基于机器学习的结直肠癌术后并发症风险预测模型的构建与验证OA

Objective:To construct and validate a risk prediction model for postoperative complications of colorectal cancer based on machine learning by retrospectively analyzing patient medical data,and to identify key features affecting the occurrence of complications.Methods:A retrospective analysis was conducted on 300 colorectal cancer patients who were treated at Cangzhou Hospital of Integrated TCM-WM·Hebei from January 2023 to October 2025.Medical records and biochemical data were collected.Patients were followed up at 30 days postoperatively and divided into complication group and non-complication group based on complication occurrence.Correlation analysis was used to screen risk factors associated with postoperative complications.Random forest(RF),XGBoost,and neural network(NN)models were constructed to predict complication risk.Five-fold cross-validation was used,and diagnostic performance was evaluated using the area under the receiver operating characteristic curve(AUC).Results:Comparison between the complication group and the non-complication group indicated that for categorical variables,significant differences were observed in tumor differentiation grade,ASA classification,intraoperative blood loss≥100 mL,smoking history,alcohol history,diabetes,hypertension,patient-generated subjective global assessment(PG-SGA)score≥4,and carcinoembryonic antigen(CEA)elevation(P<0.001).For continuous variables,the complication group had significantly higher age,neutrophil count,CRP levels and albumin levels than the non-complication group(P<0.001).Correlation analysis indicated the above variables had medium to high positive correlation with complication occurrence.The top five features by weight in the RF model were CRP(0.39),albumin(0.164),neutrophil count(0.073),intraoperative blood loss≥100 mL(0.072)and smoking history(0.057);in the XGBoost model were CRP(0.528),neutrophil count(0.132),age(0.095),ASA classification(0.049)and PG-SGA score≥4(0.042);in the NN model were CRP(0.72),PG-SGA score≥4(0.071),age(0.065),intraoperative blood los≥100 mL(0.042)and differentiation grade(0.03).CRP was the highest-weighted core feature across all three models.Age,PG-SGA score≥4,neutrophil count,and intraoperative blood loss≥100 mL were identified as important shared features,appearing in the top five of at least two models.Model validation showed that XGBoost and RF models had significantly higher net benefit than NN and achieved perfect risk stratification(TP=75,FP=0,FN=0)at a threshold of 0.24.XGBoost and RF both achieved AUC above 0.98 with sensitivity above 0.98 and specificity above 0.99.The NN model had AUC=0.951(95%CI:0.912-0.990),sensitivity=0.907,and specificity=0.996.Conclusion:Machine learning analysis identifies CRP as the optimal predictor for postoperative complication risk in colorectal cancer surgery.Age,PG-SGA score≥4,neutrophil count,and intraoperative blood loss≥100 mL are key shared features identified by the models.This provides an efficient tool for early clinical screening of high-risk patients.