C3PW: A Novel Machine Learning Method for Assessing Percutaneous Transvenous Mitral Commissurotomy Outcome in Patients with Mitral Stenosis

Abstract

Background: Patients with symptomatic moderate-to-severe mitral stenosis (MS) with pliable valves are indicated for percutaneous transvenous mitral commissurotomy (PTMC) typically over a valve replacement based on favorable anatomic characteristics. Currently, this decision is arrived based on Wilkins’ echocardiographic score which is considered to be a simplistic estimate often. In the current work, the authors present a novel machine learning (ML) framework that considers a comprehensive set of clinical and echocardiographic variables to indicate the possible outcome for choosing the transvenous procedure over an invasive surgery. Methods: Data considered for this retrospective study included demographic, clinical, and preprocedural echocardiographic variables pertaining to patients with moderate-to-severe MS without significant mitral regurgitation (MR). The success of the procedure was defined by four different postprocedural variables such as the final mitral valve area (MVA), MR, left atrial pressure, and right ventricular systolic pressure (RVSP). Three data mining tasks highlighting the importance of ML techniques to predict the outcome of the PTMC procedure in patients with MS have been demonstrated. The potential predictive value of the outcome of a PTMC procedure considering a comprehensive set of variables using ML techniques has not been investigated till date according to the authors’ knowledge. Cover coefficient-based clustering power as weights (C3PW), a novel problem transformation technique which deals with the multilabel classes, is proposed to correctly classify the patients having successful PTMC procedure. Results: Extreme gradient boosting (XGB), an ML technique, gave the best performance (accuracy: 0.79; F-score: 0.87) on the transformed single-label problem. Application of association rule mining revealed that a combination of the following parameters such as “initial MVA” <1 cm2, “RVSP” <50 mmHg, “valvular calcification” score ≤2, “leaflet mobility” score ≤2, and “normal sinus rhythm” plays a crucial role in determining the success of the PTMC procedure. The identified variables with their indicated ranges having a significant “I-rule” value can be put together as a scoring index to predict the successful outcome. Conclusions: A ML-based artificial intelligence tool has been demonstrated to serve as an alternative to the Wilkins score to select patients for a successful PTMC procedure. Of the various models tested, the C3PW approach with XGB algorithm demonstrated the best evaluation metrics. A systematically analyzed ML framework that yields highly interpretable and conclusive findings with high confidence has been demonstrated to be a useful tool in clinical decision-making.