Ensuring the path to correct AS diagnosis - is AI an adequate and promising tool?

Background Accurate diagnosis of sever aortic stenosis (AS) with 2D transthoracic echocardiography (Echo) provides prognostic information, and aids in tailoring treatment plans. Purpose We sought to validate a fully automated artificial intelligence (AI) system for AS diagnosis in Echo against clinician assessment as the reference method. Methods A total of 1,245 consecutive patients underwent routine Echo at our tertiary hospital. The left ventricular outflow tract (LVOT) and aortic valve (AV) velocity-time integral (VTI), peak velocity (Vmax), mean pressure gradient (meanPG), and the dimensionless velocity index (DVI), a synthesised measure of AS severity, were manually measured from Doppler ultrasound data by clinicians (> 5 years of experience) (Figure 1, top panel). The AI system performed fully automated view selection from the entire Echo exams and measured the parameters on appropriate LV views without operator involvement or adjustment. Pearson’s correlation coefficient (R), and mean absolute error (MAE) were used to assess agreement between AI and clinician-measured values. The AI performance of DVI in identifying true severe AS, as defined by clinician measurements (Vmax > 4.0 m/s & meanPG > 40 mmHg), was evaluated using receiver operating characteristic analysis with the area under the curve (AUC), and optimal cut-off value was established by the Youden’s method. Results Strong correlations were observed between AI and clinician-measured values for LVOT-VTI, AV-VTI, and DVI in the entire cohort (R = 0.86, MAE = 4.18; R = 0.81, MAE = 8.57; R = 0.89, MAE = 0.08, respectively) (Figure 1A). Similarly, excellent correlations were found for AV meanPG and Vmax across the entire cohort (R = 0.98, MAE = 2.01; R = 0.98, MAE = 0.15) (Figure 1B and 1C). AUC analysis demonstrated an excellent discrimination ability of the AI in identifying true severe AS (n=225), with DVI showing a sensitivity of 0.97, specificity of 0.84, positive predictive value (PPV) of 0.95, and negative predictive value (NPV) of 0.97 (Figure 1D) with an optimal cut-off value of 0.26. Conclusion This study demonstrates the excellent performance of an AI-based application for fully automated view selection and measurement of aortic parameters, with results closely matching clinician assessments. Additionally, the AI showed comparable ability to human experts in identifying true severe AS, supporting its role as a reliable diagnostic tool.