These sleep stages, gauged by brain-wave activity, remain the cornerstone of sleep studies, which also monitor changes in heart rate, temperature and muscle tone. Besides the six EEG electrodes attached to a patient’s head to measure brain waves, two electrooculograms detect the rapid eye movements of REM. Three electromyogram leads at the chin register the decreased muscle tone of REM, and two on each leg check for periodic limb movements in sleep (PLMS), a disorder characterized by repeated jerky movements that can mar the quality of sleep. Two electrocardiogram leads on the chest track the slow, even heartbeats of NREM and the more rapid, irregular ones of REM. For detecting potential breathing disorders, there’s a microphone to record snores, a pneumotachograph in the nose to measure airflow, two piezoelectric bands around the chest and abdomen to register the force of inhalation, and a meter taped to a finger to record whether oxygen saturation falls following breathing disruptions.

Although often less than pleasant for subjects—some swear they didn’t get a lick of sleep, though the monitors prove otherwise—polysomnography provides valuable diagnostic evidence. But some disorders don’t register. For example, people are frequently referred for sleep studies because of fatigue and excessive daytime sleepiness, explains Amit Verma, director of the divisions of neurophysiology and sleep disorders at the Methodist Neurological Institute in Houston. These aren’t trivial complaints—they make people prone to accidents and reduce quality of life. The first guess at a cause is usually sleep apnea, in which oxygen loss due to missed breaths rouses the sleeper as many as 100 times per hour. Apnea typically leaves a dramatic footprint on an EEG, showing patterns of wakefulness, and there’s usually a clear drop in oxygen saturation, among other telltale signs. Yet the sleep studies of some patients who feel perpetually sleepy look relatively normal.

One big problem with sleep research based on electroencephalography, according to Terrence Sejnowski, a Howard Hughes Medical Institute investigator at the Salk Institute for Biological Studies in La Jolla, Calif., is that even today sleep technicians must score EEGs manually, looking at one 30-second period, or epoch, at a time to determine that, say, a certain epoch looks like NREM stage 2. Sejnowski thinks this manual system misses many nuances that an automated system could catch. For example, Verma suspects that upper airway resistance syndrome, or UARS—essentially a milder form of sleep apnea—may be the real problem for some patients who complain of daytime sleepiness. Yet the brain-wave and breathing disturbances caused by UARS, though they may cause complications, are fleeting and don’t leave the telltale footprint that apnea does on most EEGs and other traditional sleep study measures. They may go undetected during standard scoring of EEGs.

Scientists have tried for years to automate EEG scoring, but there has always been a fatal flaw, Sejnowski says. “Previous computer systems have relied on expert EEG scoring,” he says, “but experts only agree 75% of the time.”

So Sejnowski, a neuroscientist who develops computer models that teach themselves to recognize patterns in complex signals, and graduate student Philip Low created an algorithm for analyzing low- and high-frequency EEG patterns during sleep, building in much shorter epochs for finer resolution. “When we started, we had no idea what we’d find,” Sejnowski says. “But we were amazed to see three clear clusters emerge. They’re different from traditional sleep stages, though they actually correspond pretty well to those stages.”

The first pattern is an intermediate stage between wakefulness and sleep that corresponds to traditional NREM 2. In the second, slow-wave sleep occurs as in the newly consolidated NREM 3 (now called N3), and in the third, REM appears. (They found no NREM 1, which Sejnowski believes has been a catch-all for epochs that contain abrupt transitions between stages.)

In addition to giving a clearer picture of sleep stages, the computer-analyzed EEGs also revealed information that manually scored tests miss entirely. To his surprise, looking at the new graphs, Sejnowski found that both slow-wave sleep and REM are fragmented, with the brain shifting between different wavelengths every two to three seconds. These shifts occur in the low-frequency range during REM and in high frequencies during SWS. That fragmentation could prove to be clinically important, providing diagnostic clues.