Explainable machine learning on clinical features to predict and differentiate Alzheimer's progression by sex: Toward a clinician-tailored web interface.
Journal:
Journal of the neurological sciences
PMID:
39724825
Abstract
Alzheimer's disease (AD), the most common neurodegenerative disorder world-wide, presents sex-specific differences in its manifestation and progression, necessitating personalized diagnostic approaches. Current procedures are often costly and invasive, lacking consideration of sex-based differences. This study introduces an explainable machine learning (ML) system to predict and differentiate the progression of AD based on sex, using non-invasive, easily collectible predictors such as neuropsychological test scores and sociodemographic data, enabling its application in every day clinical settings. The ML model uses SHapley Additive explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) to provide clear insights into its decision-making, making complex outcomes easier to interpret. The system includes a user-friendly graphical interface designed in collaboration with clinicians, supporting its integration into medical practice. The study extends the cohort to include healthy and Mild Cognitive Impairment subjects, aiming to support early diagnosis in AD pre-clinical stages. The ML system was trained on a large dataset of 2407 subjects from the ADNI open dataset, enhancing its robustness and applicability. By focusing on sex-specific features and utilizing longitudinal data, the system aims to improve prediction accuracy and early detection of AD, ultimately advancing personalized diagnostic and therapeutic approaches. Key findings highlight the significance of the Mini-Mental State Examination, Rey Auditory Verbal Learning Test, Logical Memory - Delayed Recall, and educational attainment in AD diagnosis and progression, with sex-based disparities. Despite performance metrics based on precision, recall, and weighted F1-score demonstrating model efficacy, future research should address the limitations of relying on a single dataset.