Automated identification of heart arrhythmias through HRV analysis and machine learning

Sudden cardiac death and arrhythmia are responsible for about 15-20% of cardiovascular disease incidences. Conventionally, the prediction and diagnosis of cardiovascular disorders (CVDs) have been mainly through the evaluation of ECG patterns by cardiologists. To improve the accuracy of and automate this process, and facilitate early detection, Heart Rate Variability (HRV) analysis has been promoted as a diagnostic and predictive tool for CVDs. In the present study, a machine learning model capable of detecting the presence of arrhythmia, using HRV indices obtained from ECG signals was built. Unlike similar works in the literature, this study deployed the developed model on Raspberry Pi with Streamlit software. Two ECG datasets from the Physionet database, one with arrhythmia patients (48 half-hour recordings) and another with healthy individuals (18 24-hour recordings), were employed. An ensemble of seven different machine learning models was used on the two sets of datasets to classify ECG recordings into Arrhythmia and Normal Sinus Rhythm (NSR). The best models were able to predict the presence of Arrhythmia in a 3-minute recording with an accuracy of 95.96%, and in a 10-minute recording with an accuracy of 96.20%. These performance measures were calculated using test dataset. The Random Forest models also had the highest precision, AUC, (Area under the Curve) recall, and F1 scores compared to the other models tested. The highest performing model (i.e., Random Forest Model) was then deployed onto a Raspberry Pi with Streamlit as the software interface for usability. This was done to facilitate a smooth user experience for faster and seamless diagnoses for cardiologists.