291 Machine Learning Predicts Spinal Cord Stimulation Responders Utilizing Preoperative Pain Scores and Patient Characteristics

INTRODUCTION: Chronic pain is a debilitating condition that impacts over 17 million people around the world. While spinal cord stimulation (SCS) is a promising therapy, patient selection remains challenging due to an inability to predict who will experience pain relief. METHODS: Patients who were scheduled for permanent SCS implant surgery for their chronic back and/or leg pain were recruited for the study. Pain outcome data (e.g., Numeric Rating Scale, Pain Catastrophizing Scale) were collected preoperatively and 3 months after surgery. Responders were defined as having a Global Improvement (GI) score =5 (scale: 0-10) within 3 months of implantation. We trained a suite of machine learning classifiers via nested leave-one-out cross validation to predict which patients will be responders using patient characteristics (e.g., age, diagnosis, medication) and preoperative pain scores. Results were averaged over 10 random seeds for robustness. RESULTS: Sixteen patients (age 64.5±11.7 years, 11 female) were included. The decision tree performed best, achieving accuracy and Cohen’s kappa of 81.88% and 0.6170, respectively. Sensitivity, specificity, area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) were 94.44%, 65.71%, 0.82 and 0.80, respectively. When examining the importance of features via a kernel-based estimation approach for Shapley values, the PCS Helpfulness score was the most important across all models trained, and the top five features were always pain scores rather than other demographics. CONCLUSIONS: We show that it’s possible to predict SCS responders before surgical implantation begins, offering an avenue to screen large numbers of chronic pain patients quickly and affordably to improve outcomes. Future work will incorporate EEG data, utilize composite pain scores to determine responders, and examine generalizability of our algorithm in larger datasets.