The project intends to develop a machine learning-based algorithm for continuous unobtrusive monitoring of life-threatening arrhythmias and integrate it into the wrist-worn device. About 38% of all end-stage kidney disease patients die from life-threatening arrhythmias, such bradycardia and tachycardia. These arrhythmias are usually detected using electrocardiogram (ECG) signals, however, continuous ECG monitoring is uncomfortable since it requires at least several electrodes attached to the body. The ability to detect bradycardia and tachycardia episodes based on photoplethysmogram (PPG) signals would make continuous tracking more convenient, since only one sensor, e.g., wrist-worn sensor, is required. This would allow continuous monitoring and early detection of asymptomatic arrhythmia episodes leading to timely prescription of appropriate treatment. One of the main problems is the lack of annotated and synchronously recorded ECG and PPG signal databases, and the sensitivity of PPG to motion-induced artefacts. However, advanced biomedical signal processing techniques, such as the continuous assessment of signal quality and deep learning algorithms, would partially overcome these problems. These methods would detect PPG segments corrupted with the artefacts thus reducing the number of false alarms and would extract and classify PPG features relevant to the bradycardic and tachycardic arrhythmias, respectively. This project aims to develop an algorithm for the detection of bradycardia and tachycardia episodes based on FPG signals by implementing it in a wrist-worn device.
KTU Research and Innovation Fund
A PPG-based bradycardia and tachycardia detector based on a dual-branch CNN was developed. The detector was trained and validated on simulated PPG signals while tested on a dataset of real annotated PPG signals (Sološenko et al., 2017). The results suggest that the detector could be used for a continuous, long-term bradycardia and tachycardia monitoring, especially in situations where sensitivity is favored over specificity. This research also demonstrated that the use of simulated PPG signals is applicable for training and validation of a CNN. This research resulted in a scientific publication: “Sološenko et al., 2022. Training Convolutional Neural Networks on Simulated Photoplethysmography Data: Application to Bradycardia and Tachycardia Detection”
Period of project implementation: 2020-04-14 - 2020-12-31