Helen Lingyun Shi
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Helen Lingyun Shi has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
- Akira Nishisaki
Department of Anesthesiology, Critical Care, and Pediatrics, University of Pennsylvania Perelman School of Medicine
1 shared publication Eamonn Tweedy
Tsui Laboratory, Department of Biomedical and Health Informatics, Children’s Hospital of Philadelphia, Philadelphia, PA
1 shared publication- Fuchiang (Rich) Tsui
Tsui Laboratory, Department of Biomedical and Health Informatics, Children’s Hospital of Philadelphia , Philadelphia, PA 19146
1 shared publication - Robert M. Md Sutton
Division of Critical Care Medicine, Department of Anesthesiology and Critical Care Medicine, The Children’s Hospital of Philadelphia, Philadelphia, PA
1 shared publication - Ryan W. Morgan
Division of Critical Care Medicine, Department of Anesthesiology and Critical Care Medicine, The Children’s Hospital of Philadelphia, Philadelphia, PA
1 shared publication - Sanjiv Mehta
Division of Critical Care Medicine, Department of Anesthesiology and Critical Care Medicine, The Children’s Hospital of Philadelphia, Philadelphia, PA
1 shared publication
Latest Publications
- Open Access
- 22 April 2026
- Published by Wolters Kluwer Health in Critical Care Explorations
- Vol. 8 (4) , e1397
Machine Learning-Based Prediction of Critical Deterioration in the PICU
BACKGROUND: Existing PICU early warning systems lack sufficient accuracy and timeliness for effective preparation. Machine learning approaches may improve prediction of critical deterioration events (CDEs), but their operational utility relative to existing tools remains unclear. OBJECTIVES: To develop a machine learning model for early detection of CDEs and evaluate operational utility against existing tools using a novel alert burden analysis. DERIVATION COHORT: PICU admissions (ages 0–24 yr, stay ≥ 24 hr) at a quaternary children’s hospital from 2014 to 2020 ( n = 12,771 patients; 21,141 admissions). CDEs (6% of patients) included cardiopulmonary resuscitation, extracorporeal membrane oxygenation initiation, dilute epinephrine administration, or unplanned intubation. VALIDATION COHORT: Temporally distinct PICU admissions from 2021 to 2022 ( n = 5144 patients; 6929 admissions; 6% CDE rate). PREDICTION MODEL: An ensemble of extreme gradient-boosted models (PICU Warning INdex [P-WIN]) trained to predict CDEs at 1–12-hour horizons using 550 features derived from demographics, medications, laboratory results, and vital signs. RESULTS: P-WIN demonstrated excellent discrimination at 2-hour (area under the receiver operating characteristic curve [AUROC], 0.95 [95% CI, 0.94–0.96] and area under the precision-recall curve [AUPRC], 0.76 [95% CI, 0.72–0.80]) and 12-hour horizons (AUROC, 0.93 [95% CI, 0.92–0.94] and AUPRC, 0.68 [95% CI, 0.64–0.72]). To alert before 80% of events, P-WIN generated 0.20 alerts per patient-day at a median 10.17 hours before CDE. Compared with the existing rule-based PICU Warning Tool (alerting before 38% of events), P-WIN generated one-third the alert burden at equivalent sensitivity (0.03 vs. 0.10 alerts per patient-day). CONCLUSIONS: P-WIN accurately predicted PICU CDEs up to 12 hours in advance with low alert burden, providing a viable opportunity for shifting care from reactive rescue to proactive, resource-intensive preparation and prevention.
- 22 April 2026
- Published by Wolters Kluwer Health in Critical Care Explorations
- Vol. 8 (4) , e1397
Open AccessMachine Learning-Based Prediction of Critical Deterioration in the PICU
BACKGROUND: Existing PICU early warning systems lack sufficient accuracy and timeliness for effective preparation. Machine learning approaches may improve prediction of critical deterioration events (CDEs), but their operational utility relative to existing tools remains unclear. OBJECTIVES: To develop a machine learning model for early detection of CDEs and evaluate operational utility against existing tools using a novel alert burden analysis. DERIVATION COHORT: PICU admissions (ages 0–24 yr, stay ≥ 24 hr) at a quaternary children’s hospital from 2014 to 2020 ( n = 12,771 patients; 21,141 admissions). CDEs (6% of patients) included cardiopulmonary resuscitation, extracorporeal membrane oxygenation initiation, dilute epinephrine administration, or unplanned intubation. VALIDATION COHORT: Temporally distinct PICU admissions from 2021 to 2022 ( n = 5144 patients; 6929 admissions; 6% CDE rate). PREDICTION MODEL: An ensemble of extreme gradient-boosted models (PICU Warning INdex [P-WIN]) trained to predict CDEs at 1–12-hour horizons using 550 features derived from demographics, medications, laboratory results, and vital signs. RESULTS: P-WIN demonstrated excellent discrimination at 2-hour (area under the receiver operating characteristic curve [AUROC], 0.95 [95% CI, 0.94–0.96] and area under the precision-recall curve [AUPRC], 0.76 [95% CI, 0.72–0.80]) and 12-hour horizons (AUROC, 0.93 [95% CI, 0.92–0.94] and AUPRC, 0.68 [95% CI, 0.64–0.72]). To alert before 80% of events, P-WIN generated 0.20 alerts per patient-day at a median 10.17 hours before CDE. Compared with the existing rule-based PICU Warning Tool (alerting before 38% of events), P-WIN generated one-third the alert burden at equivalent sensitivity (0.03 vs. 0.10 alerts per patient-day). CONCLUSIONS: P-WIN accurately predicted PICU CDEs up to 12 hours in advance with low alert burden, providing a viable opportunity for shifting care from reactive rescue to proactive, resource-intensive preparation and prevention.