(A-30) An Optimized Mobile AI Platform for Remote Cardiovascular Disease Detection and Risk Prediction: ECG-Air V3

Remote Artificial Intelligence (AI) platforms increase the accessibility of cardiovascular diagnostics and promote early therapeutic intervention. ECG-Air is an AI platform that can host AI models for cardiovascular disease (CVD) and non-CVD risk predictions. Yet, its current version has limitations in input type and deployable mobile platforms. TensorFlow Lite is a proven library for deploying AI models on edge devices. Deploying a TensorFlow Lite model can improve accuracy and enable interoperability with multiple AI models and mobile operating systems. We aim to create an ECG-based risk prediction pipeline using the TensorFlow-Lite library to improve the accessibility and interoperability of ECG-Air to power a suite of AI-based prediction models.

We reconstructed the ECG-Air application to support the execution of the TensorFlow-Lite library locally on the iOS platform. The application executes embedded AI models on ECGs received from HealthKit and transmits the results to a secure database for execution (Figure 1). The redesigned pipeline now supports analysis of multiple suits of inference algorithms: 1) ECG-AI: an already developed deep learning model that infers the likelihood of heart failure within ten years, and 2) Cox Regression Analysis: to incorporate meta-factors such as BMI, race, age, smoking status while accounting for time-to-event data to create a more comprehensive risk score. The accuracy of ECG-Air was tested against the un-optimized ECG-AI that runs TensorFlow on a laptop platform.

We installed ECG-Air on four modern iPhones and tested the TensorFlow-Lite ECG-AI and Cox Regression models against original models ran on laptop devices. We gathered 100 ECGs from 3 test users via Apple Watch before executing the embedded algorithms. The inference time of the mobile-optimized models was < 25ms, around 93% faster than previous models, which are CoreML compiled versions of TensorFlow Lite. After inference, the raw ECGs were uploaded to Google Cloud for secure storage and cross-analysis against original implementations. We validated that the current embedded models yielded identical risk scores as the originals. The modularity of the new framework allows future models, such as cascading model and LightGBM, to be easily embedded. More testing is needed to validate the accuracy of future embedded models. TensorFlow Lite can improve the performance of mobile AI applications by optimizing heavyweight AI models for edge devices. Incorporating TensorFlow Lite in ECG-Air demonstrates the feasibility of embedding deep-learning models on mobile devices that can assist telemedicine.  To our best knowledge, ECG-Air is the first mobile-optimized non-vendor third-party platform that enables remote analysis of ECG for CVD detection and prediction.

This project was supported in part by the NIH/NLM (R25 LM014214) in the Department of Biomedical Engineering and Center for Biomedical Informatics at Wake Forest University School of Medicine.