Though they show impressive empirical accuracy, machine learning methodologies have been criticized for not producing interpretable, scientific theories. In both clinical medicine and public health, the researchers aim not just to predict health outcomes, but to improve them. Hence, causal, human-interpretable models of nature hold particular value in these fields. In this dissertation, I investigate how deep learning, when integrated into scientifically-informed models and principled statistical frameworks, can be used to advance mechanistic modeling in the health sciences.

Since the widespread adoption of electronic health records (EHRs), there has been growing interest in evaluating medical interventions through large-scale observational studies of diverse patient populations. In the first chapter, I examine the opportunities and challenges that arise from applying deep neural networks to EHR data. Despite the vast scale of EHR datasets, black box predictive modeling has limited value for informing clinical care, where human judgment is indispensable. Medical researchers are often interested in estimating counterfactual treatment eff ects on patients’ time-to-event outcomes. In the second chapter, I propose the Dynamic Survival Transformer (DynST), a deep survival model that flexibly estimates hazards from both static and time-varying features typical of EHR data, and demonstrate how DynST supports robust, semiparametric inference for causal survival analysis.

Stochastic infectious disease models capture uncertainty in public health outcomes and off er mechanistic explanations of transmission patterns. However, they are often nonlinear dynamical systems with massive latent state spaces, making likelihood-based inference of model parameters difficult. In the third chapter, I develop a methodology for efficiently calibrating large-scale stochastic epidemic simulation models to observed data using Neural Posterior Estimation. In NPE, a neural network trained on simulated data learns to “invert” a stochastic simulator and returns a parametric approximation of the posterior distribution. I use NPE to calibrate a stochastic Susceptible-Infected model to a study of a healthcare-associated infection in a long-term acute care hospital and find evidence of spatially heterogeneous patient-to-patient transmission risk.