Machine learning for the real-time assessment of left ventricular ejection fraction in critically ill patients: a bedside evaluation by novices and experts in echocardiography

BACKGROUND: Machine learning algorithms have recently been developed to enable the automatic and real-time echocardiographic assessment of left ventricular ejection fraction (LVEF) and have not been evaluated in critically ill patients. METHODS: Real-time LVEF was prospectively measured in 95 ICU patients with a machine learning algorithm installed on a cart-based ultrasound system. Real-time measurements taken by novices (LVEF(Nov)) and by experts (LVEF(Exp)) were compared with LVEF reference measurements (LVEF(Ref)) taken manually by echo experts. RESULTS: LVEF(Ref) ranged from 26 to 80% (mean 54 ± 12%), and the reproducibility of measurements was 9 ± 6%. Thirty patients (32%) had a LVEF(Ref) < 50% (left ventricular systolic dysfunction). Real-time LVEF(Exp) and LVEF(Nov) measurements ranged from 31 to 68% (mean 54 ± 10%) and from 28 to 70% (mean 54 ± 9%), respectively. The reproducibility of measurements was comparable for LVEF(Exp) (5 ± 4%) and for LVEF(Nov) (6 ± 5%) and significantly better than for reference measurements (p < 0.001). We observed a strong relationship between LVEF(Ref) and both real-time LVEF(Exp) (r = 0.86, p < 0.001) and LVEF(Nov) (r = 0.81, p < 0.001). The average difference (bias) between real time and reference measurements was 0 ± 6% for LVEF(Exp) and 0 ± 7% for LVEF(Nov). The sensitivity to detect systolic dysfunction was 70% for real-time LVEF(Exp) and 73% for LVEF(Nov). The specificity to detect systolic dysfunction was 98% both for LVEF(Exp) and LVEF(Nov). CONCLUSION: Machine learning-enabled real-time measurements of LVEF were strongly correlated with manual measurements obtained by experts. The accuracy of real-time LVEF measurements was excellent, and the precision was fair. The reproducibility of LVEF measurements was better with the machine learning system. The specificity to detect left ventricular dysfunction was excellent both for experts and for novices, whereas the sensitivity could be improved. Trial registration: NCT05336448. Retrospectively registered on April 19, 2022. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s13054-022-04269-6.