Safety and Efficacy Study of AI Wall Thickness Measurement (EchoNet-LVH-RCT)

EchoNet-LVH-RCT: Blinded, Randomized Controlled Trial of Artificial Intelligence Guided Precision Assessment of Wall Thickness

To determine whether an integrated AI decision support can save time and improve accuracy of assessment of echocardiograms, the investigators are conducting a blinded, randomized controlled study of AI guided measurements of wall thickness in parasternal long axis view compared to sonographer measurements in preliminary readings of echocardiograms.

Studies will be randomized 1:1 to either sonographer preliminary report finding or AI preliminary report finding with final adjudication by the cardiologist. With AI preliminary report, the preliminary interpretations will be generated by AI (artificial intelligence) technology [a semantic segmentation model] and the PACS system's native reporting workflow will be used visualize measurements of left ventricular diameter during diastole (LVIDd), intraventricular septum thickness during diastole (IVSd) and left ventricular posterior wall thickness during diastole (LVPWd). The study team will assess how much cardiologists edit and change this preliminary interpretation is from the final interpretation.

Measurements shown in Picture Archiving and Communication System (PACS) without direct communication between sonographer and cardiologist. Annotations are shown without identifiers on how the annotations were done. Cardiologists are blinded to source of preliminary interpretation.

Currently, sonographer technicians provide preliminary interpretations prior to validation and overreading by cardiologists. This staggered, stepwise evaluation allows for the introduction of AI decision support with minimal impact on patient care. Physicians are already used to adjusting the preliminary report given the variable training of sonographers and on the lookout for changes, variation, or adjustments that need to be made.

A novel AI algorithm developed to assess measurements of left ventricular diameter during diastole (LVIDd), intraventricular septum thickness during diastole (IVSd) and left ventricular posterior wall thickness during diastole (LVPWd). The AI will provide preliminary assessments for cardiologist evaluation.

A sonographer will assess measurements of left ventricular diameter during diastole (LVIDd), intraventricular septum thickness during diastole (IVSd) and left ventricular posterior wall thickness during diastole (LVPWd).

An AI model will assess measurements of left ventricular diameter during diastole (LVIDd), intraventricular septum thickness during diastole (IVSd) and left ventricular posterior wall thickness during diastole (LVPWd).

Proportion of studies for which the difference in wall thickness between preliminary to overread is greater than 0.3cm for LVIDd

Proportion of studies for which the difference in wall thickness between preliminary to overread is greater than 0.2cm for IVSd

Proportion of studies for which the difference in wall thickness between preliminary to overread is greater than 0.2cm for LVPWd

mean absolute error (MAE) of wall thickness between cardiologist overread vs. historical clinical report wall thickness

Mean absolute error (MAE) of wall thickness between cardiologist overread with initial interpretation vs. prior clinical report wall thickness

Whether cardiologists can tell the difference between human vs. AI initial interpretation (Bang's blinding index, or the proportion of studies guessed correctly to be AI or sonographer compared to ground truth of whether it was AI or sonographer).

Inclusion Criteria: Sonographers and cardiologists at CSMC Echocardiography Lab All transthoracic echocardiogram studies performed in the CSMC echocardiogram lab Exclusion Criteria: transthoracic echocardiogram studies not able to be measured by sonographers in run in period