LELN: A Large Language Model-Dynamically Enhanced Learning Network for Patient Similarity Calculation

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Informatics
• Machine Learning

Background:

• Electronic Medical Record (EMR) data expansion fuels AI for patient similarity.
• Existing methods struggle with heterogeneous EMR formats and domain knowledge integration.
• Graph-based approaches show promise but face limitations.

Purpose of the Study:

• To propose a novel Large Language Model-Dynamically Enhanced Learning Network (LELN) for advanced patient similarity computation.
• To leverage Large Language Models (LLMs) for dynamic EMR data structuring and medical knowledge integration.
• To overcome limitations of current methods in handling heterogeneous EMR data and domain knowledge.

Main Methods:

• LELN integrates two LLM-based modules: DeepSeek-Event Extraction (DS-EE) for structured EMR event graphs and DeepSeek-Knowledge Base (DS-KB) for knowledge augmentation.
• A dual-stage spatial-temporal feature aggregation strategy uses Graph Attention Network and Bidirectional Long-Short Term Memory (BiLSTM) with attention.
• A clinical prior-guided attention mechanism enhances feature discrimination for clinical relevance.

Main Results:

• LELN achieved superior performance on heterogeneous datasets, including a Chinese EMR dataset and MIMIC-III.
• The model demonstrated high accuracy, with F1 scores of 87.66% and 85.95% on the respective datasets.
• Experiments confirmed LELN's robustness and effectiveness in patient similarity computation.

Conclusions:

• LELN effectively addresses challenges in heterogeneous EMR data handling and medical knowledge integration.
• The proposed model significantly advances AI-driven patient similarity computation for intelligent healthcare.
• LELN shows strong potential for improving clinical decision-making and patient care through enhanced EMR analysis.