Prognostic power of artificial intelligence based estimation of left ventricular ejection fraction in standard and point-of-care echocardiography
K Faierstein, R L Loutati, M F Fiman, A Y Younis, A A Amshalom, N R Rubin, I C Cohen, R R Raphael, L F Fisher, R K Klempfner, E R Raanani, E M Maor, E S SchwammenthalAbstract
Background
Artificial intelligence (AI)-based left ventricular ejection fraction (LVEF) estimation compares reasonably well with conventional readings. We assessed its prognostic value in the standard transthoracic echocardiography (TTE) and point-of-care-ultrasound (POCUS) settings.
Methods
AI-based LVEF estimation was retrospectively tested in 21,920 patients who underwent a standard TTE examination between 2012 and 2021 and prospectively tested in 555 hospitalized patients scanned by non-cardiologists with handheld POCUS devices between 2022 and 2023. Patients were divided into those with preserved (pEF), mildly reduced (mrEF), and significantly reduced LVEF (rEF) using cut-off values of 50% and 40%. Outcome measures were death and heart failure (HF) hospitalization.
Results
In the TTE cohort, patients in the AImrEF and AIrEF groups had a significantly higher risk for death (HR 1.31 [95% CI 1.20–1.43] and 1.65 [95% CI 1.53–1.77], respectively) and HF hospitalization (HR 2.15 [95% CI 1.74–2.65] and HR 3.16 [95% CI 2.63–3.79], respectively) when compared to AIpEF. Consistent results were obtained in the POCUS group. In the TTE cohort both AI-predicted and cardiologist-reported LVEF clearly discriminated three prognostic trajectories, with superposition of AI-defined and cardiologist-defined curves (p<0.001 for all pairwise comparisons). However, in the POCUS group, only AI-assisted LVEF determination could discriminate the three prognostic trajectories (p<0.001 for all pairwise comparisons), while cardiologist review failed to prognostically separate pEF and mrEF patients (p value=0.95 and 0.9 for death and HF hospitalization, respectively).
Conclusions
AI-based LVEF estimation accurately predicts outcome in standard TTE as well as in POCUS studies of hospitalized patients performed by non-cardiologists. In the context of POCUS, AI support appears valuable even for the experienced cardiologist reader.KM prognostic powerSpline all-cause mortality and HF