Objectives: Chronic postsurgical pain (CPSP) is a significant burden affecting ∼30% of patients after video-assisted thoracoscopic surgery (VATS). The introduction of machine learning (ML) might improve our prediction models of CPSP, but studies are needed to compare the different ML models. It appears likely that a multimodal ML model would be better compared with a single-modal model, but this is yet to be explored. This study evaluated different ML prediction models for CPSP after VATS using a multivariable approach. Methods: The study included patients scheduled to VATS. Preoperative assessments were performed within 4 domains included demographic variables, psychological factors, quantitative sensory testing, and inflammatory biomarkers. CPSP was assessed 1 year after surgery. Five ML techniques were applied: multiple logistic regression with backward elimination, Kernel k-Nearest Neighbors (kKNN), kKNN with variable elimination using Random Forest, Naive Bayesian Classifier, and Gradient Boosting. The models were applied across the 4 domains of predictors. Models were internally validated using leave-one-out cross-validation. Results: This study enrolled 100 patients, with 86 completing the 12-month follow-up. Results showed varying area under the receiver operating characteristics curve (ROC AUC ) across models and domains, ranging from 0.500 (95% CI: 0.500-0.500) to 0.965 (95% CI: 0.896-1.000), with Gradient Boosting demonstrating the highest ROC AUC . Discussion: The study serves as a proof-of-concept, demonstrating that different ML models can yield varying results when predicting CPSP. Among these, a prediction model based on Gradient Boosting exhibited the best fit. However, the potential risk of overfitting cannot be ruled out, necessitating further validation before clinical implementation.
Sperling et al. (2025) studied this question.