Session S72.1

Ranking Predictors of Complications following a Drug Eluting Stent Procedure Using Support Vector Machines

RK Gouripeddi*, VN Balasubramanian, J Harris, A Bhaskaran,
RM Siegel, S Panchanathan

Center for Cognitive Ubiquitous Computing
Tempe, AZ, USA

Predictive and risk stratification models using machine learning algorithms, such as Support Vector Machines (SVMs), have been used in cardiology and medicine to improve patient care and prognosis. In this work, we have used SVM based Recursive Feature Elimination (SVM-RFE) methods to select patient attributes/features relevant to the etiopathogenesis of complications following a drug eluting stent (DES) procedure. With a high dimensional feature space (145 features, in our case), and comparatively few patients, there is a high risk of ‘over-fitting’. Also, for the model to be clinically relevant, the number of patient features need to be reduced to a manageable number, to be used in patient care. SVM-RFE selects subsets of patient features that have maximal influence on the risk of a complication. Since SVMs employ kernel functions that transform the data into a different feature space, any correlations between attributes are compensated in this process to provide a more accurate ranking of the relevance of the individual patient attributes. The patient features are ranked by iteratively training the SVM, computing a ranking criterion for all the features based on the weights of the support vectors, and eliminating those features with the smallest criterion. De-identified data of 2312 patients who had a DES procedure and who had followed up during the 12 months following the procedure during the period of 2003-2007 was obtained from the Percutaneous Coronary Intervention registry maintained with the Advanced Cardiac Specialists in Arizona. A total of 145 patient features which included clinical/presentation, history, angiographic and procedural patient details were used in the initial model. The complications following a DES procedure that were considered for this model included: stent thrombosis and restenosis, which manifest as chest pain, myocardial infarction and sometimes even death. SVM-RFE was used to rank all the features. To determine the significance of the ranking, multiple SVM based predictive models were used to predict the risk of complications for patients with recursive elimination of one feature at a time based on the obtained ranking. In our results, when compared with our initial model with all the 145 features, we obtained better performance of the classifiers with 75 top ranked patient features, a 50% reduction in the original dimensionality of the data space. There was a universal improvement in performance of all SVMs with different kernels and parameters. This method of feature ranking helps to determine the most informative patient features. Use of these relevant features improves the prediction of complications following a DES procedure. Further, the high ranked patient features that contribute to the complications can be aggressively controlled and clinically managed in the patient population to prevent the occurrence of a complication.

(Abstract Control Number: 259)