A knowledge-based alarm system for intensive care monitoring was designed, built, tested on-line, and evaluated. The system is a functional prototype of a highly specific patient monitor providing alarms on hypovolemia, hyperdynamic state, left ventricular failure and hypoventilation. These intelligent alarm functions aim to maintain the quality of patient monitoring even if nurses' attention is temporarily reduced or focused elsewhere. The alarm system has an electronic access to data available in a multichannel patient monitor and the patient data management system of the intensive care unit. Median filtering, trend estimation, and rule-based reasoning are applied when processing the measured variables and estimating the patient's state.