Neural Event Prediction for Clinical Event Time-SeriesLee, Jeong Min (2022) Neural Event Prediction for Clinical Event Time-Series. Doctoral Dissertation, University of Pittsburgh. (Unpublished) This is the latest version of this item.
AbstractMassive clinical event time-series data collected in Electronic Health Records (EHR) offer great potential for improving patient care as they contain in-depth information about patient conditions, relevant diagnoses, and treatment strategies. With event prediction models, we can identify temporal associations among various types of clinical events in EHR, such as symptoms and patient management actions on one side and symptoms and outcomes with or without patient management actions on the other side. Further, we could predict the future occurrence of adverse events and help healthcare practitioners to intervene ahead of time or prepare resources to get ready for their occurrence. However, building clinical event prediction models has unique challenges posed by inherent characteristics of EHR data: (1) Different temporal characteristics. Each event in the multivariate time-series has different temporal behaviors (e.g., repetitively occurring with certain time gaps) and different temporal ranges of dependencies for precursor events. To accurately predict future events from the multiple event time series with different temporal characteristics, we need more flexible and expressive models. (2) Patient-specific variability. Based on underlying clinical conditions, each patient's sequence may consist of different sets of clinical events (observations, lab results, medications, procedures). Hence, simple population-based models learned from event sequences for many different patients may not accurately predict patient-specific dynamics of event sequences and their differences. In this thesis, we propose novel autoregressive event prediction models that can address the aforementioned issues. First, we propose new models that handle different temporal dependencies using multiple temporal mechanisms covering various time scales and temporal behaviors such as recurrence of events and multi-time-scale dependencies. Second, we develop new personalized event prediction models that let us better adjust the prediction for individual patients and their specific conditions. They pursue refinement of population-wide models to subpopulations, patient-specific model adaptation, and a meta-level model switching that can adaptively select the model with the best chance to support the immediate prediction. We evaluate our proposed models on the real-world clinical data derived from EHR of critical care patients. We show that our new models lead to improved prediction performance compared to multiple baselines. Share
Details
Available Versions of this Item
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|