Prediction of Adverse Events in Single Ventricle Physiology Infants Using Artificial Intelligence Tools

BACKGROUND: Adverse events (AEs) within the cardiac ICU (CICU) are associated with high mortality and comorbidity. Patients with single ventricle (SV) physiology are particularly vulnerable to experiencing such events before second stage surgery (bidirectional Glenn). Timely identification and management of AEs are critical for improving patient outcomes and enabling earlier, more targeted medical interventions in this vulnerable population. OBJECTIVES: To develop and evaluate machine learning (ML) models using continuous physiologic data to predict and identify AEs including cardiac arrest (CA), extracorporeal membrane oxygenation (ECMO) cannulation, and endotracheal intubation, up to 8 hours before occurrence in SV infants. DERIVATION COHORT: Retrospective cohort of 158 SV patients (324 admissions) admitted to the tertiary care CICU at Children’s National Hospital. VALIDATION COHORT: Internal validation occurred in the held-out testing group of the data (10% of the total dataset). PREDICTION MODEL: Supervised ML classifiers were (e.g., decision tree, logistic regression, support vector machine, extreme gradient boosting, random forest RF) trained and compared across four observation windows (8, 4, 2, and 1 hr) preceding each event. Model Performance was evaluated using the area under the receiver operating characteristic curve (AUROC). Fifteen physiologic and laboratory variables were extracted, of which six high-quality variables were included in the ML model. AEs were categorized into four categories (e.g., intubation, CA, ECMO, no event) for multiclass classification and ECMO-CA were also combined into a single class (ECMO-CA) to improve stability for rare events. RESULTS: A total of 256 AEs were analyzed: 157 intubations (61.328%), 42 ECMO events (16.406%), 44 CAs (17.187%), and 13 extracorporeal cardiopulmonary resuscitation events (5.078%). Across all time windows, the RF model achieved the best performance on the held-out test set for AE detection (AUROCs 0.998 at 8 hr, 0.996 at 4 hr, 0.996 at 2 hr, and 0.997 at 1 hr). For the combined-class multiclass classification RF model at 1-hour observation window, the held-out test set results showed AUROCs of 0.819 for intubation, 0.804 for ECMO-CA, and 0.840 for no-event prediction. CONCLUSIONS: A ML model can predict and discriminate between several types of AEs in SV infants before bidirectional Glenn. Accurate predictions may help perform timely interventions, potentially reducing morbidity, mortality, and healthcare costs.