Construction of machine learning-based prediction model for adverse pregnancy outcomes in pregnancy-related acute kidney injury patients
10.3760/cma.j.cn441217-20240320-00325
- VernacularTitle:基于机器学习构建妊娠相关急性肾损伤患者发生不良妊娠结局的预测模型
- Author:
Chen LU
1
;
Xuan HUANG
1
;
Runze WANG
1
;
Suhua LI
1
Author Information
1. 新疆医科大学第一附属医院肾脏疾病中心 新疆维吾尔自治区肾脏病研究所 新疆肾脏替代治疗临床医学研究中心，乌鲁木齐 830054
- Publication Type:Journal Article
- Keywords:
Pregnancy;
Machine learning;
Acute kidney injury;
Predictive model;
Pregnancy complications
- From:
Chinese Journal of Nephrology
2025;41(8):595-604
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop a predictive model for adverse pregnancy outcomes in patients with pregnancy-related acute kidney injury (Pr-AKI) using machine learning methods.Methods:This study was a single-center retrospective study. Patients with Pr-AKI in the First Affiliated Hospital of Xinjiang Medical University from January 2013 to December 2020 were included. Demographic characteristics, laboratory parameters, and fetal outcomes for comparative analysis between adverse pregnancy outcome group and favorable pregnancy outcome group were collected. Adverse pregnancy outcomes were defined as the occurrence of any one or more of the following events: stillbirth, perinatal death, preterm birth (reaching 28 weeks but less than 37 weeks), and low birth weight (< 2.5 kg). Conversely, an ideal pregnancy outcome was defined as the absence of any adverse pregnancy outcome events. The dataset was randomly divided into a training set (70%) and a validation set (30%). Logistic regression, decision tree, random forest, K-nearest neighbor, support vector machine, and lightweight gradient boosting algorithms were employed on the training set to develop predictive models for adverse pregnancy outcomes in patients with Pr-AKI. Receiver operating characteristic curves were plotted, and the area under the curves ( AUC) were calculated. Recall, precision, accuracy, and F1 scores were used to evaluate the predictive performance of each model. The optimal machine learning model was selected for subsequent analysis. Predictive model variables were screened and compressed by visualizing SHAP (SHapley additive exPlanations) with recursive feature regression. Furthermore, the efficacy of each model was evaluated through calibration curves and clinical decision curves. The optimal predictive model was selected for internal validation using the validation set, and data of in-hospital Pr-AKI patients (72 cases) in the hospital from January 2021 to June 2023 were collected for validation (time series validation set). Results:A total of 458 pregnancies in 441 patients were included in the present analysis, among which 277 cases (60.5%) resulted in adverse pregnancy outcomes. Utilizing the training set, 21 feature variables were selected for model construction. Among the 6 models, the random forest model performed the best ( AUC=0.860, recall=0.784, precision=0.813, F1-score=0.790, accuracy=0.806). With subsequent feature refinement proceeding, a total of 12 clinical indicators were selected to construct the model. Among them, proteinuria, systolic blood pressure, and the highest serum creatinine were the top three related factors, and the other related factors included: severe preeclampsia, baseline serum creatinine, serum albumin, diastolic blood pressure, aspartate aminotransferase, blood uric acid, white blood cell count, serum cystatin C, and cholesterol. Among various machine learning models, the random forest model demonstrated optimal net benefits and the widest clinical utility range, showing robust performance in both internal validation set ( AUC=0.80) and the time series validation set ( AUC=0.72). Conclusions:In this study, different machine learning algorithms are successfully applied to develop predictive models for adverse pregnancy outcomes in patients with Pr-AKI. The random forest model is translated into a clinically applicable tool, providing a reference for the convenient and rapid identification of adverse pregnancy outcomes in Pr-AKI patients.
