Global longitudinal strain (GLS) is an echocardiographic measure to detect chemotherapy-related cardiovascular dysfunction. However, its limited availability and the needed expertise may restrict its generalization. Artificial intelligence (AI)-based GLS might overcome these challenges. Our aims are to explore the agreements between AI-based GLS and conventional GLS, and to assess whether the agreements were influenced by expertise levels, cardiac remodeling and cardiovascular diseases/risks. Echocardiographic images in the apical four-chamber view of left ventricle were retrospectively analyzed based on AI-based GLS in patients treated with chemotherapy, and correlation between AI-based GLS (Caas Qardia, Pie Medical Imaging) and conventional GLS (Vivid E9/VividE95, GE Healthcare) were assessed. The agreement between unexperienced physicians (“GLS beginner”) and experienced echocardiographer were also assessed. Among 94 patients (mean age 69 ± 12 years, 73% female), mean left ventricular ejection fraction was 64 ± 6%, 14% of patients had left ventricular hypertrophy, and 21% had left atrial enlargement. Mean GLS was − 15.9 ± 3.4% and − 19.0 ± 3.7% for the AI and conventional method, respectively. There was a moderate correlation between these methods (rho = 0.74; p < 0.01), and bias was − 3.1% (95% limits of agreement: -8.1 to 2.0). The reproducibility between GLS beginner and an experienced echocardiographer was numerically better in the AI method than the conventional method (inter-observer agreement = 0.82 vs. 0.68). The agreements were consistent across abnormal cardiac structure and function categories (p-for-interaction > 0.10). In patients treated with chemotherapy. AI-based GLS was moderately correlated with conventional GLS and provided a numerically better reproducibility compared with conventional GLS, regardless of different levels of expertise.
Read full abstract