Non-invasive risk assessment and prediction of cardiac outcomes in patients with heart failure or myocardial infarction using deep learning method

Heart failure (HF) and myocardial infarction (MI) has been considered the main cause of major adverse cardiac events including ventricular arrhythmia or cardiac death.

We investigated several non-invasive values after HF or MI to predict major adverse cardiac events in a large cohort including deep learning (DL) method.

The K-REDEFINE study, a prospective, observational, multicenter analysis, investigated the prognostic implications of Holter-based variables including heart rate turbulence (HRT) and T-wave alternans (TWA) in 1,116 patients with acute MI or HF (60.812.9 years, 76.3% males). The primary composite outcome included cardiac death, and ventricular tachyarrhythmias. We trained the DL using a Holter electrocardiogram. We compared the predictive values of each variable.

Fifty-six adjudicated cardiac deaths (event rate 1.17%/yr), and 23 ventricular arrhythmias (event rate 0.49%/yr) occurred during follow-up. The area under the receiver operating characteristics curve (AUROC) of DL was 0.74 (0.70–0.77) and that of DL combined with ejection fraction (EF) was 0.77 (0.74–0.81). The AUROC of the HRT was 0.62, and TWA was 0.55. The AUROC of DL in MI group was 0.65 (0.57–0.73) and that of DL combined with EF was 0.75 (0.67–0.84). The AUROC of DL in HF group was 0.60 (0.53–0.68) and that of DL combined with EF was 0.58 (0.56–0.60). (figure)

We have demonstrated that a DL can predict cardiac death or ventricular arrhythmia events after HF or MI using 24-hour continuous ambulatory Holter monitoring. The performance of DL model outperformed when combined with EF value. The performance showed better in MI group than HF group.