Domain-adapted language model using reinforcement learning for various dementias

Area of Science:

• Artificial Intelligence in Medicine
• Computational Neuroscience
• Clinical Informatics

Background:

• Large language models (LLMs) show promise for clinical data analysis but require domain-specific adaptation.
• Alzheimer's disease and related dementias (ADRD) present complex diagnostic challenges.
• Effective LLM application in ADRD necessitates tailored approaches for advanced reasoning and data integration.

Purpose of the Study:

• To develop and validate a generative language model fine-tuned for Alzheimer's disease and related dementias (ADRD) using reinforcement learning.
• To enhance diagnostic capabilities in ADRD through a model integrating multimodal clinical data.
• To assess the clinical utility and diagnostic performance improvement offered by the ADRD-specific LLM.

Main Methods:

• A generative language model was fine-tuned using reinforcement learning with verifiable rewards and a self-certainty-aware advantage.
• Model development and validation utilized data from five ADRD cohorts, encompassing 54,535 participants.
• The framework integrated diverse data types including demographics, medical history, medications, neuropsychological tests, functional assessments, examinations, lab data, and neuroimaging.

Main Results:

• The model demonstrated robust performance on syndromic classification, primary etiological diagnosis, and biomarker prediction on held-out data from 36,688 participants.
• Model predictions were validated against postmortem-confirmed diagnoses.
• A within-subjects crossover study showed improved diagnostic performance by board-certified neurologists when assisted by the model.

Conclusions:

• Domain-specific adaptation using reinforcement learning enables LLMs to provide accurate, reasoning-driven support for ADRD evaluation.
• The developed model shows significant potential for improving diagnostic accuracy and clinical decision-making in ADRD.
• Prospective validation is crucial for translating these findings into improved patient outcomes in ADRD care.