Risk Factors for Chronic Post-Surgical Pain in the Elderly: A Single-Center Retrospective Study

Figures

• 

  Figure 1. Comparison of Model ROC Curves for three types of surgery. ROC curves of predictive models for Orthopedic (A), Urinary tumor (B), and Abdominal tumor surgery (C), illustrating the true positive rate vs false positive rate across machine learning algorithms.

• 

  Figure 2. SHAP-Based Feature Importance for Predictive Models of Different Surgery Types. Panels (A) (Orthopedic surgery), (B) (Urinary tumor surgery) and (C) (Abdominal tumor surgery) present the absolute SHAP values, which are used to indicate the extent of the influence of the corresponding features on the prediction results (without distinguishing whether the influence is positive or negative).

• 

  Figure 3. SHAP Dependence Plots for Key Features in Orthopedic Surgery Predictive Models. Panels (A)–(E) show SHAP-derived feature impacts on orthopedic surgery model predictions: (A) (“Frail state”, 0=non-frail, 1=frail) uses yellow/blue boxplots for positive/negative impacts; (B) (“Age”) has a red curve (with CI) showing a non-linear trend that rises first and drops after 75–80 years; (C) (“Postop sleep disorder”, 0=no, 1=yes) and (D) (“Chronic pain history (non-current sites)”, 0=no, 1=yes) both use yellow boxplots to indicate disorder/history increases predictions; (E) (“Intraoperative blood loss”) has a red curve showing a non-linear pattern that rises to ~1000 mL then drops. SHAP values clarify feature importance/interactions and how clinical factors shape predictions.

• 

  Figure 4. SHAP Dependence Plots for Key Features in Urinary Tumor Surgery Predictive Models. Panels (A–E) use SHAP to show how features influence urinary tumor surgery model predictions: (A) (“Frail state”, 0=non-frail, 1=frail) uses yellow/blue boxplots for distinct output impact patterns; Postoperative AST/ALT (B), Postoperative monocyte percentage (C), Postoperative monocyte percentage (D), Surgery duration (C, min) with CI depicting its non-linear relationship with predictions.

• 

  Figure 5. SHAP Dependence Plots for Key Features in Abdominal Tumor Surgery Predictive Models. Panels (A–E) use the SHAP to demonstrate how clinical features influence the predictions of machine learning models for abdominal tumor surgery (all features: 0 = no, 1 = yes), which helps clarify the contribution of clinical factors to the predictive model. These features include frailty status (A), GAD-7 ≥ 4 (B), postoperative PCA use (C), recent respiratory tract infection (D), and preoperative sleep disorder (E).

• 

  Figure 6. SHAP Waterfall Plots Illustrating Feature Contributions for Random Samples Across Models. This figure presents SHAP waterfall plot, analyzing feature contributions in orthopedics (A), urological tumor (B), and abdominal tumor (C) models by randomly selecting one sample from each. Features like orthopedic frailty state, urological preoperative lymphocyte percentage, and general surgical postoperative PCA usage have their SHAP values successively superimposed, illustrating the specific contribution direction and degree of each random single sample’s features to prediction results. Note that these SHAP values only represent the feature contribution of the corresponding random single sample, not the overall population’s feature importance.