PaperlayerView exact revision on arXiv

Source-linked AI summary

RASPER: Reward-Aligned Summarization of Clinical Notes for EHR Outcome Prediction

Arya Hadizadeh Moghaddam, Mohsen Nayebi Kerdabadi, Chen Chen, Dongjie Wang, Zijun Yao

arXiv:2610.02979v1cs.AI

TL;DR

Generic discharge-note summaries may omit task-relevant evidence, while raw notes are lengthy and noisy. RASPER trains a summarizer with downstream prediction feedback and structured-code soft prompts. It consistently improves over strong baselines across two clinical prediction tasks and two MIMIC datasets.

  • Problem

    Generic summaries may retain plausible but uninformative detail while discarding patient-specific evidence predictive for a particular task.

    Source: §1 · p. 1

  • Method

    RASPER trains a tunable LLM summarizer with reinforcement learning from prediction feedback and uses structured-code soft prompts to provide patient grounding.

    Source: §1 · p. 2, §2.4 · p. 4

  • Results

    RASPER consistently improves over strong baselines on MIMIC-III and MIMIC-IV across readmission prediction and medication recommendation.

    Source: §1 · p. 2, §5 · p. 9

  • Takeaways & Limitations

    The framework combines cohort-level structured patterns with prediction-aligned summaries that retain useful patient-specific context.

    Source: §5 · p. 9

    Evaluation is limited to MIMIC-III and MIMIC-IV, and the effects of summary length and prompt format are not deeply analyzed.

    Source: §6 · p. 9

Abstract

from arXiv · show

Unstructured discharge notes in Electronic Health Records (EHRs) often carry signal complementary to structured medical codes, holding patient-specific evidence that standardized cohort-level codes alone cannot capture. However, this evidence in notes is frequently buried in lengthy, noisy text that is not intentionally written with any specific clinical prediction in mind. Summarization is an obvious mitigation, but generic summaries, tuned for fluency rather than the outcome, routinely omit decisive evidence while retaining plausible but uninformative detail. To this end, we propose RASPER, a Reward-Aligned Summarizer for Prediction in EHR, that optimizes note summarization directly against the downstream clinical task. RASPER employs a tunable LLM-based summarizer to extract task-relevant evidence from discharge notes and trains it via reinforcement learning from prediction feedback, using a reward derived from the downstream predictor's loss. To ground the summarizer, a longitudinal encoder converts structured codes into soft prompts that incorporate each patient's clinical context into note summarization. By rewarding the quality of the resulting multimodal prediction, RASPER encourages the summarizer to retain patient-specific evidence that complements, rather than duplicates, information captured by structured codes. RASPER consistently outperforms strong baselines on both readmission prediction and medication recommendation across MIMIC-III and MIMIC-IV.

1 Introduction

Discharge notes contain patient-specific evidence, but their length, noise, and documentation-focused style make generic summarization insufficient for prediction. RASPER aligns note summarization with downstream prediction and patient context.

1 Introduction

  • Discharge notes contain patient-specific clinical information beyond standardized codes, but raw notes can be lengthy, redundant, and style-sensitive.Directly using them may dilute predictive signal and increase sensitivity to documentation artifacts.Source: §1 · p. 1
  • Generic summaries may discard details predictive for a particular task while retaining plausible but uninformative information.The most predictive information can differ across tasks, motivating alignment with both the target and patient context.Source: §1 · p. 1
  • Prior EHR-LLM approaches use raw or hard-prompted notes, fixed summaries or text features, or omit notes, rather than aligning summarization with the downstream objective.These approaches respectively inherit documentation noise, leave summaries untuned to the target, or discard the note modality.Source: §1 · pp. 1–2
  • RASPER trains a tunable LLM summarizer with reinforcement learning from prediction feedback and grounds it with structured-code soft prompts.The predictor’s loss supplies the reward; the framework targets complementary note evidence, with reported improvements over strong baselines on two tasks across MIMIC-III and MIMIC-IV.Source: §1 · p. 2

2 Methodology

RASPER combines longitudinal structured codes and discharge-note summaries for clinical prediction, then tunes the summarizer using prediction feedback. Its method specifies the patient-visit inputs, multimodal predictor, supervised initialization, PPO-based reward alignment, and staged optimization.

2.1 Problem Formulation

  • For both readmission prediction and medication recommendation, the model uses a patient’s historical visits to predict outcomes at the next visit.Each visit contains medical codes and a discharge note.Source: §2.1 · p. 2

2.2 Model Overview

  • RASPER combines structured EHR soft prompts with generated note summaries, then uses prediction feedback to tune the summarizer after supervised fine-tuning.The patient-level summary concatenates visit summaries, and the predictor combines it with soft prompts from structured codes.Source: §2.2 · p. 2, Figure 1 · p. 3

2.3 LLM-Aware Outcome Predictor

  • The predictor encodes structured codes into soft prompts and combines them with summary embeddings, allowing structured and textual modalities to complement each other.The EHR encoder uses RETAIN to capture visit- and code-level patterns, then projects its representation into continuous prompt embeddings.Source: §2.3 · pp. 2–3

2.4 Tunable LLM Summarizer

  • The summarizer standardizes each discharge note into a shorter visit summary while aiming to preserve patient-specific evidence useful for prediction.Visit summaries are concatenated with the most recent visit first, followed by earlier visits for historical context.Source: §2.4 · p. 4

2.4.1 Summarizer Supervised Fine-Tuning

  • Supervised fine-tuning first teaches the summarizer to produce structured clinical summaries from noisy notes before predictive optimization.GPT-5.4 mini generates target summaries because annotated visit-level summaries are unavailable; the summarizer is trained against those targets.Source: §2.4.1 · p. 4

2.4.2 RLPF Optimization with PPO

  • With PPO, the predictor’s loss supplies feedback to optimize summaries for downstream prediction, using patient-level rewards and reward normalization and clipping.The reward measures improvement over a patient-specific reference based on the lowest historical losses observed during predictor warm-up.Source: §2.4.2 · p. 5

2.5 Staged Predictor–Summarizer Optimization

  • Predictor warm-up precedes summarizer optimization, reducing the risk that policy updates rely on feedback from an under-trained predictor.The predictor and summarizer then use separate objectives and optimizers.Source: §2.5 · p. 6

3 Evaluation

Across MIMIC-III and MIMIC-IV, RASPER achieves the best performance among the evaluated baselines. Ablations highlight contributions from structured-code encoding and prediction-aligned summaries, while results also examine encoder and LLM backbone choices.

3.1 Datasets and Setups

  • Evaluation covers MIMIC-III and MIMIC-IV for next-visit readmission prediction and medication recommendation, excluding patients with only one recorded visit.The study uses a random 70/30 train-test split and reports means over four seeds.Source: §3.1 · p. 6

3.3.1 Performance Comparison on Baselines

  • RASPER achieves the best performance across baselines, with RePrompT the closest competing method.Table 3 reports the comparison on MIMIC-III and MIMIC-IV.Source: Table 3 · p. 7, §3.3.1 · p. 7

3.3.2 Ablation Studies on Modules

  • Removing the EHR encoder causes the largest performance drop; prediction-aligned summaries outperform raw notes and generic GPT summaries.The ablation comparison is reported in Table 4.Source: Table 4 · p. 7, §3.3.2 · p. 8

3.3.3 Ablation Studies on Encoder Selection

  • RETAIN is the best-performing encoder, with gains consistent across datasets and evaluation metrics.The encoder comparison appears in Figure 3; the authors attribute RETAIN’s performance to its multi-level modeling of patient history.Source: §3.3.3 · p. 8, Figure 3 · p. 8

3.3.4 LLM Model Selection

  • Gemma-2-2B performs best across all metrics, although its gaps over Llama-3.2-1B and Qwen2.5-0.5B are relatively small in most settings.Table 5 compares LLM backbones across two tasks on MIMIC-III and MIMIC-IV.Source: §3.3.4 · pp. 8–9, Table 5 · p. 8

4 Related Work

Prior EHR prediction work has modeled structured records, clinical knowledge, and LLM-based representations, but the described methods do not fully exploit unstructured clinical notes.

Structured EHR prediction

  • Structured-sequence methods improve prediction through attention, ontology propagation, temporal modeling, and disease-stage representations, but mainly rely on structured codes or numerical features.Examples include RETAIN [Choi et al., 2016], HAP, AdaCare, and StageNet [Gao et al., 2020].Source: §4 · p. 9

Remaining gap

  • The related methods described primarily use structured codes or numerical features and therefore do not fully exploit information in unstructured clinical notes.This limitation is stated for structured-sequence approaches; the passage also describes recent LLM work using retrieved knowledge, prompts, or structured EHR encoders.Source: §4 · p. 9

LLMs for EHR prediction

5 Conclusion

RASPER combines structured EHR codes with prediction-aligned discharge-note summaries, and experiments on MIMIC-III and MIMIC-IV show consistent improvements over strong baselines.

Conclusion

  • RASPER consistently improves over strong baselines on MIMIC-III and MIMIC-IV.The framework integrates structured codes and discharge notes using EHR soft prompts and prediction-aligned summaries.Source: §5 · p. 9
  • EHR soft prompts capture cohort-level patterns while prediction-aligned summaries retain useful patient-specific context.RASPER optimizes the summarizer with downstream prediction feedback.Source: §5 · p. 9
  • The framework produces summaries intended to be more directly useful for the target prediction task.This is the stated aim of optimizing summarization with downstream prediction feedback.Source: §5 · p. 9

6 Limitations

The study's evaluation is limited to MIMIC-III and MIMIC-IV, and it does not deeply analyze the effects of summary length or prompt format.

Limitations

  • Evaluation is limited to MIMIC-III and MIMIC-IV, so performance generalizability would be better assessed with data from other hospitals.The paper identifies additional hospital datasets as a needed evaluation.Source: §6 · p. 9
  • The study does not deeply analyze how summary lengths or prompt formats affect final results.These two design factors are specifically named as underanalyzed.Source: §6 · p. 9

7 Ethical Considerations

The study uses publicly available, de-identified EHR data without direct patient interaction and positions RASPER as research support, not a replacement for clinical judgment.

Ethical Considerations

  • The study uses publicly available, de-identified EHR datasets and involves no direct interaction with patients.The authors also state that the framework is intended to support clinical prediction research.Source: §7 · p. 9
  • RASPER is intended to support clinical prediction research rather than replace clinical judgment.The paper cautions against autonomous clinical decision-making with the framework.Source: §7 · p. 9
  • Generated summaries may omit clinically relevant information or introduce inaccuracies, so the framework should not be used autonomously for clinical decisions.This is the stated reason for the paper's autonomous-use boundary.Source: §7 · p. 9

A.1 RETAIN Encoder

The RETAIN encoder represents longitudinal medical codes using visit-level and feature-level attention. It combines the weighted signals into a patient representation, using the final visit embedding.

A.1 RETAIN Encoder

  • Visit-level attention captures the relative importance of visit embeddings, while variable-level attention assigns featurewise weights to the visit representation.Source: §A.1 · p. 12
  • RETAIN combines visit-level and variable-level attention-weighted signals to compute the patient representation.Source: §A.1 · p. 12
  • The encoder outputs visit-level embeddings and uses the final visit embedding as the patient-level EHR representation.Source: §A.1 · p. 12

A.2 Baselines

The experiments compare RASPER with established deep EHR prediction models and recent LLM-based healthcare prediction methods. These baselines span several approaches to representing longitudinal records and integrating clinical knowledge.

A.2 Baselines

  • The deep-learning comparison includes Deepr, RETAIN, GRASP, AdaCare, StageNet, ADORE, and ARCI.Source: §A.2 · p. 12
  • The deep baselines cover coded-event sequences, attention over visits and features, clinically similar patients, biomarker variation, disease stages, ontology embeddings, and temporal medical intents.Source: §A.2 · p. 12
  • The LLM-based baselines are GraphCare, LINKO, and RePrompT, which use knowledge graphs, ontology graphs, or time-aware prompt tuning with structured EHR information.Source: §A.2 · p. 12

A.3 Inference Time Comparison

The inference-time comparison considers RASPER and existing baselines for batches of 12 patients. The authors report that RASPER is more complex than RePrompT, but that its observed predictive gains justify the added complexity.

A.3 Inference Time Comparison

Related papers

Loading 2610.02979v1…