Artificial intelligence-driven analysis of peri-operative echocardiography showed good to excellent agreement with clinician measurements for most parameters (ICC 0.605-0.956, p<0.001).
Observational (n=206)
Yes
Does artificial intelligence-driven analysis of peri-operative echocardiography provide comparable measurements to manual clinician assessment in adult patients?
AI-assisted analysis of peri-operative echocardiography shows high agreement with manual clinician measurements, supporting its potential integration into clinical workflows to streamline assessment.
Effect estimate: ICC 0.605-0.956
p-value: p=<0.001
Summary Introduction Point‐of‐care transthoracic echocardiography performed by anaesthetists can influence peri‐operative management but is constrained by time, the need for measurements and advanced skill requirements. Artificial intelligence‐driven analysis may streamline this assessment and yield benefits to clinicians and patients. Methods In this prospective multicentre observational study, adult patients who were referred for pre‐operative echocardiography were enrolled. Anaesthetists certified in echocardiography acquired a predefined 12‐view protocol. The same studies were uploaded to the US2.AI cloud platform, which generated measurements and categorical classifications for 10 key echocardiographic parameters. Manual measurements by clinicians served as the reference standard. Results Of 206 enrolled patients, 202 (98%) with adequate image quality were analysed. Agreement between artificial intelligence‐ and clinician‐derived continuous measurements was good to excellent for most parameters, with intraclass correlation coefficient values 0.605–0.956 (p < 0.001). Left ventricular ejection fraction was strongly correlated (r = 0.845, p < 0.001) with a mean difference of ‐1.9%. The US2.AI software classified left ventricular systolic function correctly in 180/201 (91%) patients and left ventricular diastolic dysfunction in 193/201 (96%) patients. Correlations for right ventricular size and function, and right atrial size were strong (r = 0.860, 0.743 and 0.842, all p < 0.001) with small mean differences. The US2.AI software identified all patients with pulmonary hypertension (n = 10) and severe aortic stenosis (n = 6) correctly. Agreement for inferior vena cava collapsibility (r = 0.641) and cardiac output (r = 0.675) was moderate with low mean bias. Cohen's κ for categorical classifications was statistically significant for all parameters (p < 0.001). Discussion Using a limited predefined image sequence, anaesthetists can obtain most information essential for peri‐operative decision‐making. Agreement between US2.AI and clinicians was high for 10 echocardiographic parameters. These findings support integrating US2.AI into peri‐operative echocardiography workflows, with further studies needed to assess its impact on clinical outcomes.
Borde et al. (Tue,) conducted a observational in Pre-operative echocardiography (n=206). US2.AI cloud platform vs. Manual measurements by clinicians was evaluated on Agreement between artificial intelligence- and clinician-derived continuous measurements (ICC 0.605-0.956, p=<0.001). Artificial intelligence-driven analysis of peri-operative echocardiography showed good to excellent agreement with clinician measurements for most parameters (ICC 0.605-0.956, p<0.001).