Project for the course in APPLIED AI IN BIOMEDICINE at Politecnico di Milano
Arrhythmias are one of the most common types of Cardiovascular disease and one of the top causes of mortality in the world. The diagnosis of arrhythmias is a time-consuming task for specialists since it usually involves the manual inspection of electrocardiogram (ECG) recordings.
For this reason, computer-aided diagnosis systems are an efficient tool to provide accurate automatic diagnosis of arrhythmias from the analysis of ECG signals.
In this project, I train a Wide-ResNet model to classify ECG signals relative to single heartbeats. The model is the combination of a 1D ResNet that takes as input the ECG signal, and a FFNN that takes as input some additional features imputed from the ECG recording (e.g., distance from the peaks).
Data Augmentation and Weighting are used to account for the class imbalance in the dataset (anomalous heartbeats are rare overall).
The trained model achieved Accuracy of ~98%, Precision of ~87%, and Recall of ~95%.