Predicting Cardiopulmonary Exercise Testing Performance in Patients Undergoing Transthoracic Echocardiography - An AI Based, Multimodal Model

Abstract

Background and Aims: Transthoracic echocardiography (TTE) is a widely available tool for diagnosing and managing heart failure but has limited predictive value for survival. Cardiopulmonary exercise test (CPET) performance strongly correlates with survival in heart failure patients but is less accessible. We sought to develop an artificial intelligence (AI) algorithm using TTE and electronic medical records to predict CPET peak oxygen consumption (peak VO2) ≤ 14 mL/kg/min. Methods: An AI model was trained to predict peak VO2 ≤ 14 mL/kg/min from TTE images, structured TTE reports, demographics, medications, labs, and vitals. The training set included patients with a TTE within 6 months of a CPET. Performance was retrospectively tested in a held-out group from the development cohort and an external validation cohort. Results: 1,127 CPET studies paired with concomitant TTE were identified. The best performance was achieved by using all components (TTE images, all structured clinical data). The model performed well at predicting a peak VO2 ≤14 mL/kg/min, with an AUROC of 0.84 (development cohort) and 0.80 (external validation cohort). It performed consistently well using higher (≤18 mL/kg/min) and lower (≤12 mL/kg/min) cut-offs. Conclusions: This multimodal AI model effectively categorized patients into low and high risk predicted peak VO2, demonstrating the potential to identify previously unrecognized patients in need of advanced heart failure therapies where CPET is not available.

Ähnliche Arbeiten