Atrial fibrillation detection service for stroke prevention

This study aims to develop a cost-effective atrial fibrillation detection service that improves outcomes for patients. The service offers continuous Atrial Fibrillation (AF) detection which might help to address the real-world problem of stroke prevention. AF is the most common sustained heart rhythm disorder in adults. AF is either intermittent (paroxysmal) or permanent. Both types increase the risk ischemic stroke around fivefold. An accurate diagnosis of AF is mandatory for treatment initiation. AF treatment reduces the stroke risk and for individual patients prevent stroke. Unfortunately, current AF detection methods often fail to detect paroxysmal AF cases, because the observation duration is too short. We propose to address this problem with real time monitoring and artificial intelligence for AF detection. We developed two distinct deep learning models to detect irregular heartbeats. For the first experiment, a Long Short-Term Memory (LSTM) classifier was used to detect and differentiate AF beats and Normal beats. The data were collected from MIT-BIH Atrial Fibrillation Database. This database incorporates 10-hour Electrocardiogram (ECG) signals from 23 participants. The second experiment was based on using a ResNet algorithm to detect common arrhythmias, namely, AF, and Atrial Flutter (AFL), as well as Normal Sinus Rhythm (NSR). The algorithm was trained with data from 4051 subject. The LSTM model achieved 98.51% accuracy with 10-fold cross-validation (20 subjects) and 99.77% with blindfold validation (3 subjects). Whilst, the ResNet model achieved, the following results: accuracy = 99.98%, sensitivity = 100.00%, and specificity = 99.94%. In addition, the LSTM model was validated with five independents benchmark databases to establish the robustness and maturity through more and more varied datasets. With the LSTM validation, we established trust which enabled us to conduct a clinical trial study with Sheffield Teaching Hospital. As part of this work, an AF detection service validation tool was built for hybrid decision support where machine learning decisions are verified by a stroke consultant. This detection method makes economic sense because Heart Rate (HR) signals are costeffective to measure, transmit, and process. Having such a cost-effective solution might lead to widespread long-term observation, which can help detecting arrhythmia earlier. Detection improves the outcomes for patients and reduces healthcare cost.