Background: Wearable devices play an important role in the early diagnosis of heart diseases. However, effective management of long-term Holter measurements (1-3 weeks) by a telemedicine center (TMC) requires specifically designed software. It is required to process ECG data from multiple devices instantly; data are usually noisy since patients are recorded during usual daily activities, and they attach ECG electrodes by themselves. Still, TMC operators must be informed about life-threatening events, but they should not be distracted by false-positive cases. We aimed to develop a scalable, multiplatform application to process ECG data in a TMC.
Method: We used the multiplatform framework .NET to build the application. Deep-learning models for QRS detection, classification, and rhythm analysis were trained in the PyTorch framework; models were prepared using ECG data (N=73,450) from Medical Data Transfer, s. r. o., Czechia. The ONNX runtime libraries were used for model inference, including acceleration by graphic cards when available.
Results: We developed the "JOSEPH solver,” a multiplatform application to classify one-lead ECG signals from Holter devices. The last stable version, 0.3.110, was deployed on five virtual and one physical computer. These instances have been simultaneously analyzing ECG signals from a shared network location since Nov. 2021 without a crash. The current load is approximately 2,500x 1-hour ECG recordings per day. The pre-production bench-mark (82 patients) showed a weighted mean F1 performance of 0.97 ± 0.10 for QRS detection and classification into three classes (normal beats, premature ventricular contractions, and premature atrial contractions). It also showed a mean F1 performance of 0.96 ± 0.02 for rhythm classification into seven classes.
Conclusion: The "JOSEPH solver” is a fully automated, multiplatform, and scalable application to process incoming ECG data in the TMC. Although it is not freely accessible, we are open to processing ECG data for research and non-commercial purposes.