文献速递急性肾损伤慢性肾脏病

基于深度神经网络模型预测慢性肾脏病住院患者静脉注射碘对比剂后急性肾损伤的开发与验证:一项多队列分析

Development and validation of a deep neural network-based model to predict acute kidney injury following intravenous administration of iodinated contrast media in hospitalized patients with chronic kidney disease: a multicohort analysis.
PubMed 文献 · Nephrol Dial Transplant · 2023-01-01

该研究基于79个候选变量,用6种特征选择方法筛选出14个变量,构建深度神经网络(DNN)模型,预测慢性肾脏病住院患者静脉注射碘对比剂后对比剂后急性肾损伤(PC-AKI)。模型在内外部验证队列中的AUC分别为0.940和0.939,优于逻辑回归模型,并可将患者分为低、中、高危组。

要点

  • 14个变量构建的DNN模型在内外部验证队列中AUC分别为0.940和0.939,优于逻辑回归模型
  • 按模型风险分层,PC-AKI发生率从低危组<1.0%升至高危组>50%
  • 内外验证队列低危组未发生PC-AKI的准确率均为99%

核心结论

  • 使用常规可获得变量构建的DNN模型能较准确区分住院CKD患者静脉注射ICM后发生PC-AKI的风险。

研究设计与样本

  • 推导队列与内部验证队列来自第二湘雅医院,外部验证队列来自湘雅医院及公开数据库MIMIC-IV。
  • 共纳入4218例次就诊,PC-AKI依据KDIGO血清肌酐标准定义。
  • 使用6种特征选择方法从79个候选变量中筛选预测因子,以DNN建模并与逻辑回归分析比较,以AUC评价判别能力,并设定低危与高危切点进行分层。

主要发现

  • 推导、内部验证和外部验证队列中PC-AKI发生率分别为10.3%、10.4%和11.4%。
  • 14个变量DNN模型在内部和外部验证队列的AUC分别为0.940(95%CI:0.909-0.954)和0.939(95%CI:0.916-0.958),显著优于逻辑回归模型。
  • 亚组分析中模型判别能力较好(AUC≥0.800)。
  • 观察到PC-AKI风险从低危到中危到高危组显著升高(<1.0%至>50%),内外验证队列低危组未发生PC-AKI的准确率均为99%。

临床怎么用

  • 原文未据此给出推荐。

局限性

  • 原文未报告局限性。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND: Stratification of chronic kidney disease (CKD) patients [estimated glomerular filtration rate (eGFR) <60 mL/min/1.73 m2] at risk for post-contrast acute kidney injury (PC-AKI) following intravenous administration of iodinated contrast media (ICM) is important for clinical decision-making and clinical trial enrollment. METHODS: The derivation and internal validation cohorts originated from the Second Xiangya Hospital. The external validation cohort was generated from the Xiangya Hospital and the openly accessible database Medical Information Mart for Intensive CareIV. PC-AKI was defined based on the serum creatinine criteria of the Kidney Disease: Improving Global Outcomes (KDIGO). Six feature selection methods were used to identify the most influential predictors from 79 candidate variables. Deep neural networks (DNNs) were used to establish the model and compared with logistic regression analyses. Model discrimination was evaluated by area under the receiver operating characteristic curve (AUC). Low-risk and high-risk cutoff points were set to stratify patients. RESULTS: Among 4218 encounters studied, PC-AKI occurred in 10.3, 10.4 and 11.4% of encounters in the derivation, internal and external validation cohorts, respectively. The 14 variables-based DNN model had significantly better performance than the logistic regression model with AUC being 0.939 (95% confidence interval: 0.916-0.958) and 0.940 (95% confidence interval: 0.909-0.954) in the internal and external validation cohorts, respectively, and showed promising discrimination in subgroup analyses (AUC ≥ 0.800). The observed PC-AKI risks increased significantly from the low- to intermediate- to high-risk group (<1.0 to >50%) and the accuracy of patients not developing PC-AKI was 99% in the low-risk category in both the internal and external validation cohorts. CONCLUSIONS: A DNN model using routinely available variables can accurately discriminate the risk of PC-AKI of hospitalized CKD patients following intravenous administration of ICM.

来源:Nephrol Dial Transplant,查看原文。版权归原作者所有。

引用

  1. Yan P, Duan SB, Luo XQ, et al. Development and validation of a deep neural network-based model to predict acute kidney injury following intravenous administration of iodinated contrast media in hospitalized patients with chronic kidney disease: a multicohort analysis. Nephrol Dial Transplant. 2023. PMID: 35218197. doi:10.1093/ndt/gfac049

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。