结直肠癌切除术患者基于肌肉CT参数的麻醉诱导后低血压预测模型的建立:机器学习算法  

Establishment of predictive model for post-induction hypotension in patients undergoing colorectal cancer resection based on muscle CT parameters:machine learning algorithms

作　　者：盛崴宣[1] 高丹阳 缪慧慧 李天佐[1] Sheng Weixuan;Gao Danyang;Miao Huihui;Li Tianzuo(Department of Anesthesiology,Beijing Shijitan Hospital,Capital Medical University,Beijing 100038,China)

机构地区：[1]首都医科大学附属北京世纪坛医院麻醉科,北京100038

出　　处：《中华麻醉学杂志》2024年第11期1293-1299,共7页Chinese Journal of Anesthesiology

基　　金：国家重点研发计划(SQ2018YFC200044)。

摘　　要：目的使用机器学习算法建立结直肠癌切除术患者基于肌肉CT参数的麻醉诱导后低血压(PIH)的预测模型。方法本研究为单中心、回顾性研究。收集2018年9月1日至2021年9月30日于本院行结直肠癌切除术的318例患者的电子病历资料。预测变量包括年龄、性别、BMI、Hb、ASA分级、TNM分期、年龄调整后的Charlson合并症指数、预后营养指数、第3腰椎水平骨骼肌指数、亨氏单位平均值计算(HUAC)评估的肌肉质量。结局变量为PIH。训练集和测试集根据时间线划分(2020年9月1日之前的患者为训练集,之后的患者为测试集)。过滤法用于筛选特征变量。使用过采样法、重复交叉验证和超参数优化在训练集建立逻辑回归、贝叶斯模型、K最近邻、支持向量机、神经网络、决策树、极度梯度提升树、随机森林8个模型。选择最佳模型后,绘制特征变量重要性排序图、单变量偏依赖图和分解预测图。在测试集中计算混淆矩阵及参数,绘制受试者工作特征曲线、精度召回率曲线、校准曲线和临床决策曲线,以评价预测模型性能。结果筛选的特征变量为HUAC值、年龄、第3腰椎水平骨骼肌指数、预后营养指数、Hb和BMI。随机森林为最优模型,准确度0.9859,马修斯相关系数0.9708,受试者工作特征曲线下面积1.0,精度召回率曲线下面积1.0。校准曲线的Brier分数为0.0766;临床决策曲线显示临床净收益率最高0.6。结论本研究使用机器学习算法,确定了重要特征变量,建立了具有较高性能的、基于肌肉CT参数的PIH预测模型。Objective To establish a predictive model for post-induction hypotension(PIH)in the patients undergoing colorectal cancer resection using machine learning algorithms based on muscle CT parameters.Methods This was a single-center,retrospective study.Electronic medical records from 318 patients who underwent colorectal cancer resection from September 1,2018 to September 30,2021 at our hospital were collected.Predictive variables included age,gender,body mass index,hemoglobin,American Society of Anesthesiologists Physical Status classification,TNM staging,age-adjusted Charlson comorbidity index,prognostic nutritional index,L_(3) level skeletal muscle index,and muscle quality assessed by Hounsfield unit average calculation.The outcome variable was PIH.The training and testing sets were divided based on the timeline(patients before September 1,2020 were included in the training set,and those after that date were included in the testing set).The filtering method was used to screen the feature variables.Eight models,including logistic regression,Bayesian models,K-nearest neighbors,support vector machines,neural networks,decision trees,extreme gradient boosting trees,and random forests,were established in the training set using over-sampling technique,repeated cross-validation and hyperparameter optimization.After selecting the best model,a sorting chart of the feature variables,a univariate partial dependency profile,and a breakdown profile were drawn.In the testing set,the confusion matrix and parameters were calculated,and the receiver operating characteristic curve,precision recall curve,calibration curve,and decision curve were drawn to evaluate the performance of the predictive model.Results The screened feature variables were Hounsfield unit average calculation value,age,L_(3) level skeletal muscle index,prognostic nutritional index,hemoglobin and body mass index.The random forest was the optimal model,with an accuracy of 0.9859,a MCC of 0.9708,the area under the receiver operating characteristic curve was 1.0,and

关 键 词：低血压 麻醉诱导 机器学习 预测模型 

分 类 号：R735.34[医药卫生—肿瘤] R614[医药卫生—临床医学]