OBJECTIVES: We have initiated a project to identify statistically independent dimensions of health services utilization in epilepsy. A desirable covariate for this analysis is the type of convulsive disorder. Here we report on the classification of epilepsy patients using insurance data. METHODS: We applied an iterative classification technique to patients' sequences of insurance claims for outpatient and inpatient services, diagnostic procedures and drugs. The target population included 122,850 US commercial insurance and Medicare supplement subscribers with a physician visit diagnosis of ICD-9 345xx. We classified persons into the ICD-9 coding scheme (or as “not epilepsy”) by developing a family of rules corresponding to empirically observed claims patterns. We sampled patient histories in blocks of 50. Within each claims history, a neuroepidemiologist looked for diagnosis, procedure and treatment patterns that pointed to a clinical diagnosis, and added that pattern to the rule defining a diagnostic type. Removing classified patients, we sampled remaining claims histories and repeated the process of classification and removal and sampling until 50 claims histories suggested no new classification rules. Remaining patients were tagged as unclassifiable. Finally, we reviewed samples of the classified patients to identify disqualification criteria. RESULTS: The majority of patients are classifiable and the empirical classification rules “make sense” clinically (e.g. diagnostic changes are permitted immediately if they follow a diagnostic procedure). A significant minority of cases of epilepsy have only nonspecific treatment codes assigned. CONCLUSIONS: There are clear examples of patients with different clinical subtypes of epilepsy in claims data, and it will be possible to derive average utilization characteristics and drivers for different types of epilepsy. Analysis of the dimensions of health care utilization (combinations of drugs, procedures, physician and hospital) may yield further insight into currently unclassifiable cases and will provide sensitive measures of the cost impact of new therapies.