Purpose The aim of this study is to present the state-of-the-art of the machine learning methodologies applied for the assessment of heart failure from ECG signals. Methods We performed a multicenter retrospective cohort study in 2 hospitals. The subjects were admitted adult (age≥18 years) heart failure patients who underwent echocardiography and ECG. We performed a cross-sectional study analyzing ECG and echocardiographic data from 2209 subjects with chronic HFrEF (n=1006) HFmidEF (n=1103) compared with normal EF patients ECG (n=18196). HFrEF and HFmidEF were defined as an ejection fraction (EF) ≤ 40% and 40 Results Deep learning based analysis revealed certain ECG variables that were independent predictors of HFrEF, i.e., left atrial hypertrophy (LAH), HR > 80, QRS duration > 110 ms, right bundle branch block (RBBB), ST-T segment changes and prolongation of the QT interval (>410ms). Based on receiver operating characteristic (ROC) curve analysis, we obtained a score for HFpEF of -1 to +3, while HFrEF had a score of +4 to +6 with 98% specificity, a 95% positive predictive value. Conclusion In this study, we developed and validated a deep-learning-based model for predicting HF in ECG. Through validation, this study confirmed that the accurate performance of the deep-learning-based model was excellent for predicting heart failure and showed better accuracy than random forest model. QRS duration > 110 ms, RBBB, ST-T segment changes and prolongation of the QT interval can be used to predict the type of HFrEF with satisfactory sensitivity and specificity.