Machine learning model for predicting neuropathic pain following thoracic oncology surgery

Yu ZhangShirong WuMaolin ZhouHui PanQing FanJie XieXue XiaoTian ZhangJinjun ShuYan LuoDongmei MaQing Yang^*

• Sichuan Cancer Hospital, Chengdu, China

Objectives Neuropathic pain (NP) is a common and challenging complication following thoracic oncology surgery, characterized by complex etiological factors. However, effective predictive models for identifying high-risk patients are currently lacking. This study aims to determine the incidence and key risk factors associated with neuropathic pain following thoracic oncology surgery, and to construct and validate a series of machine learning-based risk prediction models, providing a scientific foundation for clinical decision-making. Methods This study involved 647 patients who underwent thoracic oncology surgery at a specialized cancer hospital in Sichuan Province, China (November 2022 to December 2023). An information survey was designed to collect general demographic data and influencing factors. Outcome indicators were assessed using the Numeric Rating Scale (NRS) for postoperative acute pain and the Douleur Neuropathique 4 Questionnaire (DN4) for neuropathic pain evaluation. Using stratified sampling, the patients were divided into training (80%) and testing (20%) datasets. Univariate analysis and LASSO regression were employed to identify independent risk factors for postoperative neuropathic pain, resulting in the selection of seven factors for model inclusion. Subsequently, six machine learning models were developed using Python 3.11: logistic regression (LR), K-nearest neighbors (K-NN), random forest (RF), support vector machine (SVM), XGBoost, and LightGBM (LGBM). To enhance model accuracy, parameter tuning and ten-fold cross-validation were employed, and performance was evaluated using the testing set with the Area Under the Curve (AUC) metric. A visualization analysis of the model's variable features was conducted, and the Shapley Additive Explanations (SHAP) values of the predictive models were calculated to identify the significant influencing factors and their respective impact levels on postoperative neuropathic pain in thoracic oncology. Results The incidence of postoperative NP was 24.26%. The random forest model demonstrated the highest predictive performance (AUC = 0.86). SHAP value analysis revealed that the primary determinants for the onset of neuropathic pain include the surgical approach, the surgeon's expertise, the quantity of thoracic drainage tubes, the duration of thoracic drainage tube placement, postoperative acute pain, and C-reactive protein (CRP). Conclusions The random forest model effectively predicts neuropathic pain following thoracic oncology surgery, facilitating early screening and targeted interventions to improve outcomes.