In traditional methods of spike detection, spikes are defined in absolute terms (duration, amplitude) or relative to a few seconds of background. These methods result in many false positive detections during long-term epilepsy monitoring because of numerous artefacts and non-epileptic transients. To reduce significantly false detection, we propose to render spike detection sensitive to the state of the EEG. We thus defined 5 states (active wakefulness, quiet wakefulness, desynchronized EEG, phasic EEG and slow EEG) and designed a method for automatic state classification. We then designed procedures for identification of non-epileptic transients (eye blinks, EMG, alpha, spindles, vertex sharp waves). These procedures are to be applied only in the state in which they are likely to occur (e.g., eye blinks in wakefulness). We present preliminary results from 14 recordings each lasting 100 min, which indicate a state classification reliability of 85-90%, reduction in false detection of 65-90% if state classification were perfect; true spikes lost as a result of these procedures were under 5%. These results are encouraging and validate the concept of a spike detection system which analyses a wide temporal and spatial context before deciding the significance of a wave form.