A machine learning model integrating two-dimensional ultrasound radiomics and clinical parameters for differentiating membranous nephropathy from IgA nephropathy.

Authors

Abstract

Membranous nephropathy (MN) and IgA nephropathy (IgAN) are the two most common primary glomerular diseases in China, with distinct pathophysiological mechanisms, treatment strategies, and prognoses. Although renal biopsy remains the diagnostic gold standard, its invasiveness limits routine use. Developing a non-invasive, accurate, and reproducible method for differentiating MN from IgAN is clinically important. To develop a machine learning model integrating two-dimensional ultrasound radiomics and clinical parameters for non-invasive differentiation of MN from IgAN. A total of 267 patients (145 with MN, 122 with IgAN) who underwent renal biopsy were retrospectively enrolled and randomly divided into training (n = 186) and validation (n = 81) sets at a 7:3 ratio. Radiomic features were extracted from renal ultrasound images, and independent clinical predictors were identified using multivariate logistic regression. A combined radiomics-clinical model was constructed using the Extra Trees algorithm. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Multivariate logistic regression identified PLA2R antibody (OR = 0.086, 95% CI 0.040-0.184, P < 0.001), 24-h urine protein quantification (OR = 0.874, 95% CI 0.807-0.946, P = 0.006), and age (OR = 1.019, 95% CI 1.005-1.035, P = 0.031) as independent clinical predictors. The combined radiomics-clinical model achieved an AUC of 0.920 (95% CI: 0.859-0.980) in the validation set, with an accuracy of 0.877, sensitivity of 0.946, and specificity of 0.818, significantly outperforming the radiomics model <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>A</mi> <mi>U</mi> <mi>C</mi> <mo>=</mo> <mn>0.811</mn> <mo>,</mo> <mi>P</mi> <mo>=</mo> <mn>0.001</mn></mrow> </math> and showing a marginally significant improvement over the clinical model <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>A</mi> <mi>U</mi> <mi>C</mi> <mo>=</mo> <mn>0.840</mn> <mo>,</mo> <mi>P</mi> <mo>=</mo> <mn>0.065</mn></mrow> </math> in the test cohort. The machine learning model integrating two-dimensional ultrasound radiomic features with PLA2R antibody, 24-h urine protein quantification, and age demonstrated excellent performance for non-invasive differentiation of MN from IgAN. This model offers a promising auxiliary diagnostic tool that may help optimize clinical decision-making under non-invasive conditions.

Tags

Topics