A Longitudinal Deep Learning Framework for Predicting Conversion from Mild Cognitive Impairment to Alzheimer’s Disease Using Multi-Modal Data

Abstract

Mild cognitive impairment (MCI) represents a critical intermediate stage between normal aging and Alzheimer’s disease (AD) yet predicting which individuals will progress remains challenging. Reliable risk prediction tools are needed to identify high-risk patients early and guide preventive interventions. In this study, we developed a longitudinal deep learning framework to predict conversion from MCI to AD within a 2–3-year window using a multi-modal dataset derived from the Kaggle TADPOLE challenge. The dataset included 872 baseline MCI participants with clinical, cognitive, demographic, and imaging-derived biomarkers, with 37.0% converting to AD during follow-up. After preprocessing, we trained both baseline statistical models and sequential deep learning models, with performance evaluated on an independent test set. Our best-performing model achieved an AUC of 0.848 and an overall accuracy of 76.6% on the test set, demonstrating good discrimination and calibration. Feature importance analysis identified ADAS13, CDRSB, and FAQ as the strongest positive predictors, whereas MMSE scores were negatively associated with conversion risk. Predicted risk probabilities showed clear separation between converters and non-converters, suggesting the model captured meaningful disease trajectories.