BioPulseNet: A Real-Time Convolutional Framework for Detecting Cardiac Rhythm Deviations in Mobile ECG Sensors

BioPulseNet presents a novel, lightweight convolutional framework aimed at achieving real time arrhythmia detection from wearable ECG sensors. The framework leverages highly efficient 1D convolutional layers to uncover temporal and morphological features of ECG signals, all while maintaining low computational complexity required to deploy such a system on mobile and edge devices. The framework uses adaptive feature normalization and dynamic thresholding strategies to continuously and accurately detect abnormalities in a patient’s heart rhythm, including abnormal rhythm events associated with premature beats, tachycardia and bradycardia, even under non-ideal signal quality. The framework was evaluated on published benchmark ECG datasets as well as in live settings for ECG data collection from wearable devices, demonstrating an ability to rapidly and accurately assess abnormal cardiac events in real time with high sensitivity and specificity. When compared with conventional machine learning and standard CNN architectures, BioPulseNet was able to provide fast inference and memory utilization while maintaining high detection performance. The lightweight nature of the framework will allow for continuous monitoring of a patient with low energy use, enabling for longer operation of the wearable for at-risk patients. Thus, BioPulseNet provides an efficient and scalable solution for real-time measurement of arrhythmias in a wearable ECG context, filling the need for reliable clinical grade measurement of heart rhythm in consumers using rapidly developing consumer robotics.