Explainable Machine-learning Model Based on Multimodal Ultrasound for Non-invasive Detection of Early Renal Fibrosis:...

To develop and validate an interpretable, multimodal, ultrasound-based machine-learning (ML) model for the non-invasive identification of early renal fibrosis in patients with chronic kidney disease (CKD).

In this prospective, multicenter study, 369 participants (healthy controls, n = 161; mild fibrosis, n = 208) were recruited from 16 institutions. Mild fibrosis was defined by histopathological grading. Six ML algorithms were evaluated, and multiple models were constructed using different combinations of conventional ultrasound, ultra micro angiography (UMA), shear-wave elastography (SWE), super-resolution ultrasound (SRUS), and clinical variables; performance was compared in terms of discrimination, calibration, and clinical utility. External validation was conducted using held-out centers. Model interpretability was examined using Shapley Additive exPlanations (SHAP).

In the internal dataset, the comprehensive fusion model integrating SRUS, SWE, conventional ultrasound, and clinical variables, based on Light Gradient Boosting Machine, achieved an area under the curve (AUC) of 0.948 (accuracy, 0.880; sensitivity, 0.987). In the external dataset, it maintained good generalization (AUC, 0.823; accuracy, 0.748; sensitivity, 0.883). The fusion model outperformed the clinical baseline model in both internal and external validations (internal AUC, 0.732; external AUC, 0.650). SHAP analysis identified both imaging parameters (vessel density, fractal dimension) and clinical indices (serum creatinine (Scr), estimated glomerular filtration rate (eGFR)) as key predictors, consistent with pathophysiological changes of early fibrosis.

An explainable multimodal ultrasound-based ML model showed promising performance for the non-invasive identification of early renal fibrotic changes and may support adjunctive risk stratification in CKD.