Purpose: An approach to the problem of seizure prediction aimed to provide a computationally effective method for real-time application with standard scalp EEG recordings is presented. Methods: One control record performed during steady interictal interval and 10 preseizure records lasting 15-20 minutes with standard scalp EEG 32 channels from 3 subjects suffering from pharmacoresistant medial temporal lobe epilepsy have been used. A disharmony index based on multiple abrupt changes of EEG spectral features is introduced to characterise the pre-ictal phase. The amount of changes into the time unit is considered as a predictor. A new computationally effective on-line algorithm to solve the basic problem of single change detection is proposed. Control records were used to estimate the parameters of the algorithm for each patient and preseizure records were used to evaluate the method. Results: We observed a reliable prediction of 30 seizures for 3 patients with an average anticipation time of 10.5, 12.6, 16.3 minutes (with respective variances equal to 3.1, 4.1, 2.6). The anticipation times ranged between 7.5 and 21 minutes. Only 1 false positive occurred in the control records. Conclusion: The proposed algorithm offers a computationally effective method to be applied in real time applications. Additional efforts are oriented towards decreasing the probability of false positive and further evaluation is under way.