Qian Qing, Wu Jinming, Wang Jiayang, Sun Haixia, Yang Lei
Hangzhou Normal University, Hangzhou, People's Republic of China.
Institute of Medical Information & Library, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, People's Republic of China.
Int J Gen Med. 2021 Feb 25;14:623-632. doi: 10.2147/IJGM.S289671. eCollection 2021.
To assess the performance of models for early prediction of acute kidney injury (AKI) in the Intensive Care Unit (ICU) setting.
Data were collected from the Medical Information Mart for Intensive Care (MIMIC)-III database for all patients aged ≥18 years who had their serum creatinine (SCr) level measured for 72 h following ICU admission. Those with existing conditions of kidney disease upon ICU admission were excluded from our analyses. Seventeen predictor variables comprising patient demographics and physiological indicators were selected on the basis of the Kidney Disease Improving Global Outcomes (KDIGO) and medical literature. Six models from three types of methods were tested: Logistic Regression (LR), Support Vector Machines (SVM), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Decision Machine (LightGBM), and Convolutional Neural Network (CNN). The area under receiver operating characteristic curve (AUC), accuracy, precision, recall and F-measure (F1) were calculated for each model to evaluate performance.
We extracted the ICU records of 17,205 patients from MIMIC-III dataset. LightGBM had the best performance, with all evaluation indicators achieving the highest value (average AUC = 0.905, F1 = 0.897, recall = 0.836). XGBoost had the second best performance and LR, RF, SVM performed similarly ( = 0.082, 0.158 and 0.710, respectively) on AUC. The CNN model achieved the lowest score for accuracy, precision, F1 and AUC. SVM and LR had relatively low recall compared with that of the other models. The SCr level had the most significant effect on the early prediction of AKI onset in LR, RF, SVM and LightGBM.
LightGBM demonstrated the best capability for predicting AKI in the first 72 h of ICU admission. LightGBM and XGBoost showed great potential for clinical application owing to their high recall value. This study can provide references for artificial intelligence-powered clinical decision support systems for AKI early prediction in the ICU setting.
评估在重症监护病房（ICU）环境下急性肾损伤（AKI）早期预测模型的性能。
从重症监护医学信息集市（MIMIC）-III数据库收集数据，纳入所有年龄≥18岁且在ICU入院后72小时内测量血清肌酐（SCr）水平的患者。ICU入院时已有肾脏疾病的患者被排除在分析之外。根据改善全球肾脏病预后组织（KDIGO）和医学文献，选择了17个包括患者人口统计学和生理指标的预测变量。测试了来自三种方法类型的六个模型：逻辑回归（LR）、支持向量机（SVM）、随机森林（RF）、极端梯度提升（XGBoost）、轻量级梯度提升决策树（LightGBM）和卷积神经网络（CNN）。计算每个模型的受试者工作特征曲线下面积（AUC）、准确率、精确率、召回率和F值（F1）以评估性能。
我们从MIMIC-III数据集中提取了17205例患者的ICU记录。LightGBM表现最佳，所有评估指标均达到最高值（平均AUC = 0.905，F1 = 0.897，召回率 = 0.836）。XGBoost表现次之，LR、RF、SVM在AUC上表现相似（分别为0.082、0.158和0.710）。CNN模型在准确率、精确率、F1和AUC方面得分最低。与其他模型相比，SVM和LR的召回率相对较低。在LR、RF、SVM和LightGBM中，SCr水平对AKI发病的早期预测影响最为显著。
LightGBM在ICU入院后的前72小时内显示出最佳的AKI预测能力。LightGBM和XGBoost因其高召回率而具有很大的临床应用潜力。本研究可为ICU环境下基于人工智能的AKI早期预测临床决策支持系统提供参考。
