Background: Despite spinal cord stimulation's (SCS) proven efficacy, failure rates are high with no clear understanding of which patients benefit long term. Currently, patient selection for SCS is based on the subjective experience of the implanting physician.
Objective: To develop machine learning (ML)-based predictive models of long-term SCS response.
Methods: A combined unsupervised (clustering) and supervised (classification) ML technique was applied on a prospectively collected cohort of 151 patients, which included 31 features. Clusters identified using unsupervised K-means clustering were fitted with individualized predictive models of logistic regression, random forest, and XGBoost.
Results: Two distinct clusters were found, and patients in the cohorts significantly differed in age, duration of chronic pain, preoperative numeric rating scale, and preoperative pain catastrophizing scale scores. Using the 10 most influential features, logistic regression predictive models with a nested cross-validation demonstrated the highest overall performance with the area under the curve of 0.757 and 0.708 for each respective cluster.
Conclusion: This combined unsupervised-supervised learning approach yielded high predictive performance, suggesting that advanced ML-derived approaches have potential to be used as a functional clinical tool to improve long-term SCS outcomes. Further studies are needed for optimization and external validation of these models.
Copyright © Congress of Neurological Surgeons 2022. All rights reserved.