B-po01-020 machine learning to predict recurrent events following unexplained cardiac arrest

Patients who suffer an unexplained cardiac arrest (UCA) are at risk for recurrence. We compared machine learning (ML) models to predict recurrent events among UCA survivors using baseline clinical data with and without longitudinal diagnostic data. Patients with prior UCA were enrolled in the Cardiac Arrest Survivors with Preserved Ejection Fraction Registry (CASPER) and followed for events (ICD shock or ATP). ML models predicted events using 1) baseline clinical data alone or 2) temporal data, using baseline and evolving diagnostic data. Models were created using SQL queries with a 75% derivation and 25% validation cohort. 616 patients were followed for 5.1±4.2 years. The mean age was 50±15 years (62% male), 38% had pre-arrest syncope. At follow-up, standard and high lead ECG detected Brugada pattern (9%), Early Repolarization (11%), QTc prolongation (11%), and T-wave abnormality (18%). Borderline or abnormal results were found on signal averaged ECG (14%), exercise test (8%), procainamide challenge (3%), and MRI (13%). Pathogenic and likely pathogenic variants were found in channelopathy (2%) and cardiomyopathy genes (1%). 118 patients (19%) had recurrent events (mean time-to-event 2.4±2.7 years). Baseline models yielded intermediate accuracy: regression- 68%; naïve Bayes- 63%; conditional inference- 64%; bootstrap aggregating- 72%. Temporal models yielded better accuracy: regression- 73%; naïve Bayes- 76%; conditional inference- 71%; bootstrap aggregating- 93% (Figure). In a heterogenous cohort of UCA survivors, a temporal ML model accurately predicted recurrent events, and can risk stratify and inform patient care.