Agreement between manual and AI-enhanced echocardiographic assessment of cancer therapy-related cardiac dysfunction in patients with early breast cancer

Background Cardio-oncology is a resource-intensive discipline that requires accuracy and reproducibility in serial echocardiographic examinations. Artificial intelligence (AI) enhanced echocardiographic analyses have shown reliable and accurate results in various populations and have the potential to improve precision and workflow efficiency. However, the value of AI-echo in detecting cancer therapy-related cardiac dysfunction (CTRCD) remains unknown. Aim To evaluate the feasibility and performance of AI-enhanced echocardiography by comparing agreement with manual classifications of CTRCD and assessing longitudinal trajectories of left ventricular ejection fraction (LVEF) and global longitudinal strain (GLS). Methods CARE is a prospective observational study of women with breast cancer scheduled for (neo)adjuvant anthracycline therapy and/or HER2-targeted therapy. Echocardiographic examinations were performed before, during and after completion of scheduled treatment in unselected patients by experienced cardiologists and sonographers according to standard clinical protocols. All examinations were analyzed using a commercially available AI-software platform. CTRCD was defined as mild, moderate and severe according to the echocardiographic criteria in the 2022 ESC Cardio-oncology guidelines. Agreement between manual and AI-derived classifications was assessed, and longitudinal trajectories in LVEF and GLS were analysed using linear mixed-effect models. Results In total, 2225 paired echocardiographic examinations from 515 unselected patients were included. In examinations in which manual measurements were available, AI-based measurements were feasible in 91% of examinations for LVEF and 93% for GLS. The mean±SD bias and Pearson correlation coefficient for manual vs. AI measurements were -1.2±2.2%; r=0.53 for GLS and -2.2±4.5%; r=0.57 for LVEF. The prevalence of any CTRCD assessed manually was 24%, Visit level agreement for CTRCD during follow-up was 78% for specific CTRCD class and 85% for ruling out any CTRCD (Table). The sensitivity and specificity for AI detection of any CTRCD assessed manually were 40% and 89% and positive and negative predictive values were 28% and 91%. Most discordant classifications occurred between adjacent severity categories. Longitudinal changes in LVEF and GLS measured both manually and by AI showed parallel patterns with a slight downward trend and no significant time x method interactions (Figure). Conclusion In a large real-world breast cancer cohort, AI-enhanced echocardiographic analyses demonstrated good agreement with manual classification of CTRCD. These findings support the potential role of AI-enhanced echocardiography in cardio-oncology.TableFigure