You are invited to attend Akash Choudhuri’s final exam on Wednesday, October 7, at 2:30pm.

Advisor: Bijaya Adhikari

Location: UCC 2520C. Please contact Akash, akash-choudhuri@uiowa.edu, if you plan to attend.

Title: Robust Spatio-Temporal Learning and Uncertainty Quantification over Electronic Health Records

Abstract: Patients admitted to healthcare institutions are at risk of adverse outcomes, including healthcare-associated infections, transfer to a medical intensive care unit, and even death due to comorbidities. Predicting the risk of these unfavorable events is critical in guiding clinical decision-making and maximizing the allocation of limited healthcare resources. In recent years, the widespread adoption of electronic health record (EHR) systems has simplified the collection, storage, and querying of high-dimensional multimodal patient data. As the complexity and amount of EHR data increase, more and more fine-grained details on patients' histories are collected, which could be harnessed for accurate patient risk estimation and, ultimately, improved patient care.

Machine Learning (ML) frameworks have shown significant promise in estimating patient risk by leveraging EHRs. However, existing ML approaches in the clinical domain struggle to semantically model the temporal evolution of unstructured EHR data while ignoring the inherent heterogeneity across entity interaction types and documentation viewpoints. Furthermore, patient risk is predominantly modeled as independent trajectories, neglecting higher-order interaction networks within healthcare facilities, and current models lack statistically rigorous uncertainty quantification necessary for safe clinical deployment. This dissertation aims to tackle these limitations across two primary directions. We propose frameworks that (a) imbue ML models to incorporate data-driven, semantic, and domain-specific information, and (b) quantify the overall uncertainty associated with the predictions with statistical frequentist guarantees. The dissertation describes nine projects that have been organized under three parts. The first part focuses on text encoding models for health, with the first three projects focusing on patient risk estimation using clinical notes from EHRs, while the last project focuses on developing a tool for effectively querying prior clinician and domain expert conversations. The second part focuses of developing ML frameworks to improve patient risk estimation by encompassing three projects that operate beyond longitudinal risk estimation to construct interaction networks and leverage domain information for improved risk prediction. The third part focuses on two projects that develop methods to quantify predictive uncertainty for patient interaction networks.