The Achilles heel of mitral valve repair 
Mitral valve repair is the ideal surgical option for patients with mitral valve regurgitation. When performed by expert surgeons, it provides  excellent long term outcomes with lower mortality and reduced operative risk, compared with mitral valve replacement [1][2]. One notable exception is in ischemic mitral regurgitation where recurrent regurgitation post repair is a significant limitation of surgery as it is experienced by a large percentage of patients and has significant clinical impact. For example, in the Cardiothoracic Surgical Trials Network (CTSN) randomized controlled trial of repair versus replacement for ischemic mitral regurgitation (MR), at 2 years, 35% of patients experienced recurrent moderate or greater MR as compared to 2% for mitral replacement [3, 4]. In fact, long term follow-up demonstrates that nearly 30% of patients undergoing repair for ischemic MR will have  recurrent regurgitation [5] which  is associated with a  three times greater risk of death at 10 years and increased risk of hospital readmission for CHF and reoperation [5].  Patients with recurrent MR post repair also have less cardiac reverse remodelling with smaller improvements in body-surface area indexed left ventricular end-systolic volumes, compared to those without recurrent MR [4]. At the patient level, those with recurrent MR have a lower quality of life scores as measured by the Minnesota Living with Heart Failure questionnaire compared with those who do not [4]. Overall, patients with recurrent MR after mitral valve repair for ischemic regurgitation unfortunately do not benefit from the surgical undertaking. Thus, identifying differences between those with and without a durable repair is crucial to improve the outcomes of this surgery.
In this issue of the Journal, Kachroo et al. address this issue by using the STS database to identify features that would predict an unsuccessful repair for ischemic mitral regurgitation. The authors applied Machine Learning (ML) methods to data from 173 patients to predict recurrent MR (> Grade 2) or death versus non-recurrence at one year. They trained three different ML algorithms (Support Vector Machine, Logistic Regression and Deep Neural network) and included 53 preoperative predictors relating to demographics, comorbidities, coronary artery disease architecture and bypass targeted vessels. They found that all three models performed well with AUC values between 0.72 -0.75. Important predictors for recurrent significant regurgitation identified by the models included the necessity of bypass to the ramus or the second obtuse marginal or the second diagonal coronary arteries, the presence of peripheral vascular disease, and the use of beta blockers. Through this paper, the authors have successfully sowed the seeds for the use of ML with structured, nationally collected STS data.  The combination of the use of the STS database with ML methods will also allowed for effective use of data for modelling and identification of predictors of recurrent MR using a data-driven methodology.