Advancing Population Health Management with Artificial Intelligence: Prediction, Interpretation, and Outcome Measurement in Real-World Health Care Systems

Abstract

Hundreds of millions of people worldwide live with chronic diseases that compromise their daily functioning, cause disability, and impose costs that strain health care systems at every level. Over the past decades, health care systems have progressively adopted electronic health records (EHRs) for clinical operations, administration, and billing, generating vast repositories of patient data. As computational capabilities have grown, artificial intelligence has emerged as a natural complement to this data richness, bringing within reach long-standing aspirations such as population-level health monitoring and precision medicine. Yet deploying artificial intelligence to realize its full value for human health extends well beyond the scope of conventional computational model development. It demands engagement with the specific clinical contexts, heterogeneous patient populations, diverse data modalities, and clinical workflows encountered in real-world healthcare delivery. This dissertation undertakes that engagement through four aims conducted across three large-scale U.S. healthcare systems: University of Washington Medicine (UW Medicine), Kaiser Permanente Southern California (KPSC), and the Veterans Health Administration (VA), each encompassing a distinct, large, and internally heterogeneous patient population. Spanning applications in asthma care management, population health quality measurement, and chronic pain outcome monitoring, these aims collectively demonstrate how artificial intelligence can be developed, evaluated, interpreted, and extended to serve population health in practice. Underlying all four aims is a single overarching question: what should we require of an artificial intelligence system brought into a real health care system so that its predictions, its interpretation, and its measurements answer to clinical value rather than to the typical model performance metrics we build them on and evaluate them against? In Aim 1, we evaluated a machine learning model for predicting asthma hospital encounters under real-world operational conditions at UW Medicine, examining both the timeliness of early risk warnings and the clinical significance of prediction errors. Results showed that the model could identify most at-risk patients months in advance, allowing sufficient time for early intervention, and that a substantial proportion of apparent false positives in fact exhibited signs of poor disease control that warrant preventive interventions. In Aim 2, we addressed the interpretability barrier to clinical adoption of black-box predictive models for asthma care management. Building on association rule mining, a classical data mining technique that can extract data patterns at combinatorial scale, we developed a method to rank the thousands of rule-based explanations generated for each machine learning prediction, enabling clinicians to rapidly access the most actionable insights for individual patients within fast-paced clinical workflows. In Aim 3, we developed a deep learning framework to capture temporal patterns from structured EHR data stored in large-scale relational databases, moving beyond conventional feature engineering that collapses patient history into static summaries. The framework was developed and evaluated on the KPSC data warehouse for predicting asthma hospital encounters, demonstrating practical insights into the trade-offs between sequence modeling complexity and predictive gains when applied to irregularly sampled, real-world clinical data. It was evaluated forward in time, trained on earlier years and tested on a later period whose outcomes extended into the COVID-19 pandemic, and it retained its accuracy and robustness under this temporal shift, an evaluation closer to real deployment than the conventional same-period training-test split. In Aim 4, we turned to the Veterans Health Administration to examine how artificial intelligence can support population-level outcome measurement in a real-world health care system. We first conducted a large-scale evaluation of the Whole Health System of Care, demonstrating that this patient-centered care model was associated with sustained improvements across nine routine clinical quality measures of chronic disease management and preventive care. We then proposed a research protocol for a domain-informed natural language processing (NLP) framework that leverages large language models (LLMs) to extract pain severity and functional interference outcomes from clinical narratives across 18 VA medical centers, grounded in established psychometric instruments and designed to be validated against patient-reported outcomes from a longitudinal nationwide survey. To our knowledge, this is the first study designed to validate pain outcome extraction from clinical text against patient-reported measures in a nationwide longitudinal cohort. Together, these four aims trace an arc from evaluating and interpreting predictive models on structured data to extracting clinical outcomes from unstructured narratives, progressively expanding the reach of artificial intelligence for population health management in real-world healthcare systems.