Abstract: Plausibility (truthfulness) of clinical observations in EHRs can be questionable. Gold standard cut-off ranges are the primary approach in many clinical data repositories for identifying implausible values, yet they often do not take nonlinearities of clinical observation data into account. We demonstrate the utility of semi-supervised encoders (superencoders) for outlier detection through compressing the data into an exemplar distribution learned from supposedly less noisy random samples. We found that, in general, simple superencoders can produce acceptable exemplar compressions of clinical observation data and outlier detection performances.
Learning Objective 1: At the conclusion of this activity, participants will be able to discuss how semi-supervised encoding can be applied to clinical observation data compression and outlier detection.
Hossein Estiri (Presenter)
Harvard Medical School
Shawn Murphy, Harvard Medical School