Improved prediction of sudden cardiac death in patients with heart failure through digital processing of electrocardiography

Figure 1. Discrimination and calibration of risk models for predicting SCD events. Comparison of ROC-AUC (A) shows the conventional guideline-directed indication for ICD (LVEF ≤ 35% and NYHA Class II and III [as a binary variable]; AUC = 0.59 [95% CI, 0.52–0.66]) vs. the ECG-AI combined model (ECG-AI index [as a continuous variable] + ICD indication; AUC = 0.66 [95% CI, 0.58–0.73]) for predicting 3-year composite SCD events (P = 0.017 for Delong’s test). Calibration of the ECG-AI combined model is shown (B) by dividing four bins based on the quartiles of the model-predicted risk (P = 0.11 for the Hosmer–Lemeshow test). CI, confidence interval; ECG-AI, electrocardiography-based artificial intelligence; ICD, implantable cardioverter-defibrillator; LVEF, left ventricular ejection fraction; NYHA, New York Heart Association; ROC-AUC, receiver operating characteristic area under the curve; SCD, sudden cardiac death.

Figure 2. Subgroup analyses of the discriminative abilities of the ECG-AI combined model and conventional indication for ICD. The model’s discrimination was compared to that of the Delong test. Upper bars, the ECG-AI combined model; Lower bars, the conventional indication for ICD. AF, atrial fibrillation; CI, confidence interval; ECG-AI, electrocardiogram-based artificial intelligence; ICD, implantable cardioverter-defibrillator; ROC-AUC, receiver operating characteristic area under the curve.

Figure 3. Discriminative ability of the ECG-AI index and frequency of SCD vs. non-SCD events by the LVEF category. The above forest plots show ROC-AUC with 95% confidence intervals of the ECG-AI index for predicting SCD events by LVEF. In patients with an LVEF of 35–50%, the ECG-AI index showed the best discriminative ability. The bar graphs represent the number of SCD and non-SCD patients according to LVEF. The proportion of patients with SCD vs. non-SCD steadily decreased as LVEF increased. ECG-AI, electrocardiography-based artificial intelligence; LVEF, left ventricular ejection fraction; ROC-AUC, receiver operating characteristic area under the curve; SCD, sudden cardiac death.

Figure 4. Proportional risk of SCD vs. non-SCD events based on the model-based risk. ECG-AI index (A) and ECG-AI combined model (with LVEF and NYHA class, B). The patients were divided into three groups by tertile of risk score (each P for trend < 0.05). ECG-AI, electrocardiography-based artificial intelligence; LVEF, left ventricular ejection fraction; NYHA, New York Heart Association; SCD, sudden cardiac death.