Informing Public Health Policy Design and Operations with Analytics: Methods and Applications

Abstract

Data-driven approaches hold immense potential for improving public health decision-making in complex, uncertain, and high-risk environments. Yet, there are several key challenges that stand in the way of successfully leveraging large-scale, heterogeneous, and noisy data into timely and actionable policy insights. These challenges are particularly pronounced when conventional modeling tools fall short, for instance, in settings where health risks arise from infection propagation in intricate supply chains, rapidly mutating pathogens, or delayed and fragmented surveillance systems. This thesis introduces a suite of novel methodologies and use cases at the intersection of operations research, epidemiology, and machine learning to address some of these challenges and support more informed, timely, and proactive public health decisions.

A central focus of this thesis is the management of health risks related to zoonotic viruses, which are pathogens that emerge in animals and can potentially jump to humans, then further evolve to become transmissible between humans. These viruses pose a growing global health threat. Notably, outbreaks of zoonotic viruses frequently emerge in live animal markets in developing countries, even when infection rates in the upstream farms supplying these markets remain consistently low. Motivated by this empirical observation, the first chapter of this thesis develops an innovative epidemiological model called the Transmission, Interaction, and Persistence (TIP) model. This model integrates stochastic supply chain dynamics and environmental transmission mechanisms, and sheds light on how market-level factors amplify the risk of infection outbreaks. It yields actionable insights regarding the potential effectiveness of risk mitigation strategies such as frequent market sanitation and supply consolidation.

Since March 2020, the world has experienced multiple waves of infections caused by the SARS-CoV-2 virus. Similar to past pandemics, SARS-CoV-2 has spread in waves, each driven by different genetic variants of the virus. Public health agencies have often struggled to predict in advance which variants would drive a new wave of infections. The second chapter of this thesis introduces an AI-enabled early warning system for emerging viral variants. The newly developed predictive model incorporates genetic and epidemiological features and is trained and tested on over 9 million sequenced SARS-CoV-2 variants across 30 countries. It accurately predicts whether each new variant will drive a significant wave of infections within the following 3 months.

There is ample biological and empirical evidence regarding the roles of mutating variants and population immunity in driving infection waves of respiratory viruses. Motivated by this, the third chapter of the thesis develops the first epidemiological model, called the Immunity-Variants-Epidemic (IV-Epidemic) model, that explicitly captures circulating variants and the evolving population immunity profile to more accurately reflect the long-term trajectory of variant-driven pandemics. It incorporates variant evolution and population immunity dynamics, and is able to replicate the observed multi-wave infection patterns without requiring ad hoc recalibration.

The fourth chapter of the thesis focuses on post-marketing pharmacovigilance, which is key to drug safety regulatory work. It presents PR1SM (Patients Really are 1st in Signal Management), an AI-based framework for identifying potential drug safety signals using post-marketing surveillance data. By structuring adverse event reports into parallel time series at multiple levels of clinical aggregation and adjusting for exposure trends, PR1SM complements standard disproportionality methods to detect safety signals earlier and with greater sensitivity in both real-world and synthetic settings.

Collectively, the chapters of this thesis demonstrate how operations research can be combined with domain-specific methods in biology, epidemiology, and pharmacovigilance to inform data-driven public health strategies. The proposed analytical frameworks offer interpretable, scalable, and policy-relevant tools to create more resilient public health systems.