An interpretable, clinically-aligned AI paradigm for VTE risk prediction: an approach using LLMs and compound...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Risk Prediction

Background:

• Manual Venous Thromboembolism (VTE) risk assessment presents challenges with inconsistency.
• Existing high-performing AI models for VTE risk often lack clinical auditability.
• There is a need for accurate, interpretable, and clinically aligned VTE risk stratification tools.

Purpose of the Study:

• To develop and validate an interpretable AI paradigm for VTE risk stratification.
• To utilize locally deployed LLMs for structuring Electronic Health Record (EHR) narratives.
• To employ an attention-based model (EMAX) for predicting expert-adjudicated Caprini Risk Assessment Model (RAM) scores.

Main Methods:

• A curated dataset of 14,808 eligible EHR records with expert Caprini scores was used for supervised learning.
• Locally deployed LLMs were used to structure raw EHR encounter narratives.
• An attention-based model (EMAX) was developed to predict Caprini RAM risk scores.

Main Results:

• The EMAX model achieved a high Area Under the Curve (AUC) of 0.9513 on the test set.
• The framework demonstrated an end-to-end pathway for VTE risk stratification.
• The proposed method offers an interpretable and clinically aligned approach to VTE risk assessment.

Conclusions:

• The developed AI framework provides accurate and adoptable VTE risk stratification.
• The integration of LLMs and EMAX enhances the interpretability and clinical utility of VTE risk prediction.
• This paradigm addresses the limitations of manual assessment and opaque AI models in VTE risk management.