AI-assisted echocardiographic assessment of left ventricular function in primary care: a prospective implementation study
M Illyes, R Kiss-Kovacs, L Szenasi, Z Kakonyi, A Szijarto, M Tolvaj, M Szabo, Z Sovegjarto, Z Bagyura, E Kovari, B Merkely, A Kovacs, G Agoston, M TokodiAbstract
Background
Timely detection and monitoring of left ventricular (LV) dysfunction in primary care remain challenging due to limited access to comprehensive echocardiography and variable operator expertise. Although point-of-care ultrasound (POCUS) by general practitioners (GPs) may enable earlier detection, accurate estimation of LV ejection fraction (LVEF) requires substantial experience. Artificial intelligence (AI) could support GPs in this task; however, prospective implementation studies are lacking.
Purpose
We sought to prospectively implement a previosuly validated deep learning (DL) model for the echocardiographic assessment of LV function in primary care and to compare its performance with that of GPs in estimating LVEF and detecting LV dysfunction.
Methods
Prior to study initiation, the three participating GPs completed a structured POCUS training program consisting of an online module and a two-day hands-on course, both concluding with formal examinations. Within the study framework, GPs performed transthoracic echocardiography on patients presenting to their practice and attempted to assess LV function independently. In parallel, an apical 4-chamber view (A4C) video was uploaded to a core laboratory server, where the DL model predicted LVEF. The predictions were returned to the GP for feedback. After study completion, all uploaded A4C recordings were analyzed by an EACVI-certified echocardiographer, whose LVEF measurements served as the reference standard.
Results
The GPs participating in the study uploaded 154 A4C videos for analysis, of which 133 (86%) had an ACEP image quality score ≥3 and were therefore deemed suitable for core laboratory analysis. The DL model predicted LVEF with a mean absolute error (MAE) of 4.31 percentage points and an R2 of 0.51, whereas for detecting LV dysfunction (LVEF <50%), it achieved an accuracy of 87%, specificity of 89%, and sensitivity of 79%. GPs were able to provide a visual estimate of LVEF in only 112 of the 133 videos (84%), yielding a MAE of 6.04 percentage points and an R2 of -0.07 compared with core laboratory measurements. Importantly, they detected LV dysfunction with an accuracy of 81% and a specificity of 87%, both comparable to the DL model (p=0.186 and p=0.831, respectively), but with a lower sensitivity (47%, p=0.041). During feedback, GPs reported that in 98% of cases, they would trust the DL model more than their own assessment.
Conclusions
In this prospective study, we successfully implemented a DL model for echocardiographic LV function assessment in primary care. The model estimated LVEF with lower error and demonstrated greater sensitivity for detecting LV dysfunction than GPs, while maintaining comparable specificity. Performance remained robust even when GPs were unable to estimate LVEF. These findings support the feasibility of AI-assisted echocardiography in primary care and its potential to enhance early detection and monitoring of LV dysfunction.Outline of the study protocolMain results of the study