OBJECTIVES: Approaches for predicting cost accumulation for heterogeneous samples are limited.  We employ the Grouping Algorithm for Cancer Cost Data (GACCD) to investigate cost accumulation over time and identify ‘high cost’ patients. METHODS: Two-fold cross validation was used to evaluate survival and cost accumulation using linked prostate cancer (PCa) registry and Medicare claims data from 1999-2009.  Patients were grouped according to a refined similarity metric using five patient characteristics (cancer stage, age, Charlson Comorbidity Index (CCI), performance proxy indicator, race). Cost accumulation was evaluated in the test dataset for the GACCD groups identified in the training data.  Curves using the test data plotted inverse probability weighted cumulative average total monthly costs (CATMC) for the post-diagnosis period and the proportion of people who were deemed to be ‘high cost’. RESULTS: Application of the inclusion criteria resulted in 110,824 patients. Median (mean) follow up was 48 (51) months and the mortality rate was 27.3%.  The five GACCD groups had distinguishing characteristics e.g., group2 patients were typically older, with CCI above 2 and/or diagnosed with either later stage or unstaged PCa; group3 patients were typically younger, with CCI=0 and diagnosed with early stage PCa.  Cost accumulation within the first three years varied across the groups, with the lowest (highest) rate in group3 (group2) in the training dataset.  Using the test data: at a threshold of $10,000 in CATMC, the proportion of patients that was high cost within three years following diagnosis ranged from 82% in group3 to 90% in group2; proportions ranged from 45% in group3 to over 60% in group2 at a threshold of $25,000; proportions ranged from 15% in group3 to over 30% in group2 at a threshold of $50,000. CONCLUSIONS: A grouping algorithm with a refined similarity metric can identify patient subgroups that will accumulate higher costs over time.