Interpretable and Uncertainty-Aware Spatiotemporal Learning for Healthcare

Modern healthcare increasingly relies on learning from spatiotemporal data: how anatomy develops with age, how disease reshapes anatomy, and how different patients follow different trajectories. Yet the data behind these questions is often difficult to work with and limited in the number of samples—it is 3D, high-dimensional, and highly variable from person to person. Standard deep learning methods can fit these data, but usual approaches are black boxes. They offer predictions without clearly explaining where anatomy is changing, how clinical covariates such as age or disease status matter, or how confident we should be in any given result. This dissertation addresses the gap between predictive power and clinical trust. This thesis develops a family of uncertainty-aware, interpretable representations for anatomical shapes and their temporal evolution. First, I introduce NAISR, a 3D neural additive model that brings interpretability to implicit surface representations. NAISR learns how anatomy evolves spatiotemporally, considering covariates such as age and sex, retains fine geometric details, and explicitly visualizes how individual factors locally deform the shape. Next, I present LucidAtlas, an interpretable, uncertainty-aware framework for 1D spatiotemporal signals. It models spatially varying signals and covariates, using a marginalization strategy to interpret how correlated factors drive anatomical changes. This unified framework links covariate interpretation with individualized prediction and out-of-distribution detection. Finally, PRISM advances these concepts to 3D spatiotemporal learning by modeling anatomy as a continuous, probabilistic neural field. Beyond predicting mean evolution, PRISM captures spatially varying, heteroscedastic uncertainty directly on the 3D surface. Uniquely, it leverages information geometry to derive a closed-form Fisher information metric, which analytically decomposes information (and thus uncertainty) into population dynamics and subject-specific variation. This formulation enables robust developmental stage inference and reliable anomaly detection. Together, these contributions move spatiotemporal learning for healthcare beyond simple prediction toward models that are geometrically faithful, uncertainty-aware, and clinically interpretable. By making it clearer not only what a model predicts, but also why and how confidently, this work aims to support more transparent, reliable, and individualized medical decision-making.