Written by Chris Thom
Spoon Feed
PanEcho is an open-source AI system that was found to have high accuracy in the interpretation of echocardiography images across thousands of patient studies.
AI… not just for EF anymore
This study involved the development of an AI system for automated TTE interpretation. Over 30,000 TTE studies at Yale-affiliated hospitals and clinics were utilized to train the AI algorithm. Internal validation and then external validation at alternate sites was performed to assess accuracy. The model involved an assessment of all TTE images and videos within a given study. 39 diagnostic tasks and/or parameters were calculated by the AI algorithm. A separate subset of POCUS TTE images was also included and validated, which included an assessment of 14 diagnostic tasks.
PanEcho had a median area under the curve (AUC) of 0.91 (IQR 0.88-0.93) for 18 diagnostic tasks, and it estimated 21 parameters with a median error of 0.13. Example AI measurements and associated accuracy include an AUC of 0.98 for moderate or worse LVEF, 0.92 for moderate or worse LV diastolic dysfunction, 0.98 for LV wall motion abnormality, and 0.94 for RV systolic dysfunction. Additional areas where the AI algorithm was highly accurate include AV stenosis, MV stenosis, elevated LVOT pressure gradient, and others. These diagnostic tasks were also found to be highly accurate in the POCUS cohort.
How will this change my practice?
This marks an important jump in AI-assisted ultrasound interpretation of TTE. Existing AI tools often do a single task and need a specific view for that task. PanEcho incorporates all available views to make multiple diagnostic determinations simultaneously. This is a gamechanger for POCUS in particular. We may not be far off from a novice to intermediate user acquiring moderate quality images from an Apical 4 and PSLA window, while an AI algorithm then works in the background to provide an accurate assessment of LV function, RV function, and even limited valvular assessment, such as aortic stenosis determination. Also, this AI model is open-source and available for further study, which should position it for acceleration into our daily practice.
Source
Complete AI-Enabled Echocardiography Interpretation With Multitask Deep Learning. JAMA. 2025 Jul 22;334(4):306-318. doi: 10.1001/jama.2025.8731. PMID: 40549400; PMCID: PMC12186137.
