OBJECTIVES: Health economic disease models are often build with data from clinical trials and thus do not necessarily reflect the routine care situation in hospitals. For this scenario, we outline a method to generate disease models using data mining algorithms on patient records from a regional clinical cancer registry. METHODS: Markov models are a common technique in decision making. Their structure of states and transitions reflects the progress of a disease. We define a disease state as a set of features which represents a specific state of illness, e.g., diagnosis of breast cancer in HER-2-positive (human epidermal growth factor receptor 2) women who are treated with chemotherapy and mastectomy. In particular, a feature can consist of several attributes, e.g., HER-2 status can be described with the attributes 0, 1+, 2+, 3+. States: A feature selection is executed by a modeler who decides which set of features describes a disease state or patient cohort best. For identifying the most relevant attribute combinations a cluster analysis is applied beforehand. Transitions: Patients remain in one particular state as long as they match the predetermined features and attributes. Otherwise, a change to another state occurs. Thereby, a sequence of states for each patient is defined. As a final step, these sequences are used for deriving a model structure. EXPECTED RESULTS: Markov models backed up by real-life patient records. As a result of the automatic generation process models can be used for validating hypotheses or comparing outcomes for different patient cohorts. Therefore, the usage of such models is not strictly limited to health economic analysis. A first validation indicates the feasibility of the outlined methods. It was possible to reconstruct a published disease model. CONCLUSIONS: Ongoing research is conducted with focus on data quality, i.e., accuracy, completeness and timeliness, at the regional cancer registry.