Developing an Explainable Machine Learning-Based Personalised Dementia Risk Prediction Model: A Transfer Learning...

Area of Science:

• Computational neuroscience and artificial intelligence in healthcare.
• Biomarker discovery and predictive modeling for neurodegenerative diseases.

Background:

• Alzheimer's disease (AD) onset precedes dementia by decades, necessitating early detection methods.
• Current machine learning models for AD risk prediction face limitations in generalizability and clinical utility.
• Existing models often suffer from data source over-reliance and lack of interpretability.

Purpose of the Study:

• To develop explainable, personalized risk prediction models for dementia using a transfer-learning framework.
• To enhance the generalizability and clinical applicability of predictive models for early Alzheimer's disease detection.

Main Methods:

• Employed a transfer-learning paradigm with ensemble learning algorithms.
• Trained 'source models' on a large public dataset (n=84,856) and refined them with a smaller, younger dataset (n=473) to create 'target models'.
• Utilized the SHapely Additive exPlanation (SHAP) algorithm for visualizing prediction drivers.

Main Results:

• The best source model achieved 87% geometric accuracy, 99% specificity, and 76% sensitivity.
• The target model demonstrated significant performance improvements over a baseline model, including a 16.9% increase in geometric accuracy and 19.1% in sensitivity.
• The approach showed a transfer learning efficacy rate of 20.6% and provided interpretable risk factor visualizations.

Conclusions:

• The proposed framework effectively enhances dementia risk prediction through transfer learning and ensemble methods.
• Explainable AI (SHAP) provides crucial insights into risk factors, increasing clinical utility for early AD detection.
• This approach offers a robust and clinically relevant tool for identifying individuals at risk of developing dementia.