Recognition of Brugada type 1 pattern on ECG by Echo State Networks

Abstract Background/Introduction The development of artificial intelligence (AI) and machine learning (ML) techniques, which can identify data patterns in electrocardiography, has led to the development of automated analysis tools for standard 12-lead ECGs. These tools are able to identify specific features associated with different cardiac diseases, which may be helpful in clinical practice [1]. One field of application is represented by Brugada Syndrome (BrS) diagnosis, which, despite characteristic ECG features, can be challenging to diagnose and correctly distinguish from other diseases. Purpose Within the Brugada Syndrome and Artificial Intelligence Applications to Diagnosis system study, we are developing an innovative model for diagnosing Type 1 BrS based on electrocardiographic pattern recognition through the Echo State Networks (ESN), a Recurrent Neural Network (RNN) type of architecture. Methods From patients enrolled in 5 Centers, 306 12-lead ECGs were obtained. The datasets were composed of spontaneous Type 1 pattern (coved) (group A, 98 patients); patients undergoing ajmaline test, classified as positive (group B, 47 patients) or negative (group C, 35 patients) according to test results; 220 controls with no clinical and familial history of arrhythmias (group D). ESN analyzed, as input, data collected from 1 beat of the ECG V2 lead. The datasets were used for the ESN model training and were assessed through a double cross-validation approach. The ESN performance was compared with the evaluation of 4 independent cardiologists, and sensitivity, specificity and accuracy were determined. Results The ESN model correctly classified 280/306 ECGs (91.5%), with a performance of 3.6 points superior to clinicians (269/306, 88.0%). ESN sensitivity (87.0% vs. 86.9%) and specificity (94.5% vs. 89.1%) are described in Table 1; accuracy in each group of the dataset (groups A, B, C, and D), represented in Table 2, were comparable to or greater than those of clinicians. Conclusion(s) In this work, an ESN-based system for diagnosing Type 1 BrS pattern on ECG was developed. Our results show that this ML model can detect ECG patterns associated with Type 1 BrS with accuracy, sensitivity and specificity comparable or superior to clinicians. In the future, we will use a larger dataset to increase the model performance further, with the real possibility of becoming a tool for everyday clinical practice.ESN and clinicians' diagnostic accuracyECG accuracy in each dataset group