EPIUnit - Instituto de Saúde Pública, Universidade do Porto, Porto, Portugal,
EPIUnit - Instituto de Saúde Pública, Universidade do Porto, Porto, Portugal.
Nephron. 2021;145(2):123-132. doi: 10.1159/000512080. Epub 2020 Dec 18.
In hospitalized patients, information on preadmission kidney function is often missing, impeding timely and accurate acute kidney injury (AKI) detection and affecting results of AKI-related studies.
In this study, we provided estimates of preadmission serum creatinine (SCr), based on a multivariate linear regression (Model 1) and random forest model (Model 2) built with different parametrizations. Their accuracy for AKI diagnosis was compared with the accuracy of commonly used surrogate methods: (i) SCr at hospital admission (first SCr) and (ii) SCr back-calculated from the assumed estimated glomerular filtration rate of 75 mL/min/1.73 m2 (eGFR 75).
From 44,670 unique adult admissions to a tertiary referral centre between 2013 and 2015, we analysed 8,540 patients with preadmission SCr available. To control for differences in characteristics of patients with and without SCr, we used an inverse probability weighting technique.
Estimates of SCr were likely to be higher than true preadmission SCr in a low Cr concentration and undervalued in high concentrations although for Model 2 Complete-SCr these differences were smallest. The true cumulative incidence of AKI was 14.8%. Model 2 Complete-SCr had the best agreement for AKI diagnosis (kappa 0.811, 95% CI 0.787-0.835), while surrogate methods resulted in the lowest agreement: (kappa 0.553, 0.516-0.590) and (0.648, 0.620-0.676) for first SCr and eGFR 75, respectively.
Multivariable imputation of preadmission SCr, taking into account elementary admission data, improved accuracy in AKI diagnosis over commonly used surrogate methods. Random forest-based models can serve as an effective tool in research.
在住院患者中，通常缺乏入院前的肾功能信息，这阻碍了及时、准确地发现急性肾损伤（AKI），并影响了与 AKI 相关的研究结果。
本研究使用多元线性回归（模型 1）和随机森林模型（模型 2）建立不同参数化模型，提供了入院前血清肌酐（SCr）的估算值，并比较了这些估算值与常用替代方法（i）入院时的 SCr（首次 SCr）和（ii）从假定的 75 mL/min/1.73 m2 肾小球滤过率（eGFR 75）反推的 SCr（eGFR 75-SCr）在 AKI 诊断中的准确性。
从 2013 年至 2015 年期间，在一家三级转诊中心的 44670 例独特的成人住院患者中，我们分析了 8540 例有入院前 SCr 值的患者。为了控制有和无 SCr 值患者特征的差异，我们使用了逆概率加权技术。
虽然对于模型 2 Complete-SCr，这些差异最小，但在低 Cr 浓度时，SCr 的估算值可能高于真实的入院前 SCr，而在高浓度时则被低估。AKI 的真实累积发生率为 14.8%。模型 2 Complete-SCr 对 AKI 诊断的一致性最好（kappa 0.811，95%CI 0.787-0.835），而替代方法的一致性最低：首次 SCr 和 eGFR 75 分别为 kappa 0.553（0.516-0.590）和 0.648（0.620-0.676）。
考虑到基本入院数据，对入院前 SCr 进行多变量插补，可提高 AKI 诊断的准确性，优于常用的替代方法。基于随机森林的模型可以作为研究的有效工具。
