Physics of the Medical Record

Abstract

Dr. George Hripcsak will discuss the national push for electronic health records that will make an unprecedented amount of clinical information available for research, with approximately one billion patient visits documented per year in the US. These data may lead to discoveries that improve understanding of biology, aid the diagnosis and treatment of disease, and permit the inclusion of diverse populations and rare diseases. Health record data also bring challenges, with missing and inaccurate data and substantial bias. Some of the bias comes about because the data are generated as part of the health care process. For example blood tests done at night select for patients who are sicker. We employ nonlinear time series methods borrowed from other fields to address the challenges, and we attempt to model and correct for biases due to the health care process. We incorporate mechanistic models to constrain the search space to create accurate predictions despite limited training sets and missing values.