Classifications like ICD-10 are always purpose-dependent reconstructions of a domain. The strong hierarchical, mutually exclusive divisions of classes are necessarily the result of a compromise, i.e. by using residual classes. Below the aggregation level of classes a machine manipulable level of concepts is needed to enable knowledge-based decision support or semantic integration of patient data. A set of concepts with their meanings and descriptions is provided by terminologies like LOINC. For large specialty areas, terminologies can be no longer realized by enumeration. Similar to human beings using language according to the "LEGO principle", i.e. decomposing in primitive building blocks (vocabulary) and rule-compliant composing to meaningful sentences (grammar), in almost the same manner complex concepts are dealt with in compositional terminologies like SNOMED CT by using logical approaches. Those terminologies enable a sophisticated machine-supported interpretation of exchanged patient data, e.g. for decision support modules in drug therapy.