Abstract 4365496: AI image analysis to aid echocardiographic surveillance in patients with breast cancer receiving cardiotoxic chemotherapies: A pilot proof-of-concept study

Background: Echocardiography (TTE) is used for cardiac surveillance in patients with breast cancer (BC) receiving potentially cardiotoxic therapies. There is significant variability in the assessment of TTE parameters which reduces diagnostic accuracy and may adversely affect patient care. Artificial intelligence (AI) tools have shown promise in improving TTE measurement standardization and reproducibility. Research Question: We conducted a proof-of-concept study in patients with BC who underwent serial TTEs to 1) assess the agreement between AI-generated and clinical readers’ (human) measures of left ventricular ejection fraction (EF) and global longitudinal strain (GLS), and 2) analyze differences in meeting criteria for cancer therapy-related cardiac dysfunction (CTRCD) on follow-up TTE. Methods: Using the medical record, we identified patients with BC who underwent baseline and surveillance TTE (within 6 months) from 2019 to 2024. DICOM images were analyzed using Us2.ai software. We compared Us2.ai (AI) measured EF and GLS with previously reported clinical results (human). Descriptive statistics included agreement rate, unweighted Cohen’s Kappa, Bland-Altman statistics. CTRCD was determined using the published ICOS criteria based on EF and GLS change. Results: With 162 TTE studies in 81 patients (mean age 54±4), the overall rates of agreement between AI and human for EF and GLS were 83% and 66%, respectively (Bland-Altman, p<0.00). The unweighted Cohen’s Kappa for baseline EF was 0.29 (95% CI 0.07-0.50, fair) and for GLS 0.07 (95% CI -0.08-0.22, slight). By the human read, the mean EF was 63±5% at baseline and 61±6% at surveillance. By AI read, the mean EF was 61±5% at baseline and 60±6% at surveillance (Figure). Mean GLS was 20.9±2.1% at baseline and 20.3±2.6% at surveillance by human and 18.9±2.4% at baseline and 18.7±2.5% at surveillance by AI. The per-patient interval changes in EF and GLS from baseline to surveillance TTE by human and AI are shown in the Figure. Based on the human read, 9 patients (11.1%) met CTRCD criteria (mild N=6, moderate N=3) compared to 16 patients (19.8%) by AI read (mild N=14, moderate N=2). Conclusion: In this proof-of-concept study in patients with BC undergoing surveillance TTE, automated AI measurement of EF and GLS was feasible and showed fair to slight agreement with the human read. More studies met CTRCD criteria by AI compared to human, highlighting the need for research into AI’s ability to predict outcomes.