Applications of Data Science and Artificial Intelligence to Decision Making in Healthcare and Finance

Abstract

Decision-making requires timely and accurate information in order to understand the implications of the actions and to manage the potential risk. This thesis presents computational methods to quantify risk in drug development programs, address current challenges in health economics, and investigate and predict rare events in finance. The thesis is split into three major parts.

Part I addresses a core issue in accessing the risk and value of drug development programs: the probability of success (PoS). We introduce a Markov chain model of a drug development program that allows us to fill in missing data and infer phase transitions from clinical trial metadata. We investigate the PoSs across various therapeutic areas, and then conduct further analysis for areas that are of public interest (e.g., oncology, vaccine, and anti-infective therapeutic) in order to understand the bottlenecks in the drug development process.

Part II of the thesis focuses on the use of modeling and simulations to make informed predictions and drive policy-making in healthcare. One chapter in this Part is devoted to the use of data to estimate the financial impact of gene therapy in the U.S. between 2020 and 2035, while another chapter is dedicated to estimating the cost and benefit of various clinical trial designs for the development of a vaccine to prevent COVID-19.

Part III presents a novel 'big data' analysis and machine learning prediction model of panic selling behavior by retail investors. We document the frequency and timing of panic selling, analyze the demographics of investors who tend to freak out and panic sell, and determine if panic selling is a detrimental or optimal action financially. We also develop machine learning models to predict if an investor might panic sell in the near future given the demographic characteristics of the investor, their portfolio history, and the current and past market conditions.