MODELING AND EVALUATING AN INTELLIGENT HEALTH MONITORING SYSTEM FOR ATRIAL FIBRILLATION DETECTION

Sammanfattning: The heart disease Atrial Fibrillation (AFib) has increased worldwide in recent years. Untreated AFib can lead to cardiovascular complications such as stroke and heart failure. AFib is detected by physicians using Electrocardiogram (ECG). Since this disease can occur without symptoms for some patients, it can lead to late detection. Therefore, smart solutions for continuous monitoring of ECG to detect AFib is needed. This paper presents an approach to model an low-cost intelligent health monitoring system (IHMS) to classify and detect AFib in ECG using 1D Convolutional Neural Network (CNN). The core objective of this paper were to investigate the suitability of the computing architecture, edge and cloud, for an IHMS, and how complex 1D-CNN could be deployed to an edge device. Three 1D-CNN models with increased complexity was designed, trained and tested on AFib and NSR episodes collected from 25 records of the LTAF database. Each record were noise filtered and segmented into 10 sec. The best 1D-CNN model presented an accuracy of 83.93 %, 89.83 % in AUC, 84.32 % in sensitivity (AFIb), 83.46 % in specificity (NSR), 84.81 % in F1-score and 68.23 % in MCC. Two experiments into end-to-end delay and prediction time were performed todetermine the computing architectures suitability. The end-to-end delay were measured by sending ECG segments of different sizes to both computing architectures, while the prediction time were measured by deploying the designed 1D-CNN models on both computing architectures. Both measurements were added together to form the response time of the computing architectures. The edge computing architecture produced a delay around 0.019-0.377 sec and prediction time around 0.00X sec compared to cloud’s delay around 1.32-4.43 sec and 0.000X in prediction time. Resulting, that the edge computing architecture produced a lower response time and therefore considered the more suitable architecture for an IHMS. The designed 1D-CNN models had no issues in executing on the edge device, resulting in the conclusion that the most complex model to execute had 6 convolutional layers. The presented result in this paper contributes to the development of a health monitoring system in terms of choosing computing architecture platform and model complexity for a resource constraint device.