Aims: The aim of this study is to investigate the feasibility of the features from QRS complex shape on early prediction of VF as compared to traditional HRV features. Methods: The architecture of our ANN is fully connected net-work structure consisting of three layers: an input layer with nodes representing input variables to the problem, a hidden layer con-taining nodes to help capture the nonlinearity in the data, and an output layer with a node representing the dependent variable. The hidden layer consists of six neurons which were selected by trial and error experimentation. We implemented three ANN models with three different input parameters (i) 11 HRV features, (ii) only 4 QRS shape features, and (iii) 11 HRV and 4 QRS complex shape features. The input parameters were standardized and shuffled be-fore they were used in the network. To avoid overfitting, the mod-els were evaluated 10 times with 10-fold cross validation.
Results: 11 HRV features obtained 72 % prediction accuracy. The sensitivity and specificity were 65.68 % and 98.44 %, respec-tively. When using 4 features extracted from QRS complex singed area and R-peak amplitude, the prediction performance improved dramatically. The accuracy, sensitivity, and specificity were 98.6 %, 98.4 %, and 99.04 %, respectively. By adding 4 features extracted from QRS complex singed area and R-peak amplitude to the 11 HRV features, the prediction accuracy was slightly im-proved to 99.83%, but which was not significant. Conclusion: we investigated that only QRS complex shape or combined with HRV can improve the performance of predicting VF, and our study shows that the features extracted from QRS complex shape could have significant effect on predicting VF.