- This event has passed.
October 30, 2020 @ 2:00 pm - 3:00 pm
Bias from Treatment Contamination in Clinical Decision Rules
Artificial intelligence approaches are increasingly being applied to clinical decisions. In this talk I examine challenges in producing clinically meaningful, unbiased predictions to support human clinical decision-making based on routinely collected electronic health record data. I use the case of sepsis to illustrate. Wide variation in practice patterns and outcomes for sepsis spawned national campaigns to standardize care. The quick Sepsis Related Organ Failure Assessment [qSOFA] score was included in the 2016 Sepsis-3 criteria. However, decision rules like qSOFA suffer from treatment contamination bias, because outcomes in both treatment and control groups used in their development were measured after the administration of sepsis treatment (e.g., antibiotics). Later definitions of clinical sepsis (Rhee et al, 2017) suffer from the same bias. In this talk I examine the effects of treatment contamination bias in sepsis decision rules and propose alternative approaches.
Finding Natural Experiments in Observational Data
Evidence on clinical treatments based on real world data has the great benefit of representing effectiveness in a broader population than is typically studied in prospective experiments. However, confounders that influence both the choice of treatment and the outcomes of interest can disguise the truth about a given treatment’s effectiveness. Methods such as propensity score matching and weighting offer approaches to minimize the effect of these confounders, but leave open the question of which populations of people should be compared. In this work, we explore a computational approach to identifying the subgroup from all treatment arms of an observational study that are comparable, effectively finding the natural experiment for the target treatments. This work leverages a deep generative model based on generative adversarial networks and results in the identification of a comparison distribution that minimizes the variance of estimates such as the average treatment effect.