External validation of a deep learning algorithm for automated echocardiographic strain measurements

<jats:title>Abstract</jats:title> <jats:sec> <jats:title>Funding Acknowledgements</jats:title> <jats:p>Type of funding sources: Private company. Main funding source(s): Us2ai.</jats:p> </jats:sec> <jats:sec> <jats:title>Background</jats:title> <jats:p>Echocardiographic strain imaging reflects myocardial deformation and is a sensitive measure of cardiac function and wall-motion abnormalities. Manual strain analysis is time-consuming, prone to error and requires considerable expertise. Deep learning (DL) algorithms could automate interpretation of echocardiographic strain imaging.</jats:p> </jats:sec> <jats:sec> <jats:title>Purpose</jats:title> <jats:p>To develop and externally validate automated DL-based measurement of the left ventricle (LV) global longitudinal strain (GLS) and regional strain.</jats:p> </jats:sec> <jats:sec> <jats:title>Methods</jats:title> <jats:p>We developed and validated an automated DL workflow using convolutional neural networks to automatically annotate and assess individual 2D videos of apical 4-chamber, 2-chamber and 3-chamber views. We trained the algorithm in an internal dataset and validated GLS externally in (1) a real-world general population dataset from Taiwan, (2) a core-lab measured dataset from the multinational PROMIS-HFpEF study, and regional strain in (3) the HMC-QU-MI study of patients with suspected myocardial infarction (MI).</jats:p> </jats:sec> <jats:sec> <jats:title>Results</jats:title> <jats:p>In the external datasets, the DL workflow successfully analysed 3741 (89%) studies in the real-world cohort, 176 (96%) in PROMIS-HFpEF and 158 (98%) in HMC-QU-MI. Automated GLS showed good agreement with manual measurements in the real-world dataset (mean ± SD −18.9±4.5% vs −.18.2±4.4%, respectively; bias 0.7±2.5%, root-mean-squared error [RMSE] 2.6, Pearson’s correlation 0.84; Figure 1A) and in PROMIS-HFpEF (−15.4±4.1% vs −15.9±3.6%, respectively; bias −0.6±2.7%, RMSE 2.8, Pearson’s correlation 0.76; Figure 1B). In the general population cohort, automated GLS measurements accurately identified patients with vs without heart failure (ROC-AUC 0.89 for total HF and 0.98 for HF with reduced ejection fraction). In patients with suspected MI, automated regional strain identified regional wall-motion abnormalities with average ROC-AUC 0.80 (range 0.74–0.86 for apical 4-chamber view [Figure 2A] and 0.69–0.90 for apical 2-chamber view [Figure 2B]).</jats:p> </jats:sec> <jats:sec> <jats:title>Conclusions</jats:title> <jats:p>DL algorithms can interpret cardiac strain images with similar accuracy compared with conventional measurements. These results highlight the potential of DL algorithms to democratize the use of cardiac strain measurements.</jats:p> </jats:sec>