Explainable AI framework for Alzheimer’s disease prediction using multimodal data

Accurate prediction of Alzheimer’s disease (AD) is critical in clinical research and healthcare. However, distinguishing disease progression with heterogeneous clinical symptoms and high-dimensional clinical data across different cognitive stages remains challenging. This study proposes an interpretable machine learning framework for predicting AD using baseline clinical assessments from the National Alzheimer’s Coordinating Center (NACC) cohort. SHapley Additive exPlanations (SHAP)-guided feature selection was employed to identify the most informative features. Five classifiers were applied across numerical, categorical and their combined subsets in binary and multiclass settings. Notably, the clinically important MCI-to-AD, Random Forest (RF) achieved the best performance, attaining a Balanced Accuracy of 0.9322 and a Macro F1-score of 0.9352. RF achieved ROC-AUC 0.9755 ± 0.0010 and AUC-PR 0.9924 ± 0.0005, demonstrating the discriminative capability preserved in reduced feature set. These findings demonstrate a promising foundation for future clinical decision-support systems, with integrated interpretable feature selection. Average training times computed across all evaluated classifiers showcased most prominent improvement from 1039.77s to 59.7s (~ 94%). These findings demonstrate that integrating robust interpretable feature selection with machine learning can provide a promising decision-support framework for future clinical assessment.