Optimized Gaussian Process Regression for Multimodal Spatio-Temporal Prediction of Alzheimer's Disease Progression

Accurate prediction of Alzheimer's Disease (AD) progression is essential for early intervention and personalized treatment planning. This study proposes an optimized Gaussian Process Regression (GPR) framework that integrates multimodal data—including clinical measures, genetic markers, and spatio-temporal imaging biomarkers—into a unified predictive model. A probabilistic simulated dataset was constructed based on statistical distributions derived from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Open Access Series of Imaging Studies (OASIS), with Gaussian noise injection applied to enhance variability while preserving the underlying statistical structure. Data preprocessing included min–max normalization, Multiple Imputation by Chained Equations (MICE), and dimensionality reduction using Principal Component Analysis (PCA) with ≥95% variance retention. Three kernel functions—Radial Basis Function (RBF), Matern, and Rational Quadratic—were evaluated using grid search and k-fold cross-validation. Experimental results indicate that GPR with the RBF kernel achieved the best performance, yielding a Root Mean Square Error (RMSE) of 1.220541, a Mean Absolute Error (MAE) of 0.999709, and a coefficient of determination (R2) of 0.402046. Residual analysis and Shapley Additive Explanations (SHAP)-based feature interpretation confirmed the clinical relevance of hippocampal volume, Mini-Mental State Examination (MMSE) score, and Apolipoprotein E (APOE)-ε4 allele status. The proposed approach demonstrates moderate but clinically meaningful predictive performance with robust generalization and strong interpretability, making it a promising decision-support tool for early diagnosis and monitoring of AD progression. Future work will validate the model on real-world clinical datasets and explore computational efficiency improvements through sparse approximation techniques.