基于机器学习的脑膜瘤术后短期结局预测模型

目的基于脑膜瘤患者的术前真实世界临床变量，使用机器学习算法构建术后短期预后不良的预测模型。方法回顾性地收集了2011年9月—2022年3月在同济大学附属第十人民医院神经外科进行手术切除治疗的脑膜瘤患者的临床变量和出院时的格拉斯哥预后评分（Glasgow outcome scale, GOS）。使用GOS评分将患者进行分组，≤3级的患者定义为预后不良。将患者按照7∶3的比例随机分为训练集和验证集，分别使用支持向量机（support vector machines, SVM）、随机森林（random forest, RF）、梯度提升（gradient boosting, GB）、自适应增强（adaptive boosting, AdaBoost）和多层感知器（multilayer perceptron, MLP）算法在训练集上进行建模，使用验证集检验模型的预测能力。针对预测能力较好的模型使用Shapley Additive Explanations（SHAP）算法进行模型解释。结果收集了424个脑膜瘤患者的42种临床特征数据和GOS评分，筛选后有23种临床特征纳入了训练集的机器学习模型构建。基于不同算法的机器学习模型在验证集中的预测能力表现不同，AdaBoost的表现最优，曲线下面积为0.925。SHAP算法提示在AdaBoost模型中，脑膜瘤最大径、入院时血压、术前的钙离子浓度、血尿素浓度和血肌酐浓度对模型决策的贡献度较大，提示这些术前临床特征与脑膜瘤患者术后短期预后存在相关性。结论本研究使用真实世界大数据，构建了一种可解释的基于AdaBoost算法的机器学习模型，在预测脑膜瘤患者术后短期不良结局上具有良好的效果。

Machine learning-based models for predicting short-term outcomes for meningioma patients after surgical resection

ObjectiveTo construct machine learning models for predicting short-term outcomes of meningioma patients after surgical resection. MethodsThe clinical data of 424 meningioma patients who underwent surgical resection at the Department of Neurosurgery, Shanghai Tenth People’s Hospital from September 2021 to March 2022 were retrospectively analyzed. The Glasgow outcome scale(GOS) scores of patients were evaluated at discharge, GOS score≤3 was defined as a poor prognosis. Patients were randomly divided into a training set and a validation set in a 7∶3 ratio. Machine learning models using support vector machines, random forest, gradient boosting, AdaBoost, and multilayer perceptron algorithms were developed on the training set, and their predictive abilities were tested in the validation set. The Shapley Additive Explanations(SHAP) algorithm was used for model interpretation of the better-performing model. ResultsTotal 42 clinical variables and GOS scores were collected from 424 meningioma patients. After selection, 23 clinical variables were included in the construction of machine-learning models in the training set. The predictive performance of machine-learning model based on AdaBoost algorithm was the best with an AUC value of 0.925. The SHAP algorithm indicated that in the AdaBoost model, the maximum diameter of the meningioma, blood pressure at admission, preoperative calcium ion concentration, blood urea concentration, and blood creatinine concentration contributed most to decision-making of the model, suggesting a significant correlation between these preoperative clinical features and short-term postoperative prognosis in meningioma patients. ConclusionA machine-learning model based on the AdaBoost algorithm has been constructed in the study, which demonstrates a good performance in predicting short-term adverse outcomes in meningioma patients after surgical resection.