Background: The ACCORD trial (NCT00000620) was terminated early due to an unexpected increase in mortality with intensive vs. standard glucose-lowering treatment (targeted HbA1c <6% vs. 7.0-7.9%) in type 2 diabetes (T2D) patients. We conducted exploratory analyses to identify factors predictive of mortality, using random survival forests (RSF), a machine-learning methodology.Methods: A total of 240 variables including, demographic, clinical and laboratory data, and their change from baseline during follow-up, were analyzed as potential predictors of mortality, using RSF in patients randomized to the intensive arm and standard arm. An RSF of 500 trees were constructed, with each tree constructed using a different bootstrap sample from the original data. The top 20 predictors, identified using the RSF method, were included in a multivariate analysis using a traditional Cox proportional hazards model with stepwise selection to validate the results.Results: During a 4.9 year median follow-up, 716/10251 patients died (391/5128 in intensive arm and 325/5123 in standard arm). The most important predictors of mortality in patients in the intensive arm were range and maximum change from baseline in urinary creatinine (UCr), range and minimum urinary albumin, and minimum UACR (C-statistic 0.84); and in those in the standard arm, range of UCr, range and minimum of urinary albumin, loop diuretic use, and age (C-statistic 0.84). The accuracy of the RSF in predicting mortality was markedly increased when derived variables capturing range and change over the course of trial were included in addition to baseline variables (C-statistic of 0.84 vs. 0.64). The Cox regression model provided similar results.Conclusion: Urinary biomarkers and their change during follow-up emerged as important predictors of mortality in both treatment arms; whereas loop diuretic use and age were among the top predictors in the standard arm. RSF is a rapid and flexible approach to identify potential outcome predictors among a large number of variables. This analytical approach identified data-derived associations that can be further explored to derive new insights into risk assessment.