PaperlayerView exact revision on arXiv

Source-linked AI summary

GLoC-EHR: Evidence-Cited Clinical Reasoning over Global Context and Local EHR Events

Chaiho Shin, Kwangsoo Kim

arXiv:2610.01076v1cs.LG

TL;DR

Clinical reasoning over EHRs must connect a patient’s trajectory with question-specific findings, but linking encoded records to clinical questions and supporting evidence remains challenging. GLoC-EHR combines global and local EHR memories with evidence-aware training; across three MIMIC-IV tasks, it achieved the highest macro AUROC among compared models when answering directly.

  • Problem

    EHRs capture broad clinical trajectories and specific observations, but linking encoded records to clinical questions and answer-supporting evidence remains challenging.

    Source: §1 · pp. 1–2

  • Method

    GLoC-EHR connects a contextual EHR encoder to a language model through global trajectory and local event memories, aligns both with clinical text, and trains evidence-grounded answers with correctness and support rewards.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    GLoC-EHR achieved the highest macro AUROC among compared models on three MIMIC-IV tasks when answering directly, while evidence-cited reasoning stayed close to direct prediction.

    Source: §1 · p. 2

  • Takeaways & Limitations

    0.58 to 0.44: evidence terms lowered the share of unsupported evidence at similar macro AUROC, while local memory added distinct supported findings without detectable macro-AUROC change.

    Source: §1 · p. 2, §5 · p. 9

    Evidence matching is lexical and does not assess whether cited findings are clinically correct or whether the model relied on them to answer.

    Source: §5 · p. 9

Abstract

from arXiv · show

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.

1 INTRODUCTION

GLoC-EHR connects longitudinal EHR context and selected events to language-model reasoning, while training the model to cite record-supported evidence. The introduction motivates this design by noting that clinically relevant information is dispersed across records and that prior prediction and reasoning approaches leave gaps in task transfer, evidence support, or input efficiency.

Introduction

  • EHRs provide longitudinal trajectories and question-relevant observations, but clinically useful information can be dispersed across lengthy, heterogeneous records.This dispersion makes it challenging to identify and synthesize information for patient care.Source: §1 · p. 1
  • EHR encoders provide reusable representations, but new outcomes generally require task-specific fitting, and prediction scores do not state which recorded findings support an inference (Wornow et al., 2023).Connecting encoded records, clinical questions, and supporting evidence is necessary for natural-language reasoning.Source: §1 · pp. 1–2
  • Text serialization retains explicit events but can be costly, while compressed summaries may omit supporting details and event-aligned inputs still scale with event count.In downstream cohorts, serialized records average 4.7k tokens for a 24-hour window and 12.8k for a full index admission.Source: §1 · p. 2
  • Prior EHR reasoning rewards final-answer correctness without explicitly distinguishing findings supported by the patient record from unsupported ones (Lin et al., 2025; Liao et al., 2025).EHRMIND and EHR-R1 are cited as examples of reinforcement-learning approaches with this limitation.Source: §1 · p. 2

Contributions

  • GLoC-EHR combines a global memory of the trajectory with a local memory of selected events, aligning them to clinical text and rewarding evidence supported by the record.Trajectory alignment generates hospital-course summaries; concept alignment recovers descriptions of masked concepts. Its evidence-aware objective also includes a sufficiency loss on cited findings.Source: §1 · p. 2
  • GLoC-EHR achieves the highest macro AUROC across three MIMIC-IV tasks in direct prediction, reduces unsupported evidence without detectable macro-AUROC change, and transfers to EHRSHOT and unseen laboratory questions.The local route shapes cited findings without a detectable effect on answers; evidence-cited reasoning stays close to direct prediction, and the model matches EHR-BERT on EHRSHOT while outperforming its zero-shot backbone on unseen laboratory questions.Source: §1 · p. 2

2 RELATED WORK

Prior EHR work learns contextual or predictive representations, while EHR–language research connects structured records to language models in several ways. GLoC-EHR distinguishes itself with separate global trajectory and local event memories and a reward for record-supported findings, while treating evidence support as distinct from rationale faithfulness.

Structured EHR representations

  • Contextual EHR encoders learn reusable code representations, while autoregressive and time-to-event foundation models forecast future timeline events (Choi et al., 2016; Li et al., 2020b; Rasmy et al., 2021; Pang et al., 2021; Steinberg et al., 2021; 2024; Kraljevic et al., 2024; Renc et al., 2024).GLoC-EHR builds on contextual representations to generate language grounded in the record.Source: §2 · p. 2

EHR–language alignment

Evidence support and rationale faithfulness

3 METHOD

GLoC-EHR combines global trajectory context and local event evidence, aligns both routes with clinical text, then trains the model to produce evidence-cited rationales and answers. Its pipeline progresses from supervised fine-tuning to group relative policy optimization.

3.1 DESCRIPTION-INFORMED EHR ENCODER

  • The description-informed EHR encoder combines event concepts, attributes, frozen text-description embeddings, and learned code-ID embeddings, and is pretrained to predict masked concepts and align events with descriptions.Centering, whitening, and L2 normalization address anisotropy in the LLM-derived description embeddings; code-ID dropout encourages use of descriptions, and both inputs are masked at prediction targets.Source: §3.1 · p. 4

3.2 GLOBAL AND LOCAL EHR MEMORIES

  • Global memory summarizes the trajectory, while local memory preserves selected event representations as evidence accessed through task-independent selection.The global resampler produces a fixed-size memory; the local selector is adapted for rationale generation and then fixed for direct prediction and GRPO.Source: §3.2 · p. 4

3.3 TEXT ALIGNMENT OF THE TWO ROUTES

  • Trajectory alignment trains the global route on hospital-course text, then concept alignment trains the local route to recover masked concept descriptions while replaying hospital courses.The selector progresses from forced to soft to hard selection, with masked positions excluded from candidate events during soft and hard selection.Source: §3.3 · p. 5

3.4 EVIDENCE-GROUNDED REASONING WITH GRPO

  • SFT trains evidence-cited rationales and answers from label-conditioned teacher responses; GRPO then reinforces responses using answer-correctness and record-supported evidence rewards.Evidence bullets are matched against typed descriptions for events across the input window, not only local-memory events. GRPO also uses answer and evidence-sufficiency losses; the encoder and memory constructors remain frozen while the language model and cross-attention interfaces are updated.Source: §3.4 · pp. 5–6

4 EXPERIMENTS

On MIMIC-IV outcome prediction, GLoC-EHR achieves the strongest macro AUROC among compared models when answering directly, and evidence-cited reasoning stays close to its multi-task counterpart. Evidence terms improve citation support, while the local route adds distinct supported findings without a detectable macro-AUROC change; the model also transfers to EHRSHOT without retraining.

4.1 DATA AND TASKS

  • The study uses 128,350 MIMIC-IV patients with a BHC, split 7:1:2, to predict in-hospital mortality, long length of stay, and 30-day readmission; EHRSHOT is the external cohort.Mortality and long-stay predictions use the first 24 inpatient hours; readmission uses the full index admission.Source: §4.1 · p. 6

4.2 OUTCOME-PREDICTION COMPARISON

  • GLoC-EHR (ST) leads task-specific models at 0.879 macro AUROC and 0.549 macro AUPRC, while GLoC-EHR (MT) exceeds every baseline in macro AUROC.GLoC-EHR (Reasoning) trails GLoC-EHR (MT) by 0.008 macro AUROC and 0.042 macro AUPRC.Source: Table 1 · p. 7, §4.2 · p. 7

4.3 EVIDENCE GROUNDING

  • GRPO lowers the unsupported evidence-bullet rate from 0.58 after rationale SFT to 0.44; without evidence terms, it remains 0.137 [0.133, 0.140] above GLoC-EHR.At τ = 0.85, the local route yields 13.4% [12.4, 14.5] more distinct supported findings than the global-route-only variant, with no detectable answer change under cited-slot masking.Source: §4.3 · p. 8, Figure 2 · p. 8

4.4 ABLATIONS

  • Stopping after rationale SFT lowers macro AUROC from 0.866 to 0.658, while removing evidence terms or the local route does not detectably change macro AUROC.The SFT model is nearly deterministic: 99.8% of answer probabilities are below 0.02 or above 0.98, versus 31% after GRPO.Source: §4.4 · p. 8, Table 2 · p. 9

4.5 TRANSFERABILITY

  • Without retraining on EHRSHOT, GLoC-EHR (MT) remains on par with EHR-BERT (MT), and GLoC-EHR (Reasoning) stays close to both in macro AUROC.On held-out MIMIC-IV admissions, GLoC-EHR (Reasoning) scores AUROC 0.657 for low serum sodium and 0.629 for low platelet count, versus 0.556 and 0.525 for zero-shot Qwen3-1.7B.Source: Table 3 · p. 9, §4.5 · p. 9

5 CONCLUSION · A IMPLEMENTATION AND EVALUATION DETAILS · A.1 STRUCTURED INPUT AND ENCODER PRETRAINING

GLoC-EHR combines global trajectory context and selected local events to support evidence-cited clinical reasoning, with strong MIMIC-IV results and transfer to EHRSHOT. The implementation uses bounded structured inputs and description-aligned encoder pretraining; its evidence matching and external-cohort probabilities remain limited.

5 CONCLUSION

  • Across three MIMIC-IV tasks, direct GLoC-EHR matches or exceeds fine-tuned EHR encoders in macro AUROC; evidence-cited GRPO is 0.008 macro AUROC below the direct multi-task model, while rationale SFT reaches only 0.658.At similar macro AUROC, evidence terms reduce unsupported evidence from 0.58 to 0.44; the local route adds distinct supported findings without a detectable answer change. Without retraining, GLoC-EHR transfers to EHRSHOT on par with EHR-BERT and answers unseen laboratory questions better than its zero-shot backbone.Source: §5 · p. 9
  • Evidence matching checks whether cited findings appear in the record, not whether they are clinically correct or informed the answer; one training run and cohort limit stability evidence.The authors propose semantic matching and clinician review, further runs and cohorts, and recalibration; on EHRSHOT, mean predicted long-stay probability was 0.57 against prevalence of 0.31.Source: §5 · pp. 9–10

AI USE STATEMENT

  • Generative AI assisted translation, methodological feedback, drafting and editing, code and figures, and literature searches; authors reviewed the work and take responsibility for the final content.Teacher rationales were used as SFT targets, but AI tools were not used to generate synthetic datasets.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

  • The study used de-identified EHR data under data-use agreements, collected no new data, attempted no re-identification, and sent no patient data to external services.GLoC-EHR is not intended for clinical decision-making; the authors note risks from single-system training, uncalibrated external-cohort probabilities, and dataset or temporal shift, requiring prospective validation, recalibration, and subgroup evaluation before clinical use.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • Appendix B describes cohort construction, tasks, splits, and sample sizes; Appendix A and Sections 3–4 specify methods and evaluation procedures, and code and weights are slated for release upon acceptance.GRPO training took about 55 hours on four NVIDIA H200 GPUs; credentialed-data-trained weights will be released through PhysioNet under the same access terms as MIMIC-IV.Source: REPRODUCIBILITY STATEMENT · p. 10

A IMPLEMENTATION AND EVALUATION DETAILS

  • The appendix records the current implementation configuration, with architecture and learning objectives in Section 3 and numerical settings and evaluation procedures collected for reproducibility.Source: A IMPLEMENTATION AND EVALUATION DETAILS · p. 15

A.1 STRUCTURED INPUT AND ENCODER PRETRAINING

  • Each example contains at most 2,048 task-cutoff and truncated events, and the 20,561-token OMOP vocabulary uses age, within-record date rank, event type, and value or drug-duration suffixes.Typed concept descriptions are capped at 128 tokens, encoded by Qwen3-1.7B into 2,048-dimensional vectors, then centered, whitened, and normalized; the description table remains frozen during EHR pretraining.Source: §A.1 · pp. 15–16
  • The eight-layer, 12-head bidirectional EHR Transformer uses hidden size 768; pretraining samples capped patient views, masks 30% of eligible concepts, and combines masked prediction with description contrast.Views are sampled as a single visit (0.6), most recent trajectory (0.3), or random contiguous visit window (0.1); the objective is L_EHR = L_MLM + 0.1L_con.Source: §A.1 · p. 16

A.2 BACKBONE DESCRIPTION EMBEDDINGS · A.3 LANGUAGE INTERFACE CONFIGURATION

GLoC-EHR uses whitened Qwen3-1.7B description embeddings to represent concepts and combines a global memory of the contextual EHR sequence with a selected local-event memory. The embedding comparison favors whitened final hidden states for preserving concept identity across qualified variants, while the language interface specifies how global and local memories reach the backbone.

A.2 BACKBONE DESCRIPTION EMBEDDINGS

  • Descriptions are embedded with Qwen3-1.7B so encoder states and the language model share a semantic reference, after comparing methods for converting descriptions to vectors.Each of the 20,561 vocabulary tokens receives a typed short description; representations use either mean-pooled input embeddings or final-layer hidden states, with alternative post-processing.Source: §A.2 · p. 16
  • 98.4% of qualified tokens with sibling variants find them as nearest neighbours with whitened final hidden states, versus 91.1% with whitened input-token means and 75.3% with raw input-token means.Whitened final hidden states also have the lowest same-qualifier neighbour share, 31.9%.Source: §A.2 · p. 17, Table 4 · p. 17
  • Raw input-token means have effective rank 295 of 2,048 dimensions, while raw final hidden states collapse to 70 and have mean pairwise cosine 0.89.Whitening raises effective rank to 1,793 for input-token means and 1,908 for final hidden states; removing only the first principal component leaves the space anisotropic at effective rank 116.Source: §A.2 · p. 17, Table 4 · p. 17
  • Whitened description embeddings mainly encode concept identity, leaving value and exposure distinctions chiefly to the learned code-ID residual and surrounding events.For the sodium-low example, the nearest neighbours are the high, very low, and unqualified variants of the same test.Source: §A.2 · p. 17

A.3 LANGUAGE INTERFACE CONFIGURATION

  • The global resampler converts the contextual EHR sequence into a 64 × 2,048 memory that replaces the <EMR> placeholder and supplies global cross-attention keys and values.It uses four layers and 16 attention heads initialized from 64 learned queries; inserted memory embeddings may attend bidirectionally while ordinary text attention remains causal.Source: §A.3 · p. 17
  • The local selector ranks events independently of the downstream instruction and exposes selected local vectors only through local cross-attention, not as ordinary LLM tokens.Hard selection retains up to 256 events during concept alignment and 64 in downstream stages; the selector is learned during concept alignment, adapted during rationale SFT’s soft phase, then fixed.Source: §A.3 · p. 18
  • Global cross-attention modules attach after decoder blocks 12, 14, 16, 18, 20, and 22, while local modules attach after blocks 23, 25, and 27.Each module has a learned scalar gate initialized to −2, giving initial residual weight approximately 0.119; the global interface is available before the final local injections.Source: §A.3 · p. 18

A.4 ALIGNMENT OBJECTIVES AND TRAINING SCHEDULE

GLoC-EHR aligns global trajectory representations to brief hospital courses and local representations to masked-concept descriptions. Both alignment stages freeze the pretrained EHR encoder and base LLM while training selected interface components.

Global trajectory alignment

  • Global alignment generates a brief hospital course from a structured trajectory using the global route and a response-only next-token objective.Only the resampler and global cross-attention modules are updated; discharge text is an alignment target, not a downstream prediction input.Source: §A.4 · p. 18

Local concept alignment

  • Local concept alignment masks 5% of eligible concepts, up to 32 positions, and requests descriptions for up to eight masked concepts.Concept-description examples and brief-hospital-course replay are mixed at rates of 0.4 and 0.6, respectively.Source: §A.4 · p. 18
  • In the first local-alignment phase, masked positions enter local memory and binary cross-entropy trains the selector to recognize them.After forced selection, masked positions are excluded from the selector’s candidate set, preventing the local route from simply forwarding their representations.Source: §A.4 · pp. 18–19
  • In the soft phase, full-length candidate memory vectors are weighted by α_t = softmax(s_t/τ_s), allowing generation loss to train the selector.A squared budget loss weighted 0.01 encourages mean selection probability over valid candidates to approach 0.1.Source: §A.4 · p. 19
  • At hard selection, the local projection and selector are frozen because top-k indices are non-differentiable, while local cross-attention continues adapting on the downstream hard route.The pretrained EHR encoder and base LLM are frozen throughout both alignment stages; direct prediction and rationale SFT are separate downstream branches initialized from the aligned interface.Source: §A.4 · pp. 18–19

A.5 ROUTE SPECIALIZATION AND THE LOCAL MEMORY BUDGET · A.6 DOWNSTREAM SUPERVISION AND ANSWER SCORING

The sections show that the global and local routes serve different functions and motivate a downstream local-memory budget of 64 events. They also describe direct-answer supervision and rationale-based initialization and training.

A.5 ROUTE SPECIALIZATION AND THE LOCAL MEMORY BUDGET

  • Route probes used the final Stage-2 checkpoint on 27,076 admissions from 12,835 validation patients, comparing BHC generation and masked-concept description with routes disabled by forcing gate logits to −30.Local-memory truncation to Kℓ∈{256, 128, 64} was evaluated without retraining; route-off passes did not depend on Kℓ.Source: §A.5 · p. 19
  • Removing the local route raised concept-description loss from 0.280 to 3.311 (+1084%) but left BHC loss at 1.794 to 1.792 (−0.1%); removing the global route raised both losses.The reported increases without the global route were BHC +78.6% and concept description +92.3%.Source: §A.5 · p. 19
  • The local route specializes in recovering masked concepts, while the global route supports both objectives; the trajectory summary alone cannot recover individual events needed for evidence citation.BHC replay leaves the local route free to contribute to BHC generation, yet BHC loss is unchanged when that route is off.Source: §A.5 · p. 19
  • At Kℓ=64, concept-description loss was 0.301 versus 0.280 at Kℓ=256, while BHC loss was 1.7942 versus 1.7937; downstream stages therefore use Kℓ=64.At Kℓ=256, selection was inactive in 78.9% of 24-h records and 58.5% of readmission records; at Kℓ=64, it was active in 70.4% and 79.7%, respectively. Alignment uses Kℓ=256 to provide a broad candidate set.Source: §A.5 · p. 19, Table 7 · p. 20

A.6 DOWNSTREAM SUPERVISION AND ANSWER SCORING

  • The direct branch scores Yes and No answer slots without generating a rationale; the shared multi-task interface adapts only global and local cross-attention modules, with hard top-64 selection.The EHR encoder, resampler, local projection and selector, and base LLM remain frozen; evaluation averages answer-token log-likelihoods and excludes the terminating token.Source: §A.6 · p. 20
  • Rationale initialization used about 2.5K task- and label-balanced examples with Qwen-generated, gold-outcome-conditioned post-hoc rationales, then trained the student on structured EHR input and task instructions.The rationales were not independent annotations of clinical causality; the EHR encoder stayed frozen while the base LLM, resampler, and both cross-attention interfaces were updated.Source: §A.6 · p. 20
  • Rationale SFT used a soft phase that trained the local projection and selector through event-ranking gradients, followed by hard top-64 training that froze them because top-k indices pass no gradient.The best hard-phase checkpoint initialized both the GRPO policy and its reference.Source: §A.6 · p. 20

A.7 REWARD IMPLEMENTATION AND GROUP RELATIVE POLICY OPTIMIZATION

The section specifies a GRPO training procedure that rewards correct, well-formatted answers and record-supported evidence while penalizing invalid answers and outcome leakage. It also describes lexical evidence matching, distinct-concept credit, an evidence-sufficiency loss, and the policy update configuration.

A.7 REWARD IMPLEMENTATION AND GROUP RELATIVE POLICY OPTIMIZATION

  • Each patient–task prompt yields four sampled completions, rewarded for answer correctness and format, with invalid answers or detected outcome leakage overriding the reward to −1.5.Format checks cover the reasoning and evidence sections and a single final-answer tag; explicit event identifiers reduce the format score.Source: §A.7 · pp. 20–21
  • Evidence bullets are matched strictly to candidate events from the full truncated input using lexical anchors, direction and qualifier checks, and a precision-first support rule.Candidates include events outside local memory; lexical matching can miss paraphrases but containment can credit under-specified bullets.Source: §A.7 · p. 21
  • Distinct-concept credit rewards only the first supported citation of a concept, scales credit by its frequency in training records, and excludes unsupported, repeated, and metadata citations.A floor of four bullets prevents shortening below that count from increasing the score; credit does not depend on whether the matched event is in local memory.Source: §A.7 · p. 22
  • In the final run, GRPO shortened responses from 10.1 to 7.8 bullets while distinct supported findings rose.The distinct-concept credit design was motivated by earlier runs in which duplicate bullets inflated reward or shrinking lists reduced the number of distinct supported findings.Source: §A.7 · p. 22
  • The policy update combines group-standardized advantages, a clipped policy-ratio surrogate, reference-policy KL regularization, answer loss, and evidence-sufficiency loss.With one inner update, the importance ratio is one and clipping is inactive; the EHR encoder, resampler, local projection, and selector remain frozen while the base LLM and cross-attention interfaces are updated.Source: §A.7 · p. 22

A.8 EVIDENCE INTERVENTION AND ROUTE PROBES

The intervention found that cited local-memory slots affect which findings are cited, but masking them did not make answers more sensitive than a matched control, and the initial readmission AUROC effect did not replicate.

A.8 EVIDENCE INTERVENTION AND ROUTE PROBES

  • The protocol regenerates outputs after masking cited slots before the first token, compares a matched uncited-slot mask, and measures persistence, answer changes, and AUROC with the local route on or off.With the local route off, masked evidence can reach the answer only through regenerated text or global memory.Source: §A.8 · pp. 22–23
  • Of 2,280 chains reaching the answer anchor, 2,062 cited local-memory concepts, averaging 3.2 cited concepts across 4.3 slots of about 44 local concepts.Cited and control concepts had similar informativeness (0.826 vs. 0.833), compared with 0.715 for all local concepts.Source: §A.8 · p. 23
  • Masking cited slots reduced persistence by 0.017 [0.007, 0.028] versus control; readmission cited-minus-control AUROC was +0.002 [−0.012, 0.019] on 1,000 added cases (Table 8).The initial 500-case difference was −0.010 [−0.027, 0.004], and the pooled difference was −0.003 [−0.013, 0.007]. The answer was no more sensitive to the cited mask than to the control mask.Source: Table 8 · p. 23, §A.8 · p. 23
  • The intervention measures output sensitivity, not causal effects of clinical events; remaining global memory and unmasked local slots may provide redundant evidence.A global-cross-attention-off probe also leaves the 64 inserted global embeddings in the LLM sequence, so it does not remove all global information.Source: §A.8 · p. 23

B DATA AND COHORT CONSTRUCTION · C ADDITIONAL OUTCOME-PREDICTION DETAILS AND SCORES · C.1 ALIGNMENT AND INTERFACE CONTROLS

This block specifies the MIMIC-IV and EHRSHOT cohorts and outcome-task protocols, then reports controls in which performance was lower without alignment and under LoRA.

B DATA AND COHORT CONSTRUCTION

  • The internal cohort comprised 128,350 patients with extractable BHCs, split by patient into 89,845 training, 12,835 validation, and 25,670 test patients.The data came from MIMIC-IV v2.2 in OMOP CDM format; the split was 70/10/20 with a fixed seed.Source: B DATA AND COHORT CONSTRUCTION · pp. 23, 25
  • Mortality and long-LOS tasks use the first 24 admission hours, while 30-day readmission uses the full index admission without discharge-disposition tokens.Mortality and long LOS share admissions lasting at least 48 hours; readmission positives begin with a new inpatient admission within 30 days of discharge.Source: B DATA AND COHORT CONSTRUCTION · p. 25
  • For EHRSHOT readmission, 40,627 inpatient visits were merged into 12,699 admissions because 60.9% of consecutive visits began within one day of discharge.Models were evaluated without retraining; mortality and long-LOS tasks used visit records directly.Source: Figure 3 · p. 24, B DATA AND COHORT CONSTRUCTION · p. 25

C ADDITIONAL OUTCOME-PREDICTION DETAILS AND SCORES

  • Test scores were averaged across seeds with sample SDs, while macro scores averaged the three tasks within each seed before aggregation.Supervised configurations were selected by task-specific validation AUROC or multi-task validation macro AUROC; AUPRC was also reported because positives are rare.Source: C ADDITIONAL OUTCOME-PREDICTION… · pp. 25–26
  • Zero-shot LLM baselines read deduplicated, recent-first serialized records truncated to 8,192 tokens, with direct or chain-of-thought prompts and forced answer closure when needed.The “our prompt” variants used GLoC-EHR’s evidence-then-answer format; direct prompts decoded greedily.Source: C ADDITIONAL OUTCOME-PREDICTION… · p. 26
  • Count baselines used 20,567 codes; GRU (Cho et al., 2014) and RETAIN (Choi et al., 2016) used 2,048-event inputs, while frozen-vector models paired seven feature sets with classifier heads.The frozen-vector approach was a local adaptation of Hegselmann et al. (2026), not a reproduction of its architecture or published results.Source: C ADDITIONAL OUTCOME-PREDICTION… · p. 26
  • Fine-tuned EHR-BERT and GLoC-EHR models swept learning rates and label-sampling options, with multi-task variants using task-balanced sampling and early stopping.GLoC-EHR used hard top-64 local selection and scored answer tokens renormalized over Yes/No.Source: C ADDITIONAL OUTCOME-PREDICTION… · p. 26

C.1 ALIGNMENT AND INTERFACE CONTROLS

  • Macro AUROC was 0.862 without alignment versus 0.873 with alignment; LoRA reached 0.853, below count-based LightGBM at 0.864.The no-alignment control had similar AUPRC (0.535 vs. 0.524). The LoRA result came from a single run.Source: §C.1 · p. 26

C.2 REASONING EVALUATION AND ANSWER COVERAGE

GLoC-EHR (Reasoning) is evaluated using sampled reasoning chains, with answer performance and evidence support assessed separately. Zero-shot models can match serialized record text in their evidence bullets yet answer near chance, while chain-anchor handling has little effect on macro AUROC.

Reasoning evaluation and answer coverage

  • GLoC-EHR (Reasoning) is the final checkpoint of one GRPO run; five sampled chains per case yield mean and SD across runs that reflect decoding, not training variability.Chains use T=0.8, top-p 0.95, top-k 20, and at most 1,536 reasoning tokens; a chain missing the answer anchor is excluded from its run.Source: §C.2 · p. 27
  • 99.6% of 105,870 chains reach the answer anchor; scoring unanchored chains as 0.5 changes macro AUROC by less than 0.001, while five-chain averaging raises it from 0.8656 to 0.8668.Thirteen of 21,174 cases have no chain reaching the answer anchor.Source: Table 13 · p. 29, Table 14 · p. 29, §C.2 · p. 29
  • 94.1% of Llama-3.3-70B evidence bullets and 58.4% of Qwen3-1.7B bullets match a recorded concept exactly or by containment, often by repeating serialized prompt text.Stricter matching is not compared across these systems because it penalizes wording added by the serialization that GLoC-EHR never sees.Source: §C.2 · p. 29
  • Despite support precision of 0.452 for Qwen3-1.7B versus 0.424 for GLoC-EHR (Reasoning), and 0.752 for Llama-3.3-70B, both zero-shot models answer near chance (macro AUROC 0.509 and 0.549).GLoC-EHR cites more distinct supported findings than Qwen3-1.7B: 4.16 versus 3.88.Source: Table 15 · p. 29, §C.2 · p. 29

C.3 EVIDENCE DURING GRPO TRAINING · C.4 ZERO-SHOT TRANSFER TO UNSEEN QUESTIONS

During GRPO, evidence terms improve support and distinct findings over training without those terms, while removing the local route reduces distinct findings without changing other reported metrics materially. On two unseen laboratory questions, GLoC-EHR (Reasoning) outperforms zero-shot Qwen3-1.7B in AUROC.

C.3 EVIDENCE DURING GRPO TRAINING

  • Without evidence terms, 57.7% of bullets were unsupported versus 44.2% for GLoC-EHR, and bullets-only AUROC was 0.724 versus 0.774.Balanced accuracy was similar: 0.734 without evidence terms versus 0.740 for GLoC-EHR.Source: §C.3 · p. 30
  • Without evidence terms, the model cited 3.68 distinct supported findings versus 4.18 for GLoC-EHR at the end of GRPO training.The reported endpoint covers steps 3,000–3,199.Source: §C.3 · p. 30
  • Removing the local route yielded a similar unsupported rate (0.439 vs. 0.442) and bullets-only AUROC (0.776 vs. 0.774), but fewer distinct supported findings (4.02 vs. 4.18).The global-route-only variant was trained from its own SFT checkpoint.Source: §C.3 · p. 30

C.4 ZERO-SHOT TRANSFER TO UNSEEN QUESTIONS

  • GLoC-EHR (Reasoning) exceeded zero-shot Qwen3-1.7B by 0.100 [0.068, 0.131] AUROC on sodium and 0.104 [0.071, 0.137] AUROC on platelet questions.Neither model was trained on these questions; cases came from held-out MIMIC-IV patients.Source: §C.4 · p. 30

Related papers

Loading 2610.01076v1…