Heartfailure(HF)isahighlydisablingandcostlydiseasewithahighmortalityrate.Intheprediagnosticphase (i.e., 1236 months before diagnosis), HF is difficult to detect given the insidious signs and symptoms. After diagnosis, where it isnotpossibletoreversediseaseprogression,effortsaremadetoavoidhospitaladmission and readmission, but with limited capabilities to stratify patients by risk. We propose to develop interpretable deeplearningmodelsappliedtolargescaleelectronichealthrecord(EHR)datatodetectHFrelatedeventson two different time scales. One set of modelswillbedevelopedtodetectHFdiagnosisonetotwoyearsbefore actual documented diagnosis. Separately, we propose to identify HF patients who are at risk of hospital admission and readmission.Theprojectfocusesondevelopingdeeplearningmodelsthatofferthepotentialfor greater accuracy, clinical interpretability, and utility than alternatives. The expected deliverables include comprehensive software for creating deep learning algorithms that predict HF outcomes and related software toolsformodelvisualization.
Deep learning has shown tremendous success in many domains but is yet to have similar impact in health care. The key challenges in healthcare applications are the lack of interpretation for deep learning models and limited transferability of the models across institutions. We develop interpretable deep learning algorithms for heart failure prediction that can handle large longitudinal patient records and are able to adapt across institutions.