【摘要】 目的 本研究旨在利用机器学习算法建立老年住院患者急性肾损伤（acute kidney injury,AKI）的预测模型，并对各模型的预测效能进行评价，以选出较好的老年住院患者AKI发生预测模型。方法 回顾性选取2018年1月至2022年12月于华中科技大学同济医学院附属同济医院就诊且年龄≥65岁的部分老年患者作为研究对象，利用R语言caret包中的createDataPartition函数将总数据集按7∶3的比例随机分配为训练集和测试集。通过提取目标患者人口学、合并症等指标作为特征变量，以发生AKI为结局变量。利用逻辑回归、随机森林、轻量级梯度提升（light gradient boosting machine,LightGBM）、自适应提升等七种机器学习算法建立二分类预测模型，并通过各项评价指标对模型性能进行评价，并对最优模型进行解释和简化。结果 本研究共纳入15 134例老年患者进行训练集和测试集的构建，其中6879例发生了AKI，经过数据清洗后有65项特征变量纳入预测模型中。在七种机器学习算法中，LightGBM算法具有最优的预测性能，其接收者工作特征曲线下面积（area under the receiver operating characteristic curve,AUROC）、F1指数、Brier评分分别为0.896、0.822、0.127。此外，LightGBM算法的增益重要性（gain importance,GI）指数显示使用利尿剂、中性粒细胞百分比和D-二聚体等指标与AKI发生密切相关。最后通过选取10个贡献度较大的特征构建简易模型，显示性能最好的模型仍为LightGBM(AUROC=0.884)。结论 以LightGBM为代表的机器学习模型能够较好地预测老年住院患者AKI的发生。更多还原

【Abstract】 Objective To establish a predictive model for acute kidney injury(AKI) in elderly inpatients utilizing machine learning algorithms, and to evaluate the predictive efficacy of various models, thus identifying an optimal predictive model of AKI in elderly hospitalized patients. Methods Elderly inpatients aged ≥65 years with AKI who were admitted to Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology from January 2018 to December 2022 were retrospectively selected. Then, they were randomly assigned to the training set and test set at a ratio of 7∶3 by utilizing the cerateDataPartition function in the caret package of R. Characteristic variables, such as demographic data and comorbidity were extracted, and AKI was taken as the outcome variable. Seven machine learning algorithms, including Logistic Regression, Random Forest, Light Gradient Boosting Machine(LightGBM) and Adaptive Boosting, were used to develop establish a binary classification prediction model. Then, the model performance was assessed by various evaluation indicators. The optimal model was interpreted and simplified. Results A total of 15,134 elderly patients were included for the model construction into training sets and test sets, of which 6,879 patients developed AKI. Sixty-five characteristic variables were included in the prediction model after data cleaning. Among the seven machine learning algorithms, the LightGBM algorithm exhibited the best predictive performance, with the area under the receiver operating characteristic curve(AUROC), F1 index, and Brier score of 0. 896, 0. 822, and 0. 127, respectively. In addition, the Gain Importance(GI) index of LightGBM algorithm showed that the use of diuretics, neutrophil percentage and D-dimer were closely related to the occurrence of AKI. Finally, LightGBM(AUROC=0. 884) was the optimal model by selecting 10 features with greater contribution. Conclusion Machine learning models represented by LightGBM can better predict the occurrence of AKI in elderly hospitalized patients.更多还原