A deep learning model using echocardiographic video was superior to current recommendations for estimating LVEDP (AUC 0.88 vs 0.48; p<0.001), but showed lower performance in external validation.
Cross-Sectional (n=252)
Yes
Does a deep learning model using echocardiographic video input improve the estimation of LVEDP compared to current recommendations in patients with preserved ejection fraction?
A deep learning model estimating LVEDP from echocardiographic videos showed superior performance to current guidelines in a derivation cohort but had limited generalizability in an external validation cohort.
Absolute Event Rate: 0.88% vs 0.48%
p-value: p=<0.001
Abstract Background Heart failure with preserved ejection fraction (HFpEF) is a prevalent condition with a poor prognosis. Assessment of left ventricular end-diastolic pressure (LVEDP) is important for diagnosis and management, but current non-invasive methods lack accuracy. Purpose This study aimed to train and evaluate a deep learning (DL) model for estimating LVEDP using echocardiographic video input. Methods This multicenter, retrospective, cross-sectional study included patients from 2 large referral hospitals in Belgium. Patients with an ejection fraction ≥50% were included if they had an invasive measurement of LVEDP and echocardiography performed within 7 days. Data from one hospital was used for training and internal validation of a DL model (PyTorch 3D convolutional neural network with spatial and temporal attention mechanisms). Performance metrics and an area under the receiver operating characteristics curve (AUC) were calculated and compared to current European Association of Cardiovascular Imaging recommendations using DeLong’s test. Data from the second hospital was used as external validation. Results A total of 209 patients were included in the derivation set (54% female, mean age 72 ± 15 years), of these 106 (51%) had LVEDP ≥ 16 mmHg. Current recommendations had an accuracy of 49%, a sensitivity of 19%, a specificity of 77%, and an AUC of 0.48 in estimating LVEDP. The DL model had an accuracy of 80%, a sensitivity of 94%, a specificity of 67% and an AUC of 0.88. The DL model was superior in predicting LVEDP compared to current recommendations (p0.001). A total of 43 patients were included in the validation set (56% female, mean age 75 ± 11 years), of these 14 (33%) had LVEDP ≥ 16 mmHg. The DL model demonstrated lower performance in the external validation cohort with an accuracy of 45% and AUC of 0.64. The addition of manual measurement (E, E’) did not improve the accuracy of the model. Conclusion We demonstrate that, within the limits of our small dataset, it is feasible to train a DL model to predict invasive LVEDP based on a single echocardiographic video. However, moderate performance in an external dataset underscores the need for a larger prospective study to develop the model into a clinically applicable solution.
Kellens et al. (Thu,) conducted a cross-sectional in Heart failure with preserved ejection fraction (HFpEF) (n=252). Deep learning (DL) model using echocardiographic video input vs. Current European Association of Cardiovascular Imaging recommendations was evaluated on Area under the receiver operating characteristics curve (AUC) for estimating LVEDP (p=<0.001). A deep learning model using echocardiographic video was superior to current recommendations for estimating LVEDP (AUC 0.88 vs 0.48; p<0.001), but showed lower performance in external validation.