Knowledge-Augmented Large Language Model for Multimodal Electronic Health Record-Based Risk Prediction: Development...

Area of Science:

• Computational biology and bioinformatics
• Machine learning in healthcare
• Clinical informatics

Background:

• Accurate clinical outcome prediction from electronic health records (EHRs) is vital for patient care and resource management.
• Existing machine learning models face challenges with EHR multimodality, long clinical note contexts, and imbalanced data.
• Integrating structured EHR data with unstructured clinical notes and external knowledge is needed for improved prediction.

Purpose of the Study:

• To introduce and evaluate KAMELEON (Knowledge-Augmented Multimodal EHR Learning for Outcome Prediction), a hybrid framework.
• To enhance clinical risk prediction by integrating diverse EHR modalities and external biomedical knowledge.
• To assess KAMELEON's performance against established models and large language models (LLMs).

Main Methods:

• Utilized the Medical Information Mart for Intensive Care-III (MIMIC-III) dataset (over 40,000 ICU patients).
• Evaluated 30-day readmission and in-hospital mortality prediction using patient-disjoint 80:20 train-test splits.
• Compared KAMELEON against general/medical LLMs and structured baselines using AUROC, AUPRC, and macro F1-score metrics.

Main Results:

• KAMELEON consistently outperformed all evaluated baselines across prediction tasks.
• The KAMELEON-balanced random forests model achieved an AUROC of 0.85 for 30-day readmission prediction.
• The KAMELEON-extreme gradient boosting model achieved an AUROC of 0.92 and AUPRC of 0.650 for in-hospital mortality prediction.
• LLM-generated reasoning was critical, with its removal significantly degrading performance (AUROC dropped from 0.85 to 0.7).

Conclusions:

• KAMELEON is a pioneering framework enhancing LLMs for healthcare prediction via graph-guided knowledge retrieval and structured ML.
• The framework demonstrates superior performance, validating the synergistic value of multimodal data and LLM reasoning.
• KAMELEON offers a robust approach to clinical risk estimation, addressing limitations of prior predictive models.