基于组合优化的机器学习模型预测胃癌术后感染性并发症的诊断性研究

Diagnostic study of machine learning model based on combinatorial optimization to predict postoperative infectious complications of gastric cancer

目的探讨组合优化后的机器学习算法在预测胃癌术后感染性并发症风险模型中的应用及与其他算法的准确性比较,寻找胃癌术后感染早期诊断的可靠生物标志物。方法回顾性分析2018年5月至2023年4月安徽医科大学第三附属医院420例胃癌患者的临床数据资料,利用分层随机化分组法划分为训练集和验证集。采用单因素分析确定术后感染性并发症发生的危险因素;利用训练集构建6种常规的机器学习模型:linear regression、random forest、支持向量机、梯度反向传播、light gradient boosting machine(LGBM)、XGBoost和一种组合优化的适度贪心XGBoost(modified greedy algorithm-XGBoost,MGA-XGBoost)模型。利用验证集通过准确率、精确率和受试者工作特征曲线下面积(area under curve,AUC)等评价指标对七种模型进行评估。结果术后感染性并发症与年龄、手术时间、糖尿病、手术切除范围、联合切除、分期、术前白蛋白、围手术期输血、术前预后营养指数、淋巴细胞绝对数与C-反应蛋白比值及外周血中性粒细胞与淋巴细胞比值相关(P<0.05)。在7个机器学习模型中,MGA-XGBoost模型表现最好,在验证集中的AUC为0.936、准确率为0.889、召回率为0.6、F1分数为0.682、精确率为0.79。模型内部结构中影响占比最高的是糖尿病。结论本研究表明纳入综合性炎症指标的MGA-XGBoost模型可用于预测胃癌患者术后感染性并发症,具有较高的准确性。

Objective To explore the application of combined optimized machine learning algorithm for predicting the risk model of postoperative infectious complications of gastric cancer and to compare the accuracy with other algorithms,so as to find reliable biomarkers for early diagnosis of postoperative infection of gastric cancer.Methods The clinical data of 420 patients with gastric cancer at the Third Affiliated Hospital of Anhui Medical University from May 2018 to April 2023 were retrospectively analyzed and the patients were randomly divided into training set and validation set.Univariate analysis was used to determine the risk factors of postoperative infectious complications.Six conventional machine learning models are constructed using the training set:linear regression,random forest,SVM,BP,LGBM,XGBoost,and MGA-XGBoost model.The validation set was used to evaluate the seven models through evaluation indicators such as ACC,precision,ROC and AUC.Results Postoperative infectious complications were significantly correlated with age,operation time,diabetes,extent of resection,combined resection,stage,preoperative albumin,perioperative blood transfusion,preoperative PNI,LCR and LMR.Among the seven machine learning models,the MGAXGBoost model performed best.Among the seven machine learning models,the MGA-XGBoost model performed best,with AUC of 0.936,ACC of 0.889,recall of 0.6,F1-score of 0.682,and precision of 0.79 on the validation set.Diabetes had the greatest influence on the internal structure of the model.Conclusion This study proves that the MGA-XGBoost model incorporating comprehensive inflammation indicators can predict postoperative infectious complications in patients with gastric cancer.