Interpretable machine learning model for predicting perioperative Clavien-Dindo III/IV complications of rectal cancer anterior resection after neoadjuvant therapy: a real-world study.

Objective: This study constructed machine learning (ML) models based on 10 different algorithms using the ratio of preoperative and postoperative hematological indicators to predict the risk factors for perioperative Clavien-Dindo grade III/IV complications after anterior resection in rectal cancer patients who underwent neoadjuvant therapy.

Methods: Rectal cancer patients undergoing anterior resection after neoadjuvant therapy at the Xijing Hospital of Digestive Diseases were retrospectively collected from January 2011 to December 2021. Receiver Operating Characteristic (ROC) curves, calibration curve, Decision Curve Analysis (DCA), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and F1 score were used to evaluate the model performance. The Shapley Additive exPlanation (SHAP)algorithm was used to interpret the feature importance of the best model.

Results: The dynamic hematological ratio characteristics before and after surgeries were integrated to construct the predictive models and identify the optimal one. Gradient Boosting Machine (GBM) was considered the best-performing model with the area under the test set curve (AUC: 0.939, 95 %CI: 0.905-0.972), accuracy: 0.903, sensitivity: 0.836, specificity: 0.907. The interpretable SHAP algorithm revealed that hemoglobin, red blood cell ratio and serum calcium ratio were the crucial factors influencing the predictions of the model.

Conclusion: GBM was identified as the optimal predictive model. This model, by predicting the occurrence of perioperative Clavien-Dindo III/IV complications in patients who underwent neoadjuvant therapy and anterior resection, has identified the significant role of hemoglobin, red blood cell ratio and serum calcium ratio, providing guidance for clinical practice. The integration of SHAP provides a clear interpretation of the model's predictions.