Explainable Biomarkers for Automated Glomerular and Patient-Level Disease Classification

Area of Science:

• Nephrology
• Computational Pathology
• Medical Image Analysis

Background:

• Pathologists assess renal biopsies using microscopy, but subtle changes can be missed.
• Computational approaches offer potential for quantifying subvisual clues and linking them to clinical outcomes.

Purpose of the Study:

• To demonstrate that explainable biomarkers derived from machine learning can differentiate glomerular disorders at the light microscopy level.
• To develop a computational system for classifying specific kidney diseases.

Main Methods:

• Extracted 233 explainable biomarkers (color, morphology, texture) from renal biopsy images.
• Utilized traditional machine learning to classify minimal change disease (MCD), membranous nephropathy (MN), and thin basement membrane nephropathy (TBMN).
• Employed Gini feature importance and linear discriminant analysis for the final classification model.

Main Results:

• Six morphologic and four microstructural texture features were identified as the best performing biomarkers.
• Achieved classification accuracies of 77% for glomerular-level and 87% for patient-level analysis.
• Demonstrated diagnostic value of computational methods and explainable glomerular biomarkers.

Conclusions:

• Explainable biomarkers derived from machine learning have diagnostic value in renal pathology.
• Computational methods are compatible with existing knowledge of kidney disease pathogenesis.
• The algorithm holds potential for novel prognostic and mechanistic biomarker discovery in clinical datasets.