Nutritional Biomarker-Guided Prediction of Postoperative Pain Outcomes in Elderly Patients Using a Shapley Additive Explanations (SHAP)-Informed XGBoost Approach.

Nutritional status is increasingly recognized as a critical factor influencing perioperative outcomes. In elderly surgical patients, undernutrition and inflammation may play a key role in the development of both acute and chronic postoperative pain. Despite this, few studies have explored their predictive value using machine learning approaches. This prospective, non-interventional study aimed to assess whether nutritional screening tools and biochemical biomarkers could predict postoperative pain trajectories, both acute and chronic, in elderly surgical patients, using interpretable machine learning methods. A total of 108 patients aged ≥70 undergoing elective surgery under spinal or general anesthesia were enrolled. Preoperative assessments included the Mini Nutritional Assessment-Short Form (MNA-SF), the modified Nutritional Risk in the Critically Ill score (mNUTRIC), the Acute Physiology and Chronic Health Evaluation (APACHE), the Sequential Organ Failure Assessment (SOFA), red cell distribution width (RDW), C-reactive protein (CRP), serum bilirubin, serum albumin, serum calcium, and serum ferritin. Pain was recorded at four time points: pre-surgery, immediately post-surgery, 30 days, and six months. XGBoost classifiers were trained to predict pain at each time point, using a custom ordinal-aware loss function and evaluated via accuracy, F1-scores, and confusion matrices. Feature importance was analyzed with SHAP values for model interpretability. The results were the following: Predictive accuracy varied across timepoints: 36% (pre-surgery), 52% (acute post-surgery), 42% (30 days), and 55% (six months). Misclassifications were predominantly within one ordinal level of the true pain score. SHapley Additive exPlanations (SHAP) analysis revealed APACHE, CRP, albumin, ferritin, and MNA-SF as key predictors across models. Chronic pain predictions at six months showed the highest accuracy and stability, highlighting the relevance of preoperative nutritional and inflammatory markers in long-term pain outcomes. In conclusion, this study uses interpretable machine learning in an innovative way to link nutritional screening and inflammation markers to postoperative pain in elderly patients. The findings emphasize the predictive role of nutritional and inflammatory status in pain trajectories and suggest that integrating such assessments into perioperative care may improve personalized pain management and recovery outcomes in the elderly.