Actualizing Impact of AI in Public Health: Optimization of Scarce Health Intervention Resources in the Real World

Abstract

While AI is assuming omnipresence today more than ever, its adoption is still limited in solving challenges pertaining to socially-critical problem domains such as in public health, especially among low-resource and underserved communities. Motivated by the desire to solve impactful, real-world problems that involve reasoning, strategic decision-making, or planning in uncertain, stochastic or resource-limited settings, my thesis presents novel solutions designed for two such real-world public health challenges: tuberculosis prevention and improving maternal and child healthcare. Building AI systems for realizing social impact in public health, demands solving a number of fundamental research questions. For instance, community health workers and NGOs operating with limited health resources face the challenge of optimally utilizing these resources to maximize their impact. In doing so, such NGOs must account for domain-specific considerations such as fairness or risk-averseness and plan the limited resources to serve beneficiaries at scale, in an uncertain and dynamically changing world. Even with new solution techniques for allocating limited resources in such socially critical domains now being built, their accurate evaluation through Randomized Controlled Trials (RCTs) remains difficult due to high sample variance in these settings. Towards tackling these challenges, my thesis utilizes techniques such as Restless Multi-Armed Bandits (RMAB) to solve the sequential decision-making problem of allocating scarce health intervention resources. My thesis builds computationally efficient solution algorithms to this problem, that can be adopted by non-profits without needing access to heavy computing power. Next, I also propose techniques that allow the planner to accommodate real-world considerations such as risk-averseness or fairness in planning health interventions. Furthermore, my thesis also builds solutions that can plan such health interventions while accounting for dynamically changing patient cohorts and the finite stay of patients in such health programs. Transcending the boundaries of traditional research, I have transitioned this work from the blackboard to a first-of-its-kind field evaluation of the RMAB algorithm, involving 23,000 real-world mothers over a 7-week period, results of which show a 30% improvement in the performance metric of interest. Finally, my work mitigates the challenges faced in the evaluation of such resource allocation algorithms through RCTs. Using techniques from causal reasoning, I present a novel concept that retrospectively reassigns participants to experimental groups in a trial. Using this concept, I build a new estimator, that I show, can sharply reduce sample variance.