Supervised learning techniques based on convolutional neural networks (CNNs) for the classification of mitral regurgitation severity

Abstract Background Quantification of mitral regurgitation (MR) remains challenging. Current recommendations emphasize the importance of quantitative methods, but several limitations have been identified. Artificial Intelligence (AI) tools show promise in improving accuracy and reducing variability in echocardiographic assessment of valvular disease. Purpose To evaluate the effectiveness of supervised learning tools based on convolutional neural networks in classifying MR severity. This tool will be tested based on the analysis standard deviations (SD) in velocity plots obtained from 2D colour Doppler transthoracic echocardiogram videos. Methods Prospective and single-center study including 479 patients undergoing transthoracic echocardiography: 77 controls (none or trace MR) and 402 cases with at least mild MR. Apical four-chamber 2D colour Doppler video including 10 beats were obtained. An image post-processing tool is used to obtain pixel level velocity maps from a region of interest that encompasses the mitral valve (Figure 1A), where the SD curves correlate with turbulent flow. These SD curves (Figure 1B) were converted into fixed-size images and analyzed using a supervised learning approach based on a Residual Network (ResNet-18) architecture, enhanced with spatial squeeze and excitation (sSE) modules. MR severity was determined using the multiparametric algorithm proposed by the American Society of Echocardiography (ASE-MPA). The dataset was split into training (75%) and test (25%) sets. A classification threshold of 0.5 defined significant MR. A composite clinical endpoint -death, heart failure hospitalization and mitral intervention (surgical or percutaneous)- was used for prognostic evaluation in the validation dataset (n=121). Results Mechanism of MR was secondary (SMR) in 53% of cases; eccentric jets were described in 32% of patients. According to the ASE-MPA, MR was significant (grades III or IV) in 98 patients (20.4%). The new method was feasible in all patients and demonstrated robust diagnostic performance in classifying MR severity according to the ASE-MPA: area under the ROC curve (AUC) of 0.98 and an overall accuracy of 96% (Figure 1C). The false positive rate (FPR) was 0.03, while the false negative rate (FNR) was 0.08, indicating both high sensitivity and specificity in detecting significant MR. Grad-CAM visualization revealed that the model focused on regions corresponding to the SD curve, supporting the relevance of this feature in the classification process (Figure 1D). In the validation set, 36% of patients experienced the composite endpoint over a median follow-up of 23 months. Kaplan-Meier analysis showed nearly identical event-free survival curves between model predictions and ASE-MPA classification (Figure 2A-B, respectively). Conclusions We describe a novel CNN-based MR grading tool with high diagnostic accuracy and prognostic equivalence to ASE-MPA criteria, supporting its potential clinical utility.