Constrained clustering and cognitive decline detection

This thesis addresses two topics: constrained clustering and cognitive decline detection. First, we develop probabilistic clustering methods that incorporate pairwise constraints expressing prior beliefs about sample relationships. We introduce Penalized Probabilistic Clustering and a Gaussian process–based approach that more efficiently leverages constraint information. Second, we study detection of cognitive decline from longitudinal clinical data using mixed‑effect models and both generative and discriminative classifiers, including support vector machines with kernels designed for time‑series data. Results demonstrate improved predictive performance, particularly when individual‑specific effects are modeled explicitly.