Introduction and aim. The differential diagnosis of wide QRS complex arrhythmias, including ventricular and supraventricular arrhythmias, relies on the analysis of 12-lead ECG. However, the complexity and subjectivity of manual ECG analysis has prompted the exploration of machine learning methods for automated analysis. In this study, we aimed to develop a machine learning-based methodology for differential diagnosis of wide QRS complex arrhythmias.
Materials and methods. We analyzed 14,306 individual wide QRS complexes in 100 patients (61 males, median age – 44.5 years) with QRS duration between 120 ms to 230 ms. In total, 77% ventricular arrhythmias and 23% supraventricular arrhythmias with wide QRS aberrant conduction were verified using electroanatomical mapping. All analyzed wide QRS complex were represented by 145 different QRS forms (mean - 98 complex for each form). The differential diagnosis algorithms for wide QRS arrhythmias were modeled using various neural networks that were developed in an environment for automated generation and analysis of neural network models called STATISTICA Automated Neural Networks (SANN). After manual selection and cross-validation, the models were transferred to the original proprietary computer program with a graphical user interface. To assess the diagnostic significance qualitatively and quantitatively, ROC analysis was used to determine the informative value of a diagnostic test based on sensitivity (Sn), specificity (Sp), and diagnostic accuracy (Acc).
Results. Cross-validation of the developed neural network for differential diagnosis of wide QRS complex arrhythmias on bootstrapping-generated samples of virtual patients demonstrated the high efficiency of the neural network with a Sn of 97.1%, Sp of 99.4%, and Acc of 97.6%.
Conclusion. Automatic analysis of wide QRS complex arrhythmias can be based on a combination of automated generation and analysis of neural networks with their subsequent manual selection, cross-validation and integration into the graphical user interface.