Development of a Novel Deep Learning Algorithm to Detect Cardiac Arrhythmias from Electrocardiogram Signals
Principal Investigator:
Mr Hiral Radia
Approved Research ID:
47100
Approval date:
October 2nd 2019
Lay summary
We aim to develop a deep learning-based algorithm to automatically detect abnormal heart rhythms (cardiac arrhythmias) using signals from an electrocardiogram, a simple test that measures the rhythm and electrical activity of the heart. Cardiac arrhythmias occur when abnormal electrical activity in the heart causes it to beat too quickly, too slowly, or with an irregular rhythm. Such conditions are experienced by over 2 million people a year in the UK, and are significant risk factors for many potentially life-threatening conditions. For example, the most common heart rhythm disturbance in the UK, Atrial Fibrillation, is the cause of one third of strokes. Fortunately these arrhythmias are often treatable, and many patients are able to lead normal lives following their diagnosis. Timely diagnosis is therefore critical to the ongoing health and wellbeing of sufferers. Cardiac arrhythmias can be identified using electrocardiogram (ECG) recordings. Different arrhythmias produce characteristic changes to the ECG waveform, which may be detected using automated methods. This has the potential to increase the efficiency of clinical diagnosis, and also shows promise for ambulatory detection using mobile devices such as smart watches. Traditional arrhythmia detection involves extracting predetermined features from the ECG waveform, for example the intervals between the tallest peaks, and using these to evaluate the most likely rhythm classification for each recording. More recently, deep learning techniques have been implemented to both identify the most important features and deduce the appropriate classification. We are developing a novel deep-learning based classification algorithm that aims to exceed the accuracy of current models, with the aid of the UK Biobank's ECG databases. In order for deep learning models to produce reliable results, they must be trained and tested using as much data as possible. We will therefore test our algorithm using the Biobank's database of resting ECGs, which is significantly larger than most currently available to researchers. Testing will be carried out using samples from both the Imaging Study and assessment centre Exercise Tests, with the goal of improving the classifier's ability to distinguish between normal and abnormal ECG samples.