Association of echocardiographic findings with mortality: human assessment vs. automated deep learning analysis

Artificial intelligence (AI) has emerged as a promising tool for echocardiographic image analysis, potentially improving efficiency and reducing inter-observer variability. Real-world comparisons between AI-based analysis and human expert interpretation, and their correlation to clinical outcomes, remain limited. This study aimed to evaluate the correlation between AI-based echocardiographic analysis and human expert interpretation and to compare their association with one-year mortality in hospitalized patients.

We conducted a retrospective analysis of 889 consecutive hospitalized patients who underwent a clinically indicated echocardiographic exam. All studies were read and analysed by both human echocardiographic experts and by commercially available AI software (Us2.ai). We performed correlation analysis of common echocardiographic variables obtained by human vs. AI and compared their performance in the prediction of 1-year mortality. Of the 889 patients, 731 (82%) patients (mean age 68 ± 16, 46% Females) had sufficient echocardiographic data to be included in the analysis. Most parameters exhibited a strong correlation between human and AI-derived measurements. AI-derived LVEF values were significantly higher than human estimates (mean difference 5.8%, P < 0.001). In a multivariable model, AI- and human-based mortality prediction were comparable (AUC 0.67 vs. 0.66, P = 0.86). When including AI-obtained automated left ventricular strain analysis, the AI-based model was superior to humans in predicting 1-year mortality (AUC 0.73 vs. 0.66, P = 0.048).

AI-based echocardiographic analysis shows excellent correlation with human-derived measurements. Incorporating automated strain analysis resulted stronger association with mortality of the AI analysis compared to standard human analysis.