Limited Echocardiogram Acquisition by Clinicians Aided with Deep Learning: A Randomized Controlled Trial

BackgroundDeep learning (DL) programs can aid in the acquisition of echocardiograms by medical professionals not previously trained in sonography, potentially addressing access issues in underserved communities. This study evaluates whether DL-enabled devices improve limited echocardiogram acquisition by novice clinicians not trained on sonography. MethodsIn this single-center randomized controlled trial (2023-2024), internal medicine residents (N=38) without sonography training received a personal ultrasound device with (N=19) or without (N=19) DL capability for two weeks while caring for patients on a hospital ward. Participants were allowed to use the devices at their discretion for patient-related care. The DL software provided real-time guidance for probe placement and image quality assessment. The primary outcome was time to acquire a five-view limited echocardiogram. Measurements occurred at randomization and after two weeks, with all scans performed on the same standardized patient. Secondary outcomes included image quality using the modified Rapid Assessment for Competency in Echocardiography (RACE) scale and participant attitudes. ResultsAt baseline, both groups had comparable scan times and image quality scores. At follow-up, the DL group demonstrated significantly faster total scan times (152 seconds [IQR 115-195] vs. 266 seconds [IQR 206-324]; p<0.001; Cohens D 1.7) and better image quality with higher RACE scores (15 [IQR 10-18] vs. 11 [IQR 7-13.5]; p=0.034; Cohens D 0.84). Trust in the AI features did not differ between the groups post-intervention. ConclusionsUltrasound machines with DL features may improve image acquisition times and image quality by novices not trained in sonography. These findings suggest DL algorithms could help address critical gaps in image acquisition by healthcare professionals.