The program, called ARK (Automated Rechtshaffen-Kales) incorporates novel waveform extraction techniques in parallel with more conventional period/amplitude and spectral approaches, and runs in two independent modes. Event Collection traverses a synchronized multi-channel EEG/EOG/EMG data set and abstracts to an "event file" both discrete events (transient waveforms such as K complexes, sleep spindles and saturation artifacts), and statistical information from each epoch: characteristic voltages, frequencies and durations. Identified discrete events are optionally displayed for analyst review on-screen as graphs "clipped" from the data. During a typical session there may be a hundred or so waveforms identified, as Stage 2 sleep contains a K complex or spindle every minute or so, and Stage 2 is typically 20-25% of a session. As data artifacts can be interpreted as having meaning in sleep studies, but may also be simply due to amplifier saturation, they are treated as discrete events and may be individually reviewed.

Conventional time- and frequency-domain techniques are used to form many commonly measured sleep parameters (55 at last count), and to guide the program toward data segments potentially containing characteristic waveforms and away from featureless or ill-suited stretches of data. Event Collection for a full night session takes about 45 minutes on a Pentium 166 PC. Staging analysis of collected data read back from an event file takes less than 5 minutes.

During Epoch Formation, events collected in the first phase are read from the event file, sorted into epochs and correlated, then used according to rules in Rechtschaffen & Kales and other sources to construct a sleep stage diagram. For each epoch, weights which pivot about unity are assigned to each decision (e.g. "stage 1 EMG amplitude usually below stage W"). Individual weights are multiplicatively combined in a "continuous (fuzzy) logic" manner to form a stage weight factor:

weight = 1 + (value - nominal) / reference.

In the stage 1 EMG decrease example, the value is the epoch mean EMG voltage, the nominal is that value from the preceding waking stage, and the reference is the session-long standard deviation of EMG mean voltage:

wEMG_1 = 1 + (vEMG_1 - vEMG_W) / vSD_EMG.

After all criteria have been evaluated, a stage decision is reached subject to additional rules governing transitions between stages, some of which are explicitly given in R&K. All parameter and weight calculations can also be made on a sub-epoch basis, where there are (by default) 5 sub-epochs per epoch. Several multiplicative factors (with unity default values) unstated here are also programmed, but not shown for this report, to allow tuning the algorithms against other standards or for individual preferences.

Over fifty plots of possible interest in sleep studies are optionally produced and can be manipulated (sized, smoothed, filtered, clipped, etc.) in dozens of ways, printed or saved as data files or bitmaps for inclusion in documents. Epoch Formation can be run as a script command in only a few minutes from any saved event file, so that the raw data need not be accessed again to produce a stage analysis.

An event can be the identification and extraction from raw data of an entire K complex or sleep spindle, or simply a few contiguous cycles of slow-wave EEG. REMs and blinks are dual-channel simultaneous EOG events, and Movement Arousals are EMG events concurrent with corresponding EEG or EOG changes. Low-Voltage Mixed-Frequency intervals are derived from Fast Fourier Transform spectra over short sections of data in the 2-7 Hz band, whereas High-Amplitude Slow Waves derive from Period/Amplitude Analysis as the band is too narrow (<2 Hz) and too near the origin for safety and utility. Altogether, at this time there are 23 classes of events which enter into the stage scoring process. There are about 50,000 events of all types in the event file collected from the tutorial data set supplied as part of this project.

ARK runs analyses of sleep sessions from an editable script file (a text editor is part of the program). In the script file, the analyst specifies data files, their locations and characteristics, and lists "templates" to be used for waveform extraction. Templates are short data segments "clipped" from real data (by dragging the mouse across a section of an EEG graph, for example) which contain a single waveform, like a K complex, spindle, sawtooth or artifact. A mathematical algorithm of considerable flexibility and robustness finds all occurrences of similar events as the program traverses the data. The basic algorithm, Fast Orthogonal Search (FOS) was developed by Canadian researcher Michael J. Korenberg, about a decade ago. It was powerful but slow on the era's machines, and the periodicities it sought to find in data had to be specified in advance. ARK now automates the selection of candidate periods as a recommended option.

Complete, visually scored analysis results (sleep hypnograms) and the raw data files were supplied by the collaborating centers, both on media (Iomega ZIP Drive disks) and in printed form. Where numerical tables of scoring results were unavailable, an image processing program written for the purpose by CoDebris enabled scanning and automated digitization of printed sleep stage diagrams for internal comparison with the computer-generated sleep stage diagram.