Saurabh Mathur: “Human-allied probabilistic AI for Healthcare”

I work on probabilistic AI and AI for healthcare. Specifically, I am interested in developing methods to reconcile information from multi-source data with different forms of domain knowledge into human-understandable forms. Since this scenario is frequently encountered in knowledge-rich but data-poor domains such as healthcare, I am also interested in adapting these methods for complex clinical problems such as modeling the risk of adverse pregnancy outcomes (e.g., preterm birth and gestational diabetes) and understanding the causes of neurological injury in children on life-support (ECMO). To this effect, my work has considered clinical domain knowledge about relationships between two or more variables including causal influence, monotonicities, and independence. We have developed methods to combine these diverse forms of domain knowledge with noisy, sparse, and uncertain data to construct explainable and interpretable probabilistic models including Bayesian and credal networks, tractable probabilistic models such as sum-product networks and cutset networks, and neurosymbolic models.