Grading left ventricular diastolic function and filling pressure using artificial intelligence-enabled echocardiography
R Loutati, M Fiman, A Am-Shalom, N Rubin, R Kuperstein, E Raanani, E Schwammenthal, R Klempfner, E MaorAbstract
Background
Assessment of left ventricular diastolic dysfunction (LVDD) and filling pressure is standard in echocardiography but remains time- and expertise-dependent. We evaluated the performance and prognostic significance of an artificial intelligence (AI)-enabled echocardiography for LVDD grading.
Methods
An AI model was developed to classify four grades of LVDD using only apical four-chamber transthoracic echocardiography (TTE) videos without Doppler imaging acquired between 2012 and 2023. Model performance was assessed against contemporary echocardiographic grading guidelines and invasive measurements, and further evaluated in a prospective point-of-care ultrasound (POCUS) cohort acquired by non-cardiologists (Trial number NCT05455541). Associations between model outputs and mortality and heart failure (HF)-related events were examined.
Results
The study included 120,649 TTEs from 90,967 patients (mean age 62.9±17.2 years; 43.1% female). In a hold-out test cohort of 23,952 patients, the model achieved an area under the receiver operating characteristic curve (AUC) of 0.888 for detecting both ≥grade 1 and ≥grade 2 LVDD (elevated filling pressure), and 0.960 for grade 3 LVDD. Validation against invasive measurements in 124 patients yielded an AUC of 0.851. Over a median follow-up of 7 years (interquartile range 4.5-10), 35% died and 12.6% experienced HF-related events. Model-predicted elevated filling pressure was independently associated with higher all-cause mortality (hazard ratio 1.55, 95% CI 1.46–1.64, p<0.001) and HF-related events (subdistribution hazard ratio 2.65, 95% CI 2.42–2.89, p<0.001). Consistent diagnostic and prognostic performance was observed in the prospective POCUS cohort.
Conclusions
AI-enabled echocardiography demonstrated accurate LVDD grading and showed strong associations with clinical outcomes, supporting its potential integration into routine TTE and POCUS practice.Kaplan-Meier survival curvesSpline figure of adverse outcomes