Cloud-free intelligent ECG patch for the real-time classification of congestive heart failure and normal sinus rhythm

Abstract

Visual interpretation of ECG signals by experienced cardiologist, has played a central role in the assessment of cardiovascular diseases (CVDs).1, 2 ECG signal’s Morphological contain rich temporal information that can support early pathological changes of heart.3 Congestive heart failure (CHF) is associated with alterations in T-wave and reduced QRS amplitude.3,4 On the other hand, Remote monitoring (RM) is recognised as frontier in healthcare management. However, RM systems rely on cloud-based data analytics, requiring continuous transmission of ECG data5, which increases latency and wireless energy consumption and require stable internet. These may delay clinical decision-making and timely intervention, particularly in time-sensitive CVDs scenairo. This study aim to develop an intelligent ECG patch with an on-chip signal processing and classification pipeline for real-time classification of CHF and normal sinus rhythm (NSR). Wearable ECG Patch: We designed a fully customisable single-lead wearable ECG patch incorporating a clinical-grade ECG analog front end, and an ESP32 microcontroller. The firmware is developed to support continuous ECG acquisition and analysis. Machine learning (ML) model: We used the publicly available CHF database from PhysioNet, 15 severe CHF patients and a NSR (had no significant arrhythmias) database with 18 subjects to develop light weight AI model. We extracted 3,300 ECG segments of 5 s (1,500 CHF NYHA Class 3–4; 1,800 NSR) and extracted six features. The identified features were split into 80% and 20% for training and testing. Finally, ML model were deployed onto EPS32 controller for classification purpose. AI Enable Wearable ECG Patch The developed device can capture ECG signal (P, QRS, and T wave) at 256Hz samples per seconds. The results showed that, out of the six extracted features, Peak-to-Average Power Ratio (PAPR) and Hjorth mobility were the most important features. Further, we found logistic regression as a lightweight model that classifies CHF and NSR with 99.26% accuracy, 99.18% sensitivity, and 99.36% specificity on unseen datasets, achieving a total latency of ~1.11 ms per 5 s segment, with model inference time below 0.6 ms. The trained model was deployed onto the ESP32 microcontroller to enable real-time classification of CHF and NSR. Figure 1 depicts the signal processing and classification pipeline used to determine the outcome and Figure 2 illustrates the ECG deviceholded by a participants and mobile App. We have devleoped an intelligent single-lead wearable ECG patch and deployed a trained model on the ESP32 to allow on-device processing and classification of CHF and NSR in real time. The preliminarily finding illustrate the feasibility of low-power, on-device cardiac disease classification and lay the groundwork for the development of continuous, home-based HF management solutions in the future.Signal processing and classification pipelineECG device and smartphone-based app.