Machine learning approaches for early detection of Cognitive decline: A comparative study using the OASIS-3 dataset
TL;DR
A comparative study of machine learning algorithms applied to early detection of cognitive decline using the OASIS-3 dataset, a longitudinal multimodal neuroimaging and clinical cohort of 1378 participants aged 42-95 years, demonstrating the added value of logic-augmented classifiers for early risk stratification.
Abstract
Cognitive decline associated with Alzheimer's disease (AD) and mild cognitive impairment (MCI) represents one of the most critical public health challenges of the twenty-first century, affecting tens of millions of people worldwide. This paper presents a comparative study of machine learning algorithms applied to early detection of cognitive decline using the open access series of imaging studies (OASIS-3) dataset, a longitudinal multimodal neuroimaging and clinical cohort of 1378 participants aged 42-95 years. To avoid data leakage, all features directly encoding a clinical diagnosis, including clinical dementia rating (CDR) scores, diagnostic (DX) codes, and mini-mental state examination (MMSE) results, are excluded; instead, fifteen structural neuroimaging features describing brain morphology (e.g., hippocampal and entorhinal measurements) and neuropsychological sub-test features are selected via a two-stage recursive feature elimination and greedy forward-search procedure. We evaluate six supervised models: logistic regression, support vector machine (SVM), random forest, gradient boosting, K-nearest neighbours (KNN), and CatBoost. Gradient boosting achieves the highest weighted F1-score (0.856) on the held-out test set, followed by CatBoost (0.8482); CatBoost offers approximately $2\times$ lower inference latency. A neurosymbolic post-processing stage using the Scallop probabilistic logic engine identifies 166 participants with measurable structural brain atrophy (reduction in hippocampal volume) prior to a formal clinical diagnosis, demonstrating the added value of logic-augmented classifiers for early risk stratification. Feature-importance analysis identifies white-matter hyperintensity volume (MRI-visible lesions associated with vascular and neurodegenerative damage) and entorhinal/hippocampal measurements as the dominant predictors, consistent with established neurobiological evidence.