OBJECTIVES: The study objective was to compare the incremental predictive value of a probability-based classification versus a deterministic grouping of previous healthcare use for subsequent healthcare outcomes. METHODS: Administrative databases from Saskatchewan, Canada (population 1.1 million) were used to identify an incident chronic obstructive pulmonary disease (COPD) cohort (35+ years). All episodes of healthcare use for COPD exacerbations between 2007/08 and 2011/2012 fiscal years were identified. Latent class analysis (LCA) probabilistically grouped the cohort based on healthcare utilization in the index episode using visits to general practitioners (GP), specialists, emergency departments (ED), inpatient hospitalizations (IH), and outpatient medication dispensations. Individuals were also deterministically grouped as institutionalized (IH/ED) or ambulatory (GP/specialist) based on the initiating service of the index episode. Time to a subsequent episode and the cost of the subsequent episode were modelled with Cox proportional hazard regression and generalized linear models, respectively; using only demographic and disease characteristics (model 1), model 1 + LCA-defined groups (model 2), and model 1 + initiating-service groups (model 3). Optimism-corrected C-statistics and root mean square error (RMSE) were used to compare model prediction accuracy. RESULTS: The cohort consisted of 3,105 individuals; of which 38.3% had a recurrent episode, with an average cost of $3,803CAD (SD=$11,959). LCA identified two groups: high user (15.2%) and low user (84.8%), while about 51.6% of the cohort was grouped as institutionalized. Model 2 predicted time to a subsequent episode better (C-statistic = 0.94) than model 3 (C-statistic = 0.91), and model 1 (C-statistic = 0.87). Model 3 predicted cost of the subsequent episode better (RMSE=11934.5) than model 2 (RMSE=11999.8) and model 1 (RMSE=12075.3). CONCLUSIONS: The findings suggest that both LCA-defined groups and initiating-service groups could be used to summarize past utilization for risk prediction modeling; the choice may depend on the outcome of interest.