Department of Emergency Medicine, Johns Hopkins University, Baltimore, MD; Division of Health Sciences Informatics, Johns Hopkins University, Baltimore, MD.
Department of Emergency Medicine, Johns Hopkins University, Baltimore, MD; Center for Disease Dynamics, Economics and Policy, Washington, DC.
Ann Emerg Med. 2020 Oct;76(4):501-514. doi: 10.1016/j.annemergmed.2020.05.026. Epub 2020 Jul 24.
Acute kidney injury occurs commonly and is a leading cause of prolonged hospitalization, development and progression of chronic kidney disease, and death. Early acute kidney injury treatment can improve outcomes. However, current decision support is not able to detect patients at the highest risk of developing acute kidney injury. We analyzed routinely collected emergency department (ED) data and developed prediction models with capacity for early identification of ED patients at high risk for acute kidney injury.
A multisite, retrospective, cross-sectional study was performed at 3 EDs between January 2014 and July 2017. All adult ED visits in which patients were hospitalized and serum creatinine level was measured both on arrival and again with 72 hours were included. We built machine-learning-based classifiers that rely on vital signs, chief complaints, medical history and active medical visits, and laboratory results to predict the development of acute kidney injury stage 1 and 2 in the next 24 to 72 hours, according to creatinine-based international consensus criteria. Predictive performance was evaluated out of sample by Monte Carlo cross validation.
The final cohort included 91,258 visits by 59,792 unique patients. Seventy-two-hour incidence of acute kidney injury was 7.9% for stages greater than or equal to 1 and 1.0% for stages greater than or equal to 2. The area under the receiver operating characteristic curve for acute kidney injury prediction ranged from 0.81 (95% confidence interval 0.80 to 0.82) to 0.74 (95% confidence interval 0.74 to 0.75), with a median time from ED arrival to prediction of 1.7 hours (interquartile range 1.3 to 2.5 hours).
Machine learning applied to routinely collected ED data identified ED patients at high risk for acute kidney injury up to 72 hours before they met diagnostic criteria. Further prospective evaluation is necessary.
急性肾损伤（AKI）较为常见，是导致住院时间延长、慢性肾脏病发生和进展以及死亡的主要原因。早期 AKI 治疗可以改善预后。然而，目前的决策支持并不能发现处于 AKI 风险最高的患者。我们分析了常规收集的急诊科（ED）数据，并开发了预测模型，以早期识别 ED 患者发生 AKI 的高风险。
本研究为多中心、回顾性、横断面研究，于 2014 年 1 月至 2017 年 7 月在 3 家 ED 进行。纳入所有因住院且入院时和入院后 72 小时均检测血清肌酐水平的成年 ED 就诊患者。我们构建了基于机器学习的分类器，这些分类器依赖于生命体征、主要症状、病史和现病史以及实验室结果，根据基于肌酐的国际共识标准，预测在接下来的 24 至 72 小时内发生 AKI 1 期和 2 期的可能性。通过蒙特卡罗交叉验证评估样本外的预测性能。
最终队列包括 91258 次就诊，涉及 59792 位不同的患者。72 小时时，AKI 发生率大于或等于 1 期为 7.9%，大于或等于 2 期为 1.0%。AKI 预测的接受者操作特征曲线下面积范围为 0.81（95%置信区间 0.80 至 0.82）至 0.74（95%置信区间 0.74 至 0.75），从 ED 就诊到预测的中位数时间为 1.7 小时（四分位距 1.3 至 2.5 小时）。
应用于常规收集的 ED 数据的机器学习可在符合诊断标准前 72 小时识别 ED 患者发生 AKI 的高风险。还需要进一步的前瞻性评估。
