Detecting irregular heartbeat using deep forest with multilevel discrete wavelet transforms

One of the leading causes of death is cardiovascular disease (CVD). This disease is the cause of 31% of deaths worldwide in 2016, and 85% of them are heart attacks. The traditional way to detect CVD is based on medical records and clinical analysis of the patient. Electrocardiogram (ECG) analysis is one way to determine irregular heartbeat or arrhythmia. Computer assistance with implementing specific machine learning algorithms can help recognize irregular heartbeats automatically. However, raw ECG data may contain noise that affects the accuracy of irregular heartbeat detection. In this study, the ECG data used was from the Massachusetts Institute of Technology–Beth Israel Hospital (MIT-BIH) database. The data has four categories: Normal, Atrial Fibrillation, PVC Bigeminy, and Ventricular Tachycardia. ECG raw data processing using multilevel discrete wavelet transforms (DWT) based on Haar and Daubechies wavelet. The process uses various values of mode (i.e. db1 until db10), level (i.e. level 1 and level 2), and filter (low and high pass filter), and the result is 20 data processed. Each data is used to create a model with several classification algorithms, i.e. K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Naïve Bayes, Decision Tree, Random Forest, and Deep Forest. The validation process uses 10-fold cross-validation. The results of this study indicate that Multilevel Discrete Wavelet Transforms improve irregular heartbeat detection accuracy when compared to raw ECG and processed data using a single DWT. While the best detection accuracy is the Deep Forest model, with an accuracy value of 63.57% using processed data with db1 mode values, level 2 and combining high and low pass filters.