危重症急性肾损伤患者预后预测模型
Predictive Model for Prognosis of Critically Ill Patients with Acute Kidney Injury

目的：建立并验证基于机器学习方法的危重症急性肾损伤患者可解释死亡预测模型。方法：本研究提取了美国大型公开重症数据库MIMIC-IV的2008年至2019年、急性肾损伤(AKI)患者诊断AKI当天的临床数据，随机将数据分为训练队列和验证队列。提取了MIMIC-III数据库3005例AKI患者数据作为外部验证集。通过多轮特征选择(低方差过滤、高相关性过滤、互信息筛选、SHAP值分析和递归特征消除)选择出12个最佳特征组合，10种机器学习方法被用来开发评估住院死亡率的模型。根据其曲线下面积(AUC)选择最优模型。采用SHapley Additive exPlanation (SHAP)值来解释最优模型。结果：本研究共计急性肾损伤3701例最终纳入患者(中位年龄，65岁，女性41.3%)。他们被随机分为一个培训队列(2591人，70%)和一个验证队列(1110人，30%)。10个机器学习模型中随机森林(RF)模型具有最好的判别能力，并采用SHAP方法解释了模型。最终的模型在内部(AUC = 0.807)和外部(AUC = 0.720)验证中都能较准确预测AKI，将有助于临床医生判断重症监护病房ICU住院患者的预后，并进行早期干预。
Objective The aim is to create and confirm a machine learning-based model for predicting mortality in critically ill patients with acute kidney injury (AKI). Methods: Patient clinical records for AKI as of its diagnosis day, spanning 2008 to 2019, were retrieved from the extensive MIMIC-IV public critical care database in the United States. The dataset was arbitrarily segmented into two groups: one for training and another for validation. Furthermore, the MIMIC-III database provided data from 3005 AKI patients, serving as an external validation dataset. A dozen ideal features were chosen after several stages of feature selection, including low variance filtering, high correlation filtering, mutual information screening, SHAP value analysis, and recursive feature elimination. A total of ten machine learning techniques were employed to create models for evaluating mortality rates within hospitals. The selection of the best model was guided by the area beneath the receiver operating characteristic curve (AUC). The optimal model was analyzed using SHAP (SHapley Additive ExPlanation) values. Results: The study eventually incorporated 3701 patients with AKI, averaging 65 years in age and 41.3% female. The subjects were arbitrarily split into two groups: a training group (2591 patients, 70%) and a validation group (1110 patients, 30%). Within the group of 10 machine learning models, the random forest (RF) model stood out as the most effective in discrimination and was analyzed through the SHAP technique. The ultimate model successfully forecasted AKI in both internal (AUC = 0.807) and external (AUC = 0.720) validations. This can aid medical professionals in evaluating the future outlook of patients in the intensive care unit (ICU) and enable prompt medical actions.