DeepAlerts: Deep Learning Based Multi-Horizon Alerts for Clinical Deterioration on Oncology Hospital Wards

摘要: Machine learning and data mining techniques are increasingly being applied to electronic health record (EHR) discover underlying patterns make predictions for clinical use. For instance, these may be evaluated predict deterioration events such as cardiopulmonary arrest or escalation of care the intensive unit (ICU). In practice, early warning systems with multiple time horizons could indicate different levels urgency, allowing clinicians decisions regarding triage, testing, interventions patients at risk poor outcomes. These horizon alerts related have intrinsic dependencies, which elicit multi-task learning. this paper, we investigate approaches properly train deep models predicting via generating multi-horizon hospitalized outside ICU, particular application oncology patients. Prior knowledge is used a regularization exploit positive effects from task relatedness. Simultaneously, propose task-specific loss balancing reduce negative when optimizing joint function models. addition, demonstrate effectiveness feature-generating prediction outcome interpretation. To evaluate model performance in real world scenario, apply our EHR 20,700 hospitalizations adult patients' baseline high-risk status provides unique opportunity: an accurate enriched population produce improved predictive value false alerts. With dataset, applying all proposed achieves best compared common previously developed warning.