FUSION-AD: interpretable AI framework for risk assessment and subgroup discovery in Alzheimer's disease

Area of Science:

• Artificial Intelligence
• Neuroscience
• Computational Biology

Background:

• Alzheimer's disease (AD) presents complex challenges in treatment due to multifactorial causes and varied progression.
• Existing diagnostic and risk assessment tools often lack interpretability and personalization.

Purpose of the Study:

• To introduce FUSION-AD, an interpretable artificial intelligence (AI) framework for Alzheimer's disease (AD) risk assessment and subgroup discovery.
• To integrate diverse AI methodologies for enhanced predictive accuracy and clinical insight into AD.

Main Methods:

• Developed FUSION-AD integrating tree-based models, neural networks, rule mining, and subgroup discovery.
• Utilized a synthetic dataset of 2,149 clinical records (ages 60-90) with preprocessing, benchmarking, and explanation techniques (SHAP, attention analysis).
• Applied association rule mining and subgroup discovery to identify risk factors and clinical clusters.

Main Results:

• TabNet achieved the highest performance (AUROC 0.95), with MMSE, Functional Assessment, and Memory Complaints identified as key predictors.
• Diabetes and high BMI emerged as significant risk factors through association rule mining.
• Four distinct clinical subgroups were identified, with differing functional decline and behavioral symptom profiles.

Conclusions:

• FUSION-AD effectively balances predictive performance and interpretability for AD risk assessment.
• Identified subgroups suggest tailored preventive strategies for lifestyle-driven profiles and closer monitoring for cognitively impaired groups.
• Findings support the development of clinical decision-support systems, requiring further validation on real-world data.