Electrocardiography-Based Prediction of Sudden Cardiac Death in Heart Failure Patients: Application of Artificial Intelligence

ABSTRACT Background Although predicting sudden cardiac death (SCD) in patients with heart failure (HF) is critical, the current predictive model is suboptimal. Electrocardiography-based artificial intelligence (ECG-AI) algorithms may better stratify risk. We assessed whether the ECG-AI index established here could better predict SCD in HF and whether the ECG-AI index and conventional predictors of SCD can improve SCD stratification. Methods In a prospective observational study, four tertiary care hospitals in metropolitan Tokyo that enrolled 2,559 patients hospitalized with HF who were successfully discharged after acute decompensation. ECG data collected during the index hospitalization were extracted from the hospitals’ electronic medical record systems. The ECG-AI index is the output from an AI model that was trained to predict the risk of SCD based on ECG input. The association between ECG-AI index and SCD was evaluated with adjustment for left ventricular ejection fraction (LVEF), New York Heart Association (NYHA) class, and competing risk of non-SCD. The outcome measure was a composite of SCD and implantable cardioverter-defibrillator activation. The ECG-AI index was established using a derivation (hospital A) and validation cohort (hospital B), and its ability was evaluated in a test cohort (hospitals C and D). Results The ECG-AI index plus classical predictive guidelines (i.e., LVEF ≤ 35%, NYHA class II–III) significantly improved the discriminative value of SCD (area under the receiver operating characteristic curve, 0.66 vs. 0.59; p=0.017; Delong’s test) with good calibration (p=0.11; Hosmer–Lemeshow test) and improved net reclassification (36%; 95% confidence interval, 9%–64%; p=0.009). The Fine-Gray model considering the competing risk of non-SCD demonstrated that the ECG-AI index was independently associated with SCD (adjusted sub-distributional hazard ratio, 1.25; 95% confidence interval, 1.04–1.49; p=0.015). An increased proportional risk of SCD vs. non-SCD with increasing ECG-AI index was also observed (low, 16.7%; intermediate, 18.5%; high, 28.7% risk; p for trend = 0.023). Similar findings were observed in patients aged ≤75 years with a non-ischemic etiology and an LVEF >35%. Conclusions Among patients with HF, ECG-based AI significantly improved the SCD risk stratification compared to the conventional indication for implantable cardioverter-defibrillators inclusive of LVEF and NYHA class.