Exploration of Different Machine Learning Methods and Domains of Predictors for Chronic Postsurgical Pain After...

Area of Science:

• Medical Informatics
• Surgical Oncology
• Pain Medicine

Background:

• Chronic postsurgical pain (CPSP) affects approximately 30% of patients following video-assisted thoracoscopic surgery (VATS).
• Machine learning (ML) offers potential for improved prediction of CPSP.
• Multimodal ML models may outperform single-modal approaches, but require investigation.

Purpose of the Study:

• To evaluate and compare different ML models for predicting CPSP after VATS.
• To assess the efficacy of a multivariable approach in CPSP prediction.
• To determine if multimodal ML models improve prediction accuracy.

Main Methods:

• A cohort of 100 patients undergoing VATS were assessed preoperatively across four domains: demographics, psychological factors, quantitative sensory testing, and inflammatory biomarkers.
• Five ML techniques were employed: multiple logistic regression, Kernel k-Nearest Neighbors (kKNN), kKNN with Random Forest variable elimination, Naive Bayesian Classifier, and Gradient Boosting.
• Models were validated using leave-one-out cross-validation, with CPSP assessed one year post-surgery.

Main Results:

• The study included 86 patients who completed the 12-month follow-up.
• Area under the receiver operating characteristics curve (ROCAUC) varied significantly across models and domains, ranging from 0.500 to 0.965.
• The Gradient Boosting model achieved the highest ROCAUC, indicating superior predictive performance.

Conclusions:

• Different ML models demonstrate varying predictive capabilities for CPSP after VATS.
• Gradient Boosting emerged as the most effective ML model in this study.
• Further external validation is recommended to mitigate the risk of overfitting before clinical application.