Discharge Instructions are not One Task: Grounding Differences Between Surgical and Non-Surgical Admissions

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Health & Wellbeing — Clinical Care & Medical Practice, Health & Medical Research · Depth: Expert, quick

Summary

This work investigates discharge instructions (DIs) within the MIMIC-IV dataset, focusing on their complexity and grounding against Electronic Health Records (EHR). Researchers decomposed DIs into medically relevant statements using two large language models and verified each against the EHR. A key finding reveals that DIs for Surgical admissions are significantly longer, averaging 24–25 statements per admission, compared to 11–12 statements in Non-Surgical cases. While the absolute amount of EHR-supported content remains similar across both types, the additional length in Surgical DIs stems primarily from information not directly stated in the record or requiring clinically plausible extrapolation. This analysis highlights the need for improved fine-grained grounding and completeness evaluations to enhance the safety and reliability of automated discharge instruction generation.

Key takeaway

For NLP Engineers developing clinical text generation systems, understanding the distinct complexities of discharge instructions is crucial. You should prioritize developing models that can not only synthesize EHR data but also perform clinically plausible extrapolation, especially for surgical cases where much content is not verbatim in the record. Implement fine-grained grounding and completeness evaluations to ensure generated instructions are both accurate and safe for patients.

Key insights

Discharge instructions vary significantly by admission type, with surgical cases requiring more extrapolated content.

Principles

• Discharge instructions are not a monolithic task.
• Clinical text generation needs fine-grained evaluation.
• Surgical DIs demand more inferred content.

Method

Decomposed discharge instructions into medically relevant statements using two LLMs, then verified each statement against the Electronic Health Record (EHR) for grounding.

In practice

• Develop LLMs capable of clinically plausible extrapolation.
• Design evaluation metrics for grounding and completeness.
• Tailor DI generation models by admission type.

Topics

• Discharge Instructions
• Clinical Text Generation
• Large Language Models
• Electronic Health Records
• MIMIC-IV
• Grounding Evaluation
• Surgical Admissions

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Systematic Evaluation of the Quality of Synthetic Clinical Notes Rephrased by LLMs at Million-Note Scale — LLM-rephrased clinical notes retain coarse-grained utility but lose fine-grained detail, with chunking offering mitigation.
• Before the Labels: How Dataset Construction Shapes Suicidality Detection in Clinical Text — EHR-based suicidality datasets encode specific operationalizations of suicidality, shaped by construction choices, not objective ground truth.
• Deployment-Centered Evaluation: Predicting Query-Level Rejection Risk in a Clinical LLM System — Predicting clinical LLM user rejection is feasible by integrating deployment-specific context with query content.
• OpenMed: Running Medical NLP Locally Without Sending Patient Data to the Cloud — OpenMed enables secure, local medical NLP by processing sensitive patient data entirely within hospital systems.
• Language Models as Interfaces, Not Oracles: A Hybrid LLM-ML System for Pediatric Appendicitis — Hybrid AI systems improve clinical decision support by using LLMs for feature extraction and validated ML models for reliable, auditable prediction.
• Possible or Definite? A Benchmark for Evaluating Diagnostic Uncertainty Preservation in Clinical Text — Large language models poorly preserve diagnostic uncertainty in clinical text, posing a significant safety risk.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.