The data of the early self-experiment nights were useless for any sort of advanced statistical analysis, since the experimental settings, the recording data structure and the stimuli tried out changed steadily due to continuous development of the software and methods. Nevertheless, there were still several (rather subjective) experiences gathered and lessons learned. The most important of these can be summarized as follows:
Many of the recorded nights revealed bugs and malfunctioning of the software, for example the need to not store the night’s sleep data in RAM but on hard disk. Storing the data in RAM led to severe performance issues especially at the end of the night (the data alone reach magnitudes of up to a Gigabyte for one night). Moreover, Windows power options and Windows updates shutting down the computer after some hours or program crashes during the night led to data loss, if data were stored in RAM.
The time-frequency plot of Zeo’s EEG signal showed distinct patterns, corresponding to the Zeo’s sleep stage classification. Taking also sleep stage features of the AASM manual (Iber & American Academy of Sleep Medicine, 2007) into consideration, as for example large delta wave activity for N3 sleep or sleep spindle activity for N2 sleep, sleep stages can be classified with the Zeo’s raw data at least up to some degree. A more detailed analysis including comparison to parallel polysomnographic recordings are describes below.
An idea which came up during the early self-experiments was to simplify the sleep communication process by excluding the REM detection step and to play the stimuli during the whole night instead. During the early self-experiments this option was tested with visual light stimuli (the LED half-circle), and, later, with acoustic stimuli. The effect observed was that the sleep quality suffered dramatically due to the following reasons: a) falling asleep with ongoing stimulus presentation is much harder, b) stimulus generation leads to numerous awakenings during the night, especially during the light sleep periods and after micro-awakenings when turning around. As a result, continuous stimuli played throughout the night are not considered to lead to advantages for sleep communication.
Starting stimulus generation immediately after REM detection led to the problem that the dreams of the sleeper were just about to start and no complete dream scenery was built up in which the stimuli could be incorporated. Thus, a waiting period before stimulus generation was found to be useful.
Using auto-increasing stimulus intensity leads to the problem that the intensity increases until the sleeper wakes up. This problem can be solved by implementing a stimulus control for example with eye movements with which the sleeper can modify the stimulus intensity from within the dream.
In order to prevent dreaming of a non-existent stimulus and thinking it was a real incorporation, a random factor could be used in order to make it impossible to know beforehand what message is sent into the dream world and to be able to distinguish between made-up and real incorporations. One possible solution: random, computer generated numbers transformed into stimuli via Morse code make it impossible to know beforehand which number to expect. Thus, if dreaming of a stimulus and the dream report reveals another number as generated by the computer, sleep communication was not successful. This idea can be extended in such a way that the sleeper directly answers the random number back to the wake world using eye movements, so that is becomes clear that the sleeper knew about the incorporation and identified it as being a message from the wake world. Another further extension would be to include a simple transform of the stimulus, so that the response of the sleeper is improbable to be automatic and subconscious due to extensive training during wake state. This idea was used in the further experiments (see below).