From Discharge Letters to Process Traces with LLMs: A Human-in-the-Loop Pipeline.

Area of Science:

• Clinical Informatics
• Process Mining
• Artificial Intelligence in Healthcare

Background:

• Discharge letters contain valuable clinical information but are often unstructured.
• Analyzing unstructured clinical data for process improvement is challenging.
• Existing methods lack efficient ways to convert narrative text into analyzable event logs.

Purpose of the Study:

• To present a novel human-in-the-loop pipeline for converting unstructured discharge letters into process-mining-ready event traces.
• To leverage Large Language Models (LLMs) for automated extraction of clinical events.
• To enable robust process mining (PM) analyses of hospital care pathways.

Main Methods:

• A modular pipeline combining LLM prompting with human expert validation.
• Task-specific LLM prompting to extract temporally ordered clinical events from discharge letters.
• Iterative refinement and expert validation of extracted event traces.
• Storage of validated traces for subsequent PM analyses.

Main Results:

• Successfully converted unstructured discharge letters into process-mining-ready event traces.
• The pipeline demonstrated effectiveness in extracting temporally ordered clinical events.
• The approach was evaluated on 466 Stroke Unit discharge letters, showing feasibility.

Conclusions:

• The proposed human-in-the-loop pipeline offers an effective method for transforming unstructured clinical text into structured data for process mining.
• LLM-driven extraction followed by expert validation ensures accuracy and utility of event traces.
• This approach facilitates deeper insights into healthcare processes, exemplified by its application to stroke unit data.