Validation of a deep learning–based workflow for echocardiographic interpretation in a cardio-oncology population

Background Echocardiography is central to cardiovascular risk assessment and surveillance in patients with cancer receiving potentially cardiotoxic therapies. However, conventional analysis is operator-dependent and subject to variability, potentially influencing clinical decision-making and contributing to inefficiencies amid growing cardio-oncology demand. Deep learning (DL)–based workflows may improve reproducibility and reporting efficiency, yet the utility in oncology populations remains limited. Aims To evaluate the agreement, reliability, diagnostic performance and variability of an artificial intelligence (AI)-echo software for assessing left ventricular systolic function in a cardio-oncology cohort. Methods This study incorporated retrospective and prospective cohorts at a tertiary cardio-oncology centre. Echocardiographic studies were analysed using a validated DL platform (FDA/CE cleared), with standard sonographer-derived measurements as reference. In the prospective arm, image acquisition followed routine practice, with operators blinded to AI outputs. The primary endpoint was agreement between DL-derived and conventional 2D left ventricular ejection fraction (LVEF). Secondary analyses included comparison with 3D LVEF, temporal variability, interstudy repeatability and discrimination of LVEF <50%. Agreement and correlation were assessed using Bland–Altman analysis, intraclass correlation coefficients (ICC), and Spearman’s rank test. Variability was quantified using the standard error of measurement (SEM) and compared with Levene’s Test. Results A total of 384 patients were included (retrospective n = 282; prospective n = 102). Mean age was 62 ± 16 years, and 61% were women. Breast cancer was the most frequent malignancy (28.8%), followed by haematological (19.2%). Median manual 2D LVEF was 60% (IQR 54–64) and DL-derived LVEF was 58.8% (IQR 53–64), demonstrating good agreement and correlation (ICC: 0.774, 95% CI: 0.729–0.812; bias −0.09; SD 5.59; 95% CI: −10.9 to 11; Spearman ρ: 0.729). Comparison of DL-derived LVEF with 3D LVEF showed similar agreement (ICC 0.80; bias 0.04; 95% limits −9.97 to 10.0). Subgroup analyses showed consistent performance across clinically relevant cohorts. The DL algorithm accurately identified LVEF <50% (ROC-AUC 0.91, 95% CI 0.871–0.945). Interstudy variability was lower for DL repeated measurements compared to manual (SEM 0.21 vs 3.14, p <0.01) and temporal variability was comparable between methods (manual SEM 3.97 vs DL-SEM 5.93; p = 0.112). Conclusions In a real-world cardio-oncology population, a DL-based echocardiography workflow showed good agreement with conventional assessment of LVEF, reliable discrimination of reduced systolic function, comparable longitudinal variability and improved interstudy repeatability. These findings support the potential role of AI-echocardiography in cardio-oncology.