胸腔镜单肺叶切除术患者术后住院时间延长预测模型的构建

Construction of prediction models for prolonged length of postoperative hospital stay in patients undergoing thoracoscopic lobectomy

目的构建胸腔镜单肺叶切除术患者术后住院时间延长的预测模型。方法回顾性收集2016年3月至2019年2月于本院胸外科择期气管插管全麻下行胸腔镜单肺叶切除术患者的病历资料,性别不限,年龄≥18岁,ASA分级Ⅰ~Ⅲ级,术后病理诊断为非小细胞肺癌。收集患者基本信息(性别、年龄、吸烟史)、既往病史(血脂异常、高血压、糖尿病、心脑血管疾病、周围血管疾病、慢性阻塞性肺疾病)、过敏史、其他肿瘤史、手术切除部位、麻醉因素(术中非甾体抗炎药和糖皮质激素使用情况、麻醉时长、术中硬膜外麻醉+术后硬膜外镇痛情况)和术后并发症发生情况(胸腔积液、气胸、肺不张)。根据术后住院时间是否延长分为术后住院时间正常组(≤7 d)和术后住院时间延长组(>7 d)。采用logistic回归分析,筛选胸腔镜单肺叶切除术患者术后住院时间延长的预测因子。基于TensorFlow深度学习框架构建患者术后住院时间延长的回归预测模型,评价模型预测效果;进一步基于TensorFlow框架搭建深度神经网络,构建患者术后住院时间延长的分类预测模型,评价模型预测效果,并与传统机器学习算法构建的预测模型进行比较。结果最终共纳入428例患者。多因素logistic回归分析结果显示,年龄和麻醉时长是胸腔镜单肺叶切除术患者术后住院时间延长的危险因素,女性、其他肿瘤史和右肺中叶切除是保护因素(P<0.05)。回归预测模型在训练集和测试集上的平均绝对误差分别为2.16和2.14,均方误差分别为11.05和11.73,模型拟合效果欠佳。分类预测模型在测试集上的准确率、F1值、受试者工作特征曲线下面积分别为75.58%、0.553和0.702,模型预测效果尚可,然而预测效果不优于基于逻辑回归、随机森林、梯度提升和支持向量机的4种传统机器学习方法构建的预测模型。结论性别、年龄、手术切除部位、其他肿瘤史和麻醉时长可作为预测因子,基于深度神经网络构建胸腔镜单肺叶切除术患者术后住院时间延长的分类预测模型。

Objective To construct the prediction model for the prolonged length of postoperative hospital stay in the patients undergoing thoracoscopic lobectomy.Methods The patients of both sexes,aged≥18 yr,of American Society of Anesthesiologists Physical Status classificationⅠ-Ⅲ,who received elective thoracoscopic lobectomy with general anesthesia from March 2016 to February 2019 in our hospital,were selected,their clinical data were collected,and the patients were pathologically diagnosed with non-small-cell lung cancer after operation.Basic information(sex,age,smoking history),previous history(dyslipidemia,hypertension,diabetes,cardiovascular and cerebrovascular diseases,peripheral vascular diseases,chronic obstructive pulmonary diseases),allergy history,other tumor history,surgical resection site,anesthetic factors(intraoperative use of non-steroidal anti-inflammatory drugs and glucocorticoids,duration of anesthesia,intraoperative epidural anesthesia+postoperative epidural analgesia)and postoperative complications(pleural effusion,pneumothorax,atelectasis)was collected.The patients were divided into 2 groups according to whether the length of postoperative hospital stay was prolonged:normal group(≤7 days)and prolonged group(>7 days).Logistic regression analysis was used to identify the predictors for prolonged length of postoperative hospital stay.The regression model for prediction of prolonged length of postoperative hospital stay was constructed based on the TensorFlow deep learning framework,and the efficacy of prediction was evaluated.A deep neural network was further established based on the TensorFlow framework to construct a classification prediction model for prolonged length of postoperative hospital stay,and the efficacy of prediction was assessed,further comparing it with the prediction model constructed by the traditional machine learning method.Results A total of 428 patients were finally enrolled in the study.The results of multivariate logistic regression analysis showed that age and anesthesia duration were the risk factors for the prolonged length of postoperative hospital stay,and female,other tumor history and resection of right middle lobe were the protective factors(P<0.05).The performance of the regression model proved ineffective,getting 2.16 mean absolute error and 11.05 mean square error on the training set,2.14 mean absolute error and 11.73 mean square error on the test set.The classification model achieved better score with accuracy 75.58%,F1-measure 0.553 and area under the receiver operating characteristic curve 0.702 on the test set,however,it showed no better performance than that of 4 other prediction models established by 4 traditional machine learning methods,specifically Logistic Regression,Random Forest,Gradient Boosting and Support Vector Machine.Conclusions Sex,age,surgical resection site,other tumor history and duration of anesthesia can serve as the predictors,and a classification prediction model for prolonged length of postoperative hospital stay is constructed based on a deep neural network in the patients undergoing thoracoscopic lobectomy.