An intelligent alarm system for the postoperative monitoring of cardiac surgery patients, which did not require any manual data entries, was tested in two phases. A clinician monitored at bedside the patients' recovery and verified clinically abnormal physiological states. After the first test with ten patients, the system's rulebase was upgraded and then tested with an additional 15 patients. The alarm system employed two PC/ATs and was programmed to give notice of four pathological states (hyperdynamic state, hypovolemic state, hypoventilation and left ventricular failure) at two levels of urgency (alarm and alert levels). The monitoring lasted 5.4 +/- 1.7 hours per patient (mean +/- S.D.), totalling 134.7 hours. The system alarmed 27 times during the first and 73 times during the second phase of the testing. The sensitivity of the alarms was 100% in both phases, and the specificities increased from 20.0% to 73.9% and from 59.1% to 70.0% for the alarms and the alerts, respectively. This computerized decision support system based exclusively on data available in the automatically collected data base had a low false positive rate and gave early warnings about pathological states in the homogeneous group of adult postoperative cardiac patients.