September 13, 2022

12:00 pm / 1:00 pm

Venue

Clark Hall, Room 110

Join Zoom Meeting:

https://wse.zoom.us/j/98624413365

Meeting ID: 986 2441 3365

Guest speaker:  Harsh Parikh

Ph.D. Candidate

Duke University

?Interpretable Causal Inference for High-Stakes Decision Making?

Abstract: Many fundamental problems affecting the care ofcritically ill patients lead to similar analytical challenges: physicianscannot easily estimate the effects of at-risk medical conditions or treatments(which is problematic for treatment decisions) because the causal effects ofmedical conditions and drugs are entangled. They also cannot easily performstudies: there are not enough critically ill patients for high-dimensionalobservational causal inference analysis, and randomized controlled trials oftencannot ethically be conducted. Our work introduces a general framework that canhelp estimate heterogeneous causal effects from high-dimensional patient-leveldata under these conditions. Each step of our framework is designed to beinterpretable. Importantly, we leverage established mechanistic models todescribe personalized decision-response interactions, allowing us to identifyindividuals who might react similarly to treatments. We learn a flexibledistance metric on the space of covariates to perform almost exact matching forestimating the medium and long term causal effects. The learned distance metricstretches the covariate space according to each covariate’s contribution toprognosis: this stretching means that mismatches on important covariates carrya larger penalty than mismatches on irrelevant covariates. The matched group weconstruct for each patient can be validated, or possibly, criticized. In thecontext of medical data, this validation can be performed via a chart reviewthat provides a qualitative assessment of the matches in terms of informationthat was not directly used in the matching procedure.

Biography: Harsh Parikh is a Ph.D. Candidate at Duke Universityworkingwith Dr. Cynthia Rudin, Dr. Alexander Volfvosky, and Dr. Sudeepa Roy asapart of Almost Matching Exactly Lab. His research interest includes workingon causal inference methodology with applications in critical care, publichealth, or education. His current research work includes (i)interpretable-and-accurate matching methods for high-stakes scenarios, (ii)methods for causal inference on social network/relational data, and (iii)frameworks tocombine experimental and observational data. He has ongoingactive collaboration with neuro-physicians at MGH and researchers at AmazonScience. He also received the Amazon Graduate Research Fellowship (Sept 2020 -Jan 2022). He received B.Tech in Computer Science from IIT Delhi (2015)and M.S inEconomics and Computation from Duke University (2018).