Electrocardiogram Analysis Using Machine Learning and Digitization Techniques for Enhanced Clinical Diagnosis

The leading cause of death worldwide is Cardiovascular Disease (CVDs), accounting for 17 million deaths worldwide in 2019. CVDs are vital to diagnose early, as they can lead to major heart problems. The predominant method to diagnose CVDs is an Electrocardiogram (ECG). Currently, CVDs are primarily diagnosed by human interpretation of an ECG. While this is the most effective method, it has shown insufficient accuracy of diagnosis due to human factors. The emergence of machine interpretation of an ECG has not shown any better accuracy than human diagnoses. ECGs are commonly represented as a printed paper strip with 12 leads. This makes it very difficult to access machine interpretation ECG methods, as they would require a digital ECG. To address this gap, this project presents an approach that digitizes paper ECGs and applies artificial intelligence techniques to diagnose CVDs from scanned paper ECG strips. The CVDs being detected are Atrial Fibrillation, Tachycardia, Bradycardia, and Sinus Rhythm, as well as STEMI (heart attack). Using machine-learning-based image processing, a 12-lead paper ECG is split into digitized individual leads. QRS complexes in an ECG are detected by the Pan-Tompkins algorithm. Labeled QRS complex statistics were used to train various machine learning models. Results show the highest accuracy from a Random Forest Classifier model at 91% to detect Atrial Fibrillation, Tachycardia, and Bradycardia. A CNN-LSTM was used to detect STEMI, which had a 90.80% accuracy. In the final application, users can upload a scanned image of an ECG strip, which the system processes using machine learning algorithms to classify the ECG into various CVDs.