Reinforcement Learning and Spatio-Temporal Graph Neural Networks for Alzheimer's Disease Progression Prediction with Variational Autoencoder-Based Data Imputation

Early detection and accurate modeling of Alzheimer's disease (AD) progression are still significant challenges because of its intricate neurodegenerative patterns. Traditional deep learning methods are not effective in capturing temporal dependencies, which restricts their predictive capabilities. In this work, we introduce a reinforcement learning-augmented spatio-temporal graph neural network (RL-STGNN) coupled with variational autoencoders (VAE) to model disease progression from longitudinal MRI scans. This model successfully captures spatial and temporal relations in neuroimaging data and enhances feature representation in small, unbalanced datasets with self-supervised contrastive learning (SimCLR). The suggested method was tested on 2,149 ADNI dataset MRI scans using five-fold cross-validation. The RL-STGNN model performed at a 98.20% accuracy level, outperforming baseline deep learning models, and the SimCLR variant attained a 93.95% accuracy level, illustrating the effectiveness of contrastive learning to enhance feature extraction. To achieve clinical usability, we integrated SHapley Additive exPlanations (SHAP) to rank feature importance, providing additional interpretability in biomarker discovery. This framework has major clinical applications: it facilitates the early diagnosis of AD, enables disease monitoring, and guides tailored treatment plans. By marrying powerful deep learning with interpretability tools, the model offers a strong, explainable, and data-driven decision-support system for neurologists, potentially enhancing patient outcomes and facilitating clinical trials.