Detection of Life-threatening Malignant Cardiac Arrhythmias Using Machine Learning Methods

Malignant cardiac arrhythmias such as ventricular tachycardia (VT) and ventricular fibrillation (VF) cause Cardiac arrest (CA) or Sudden cardiac death (SCD); which accounts for 50% of all cardiac deaths. In this study, we proposed a patient-general system that analyzes and detects VT/VF arrhythmic episodes in ECG signals by classifying instances as either normal or arrhythmic. A great deal of experimental work was conducted to choose the best algorithms for preprocessing, the extraction of a comprehensive set of time-and frequency-domain features from a specific window length of the ECG signals to capture information about their amplitude, frequency, and morphology, and for feature selection, and classification. Three different window sizes were tested 10s, 20s, and 60s. On testing data with 10 s-window-size, the Random Forest and kNN ensembles obtained the best results with an accuracy (acc) of 99.11%, a sensitivity (se) of 98%, and a specificity (sp) of 99.6%. We also show through comparative results that the Minimum Redundancy Maximum Relevance (mRMR) feature selection algorithm outperforms the Principal Component Analysis (PCA) algorithm in choosing the optimal feature combinations. Our results demonstrate the potential of an automated, accurate, and robust detection system and how it could contribute significantly to the enhancement of the diagnostic process of arrhythmias and the prediction/prevention of cardiac arrest.