Abstract
AIMS: Heart rate variability (HRV), a non-invasive measure of autonomic function, may offer prognostic value after pediatric cardiac arrest. We used machine learning models to determine whether HRV features within the first 24 h after return of spontaneous circulation can predict outcomes in children following cardiac arrest, and whether adding clinical cardiac arrest characteristics improves model performance. METHODS: Retrospective study of children who received post-arrest care in the PICU at the Children's Hospital of Philadelphia from 2020 to 2023. Thirty-six HRV features were extracted from ECG recordings and Extreme Gradient Boosting (XGB) models were trained to predict unfavorable neurological outcome, defined as Pediatric Cerebral Performance Category 4-6 and an increase >1 from baseline. Models were evaluated by cross-validation across the entire 24-h period and within sequential 6-h epochs. Additional models included clinical arrest characteristics. Performance was assessed by area under the receiver operating characteristic curve (AUROC). RESULTS: Of the 75 patients who met inclusion criteria (median age 6.8 [IQR 10.4] years), 51% had an unfavorable outcome. Model considering HRV features and age achieved an AUROC of 0.80 (95% CI: 0.68-0.88). Top HRV predictors included standard deviation (SDNN), power at very low and low frequency bands, entropy, and fractal scaling. Performance was similar across the 6-h epochs (p's > 0.1). Adding cardiac arrest characteristics did not improve model performance (AUROC 0.83 [0.73-0.92], p > 0.41). CONCLUSION: Using machine learning, HRV features within 24 h after pediatric cardiac arrest predict unfavorable outcome with AUROC 0.8. Adding clinical variables did not improve model performance.