Construction of a risk factor prediction model for postoperative complications in elderly patients with colorectal cancer using machine learning
摘要
This study aimed to assess the impact of various clinical and surgical factors on the occurrence of postoperative complications in elderly patients with colorectal cancer and to use various machine learning models to predict the risk of these complications. This study included 109 elderly patients with colorectal cancer who underwent robot-assisted surgery and natural orifice specimen extraction (NOSE) at the Second Xiangya Hospital of Central South University from March 2016 to March 2024. The patients were divided into a no-complication group (77 patients) and a complication group (32 patients) based on whether complications occurred within 30 days after surgery. The clinical and surgical data were collected and analysed using various machine learning models to evaluate the importance of each variable, and SHAP values were used to interpret the model’s predictions. The average age of patients in the complication group was significantly greater than that in the no-complication group (77.75 vs. 72.55 years, P < 0.001), and their BMI was also greater (24.22 vs. 22.51 kg/m2, P = 0.010). The preoperative haemoglobin levels and albumin levels were lower in the complication group. The duration of surgery and amount of intraoperative blood loss significantly longer and greater, respectively, and the postoperative hospital stay and time to first normal diet were longer in the complication group. Multivariate logistic regression analysis revealed that lymph node metastasis (N1 stage: OR = 20.64, P < 0.001; N2 stage: OR = 6.31, P = 0.002); and a history of abdominal surgery (OR = 4.74, P = 0.001), hypertension (OR = 4.07, P < 0.001), diabetes (OR = 7.70, P < 0.001), or coronary heart disease (OR = 19.07, P < 0.001) significantly increased the risk of complications. The XGBoost and logistic regression models performed best in terms of clinical applicability and prediction accuracy. SHAP interpretation of the XGBoost model revealed that age, a history of coronary heart disease, preoperative haemoglobin level, and ASA grade were the main influencing factors. Various clinical and surgical factors significantly affect the occurrence of postoperative complications in elderly patients with colorectal cancer. The XGBoost model performed excellently in predicting postoperative complications and has high clinical application potential. The results can provide a basis for preoperative risk assessment and postoperative management.