Background and hypothesis: Specific urinary peptides hold information on disease pathophysiology, which, in combination with artificial intelligence, could enable non-invasive assessment of chronic kidney disease (CKD) aetiology. Existing approaches are generally specific for the diagnosis of single aetiologies. We present the development of models able to simultaneously distinguish and spatially visualize multiple CKD aetiologies.
Methods: The urinary peptide data of 1850 healthy control (HC) and CKD [diabetic kidney disease (DKD), immunoglobulin A nephropathy (IgAN) and vasculitis] participants were extracted from the Human Urinary Proteome Database. Uniform manifold approximation and projection (UMAP) coupled to a support vector machine algorithm was used to generate multi-peptide models to perform binary (DKD, HC) and multiclass (DKD, HC, IgAN, vasculitis) classifications. This pipeline was compared with the current state-of-the-art single-aetiology CKD urinary peptide models.
Results: In an independent test set, the developed models achieved 90.35% and 70.13% overall predictive accuracies, respectively, for the binary and the multiclass classifications. Omitting the UMAP step led to improved predictive accuracies (96.14% and 85.06%, respectively). As expected, the HC class was distinguished with the highest accuracy. The different classes displayed a tendency to form distinct clusters in the 3D space based on their disease state.
Conclusion: Urinary peptide data present an effective basis for CKD aetiology differentiation using machine learning models. Although adding the UMAP step to the models did not improve prediction accuracy, it may provide a unique visualization advantage. Additional studies are warranted to further validate the pipeline's clinical potential as well as to expand it to other CKD aetiologies and also other diseases.
Keywords: UMAP; chronic kidney disease; differential diagnosis; peptides; urine.
© The Author(s) 2023. Published by Oxford University Press on behalf of the ERA.