Session M1.1

Support Vector Machine Based Conformal Predictors for Risk of Complications following a Coronary Drug Eluting Stent Procedure

VN Balasubramanian*, G Ramkiran, S Panchanathan, J Vermillion,
A Bhaskaran, RM Siegel

Arizona State University
Tempe, AZ, USA

Machine learning algorithms such as Support Vector Machines (SVMs) have been used in cardiology to improve the quality of care, stratify risk and provide prognostications. Conformal Predictors are a recently developed set of machine learning algorithms that, unlike many conventional classification systems, allow not just risk classification on new patients, but add “valid” measures of confidence in our predictions for individual patients. In this work, we have applied a novel SVM-based conformal prediction framework to predict the risk of complications following a coronary Drug Eluting Stent (DES) procedure, using patient data provided by Advanced Cardiac Specialists (ACS), Phoenix, Arizona. In addition to standard Major Adverse Cardiac Events and generic procedural complications, DES procedures (which are currently the de facto option for Percutaneous Coronary Intervention (PCI)) have resulted in additional complications including stent thrombosis and restenosis, which could result in myocardial infarction or death. Data was obtained from the central PCI registry maintained at ACS. 2312 patient cases (de-identified) who had a DES procedure performed during 2003–07, and who had followed up during the 12 months after the procedure, were selected. A novel conformal prediction approach based on SVMs was used to classify a patient with the potential risk of complications, based on history, clinical presentation, angiographic and procedural attributes. The datasets were randomly divided into 75% as training and 25% as testing data. The SVM algorithm provided a high accuracy of 94% on unseen test patients, whose outcomes were already known. Our results also accurately reflected the validity of the SVM-based conformal predictors. For example, at the 95% confidence level, the error rate was consistently less than or equal to 5%; at the 90% confidence level, the error rate was always bounded within 10%. This predictive model helps to risk stratify a patient for post-DES complications. The valid measures of confidence can be used by the physician to make an informed, evidence-based decision to manage the patient appropriately. In the future, this approach can be broadened in its application to other areas in cardiology and medicine.

(Abstract Control Number: 232)