Proteomic Phenotyping with Machine Learning for Cardiovascular Outcomes in Hemodialysis: Insights from the AURORA Trial

Abstract Aims Cardiovascular (CV) trials have yielded neutral results in hemodialysis. A better understanding of patient profiles is needed to personalize treatment strategies in order to improve CV outcomes in this setting. This study sought to identify biological phenotypes based on proteomic data using machine learning approaches in patients undergoing hemodialysis. Methods and results A clustering analysis using 253 plasma protein biomarkers was performed in 382 patients (machine learning derivation analysis) from the AURORA trial which tested the effect of rosuvastatin on CV outcomes in patients on hemodialysis. A decision tree was subsequently constructed to predict cluster membership and assess its association with CV outcomes in another subset of the trial (n=389 patients, validation analysis). Four phenotypes were identified, namely ‘cytokine storm signaling’, ‘toll-like receptors (TLRs) signaling’, ‘multiple pathways related to inflammation and fibrosis’ phenotypes, as well as a ‘reference phenotype’ which exhibited the least biological abnormalities. In multivariable analysis of the validation study, after adjusting for key prognostic factors, the TLRs phenotype was significantly associated with CV death, all-cause mortality, and MACE (HR=1.65 [1.13 - 2.41], 1.43 [1.03 - 1.98], and 1.48 [1.04 - 2.10], respectively). Conclusion Using unsupervised machine learning on proteomic data, we identified 4 mechanistic biological phenotypes involving cytokine storm and TLRs signaling, inflammation and fibrosis. These biological phenotypes may contribute to CV prognosis and pave the way for personalized therapy in hemodialysis.