To evaluate the clinical efficacy and mechanisms underlying EEG feedback training of epileptic patients, 5 adult patients with poorly controlled seizures were studied for 4--10 months during which quantitative analysis of seizures, the EEG, and serum anticonvulsant levels was conducted. Sustained seizure reduction did not occur during the first 4--5 weeks in which feedback signals were presented randomly in relation to the EEG. When feedback was then made contingent upon central 9--14 c/sec activity, seizures declined by 60% in 3 patients. Power spectral analysis showed upward shifts in EEG frequency, decreases in abnormal slow activity, and enhancement of alpha rhythm activity as a function of contingent training, but no specific EEG change was associated with seizure reduction in all patients. No evidence was obtained for the hypothesized involvement of a 'sensorimotor rhythm' or motor inhibition in the training effects. The lack of effect in two patients could not be attributed to insufficient training, lack of motivation, or to differences in seizure classification. A second phase of research showed that continued laboratory training was both sufficient and necessary for maintaining clinical and EEG effects. Results indicate that: (1) significant seizure reductions can occur with EEG feedback training which are not related to placebo effects, non-specific factors or to changes in medication; (2) EEG changes associated with such training can best be described as 'normalization'; (3) continued clinical investigation of EEG feedback training as a non-pharmacological adjunct to conventional therapy appears justified.