Heartisan: An Incremental Learning Based Arrhythmia Detection, Data Collection, and Monitoring System

Cardiac disease is the leading cause of death worldwide, which is why the importance of early heart disease prediction is rising daily. Patient data from modern ECG systems can be utilized to improve such machine-learning models. Here, a system has been proposed that aids in early arrhythmia prediction using a convolutional neural network and continuously improves the model using incremental learning utilizing patient data from a web application. The web app comes with a patient and a doctor's portal. Patients can view heart conditions and send ECG beats and predictions for verification. Whereas the doctor's portal is used to annotate the model's falsely predicted heartbeats. The system continuously updates the model using newly annotated data following an incremental learning approach. The proposed incremental learning strategy was simulated using the MIT-BIH dataset, and the approach demonstrated a promising result as the overall accuracy, and AUC improved as well as the F1 score of individual classes showed a notable shift. The system is expected to contribute to building a novel large arrhythmia dataset in an efficient strategy, as well as provide patients with a heart condition monitoring system employing a highly accurate arrhythmia classifier in the long run.