Predicting the response of high frequency spinal cord stimulation in patients with failed back surgery syndrome: A retrospective study with machine learning techniques

Abstract

Despite the proven clinical value of spinal cord stimulation (SCS) for patients with failed back surgery syndrome (FBSS), factors related to a successful SCS outcome are not yet clearly understood. This study aimed to predict responders for high frequency SCS at 10 kHz (HF-10). Data before implantation and the last available data was extracted for 119 FBSS patients treated with HF-10 SCS. Correlations, logistic regression, linear discriminant analysis, classification and regression trees, random forest, bagging, and boosting were applied. Based on feature selection, trial pain relief, predominant pain location, and the number of previous surgeries were relevant factors for predicting pain relief. To predict responders with 50% pain relief, 58.33% accuracy was obtained with boosting, random forest and bagging. For predicting responders with 30% pain relief, 70.83% accuracy was obtained using logistic regression, linear discriminant analysis, boosting, and classification trees. For predicting pain medication decrease, accuracies above 80% were obtained using logistic regression and linear discriminant analysis. Several machine learning techniques were able to predict responders to HF-10 SCS with an acceptable accuracy. However, none of the techniques revealed a high accuracy. The inconsistent results regarding predictive factors in literature, combined with acceptable accuracy of the currently obtained models, might suggest that routinely collected baseline parameters from clinical practice are not sufficient to consistently predict the SCS response with a high accuracy in the long-term.