AI Reads Heart Ultrasound Videos to Predict Mitral Valve Repair Success

Deep learning models trained on routine ultrasound videos of the mitral valve can predict, before a single clip is placed, whether a transcatheter mitral valve edge-to-edge repair will succeed, according to a retrospective single-center study published in Echo Research & Practice. Among 467 patients treated at St Michael’s Hospital in Toronto between 2014 and 2022, the best-performing architecture, called Xception, reached an area under the receiver operating characteristic curve of 0.76 for predicting procedural success, outperforming both conventional logistic regression and a panel of classical machine learning algorithms built from structured echocardiographic measurements.

Transcatheter mitral valve edge-to-edge repair, or MTEER, has become an established option for patients with significant mitral regurgitation who are at elevated surgical risk. In the procedure, a clip device such as the MitraClip is guided by transesophageal echocardiography to grasp the mitral leaflets and reduce the leak. The researchers defined procedural success as residual mitral regurgitation of mild severity or less together with a mean mitral valve gradient below 5 mmHg at the end of the procedure. Success was achieved in 424 of the 467 patients, a rate of 90.7 percent, with a single device implanted in 56.0 percent of cases and a mean post-procedural gradient of 1.7 ± 0.9 mmHg.

The clinical motivation is that patient selection remains difficult, particularly in people with complex mitral valve anatomy. Existing surgical and transcatheter risk scores rely on a limited number of demographic and clinical variables, were largely derived from surgical populations to estimate perioperative mortality rather than technical success of transcatheter repair, and do not incorporate the detailed echocardiographic anatomical features that drive procedural planning. The study team, led by Mathias Claeys and Géraldine Ong with collaborators including Guanghui Wang, set out to develop and evaluate machine learning and deep learning models that integrate clinical and echocardiographic data to predict procedural success.

The cohort consisted of all consecutive patients who underwent MTEER at the single tertiary center during the eight-year window, with fourteen of 481 patients excluded for incomplete clinical or echocardiographic data. The mean age was 77 ± 11 years and 44.7 percent of patients were female. Hypertension, atrial fibrillation, and coronary artery disease were highly prevalent, and nearly 40 percent had a prior hospitalization for heart failure. Most patients had moderate-to-severe or severe mitral regurgitation, with an equal distribution of degenerative and functional etiologies, and the mean left ventricular ejection fraction was 50 ± 17 percent. The authors note that this real-world population is considerably broader and more heterogeneous than those enrolled in pivotal randomized trials, with many centers now extending the procedure to anatomically challenging cases such as extensive leaflet or annular calcification, significant tethering, or prior interventions.

The two modeling tracks differed fundamentally in what they were allowed to see. The machine learning arm used five supervised algorithms—Decision Tree, Random Forest, Gradient Boosting, k-Nearest Neighbors, and Support Vector Machine—fed with 34 predetermined echocardiographic measurements chosen for their known or potential association with procedural outcomes. These parameters spanned leaflet morphology such as leaflet lengths, coaptation and flail gap, tethering angles, and tenting area; annular geometry including intercommissural and anteroposterior diameters; ventricular geometry such as interpapillary muscle distance; and Doppler-derived measures including regurgitation severity and mean gradient. Continuous variables were standardized, hyperparameters were tuned by grid search with cross-validation, and the dataset was randomly split into training and testing cohorts.

The deep learning arm, by contrast, worked directly from images. The researchers extracted 1,374 pre-procedural transesophageal echocardiographic videoclips covering standard mitral valve two-chamber, three-chamber, and four-chamber views with and without color Doppler, converted them into individual frames, and assembled a dataset of 72,952 images. Images were labeled at the patient level as procedural success or failure based on post-procedural echocardiographic outcomes. To prevent data leakage, all images from a given patient were assigned exclusively to either the training or the testing set. Three convolutional neural network architectures—Inception, Xception, and MobileNet—were fine-tuned from pre-trained weights for binary classification, with resizing, rotation, flipping, and normalization applied to mitigate overfitting, and early stopping and batch normalization used to stabilize training. A 10-fold cross-validation strategy was employed, with models trained on nine subsets and tested on the remaining one under strict patient-level separation.

The results showed a clear hierarchy. Logistic regression, built as a benchmark classifier on the same predefined echocardiographic variables, achieved a training accuracy of 66 percent and a testing accuracy of 69 percent. Among the machine learning algorithms, k-nearest neighbors performed best with a testing accuracy of 66 percent, followed by gradient boosting at 64 percent, random forest at 63 percent, and support vector machines at 61 percent, while decision tree trailed at 49 percent. The deep learning models consistently outperformed logistic regression, and Xception achieved the highest discriminatory ability with an AUC of 0.76 alongside favorable accuracy, sensitivity, and specificity. Because the study was designed primarily as a feasibility study, the authors did not perform formal statistical comparisons between algorithms and did not report confidence intervals for the performance metrics, reserving those analyses for future, adequately powered multicenter work.

The authors interpret this performance gap as evidence that much of the predictive information relevant to MTEER outcomes resides in the imaging data itself rather than in manually derived measurements. The machine learning models depended exclusively on 34 measurements selected by human operators, which, while established markers of valve anatomy and disease severity, may not fully capture the complex spatial relationships, leaflet morphology, tissue characteristics, and dynamic motion patterns that influence whether a clip will hold. Deep learning, trained directly on the ultrasound videos, can automatically extract such features without relying on predefined measurements, potentially capturing anatomical characteristics such as calcification or prior valve intervention indirectly through the images even when they were not explicitly encoded as variables.

The clinical stakes of getting this prediction right are substantial. Data from the STS/ACC TVT Registry, which captures all commercial MTEER procedures performed in the United States, show that patients who meet both success criteria—mild-or-less residual regurgitation and a low post-procedural transmitral gradient—experience substantially better one-year outcomes than those with suboptimal results, with lower mortality (14 percent versus 27 percent) and fewer heart failure readmissions (8 percent versus 17 percent). The authors emphasize, however, that their endpoint was the immediate post-procedural echocardiographic result, a technical endpoint that can be influenced by transient loading conditions, general anesthesia, and intraprocedural hemodynamics, and may not fully reflect longer-term valve performance or clinical status.

The study carries important limitations that the authors themselves detail. It was conducted at a single center, external validation in independent multicenter cohorts was not performed, and the reported metrics should be interpreted as preliminary and may overestimate performance when applied to external populations with different characteristics, imaging protocols, or procedural practices. The commercially available deep learning framework used did not readily support explainability analyses such as saliency mapping or Gradient-weighted Class Activation Mapping, so the specific anatomical regions driving the classifications cannot be identified. Interobserver variability for echocardiographic measurements was not formally assessed, procedural outcomes were not adjudicated by an independent core laboratory, and the eight-year study period spanned evolving device generations and techniques that may have influenced both outcomes and model performance. The authors conclude that further validation in larger multicenter cohorts is warranted to determine whether these artificial intelligence tools can genuinely improve patient selection, procedural planning, and Heart Team decision-making in contemporary MTEER practice.

Subject of Research: Machine learning and deep learning prediction of procedural success in transcatheter mitral valve edge-to-edge repair from echocardiographic imaging and clinical data

Article Title: Artificial intelligence-based prediction of procedural success in transcatheter mitral valve edge-to-edge repair

Article References: Claeys, M., Fam, N., Xiao, X., Vyas, T., Chowdhury, M., Moe, G., Alnasser, S., Prud’homme, P., Merdad, A., Bagai, A., Connelly, K. A., Edwards, J., Wang, G., & Ong, G. (2026). Artificial intelligence-based prediction of procedural success in transcatheter mitral valve edge-to-edge repair. Echo Research & Practice, 13(1), Article 33. https://doi.org/10.1186/s44156-026-00134-2

Image Credits: AI Generated

DOI: 10.1186/s44156-026-00134-2

Keywords: artificial intelligence, deep learning, machine learning, echocardiography, mitral regurgitation, transcatheter edge-to-edge repair, MitraClip, procedural success, risk prediction, structural heart disease, Xception, patient selection