Background: With the improved life expectancy in people living with HIV, predicting individual frailty is important for clinical care. DNA methylation (DNAm) has been previously linked to aging and mortality in non-HIV populations. We aim to establish a panel of DNAm biomarkers to predict frailty in HIV-positive population.
Methods: Our samples consist of HIV-positive participants (Ntotal=1,081) from the Veterans Aging Cohort Study (VACS) and were divided into the training set (Ntraining=460), the validation set (Nvalidation=114), and the testing set (Ntesting=507). We used a well-established score, VACS Index, as a measure of frailty. Model training and fine-tuning were conducted using the ensemble method that aggregated prediction results of four base models. The final number of features were determined in the validation set and the model performance was assessed in the testing set. We conducted a survival analysis to assess whether the predicted frailty was associated with 10-year mortality and gene ontology enrichment analysis of the predictive CpG sites was performed.
Results: We selected a panel of 393 CpGs to build the ensemble machine learning model. The prediction model showed an excellent performance on predicting frailty with Area Under Curve of 0.809 (95%CI: 0.767-0.851) and balanced accuracy of 0.653 in the testing set. The predicted frailty based on our model was significantly associated with 10-year mortality (hazard ratio=1.79, p=4E-05). These 393 CpGs were located within or near 280 genes that were enriched in biological pathways including immune and inflammation response.