Machine learning based electrocardiogram peaks analyzer for Wolff-Parkinson-White syndrome

Abstract

The slurred upstroke named Delta in an Electrocardiogram (ECG) signal is known as the main indicator of the existence of the Wolff-Parkinson-White (WPW) pattern. Previous studies propose methods of detection based only on the position of the well-known peaks (P, Q, R, S, and T) where it might suffer from confusion with other pathologies patterns such as complete bundle branch block. This work introduces a new robust method for precisely detecting the actual location of Delta wave using signal peaks scan cooperating with machine learning algorithms that could be trained for any ECG lead. The method is sequenced in seven steps which include the detection of P, QRS, and T waves. Each step represents a machine learning model which considers the statistical parameters of the signal for either selecting an appropriate feature of the signal or manipulating a function for peaks scan or tangent deviation scan. The method is performed using Neural Networks, Naïve Bayes, and K-Nearest Neighbors (KNN) algorithms on 122 signals from the MIT-BIH Arrhythmia database. The performance is evaluated in different cross-validation configurations, where Neural Networks and KNN recorded the highest precision in holdout cross-validation with mean accuracy, sensitivity, specificity, and MASE of the seven steps of 99.25%, 97.16%, 74.37%, 2.33 respectively from Neural Networks performance, and of 99.11%, 94.64%, 74.37%, 1.26 respectively from KNN performance. Naïve Bayes obtained the same evaluation metrics results in 3-fold cross-validation with 81.23%, 46.91%, 57.84%, and 5.63 respectively.