Health

Stable Learning Establishes Some Common Ground between Causal Inference and Machine Learning

Observational studies of the efficacy of medical treatments are often plagued by selection bias, and expert-selected covariates may insufficiently adjust for confounding. We develop a framework for using unstructured clinical text based on natural language processing to uncover interpretable potential confounders from text. We validate our method by comparing the estimated hazard ratio (HR) with and without the confounders against established RCTs. We apply our method to cancer datasets from the Stanford Cancer Institute. The uncovered terms can also be interpreted by oncologists for clinical insights. The method can enable more credible causal inference using observational data, uncover meaningful insights from clinical text, and inform high-stakes medical decisions.

Uncovering interpretable potential confounders in electronic medical records.

Evaluating the Benefits of Commonly Prescribed Drugs to Prevent Poor Outcomes from Respiratory Distress

Applying recently developed machine learning and causal inference methods to historical health claims data allows the evaluation of the impact of certain medications on outcomes for patients hospitalized with respiratory conditions. The methods incorporate new techniques to estimate causal effects across distinct, proprietary data sources without merging the datasets. The results can be used to suggest candidates for clinical trials in the fight against COVID-19.

AI & Machine Learning

Accelerating Health Technology Innovation to Address COVID-19

Accelerating the development of health technology such as treatments, diagnostics, and vaccines is crucial to ending the COVID-19 crisis as quickly as possible. This project evaluates alternative methods for speeding up the process, including the design of incentives for development and manufacturing, as well as approaches for optimizing clinical trials using shared control groups and adaptive experiments.

Incentive Design

Adaptive Experiments to Help Patients Make Informed Choices About Contraception

Lab researchers are working with Yaounde Gynecology, Obstetrics and Pediatrics Hospital in Cameroon to help women make informed choices about contraceptives. Adaptive experiments identify effective and efficient strategies, including price subsidies, for providing information through a tablet application. The experiment accelerates learning in an environment of many potential alternatives and improves expected patient outcomes by allocating more and more patients to the tablet design that works best.

Adaptive & Iterative Experimentation, Social Sciences & Behavioral Nudges, AI & Machine Learning