Machine learning-based diagnostic prediction of IgA nephropathy: model development and validation study.

Area of Science:

• Nephrology
• Medical Informatics
• Machine Learning

Background:

• Immunoglobulin A (IgA) nephropathy is a leading cause of kidney failure.
• Current diagnosis relies on invasive kidney biopsy, limiting early detection.
• Developing non-invasive methods is essential for timely IgA nephropathy management.

Purpose of the Study:

• To develop and validate non-invasive prediction models for IgA nephropathy.
• To utilize machine learning algorithms for enhanced diagnostic accuracy.
• To identify key clinical predictors for IgA nephropathy.

Main Methods:

• Retrospective analysis of demographic, blood, and urine test data from 1268 participants.
• Development and temporal validation of five machine learning models (XGBoost, LightGBM, Random Forest, ANN, 1D-CNN) and logistic regression.
• Performance evaluation using Area Under the Receiver Operating Characteristic Curve (AUROC) and variable importance analysis (SHAP).

Main Results:

• Machine learning models achieved high predictive performance, with XGBoost showing the highest AUROC of 0.894 in the validation cohort.
• LightGBM achieved an AUROC of 0.913 in the derivation cohort, outperforming several other models.
• Key predictors identified include age, serum albumin, IgA/C3 ratio, and urine red blood cells.

Conclusions:

• Machine learning offers a promising non-invasive approach for predicting IgA nephropathy.
• These models can aid in early detection and potentially reduce the need for kidney biopsies.
• The identified predictors align with established clinical understanding of IgA nephropathy progression.