Hu Zhengbiao, Bu Shuangshan, Wang Kai, Cao Qianqian, Zheng Huanhuan, Yang Jie, Chen Shanshan, Wu Yuemeng, Ren Wenkai, He Chenlei
Department of Ultrasound Medicine, Affiliated Dongyang Hospital of Wenzhou Medical University, No. 60 Wuning West Road, Dongyang City, Zhejiang Province, 322100, China.
Department of Nephrology, Affiliated Dongyang Hospital of Wenzhou Medical University, Dongyang City, Zhejiang Province, 322100, China.
BMC Nephrol. 2025 Jun 23;26(1):289. doi: 10.1186/s12882-025-04253-6.
IgA nephropathy (IgAN) is the most common form of glomerulonephritis worldwide, characterized by immune complex deposition in the glomerular mesangium, leading to mesangial hypercellularity, persistent microhematuria, proteinuria, and progressive renal impairment. Given its common occurrence, diagnosis normally involves renal biopsy, with its accompanying risks of bleeding and infection. In this study, multiple machine learning algorithms were used to develop a non-invasive and improved model for the diagnosis of IgAN.
This retrospective study included 292 patients with IgAN and 310 individuals with different nephropathies, utilizing 82 clinical variables, with kidney pathology results serving as ML labels. A random forest (RF) regression model addressed missing values. Subjects were divided into a development set (n = 542) and a test set (n = 60). The RF method was applied to select 17 key features for building diagnostic models, including the RF model, support vector machine (SVM), adaptive boosting (ADB), and traditional doctor judgment. Performance was evaluated using accuracy, sensitivity, specificity, and area under the curve (AUC) from receiver operating characteristic (ROC) analyses.
The random forest model performed best with an accuracy of 82.3% and an AUC of 0.89 on the test set, outstripping SVM with an AUC of 0.82 and ADB with an AUC of 0.88. High urinary protein, low serum albumin, and elevated IgG levels were the primary features correlated with IgAN.
In this study, a non-invasive diagnostic model for IgAN was developed, with RF line and superior accuracy and clinical applicability. This further highlights the potential of ML-based approaches in reducing reliance on invasive procedures and providing opportunities for early IgAN diagnosis.
IgA 肾病（IgAN）是全球最常见的肾小球肾炎形式，其特征是免疫复合物沉积于肾小球系膜，导致系膜细胞增生、持续性镜下血尿、蛋白尿及进行性肾功能损害。鉴于其常见性，诊断通常需要进行肾活检，而肾活检存在出血和感染风险。在本研究中，使用了多种机器学习算法来开发一种用于 IgA 肾病诊断的非侵入性且改进的模型。
这项回顾性研究纳入了 292 例 IgA 肾病患者和 310 例患有不同肾病的个体，利用 82 个临床变量，将肾脏病理结果作为机器学习标签。随机森林（RF）回归模型处理缺失值。受试者被分为一个开发集（n = 542）和一个测试集（n = 60）。应用 RF 方法选择 17 个关键特征用于构建诊断模型，包括 RF 模型、支持向量机（SVM）、自适应增强（ADB）以及传统医生判断。使用受试者工作特征（ROC）分析的准确性、敏感性、特异性和曲线下面积（AUC）来评估性能。
随机森林模型在测试集上表现最佳，准确率为 82.3%，AUC 为 0.89，超过了 AUC 为 0.82 的 SVM 和 AUC 为 0.88 的 ADB。高尿蛋白、低血清白蛋白和升高的 IgG 水平是与 IgA 肾病相关的主要特征。
在本研究中，开发了一种用于 IgA 肾病的非侵入性诊断模型，RF 模型具有更高的准确性和临床适用性。这进一步凸显了基于机器学习的方法在减少对侵入性程序的依赖以及为 IgA 肾病早期诊断提供机会方面的潜力。
