Clinical Electrophysiology -> SCA Risk Assessment: -> Clinical Risk Assessment Techniques D-AB11 - Emerging Innovations to Predict Sudden Death (ID 18) Abstract Plus


Background: Patients who suffer an unexplained cardiac arrest (UCA) may be at risk for a life-threatening recurrence. However, identifying those at risk for recurrence in this heterogenous population remains challenging using conventional models for risk prediction.
Objective: To evaluate the accuracy of a machine learning model to predict recurrent events among UCA survivors.
Methods: Patients with prior UCA (LVEF>50%, no coronary stenosis>50%) were enrolled in the Cardiac Arrest Survivors with Preserved Ejection Fraction Registry (CASPER). A comprehensive evaluation including ECGs, exercise treadmill testing, cardiac imaging, provocative drug challenge, and genetic testing were performed where appropriate. Patients were followed for recurrent events, including ICD shock or ATP. Machine learning models were employed to predict recurrent events using baseline demographics and investigations to overcome challenges with variable investigations and/or missing data.
Results: 597 UCA patients were followed for 3.15±2.28 years. The mean age was 46.9±14.4 years, 40% were female, and 39% had a history of syncope. Baseline ECG findings included resting heart rate (mean 71±17 bpm), Brugada pattern ST elevation (4%), Early Repolarization pattern (12%), T-wave abnormality (27%), and QTc prolongation (10%; >480ms in men, and >460ms in women). Investigations included borderline/abnormal signal-averaged ECG (45%), exercise treadmill test (42%), procainamide challenge (13%), and cardiac MRI (28%). Genetic testing identified pathogenic variants associated with channelopathy (3%) and cardiomyopathy genes (3%).
137 patients (23%) had recurrent events, with a mean time-to-event of 2.60±2.24 years. 30 machine learning models were tested using the aggregate data, with the bootstrap aggregating model yielding the best predictive value (99.8% accuracy, kappa 0.995). Cross-validation through re-sampling demonstrated high accuracy (87.1%). Other machine learning models were less accurate, including decision tree (97.2%), random forest (98.2%), and logistic regression (84.3%).
Conclusion: A machine learning model accurately predicted recurrent events across a large cohort of UCA survivors, despite variable investigations and a heterogenous cohort.