An Enhancement Performance Study to Predict Premature Ventricular Contraction Using Machine Learning Techniques

Abstract

Premature ventricular contractions (PVC) are common cardiac arrhythmias in which heartbeats are initiated from the ventricles instead of the sinoatrial (SA) node. Detecting abnormal PVC beats is essential for the early detection and prevention of life-threatening heart diseases. This paper focuses on the detection of PVC arrhythmia using machine learning (ML) with the MITBIH arrhythmia database. Since the electrocardiogram (ECG) signal is inherently non-stationary, noise reduction is necessary to denoise the raw data during the pre-processing stage. The process of PVC arrhythmia detection involves several stages, including ECG signal feature extraction and selection. Features are extracted and selected from the ECG signal in the morphological stage as input for three different classifiers and the final classification stage. The focus of this paper is on the classification stage, which employs support vector machines (SVM), random forest (RF), and quadratic discriminant analysis (QDA). After tuning different parameters using GridSearchCV, the performance metrics of various models for each classification method were compared. The performance measures considered for comparative analysis are specificity (Spe), precision (Pre), F1-score, sensitivity (Sen), and accuracy (Acc). The SVM classifier achieves Spe 90.3%, Pre 90.7%, Sen 97.3%, Acc 93.8%, and F1-score 94%. The RF classifier results are: Spe 99.1%, Pre 99.1%, Sen 100%, Acc 99.5%, and F1-score 99.5%. As for the QDA classifier results, they are as follows: Spe 89.7%, Pre 89.9%, Sen 91.3%, Acc 90.5%, and F1-score 90.6%.

It is found from the experimental results that the RF classifier achieves a higher accuracy rate.