Individualized Prediction of Postoperative Nausea and Vomiting Following Thoracoscopic Lung Surgery Through a Penalized Multivariable Nomogram

Technical Summary: Individualized Prediction of Postoperative Nausea and Vomiting Following Thoracoscopic Lung Surgery

Problem Statement
Postoperative nausea and vomiting (PONV) remains a frequent and clinically significant complication following video-assisted thoracoscopic surgery (VATS), with incidence rates potentially approaching 50%. While minimally invasive techniques reduce surgical trauma, specific procedural factors—such as double-lumen intubation, one-lung ventilation, and extensive intrathoracic manipulation—induce physiological stress that predisposes patients to PONV. Persistent symptoms hinder early mobilization and oral intake, negatively impacting recovery quality (QoR-15).

Current clinical risk stratification relies heavily on the Apfel simplified score. However, the authors argue that this general scoring system fails to capture procedure-specific stressors (e.g., mediastinal lymph node dissection, operative duration) and perioperative physiological alterations unique to thoracic surgery. Furthermore, existing predictive models often incorporate postoperative variables (e.g., pain scores, recovery assessments) as predictors. While statistically potent, these variables violate temporal causality, rendering such models impractical for preoperative prophylactic decision-making.

Methodology
The study employed a retrospective observational design involving 121 consecutive patients undergoing VATS at Anhui Chest Hospital between August 2024 and May 2025.

• Data Partitioning: The cohort was randomly split into a training set ($n=84$) and an independent testing set ($n=37$) at a 7:3 ratio.
• Predictor Selection: To ensure temporal validity, the model was constructed using exclusively preoperative and intraoperative variables. Postoperative indicators were strictly excluded. Fourteen candidate predictors were incorporated:
  • Demographics/Lifestyle: Age, sex, BMI, smoking history.
  • Baseline Assessments: ASA classification, Apfel score.
  • Laboratory Biomarkers: Hemoglobin, albumin, globulin, C-reactive protein (CRP).
  • Intraoperative Factors: Surgical procedure type, mediastinal lymph node dissection, operative duration, and intraoperative blood loss.
• Algorithm: Given the relatively small sample size and the risk of multicollinearity among the 14 predictors, the authors utilized L2-regularized logistic regression (Ridge regression). This approach was chosen to constrain coefficient magnitudes, reduce overfitting, and improve model stability. The regularization parameter ($C$) was optimized via cross-validation, resulting in an optimal coefficient of 0.0518.
• Validation: The model was evaluated using:
  • Discrimination: Area Under the Receiver Operating Characteristic Curve (AUC).
  • Calibration: Brier score and calibration curves.
  • Clinical Utility: Decision Curve Analysis (DCA).
  • Stability: Five-fold cross-validation.

Key Results

• Model Performance: The final penalized nomogram integrated all 14 variables.
  • Training Cohort: Achieved an AUC of 0.9132.
  • Independent Testing Cohort: Maintained strong discriminative ability with an AUC of 0.8743 (95% CI: 0.6638–1.0000).
  • Cross-Validation: Five-fold cross-validation yielded a mean AUC of 0.8615 ± 0.0508, indicating stable performance across data partitions.
• Classification Metrics: Using an optimal probability threshold of 0.5792 (determined by the Youden index), the model achieved:
  • Sensitivity: 72.22%
  • Specificity: 89.47%
  • Positive Predictive Value (PPV): 86.67%
  • Negative Predictive Value (NPV): 77.27%
• Calibration and Utility: The testing Brier score was 0.1648, indicating close agreement between predicted probabilities and observed outcomes. Decision Curve Analysis demonstrated that the nomogram provided a higher net clinical benefit than the strategies of "treat all" or "treat none" across a broad range of threshold probabilities.

Key Contributions

1. Temporal Validity: The study successfully developed a prediction tool using only variables available before or during surgery, avoiding the temporal causality pitfalls of models that rely on postoperative data.
2. Procedure-Specific Integration: The model moves beyond the generic Apfel score by integrating thoracic-specific stressors (e.g., lymph node dissection, operative duration) and biological markers (e.g., CRP, globulin) that reflect the patient's inflammatory and nutritional status.
3. Regularized Approach: By applying Ridge regression, the study addressed the challenge of overfitting in a small-sample, high-dimensional dataset, ensuring the model's generalizability despite the modest cohort size.
4. High Specificity: The model demonstrated particularly high specificity (89.47%), suggesting it is highly effective at identifying patients who are unlikely to develop PONV, thereby potentially sparing them from unnecessary prophylactic interventions.

Significance and Claims
The authors posit that this penalized nomogram serves as a practical decision-support tool for Enhanced Recovery After Surgery (ERAS) pathways. By enabling individualized risk stratification, the model facilitates:

• Targeted Prophylaxis: High-risk patients can be identified for intensive multimodal antiemetic strategies.
• Resource Optimization: Low-risk patients (identified by the model's high specificity) may avoid unnecessary medication costs and potential side effects.

The paper concludes that while the model shows strong internal validity and calibration, its widespread clinical application requires further external validation in larger, multicenter populations. The authors explicitly note limitations regarding the single-center nature of the study and the exclusion of detailed intraoperative anesthetic data (e.g., specific opioid consumption) due to electronic record constraints.