(210) Predicting Deterioration of Single-Ventricle Patients in Intensive Care Using Bedside Monitor Data

Sunday, February 23, 2025

11:00am – 12:00pm Eastern Time

Location: Connections Central - RST 12

First Author(s)

FT

Fuchiang Rich Tsui, MS, PhD, FAMIA

Associate Professor, Endowed Chair in Biomedical Informatics and Entrepreneurial Science
n/a, United States

Co-Author(s)

VR

Victor M. Ruiz, PhD

Data Scientist
Department of Biomedical and Health Informatics, Children's Hospital of Philadelphia, United States
LS

Lingyun Shi, MS

Data Scientist
Department of Biomedical and Health Informatics, Children's Hospital of Philadelphia, United States
RW

Renee Willet, MD

Attending Physician
Children's Hospital of Philadelphia, Pennsylvania, United States
MG

Michael Goldsmith, MD

Attending Physician
Children's Hospital of Philadelphia, Pennsylvania, United States

Introduction: Infants with single-ventricle (SV) physiology are a complex subpopulation with elevated mortality and morbidity risk before stage-2 palliation. Early detection of deteriorating physiology in these patients may help decrease critical decompensation events and allow timely interventions. Previous studies used either electronic health records or high-frequency waveforms for deterioration prediction; however, there is a lack of studies evaluating the value of low-frequency vital-sign data sampled at the second level from bedside monitors for deterioration prediction in this SV population.

Methods: We conducted a retrospective single-center cohort study with 3+ years of data from a cardiac intensive care unit (CICU) at a tertiary children’s hospital. We included all single ventricle patient admissions prior to Stage II surgery during the study period between 2019-11-01 and 2023-04-23. We excluded admissions with length of stay < 24 hours or deterioration events in the first 8 hours of CICU admission. The primary outcome was deterioration with any of the following critical events: unplanned intubation, cardiac arrest, and ECMO cannulation. We collected 22 vital-sign variables from bedside monitors sampled at 0.5Hz. We developed a random forest (RF) model at one hour before the event. The evaluation metrics included area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC).

Results: 45 deterioration events from 33 SV patients and 48 randomly sampled non-deterioration time points from 36 patients. At 1 hour prior to deterioration, the RF model had an AUC of 0.83 (95% C.I.: 0.52-1.00) and an AUPRC of 0.86 (0.65-1.00).

Conclusions: We demonstrated that a machine-learning random forest model solely based on low-frequency data successfully predicted deterioration in the high-risk single-ventricle population. Future work includes integrating low-frequency vital-sign and EHR data with external validation before clinical use.