Machine learning-based unsupervised phenotypic clustering analysis of patients with IgA nephropathy: Distinct therapeutic responses of different groups.

Abstract

Background: Immunoglobulin A nephropathy (IgAN) has a heterogeneous clinical presentation. Comparison of different IgAN subgroups may facilitate the application of more targeted therapies. This study was aimed to distinct disease phenotypes in IgAN and to develop prognostic models for renal composite outcomes.

Methods: Clinical and pathological data were from 2000 patients with biopsy-proven primary IgAN from four centers, including the First Affiliated Hospital of Sun Yat-sen University (SYSU), the Fifth Affiliated Hospital of Sun Yat-sen University, the Huadu District People's Hospital of Guangzhou, and Jieyang Affiliated Hospital of SYSU in China between January 2009 and December 2018 (training cohort: 1203 patients, validation cohort: 797 patients). Components from principal components analysis (PCA) were used to fit a k-means clustering algorithm and identify distinct subgroups. A subgroup-based prediction model was developed to assess prognosis and therapeutic efficacy in each subgroup.

Results: The PCA-k-means clustering algorithm identified four subgroups. Subgroup 1 had significantly better long-term renal survival upon administration of a renin-angiotensin system blocker (adjusted hazard ratio [aHR]: 0.16, 95% confidence interval [CI]: 0.10-0.27, P <0.001). Subgroup 2 had a significant improvement from corticosteroid therapy (aHR: 0.19, 95% CI: 0.06-0.61, P = 0.005). Subgroups 3 and 4 had milder pathological changes and relatively stable kidney function for several years. Subgroup 3 (predominantly males) had a high incidence of metabolic risk factors, necessitating more intensive monitoring; subgroup 4 (predominantly females) had a high incidence of recurrent macroscopic hematuria. These patterns were similar in the validation cohort. A subgroup-based prognosis prediction model demonstrated an area under the curve of 0.856 in the validation dataset.

Conclusion: The unsupervised clustering method provided reliable classification of IgAN patients into different subgroups according to clinical features, prognoses, and treatment responsiveness. Our subgroup-based prediction model has significant clinical utility for the assessment of risk and treatment in patients with IgAN.