|
| 机器学习方法构建老年患者术后急性肾损伤的预测模型 |
| Developing a prediction model for postoperative acute kidney injury in elderly patients by using machine learning methods |
| |
| DOI：10.12089/jca.2023.12.003 |
| 中文关键词: 老年 急性肾损伤 机器学习 风险预测模型 |
| 英文关键词: Aged Acute kidney injury Machine learning Risk prediction model |
| 基金项目:中国医学科学院医学与健康科技创新工程项目（2019-I2M-5-011） |
|
| 摘要点击次数: 645 |
| 全文下载次数: 271 |
| 中文摘要: |
目的 采用机器学习方法构建老年患者术后急性肾损伤（AKI）的预测模型。
方法 收集2019年6月至2020年7月行手术治疗的老年患者的术前信息和术后随访信息，提取实验室检查结果，共纳入115个术前变量，应用极端梯度提升（XGB）、梯度提升（GBM）、随机森林（RF）、支持向量机（SVM）和弹性网络逻辑回归（ELA）5种方法构建术后AKI预测模型，采用受试者工作特征曲线下面积（AUROC）、精确度-召回率曲线下面积（AUPRC）和Brier评分评估模型性能。为简化模型以便于临床实践应用，得到原始模型后剔除部分相关性较低的变量，并再次用上述方法对模型进行评估。
结果 本研究最终纳入老年患者5 929例，男3 359例（56.7%），女2 570例（43.3%），年龄65～99岁，其中154例（2.6%）发生术后AKI。在应用5种机器学习方法构建的预测模型中，XGB模型的AUROC和AUPRC最高，分别为0.798（95%CI 0.705～0.888）和0.230（95%CI 0.079～0.374）；Brier评分最低，为0.023（95%CI 0.014～0.029）。对XGB模型进行简化后，保留72个变量，简化模型的AUROC、AUPRC和Brier评分分别为0.790（95%CI 0.711～0.861）、0.176（95%CI 0.070～0.313）和0.024（95%CI 0.017～0.033），与原模型差异无统计学意义。
结论 在应用5种机器学习方法构建的术后急性肾损伤预测模型中，XGB的预测效能最佳，经简化的XGB预测模型仍保留较高的预测效能，且更容易在临床上推广使用。 |
| 英文摘要: |
Objective To develop a predictive model for postoperative acute kidney injury (AKI) in elderly patients using machine learning methods.
Methods The preoperative information and postoperative follow-up information of elderly patients who underwent surgery from June 2019 to July 2020 were collected, and the laboratory examination results were extracted. A total of 115 preoperative variables were included. A model of postoperative AKI was constructed using five methods: extreme gradient boosting (XGB), gradient boosting machine (GBM), random forest (RF), support vector machine (SVM), and elastic net logistic regression (ELA). The performance of the model was evaluated using area under the receiver operating characteristic curve (AUROC), area under the precision recall curve (AUPRC), and Brier score. To simplify the model for clinical application， the original model was obtained and some variables with low correlation were removed, and the model was evaluated again using the above method.
Results This study ultimately included 5 929 elderly patients, 3 359 males （56.7%） and 2 570 females (43.3%), aged 65-99 years. Among them, 154 patients (2.6%) experienced postoperative AKI. Among the prediction models constructed using five machine learning methods, XGB has the highest AUROC and AUPRC, with values of 0.798 (95% CI 0.705-0.888) and 0.230 (95% CI 0.079-0.374), respectively. Its Brier score is the lowest among all models, the score is 0.023 (95% CI 0.014-0.029). After simplifying the XGB model, 72 variables were retained. The AUROC of the simplified model was 0.790 (95% CI 0.711-0.861), slightly lower than that of the original model. The AUPRC was 0.176 (95% CI 0.070-0.313), and the Brier score was 0.024 (95% CI 0.017-0.033), and there was no significant statistical difference, indicating that there was no significant difference in the predictive ability of the simplified model compared to the original model.
Conclusion Among the five machine learning methods used to construct postoperative AKI prediction models, XGB has the best predictive performance. The simplified XGB prediction model still retains high predictive performance and is easier to be promoted in clinical practice． |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
