Health as a Timeline: Machine Learning and Electronic Health Records to Characterize, Predict, and Intervene
Electronic health records (EHRs) now serve health care professionals in 95 percent of hospitals nationwide, a 9-fold increase from a decade ago. Computerization is causing a renovation throughout health care, and analytical techniques are needed to bring the potential of this readily growing data source to full potential, from mortality reduction and early warning detection, to patient education and informed consent. A leading representation of patient data is representation as timelines. The work I will present describes statistical timeline analysis--specifically the development of tractable learning and inference algorithms in continuous-time Bayesian networks and point processes--to characterize and predict temporal health trajectories. These methods are explored both in simulation and in 50 years of EHR data at the Marshfield Clinic in Wisconsin, producing personalized forecasts and recommendations for optimal outcomes. Health data and big data frameworks are here now--our next step is to combine these data, frameworks, and analysis to help fulfill the promise of improving health outcomes by moving beyond paper.