Dissertation Title:  "Topics in Machine Learning for Health Services Research"

This dissertation explores topics where machine learning can be used to improve or expand the scope of health services research.

Chapter 1: This study demonstrates the viability of using deep learning to identify people who are likely to benefit from osteoporotic or fragility fracture risk screening using chest radiographs. Previous work has shown that deep learning algorithms are capable of identifying osteoporotic individuals across a wide range of imaging modalities. However, the vast majority of studies have focused on patients that have been screened for osteoporosis and the study populations are almost entirely comprised of patients who are already recommended for screening. We develop and validate an algorithm on a large dataset of chest radiographs consisting of 59,737 individuals that is significantly more diverse across both age and sex than the data used in previous work. Using an ensemble of image classification models and conformal prediction, our algorithm was able to identify individuals without an osteoporosis or osteopenia diagnosis that were 1.93 (95% CI: 1.66, 2.26) times more likely to have experienced any fracture and more than twice as likely to experience common osteoporotic fractures, such as hip and pelvis fractures (Est: 2.19, 95% CI: 1.67, 2.88) or spine and rib fractures (Est: 2.49, 95% CI: 1.99, 3.11), after adjusting for age, sex, and BMI. Approximately 45% of individuals identified were not currently recommended for screening. We further outline how conformal prediction can be used to adjust the size of the flagged patient population to account for important implementation factors, such as a health centers’ capacity to screen additional patients or requirements for a higher standard of evidence when making recommendations that fall outside of current clinical guidelines. The size of the flagged patient population could also be set to match the projected cost-benefit tradeoff of screening additional people.

Chapter 2:  Early entry by firms in markets for generic pharmaceuticals is paramount for price declines and increasing consumer welfare. The first-mover advantage (FMA), defined as the additional market share a firm earns by entering first compared to entering later, is considered one of the most important incentives for prompt entry into generic markets. Inherent in this definition of FMA is a notion of causality, where, in an ideal but infeasible experiment, we could observe the same firm entering at different times in the same market and compare market shares across different entry timing decisions. This paper will exploit unique characteristics of generic drug markets and advances in doubly robust methods for causal inference to estimate the FMA. Our findings suggest entering first results in a significant advantage when compared to if that same firm entered later. This advantage is mainly accrued through sales in periods where competition is limited and, to a lesser extent, higher market shares during the first two years after the start of competition between generics. The latter point contradicts a number of previous studies that found the FMA is sustained for up to six years after the start of competition between generics. We further characterize the impact that important regulatory features, such as exclusivity periods and the presence of authorized generics, have on the market shares of first entrants and provide evidence to suggest our estimates are unlikely to be explained by unobserved confounding.

Chapter 3: COVID-19 interrupted delivery of mental health care in the US. Symptoms associated with various mental disorders increased in prevalence at the same time. The expectation is that treatment would increase with measured need. Departures from that expectation serve to index the degree of disruption in the delivery of mental health care to the US population. We conducted a retrospective observational analysis using prescription claims data covering 89 percent of all prescriptions in the US that compared observed new-starts of common psychotropic medications to forecasted new-starts. Forecasts were generated using the Prophet forecasting model. During the initial course of the COVID-19 pandemic new starts of antidepressants declined by 7.5 percent, anxiolytics by 5.6 percent, and antipsychotics by 2.6 percent compared with expected levels. Declines were more pronounced among children and adolescents, with declines in new starts ranging from 20 to 30% over the same period for the three drug classes. Our findings suggest that there was a large unmet need for mental health treatment in the US attributable to COVID-19 over this period.