Non-Invasive Monitoring of Transcatheter Heart Valve Using Photoplethysmography and Machine Learning

Abstract

Transcatheter aortic valve implantation (TAVI) has become the preferred treatment for aortic stenosis in the elderly. However, the durability of transcatheter heart valves (THV) and the risk of leaflet thrombosis preclude the extension of TAVI in young people. This study sought to formulate a proof-of-concept solution for non-invasive, continuous monitoring of THV function using photoplethysmography (PPG) sensors and machine learning models. An in vitro mock circulatory loop was developed using a compliant aortic phantom and an implanted self-expanding Evolut FX device. Two PPG sensors were attached to the phantom surface to record flow signals during various hemodynamic scenarios generated by Latin Hypercube Sampling. Endoscopic imaging measurements of the geometric orifice area (GOA) were used to assess valve performance. PPG-derived metrics and flow variables were used to train regression and classification models that predicted healthy versus decreased leaflet motion. The regression model showed an R2 of 0.83, RMSE of 7.18 mm2, and MAE of 5.58 mm2. The classifier correctly identified reduced leaflet motion (95% accuracy, 0.89 precision, and 0.91 recall). This study demonstrates the efficacy of PPG sensors and machine learning for non-invasive THV monitoring. The approach offers a promising tool for early detection of leaflet dysfunction, thereby improving the management of TAVI patients.