基于随机森林算法构建腰椎后路椎间融合术后感染预测模型

Development of a prediction model of surgical site infection after posterior lumbar interbody fusion in patients based on random forest algorithm

中文关键词:  腰椎后路椎间融合术  手术部位感染  危险因素  预测模型  随机森林

英文关键词:Lumbar posterior interbody fusion  SSI  Risk factors  Predictive model  Random forest

中文摘要:

  【摘要】 目的：利用随机森林算法对腰椎后路椎间融合术后发生感染的相关危险因素进行分析并制定预测模型，为临床预防腰椎后路椎间融合术后手术部位感染（surgical site infection，SSI）的发生提供参考依据。方法：回顾性研究北京中卫云医疗数据分析与应用技术研究院经过数据处理分析提供的2019年6月～2021年6月在河北医科大学第一医院、第二医院、第三医院等河北省及北京市共15家三级甲等医院脊柱外科住院接受腰椎后路椎间融合术治疗患者的脱敏数据资料。统计分析比较感染组（SSI）和非感染组（non-SSI）的分类数据，得到对术后感染具有显著影响的变量，使用SPSS Modeler 20数据建模系统作为工具，采用随机森林（RF）算法进行分析，得到术后感染的患者特征，即感染模型。结果：本研究共纳入8764例患者数据，其中373例患者被诊断为SSI，发病率为4.4%（95%CI，2.2%～6.5%）。经过Logistic回归模型分析多个自变量与因变量的相关性，确定六个变量[包括肥胖、美国麻醉师协会（American Society of Anesthesiologists，ASA）分级Ⅲ级及以上、手术时间延长、慢性心脏病、糖尿病和肾功能不全]与SSI独立相关。以随机森林模型进行分类可获得较高的精度，为90.6％，腰椎后路椎间融合术后易发生感染的患者特征，即两种感染模式：[（BMI=1） and （SD=1） and （ASA=1） and （RI=1）] or [（BMI=0） and （SD=1） and （DM=1） and （RI=1）]。结论：随机森林分类算法应用于本研究可获得90.6％的平均精度，并得到两种感染模型，（1）患者肥胖，肾功能不全，ASA分级Ⅲ级及以上，且手术时间≥3h；（2）患者无肥胖，但同时患糖尿病、肾功能不全，且手术时间≥3h。

英文摘要:

  【Abstract】 Objectives: To analyze the risk factors related to infection after posterior lumbar interbody fusion(PLIF) by random forest algorithm and develop a prediction model, providing a certain reference for clinical prevention of surgical site infection(SSI) after PLIF. Methods: A retrospective study was conducted on the masked data of patients hospitalized for PLIF in the spinal surgery department of some third-level grade A hospitals in Beijing municipality and Hebei Province from June 2019 to June 2021 provided by Beijing Zhongwei Cloud Medical Data Analysis and Application Technology Research Institute through data processing and analysis. The classification data were analyzed and compared between SSI group and non-SSI group to obtain variables that significantly impacted the postoperative infection. SPSS Modeler 20 system was used as the tool for model development, and random forest algorithm was applied to analyze, obtaining the patient characteristics of postoperative infection, namely the infection model. Results: A total of 8,764 patients were included in study, and 373 patients were diagnosed with SSI, with an incidence rate of 4.4%(95%CI: 2.2% to 6.5%). After statistical analysis, six variables, including obesity, ASA Ⅲ and above, prolonged operative time, chronic heart disease, diabetes and renal dysfunction, were independently associated with SSI. Classification with a random forest model yielded a high accuracy of 90.6%. The characteristics of patients prone to infection after PLIF(two models of infection) was: [(BMI=1) and (SD=1) and (ASA=1) and (RI=1)] or [(BMI=0) and (SD=1) and (DM=1) and (RI=1)]. Conclusions: The random forest algorithm applied in this study could obtain an average accuracy of 90.6%, and two infection models were obtained as: (1)Patients with obesity, renal insufficiency, ASA grade Ⅲ or above, and operative time≥3h; (2)Patients who are not obese, but with diabetes, renal insufficiency, and the operative time≥3h.