ECG-AIR: AN AI PLATFORM FOR REMOTE SMARTWATCH ECG-BASED CARDIOVASCULAR DISEASE DETECTION AND PREDICTION

Abstract

BackgroundAI-based analysis of digital single lead ECG data can detect and predict risk for cardiovascular disease (CVD). Remote application of AI models via ECG enabled smartwatch may facilitate early therapeutic interventions.ObjectivesCreate a platform, ECG-AI, enabling remote execution of AI models on smartwatch ECG.MethodsWe developed an IOS application (ECG-Air) enabling retrieval of digital ECG from Apple Watch and its AI-based analysis. ECG-Air supports two sub-applications: 1) Gather: to collect smartwatch ECGs from users and 2) Run: to execute AI models. ECG-Air Run executes AI models embedded within the iPhone and/or transmits digital ECG data to cloud hosting AI models for execution (Figure). We tested ECG-Air using our previously developed deep learning model (ECG-AI) for heart failure (HF) risk prediction.ResultsWe installed ECG-Air application on four iPhones running on IOS 15+ versions. We tested Gather by successfully recording Apple Watch ECGs and securely transmitting ECGs to Firebase platform. To test Run, we embedded our ECG-AI model on ECG-Air. We then recorded 100 ECGs from 3 test users via Apple Watches. The time between completion of ECG on Apple Watch and Execution of ECG-AI model was < 500ms. We also embedded our ECG-AI application on Google Cloud so that the smartwatch ECGs could also be analyzed on cloud rather than smartphone. The time between recording of ECG recording on smartwatch and the execution of the ECG-AI on cloud was <5s. We also validated that the original TensorFlow version of ECG-AI model and its TensorFlow-Light version for smartphone provided the exact same risk prediction.Conclusion BackgroundAI-based analysis of digital single lead ECG data can detect and predict risk for cardiovascular disease (CVD). Remote application of AI models via ECG enabled smartwatch may facilitate early therapeutic interventions. AI-based analysis of digital single lead ECG data can detect and predict risk for cardiovascular disease (CVD). Remote application of AI models via ECG enabled smartwatch may facilitate early therapeutic interventions. ObjectivesCreate a platform, ECG-AI, enabling remote execution of AI models on smartwatch ECG. Create a platform, ECG-AI, enabling remote execution of AI models on smartwatch ECG. MethodsWe developed an IOS application (ECG-Air) enabling retrieval of digital ECG from Apple Watch and its AI-based analysis. ECG-Air supports two sub-applications: 1) Gather: to collect smartwatch ECGs from users and 2) Run: to execute AI models. ECG-Air Run executes AI models embedded within the iPhone and/or transmits digital ECG data to cloud hosting AI models for execution (Figure). We tested ECG-Air using our previously developed deep learning model (ECG-AI) for heart failure (HF) risk prediction. We developed an IOS application (ECG-Air) enabling retrieval of digital ECG from Apple Watch and its AI-based analysis. ECG-Air supports two sub-applications: 1) Gather: to collect smartwatch ECGs from users and 2) Run: to execute AI models. ECG-Air Run executes AI models embedded within the iPhone and/or transmits digital ECG data to cloud hosting AI models for execution (Figure). We tested ECG-Air using our previously developed deep learning model (ECG-AI) for heart failure (HF) risk prediction. ResultsWe installed ECG-Air application on four iPhones running on IOS 15+ versions. We tested Gather by successfully recording Apple Watch ECGs and securely transmitting ECGs to Firebase platform. To test Run, we embedded our ECG-AI model on ECG-Air. We then recorded 100 ECGs from 3 test users via Apple Watches. The time between completion of ECG on Apple Watch and Execution of ECG-AI model was < 500ms. We also embedded our ECG-AI application on Google Cloud so that the smartwatch ECGs could also be analyzed on cloud rather than smartphone. The time between recording of ECG recording on smartwatch and the execution of the ECG-AI on cloud was <5s. We also validated that the original TensorFlow version of ECG-AI model and its TensorFlow-Light version for smartphone provided the exact same risk prediction. We installed ECG-Air application on four iPhones running on IOS 15+ versions. We tested Gather by successfully recording Apple Watch ECGs and securely transmitting ECGs to Firebase platform. To test Run, we embedded our ECG-AI model on ECG-Air. We then recorded 100 ECGs from 3 test users via Apple Watches. The time between completion of ECG on Apple Watch and Execution of ECG-AI model was < 500ms. We also embedded our ECG-AI application on Google Cloud so that the smartwatch ECGs could also be analyzed on cloud rather than smartphone. The time between recording of ECG recording on smartwatch and the execution of the ECG-AI on cloud was <5s. We also validated that the original TensorFlow version of ECG-AI model and its TensorFlow-Light version for smartphone provided the exact same risk prediction. Conclusion