwen3.5-397B-A17B) used for rationale SFT, whose outputs are not fixed or independently audited.
Structured electronic health records (EHRs) contain a patient's clinical trajectory as a sequence of clinical codes. Answering clinical questions from such records requires both the context of the whole trajectory and the specific events that support the answer. We introduce GLoC-EHR, a multimodal language model that reads a contextual encoding of the record through a fixed-size global memory of the trajectory and a local memory of selected events. The model learns to generate hospital-course summaries from the global memory and descriptions of masked concepts from the local memory, aligning both with clinical text. It is then trained to cite evidence before answering, through rationale fine-tuning followed by group relative policy optimization (GRPO) with rewards for correct answers and record-supported evidence. On three MIMIC-IV outcome tasks, GLoC-EHR attains the highest macro AUROC among the compared models when it answers directly, whereas zero-shot LLMs reading the serialized record fall far behind. With evidence-cited reasoning, it stays close to its direct multi-task counterpart in macro AUROC, and the evidence terms of the objective reduce unsupported evidence at a similar macro AUROC. The local memory adds distinct supported findings, particularly under strict matching, without a detectable change in macro AUROC. Without retraining, GLoC-EHR transfers to EHRSHOT on par with EHR-BERT and answers two unseen laboratory questions better than zero-shot prompting of its own backbone.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.