Using Artificial Intelligence to Understand and Improve Doctors' Decisions
主講人：Ziad Obermeyer, MD(UC Berkeley)
簡介：Low-value health care—care that provides little health benefit relative to its cost—is a central concern for policymakers.Identifying exactly which care is likely to be of low-value ex ante, however, has proven challenging. Here we apply machine learning tools to study an iconic decision, widely thought to epitomize low-value care: advanced testing for heart attack (acute coronary syndromes) in the emergency setting. By comparing doctors' decisions to individualized, prospective risk estimates, we show that mis-prediction of risk is a major driver of low-value care, in two ways. First, we find substantial over-testing: patients with very low model-predicted risk, whom doctors nonetheless decide to test. These tests are low yield, in that few patients go on to benefit from interventions to treat heart attack in their wake. Second, we also find evidence of a second kind of low-value care, under-testing: large numbers of patients at high model-predicted risk, whom doctors choose not to test. We find serious complications (or death), consistent with untreated heart attack, at remarkably high rates in highrisk patients. These results suggest that both under- and over-testing are prevalent, and that targeting misprediction is an important but understudied policy priority.