基于电子病历的重症老年患者急性肾损伤连续风险预测研究

目的 探索重症老年患者(≥60岁)急性肾损伤早期连续风险预测的可行性,促进机器学习在临床决策支持中的应用。具体实现以6 h为单位连续预测重症老年患者在未来48 h的急性肾损伤发病风险,并探索可实现何种程度的早期预测,以及比较当前数据和累积数据的预测效果。方法 基于重症监护医学信息数据库(Medical Information Mart for Intensive Care,MIMIC)-Ⅲ,应用逻辑回归、支持向量机、随机森林和轻量梯度提升机(light gradient boosting machine,LightGBM)建模预测。基于曲线下面积(area under curve,AUC)、精确度和召回率进行结果评估。结果 共11 261条重症老年患者记录纳入研究。基于当前6 h数据预测时,LightGBM的AUC达0.845~0.925,随机森林、支持向量机和逻辑回归的最高AUC均低于0.73。基于入重症监护病房最初6 h数据,LightGBM效果最好,AUC达0.845。LightGBM应用当前数据比累积数据获得更高的AUC、精确度和召回率,随机森林、支持向量机和逻辑回归反之。结论 利用LightGBM对重症老年患者进行急性肾损伤早期连续预测切实可行,仅基于重症监护病房前6 h数据的预测结果就可以达到24 h积累数据的预测效果。此外,不同模型对数据的接收能力和适用性不同,LightGBM在当前数据中表现优于累积数据,其他3种模型在累积数据中表现优于当前数据。

Continuous risk prediction of acute kidney injury in elderly critically-ill patients based on electronic medical records

Objective To explore the feasibility of early continuous risk prediction of acute renal injury in severe elderly patients (≥60 years old) and promote the application of machine learning in clinical decision support. Methods The data were collected from the Medical Information Mart for Intensive Care (MIMIC)-Ⅲ database. Logistic regression (LR), support vector machine (SVM), random forest (RF), and light gradient boosting machine (LightGBM) were applied to predict the risk of acute kidney injury (AKI). The prediction results were evaluated based on area under curve (AUC), accuracy, and recall. Results A total of 11 261 intensive care unit (ICU) records were included. When the data of every six hours was used for continuous prediction, AUCs yielded with LightGBM were 0.845－0.925, and those with RF, SVM, and LR were all less than 0.73. As for using the data of the first 6 hours, LightGBM reached AUC 0.845. Compared current data with the cumulative data of ICU, LightGBM yielded higher AUC, accuracy, and recall, whilst it was opposite in RF, SVM, and LR. Conclusion LightGBM completed AKI continuous prediction task with acceptable performance. It is practical to use the data of the first 6 hours on ICU admission for AKI early prediction, which achieve prediction effect of 24－hour accumulated data. In addition, different models have different data applicability. LightGBM performed better based on current data while the other three models favored cumulative data.