Research Project
9620967
Abstract Significance: In this SBIR project, we propose to predict and detect pediatric severe sepsis by developing a machine-learning-based clinical decision support system for electronic health record (EHR) pediatric sepsis screening. The pediatric population is underserved, with fundamental research and understanding of pediatric sepsis syndromes lagging behind that of the adult population. The proposed work will develop machine learning sepsis predictions on the highly heterogeneous pediatric population, by combining multi-task learning methods with expert clinical knowledge of how pediatric sepsis presentation is dependent on age and on pre- existing conditions. The multi-task learning approach will use age or comorbidities to define ?tasks,? each one associated with prediction on a particular subpopulation, and then link the learning process together between tasks. Research Questions: Which methods of using these task-defining parameters are most effective? How can we most effectively learn the degree of similarity between pediatric subpopulations and leverage this to improve classification performance? Prior Work: InSight was originally developed to predict sepsis and septic shock from adult EHR data. After retraining on pediatric cases, in preliminary experiments with a retrospective set of pediatric (2-17 yr) inpatient encounters (n = 11,127;? 103 [0.9%] severely septic), at the University of California San Francisco (UCSF), InSight achieved an AUROC 0.912 and 0.727 for the detection and 4-hour pre-onset prediction of sepsis. This performance can be improved for better pediatric sepsis prediction. Specific Aims: To empirically evaluate different learning schemes using age with and without multi-task methods, within the UCSF pediatric severe sepsis data set (Aim 1). To exploit an expert-proposed network graph structure for comorbidity-described pediatric subpopulations that provides superior predictive performance over naïve methods and graphs, both for the overall population and for underserved subpopulations (Aim 2). Methods: We propose to use multi-task methods that penalize deviations between classifiers on neighboring tasks and that iteratively learn the strength of these links. These methods will be compared with total task independence, or passing age into classifier training as an ordinary input. Criteria for Success: Success will be shown by 4-hour pre-onset AUROC gains of 0.02 (overall population) and 0.03 for the previously weakest of three age subpopulations (2-5, 6-12, and 13-17 yrs;? 4-hour pre-onset prediction, p < 0.05, McNemar?s test, 4-fold cross-validation). The best structure for mapping similarities between comorbidity tasks will improve the overall AUROC by 0.03 (p < 0.05) and by 0.07 for ? 2 comorbidity subpopulations of ? 100 patients (p < 0.05). Outcome: These improvements will enable InSight to deliver strong sepsis predictive performance across the widely heterogeneous pediatric population.
