Comparing machine learning classifier models in discriminating cognitively unimpaired older adults from three clinical cohorts in the Alzheimer's disease spectrum: demonstration analyses in the COMPASS-ND study.

Abstract

Background: Research in aging, impairment, and Alzheimer's disease (AD) often requires powerful computational models for discriminating between clinical cohorts and identifying early biomarkers and key risk or protective factors. Machine Learning (ML) approaches represent a diverse set of data-driven tools for performing such tasks in big or complex datasets. We present systematic demonstration analyses to compare seven frequently used ML classifier models and two eXplainable Artificial Intelligence (XAI) techniques on multiple performance metrics for a common neurodegenerative disease dataset. The aim is to identify and characterize the best performing ML and XAI algorithms for the present data.

Method: We accessed a Canadian Consortium on Neurodegeneration in Aging dataset featuring four well-characterized cohorts: Cognitively Unimpaired (CU), Subjective Cognitive Impairment (SCI), Mild Cognitive Impairment (MCI), and AD (N = 255). All participants contributed 102 multi-modal biomarkers and risk factors. Seven ML algorithms were compared along six performance metrics in discriminating between cohorts. Two XAI algorithms were compared using five performance and five similarity metrics.

Results: Although all ML models performed relatively well in the extreme-cohort comparison (CU/AD), the Super Learner (SL), Random Forest (RF) and Gradient-Boosted trees (GB) algorithms excelled in the challenging near-cohort comparisons (CU/SCI). For the XAI interpretation comparison, SHapley Additive exPlanations (SHAP) generally outperformed Local Interpretable Model agnostic Explanation (LIME) in key performance properties.

Conclusion: The ML results indicate that two tree-based methods (RF and GB) are reliable and effective as initial models for classification tasks involving discrete clinical aging and neurodegeneration data. In the XAI phase, SHAP performed better than LIME due to lower computational time (when applied to RF and GB) and incorporation of feature interactions, leading to more reliable results.