ExpenSystems WtthApphcatton~. Vol 6, pp 399-409, 1993 0957-4174/93 $600 + 00 Pnnted m the USA © 1993Pergamon PressLtd Information Processing Models for Sleep Staging JOSE C. PRINCIPE Umverslty of Honda, Gamesvllle,FL TAE-GYU CHANG Chung-Ang University, Seoul, Korea SUNIT GALA UmSQL, Inc., Austin, TX ANA TOME INESC/Umversmdadde Avelro, Avelro, Portugal Abstract--Three dtfferent methods of automated sleep staging are described and compared in the paper The interesting aspect of the comparison ts that the inputs to the three dtfferent Information processing models (an expert system, a behef automaton, and a neural network) are the outputs of the same frontend processor that extracts EEG features We found out that the level of man-machme agreement ts very stmdar among the systems (around 85%) approaching the level of agreement required among human experts However, the stmdanty of performance m such a dlverstfied set of approaches pomts out the need for enhancing the accuracy of the first level of analysts (EEG feature extraction). 1. INTRODUCTION THIS PAPER describes and compares three different models of information processing applied to the prob- lem of automated sleep staging. Sleep has been cate- gorized in stages by Rechtschaffen and Kales (R + K) (1968). Although sleep staging has its drawbacks (in essence any taxonomy overlooks minor differences), it is unquestionable that the rise of sleep as a scientific discipline was largely due to the R + K sleep scoring criteria. Sleep scoring requires the visual examination of at least one channel of electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG), and classifies each 30 seconds (or 60 seconds) in one out of six stages: Stage 0 or awake, Stage 1 (light sleep), Stage 2, Stages 3 and 4 (deep sleep), and Stage 5 or Requests for repnnts should be sent to J C Pnnclpe, Electrical En- gmeenng Department, Umversltyof Honda, Gamesvllle,FL 32611 rapid eye movement (REM) sleep. The definitions of each sleep stage are contained in a profusely illustrated manual, and they represent a coherent but incomplete set of rules. The difficulties in applying the R + K definitions are several: First, the definitions are many times imprecise (how to measure the period of a wave?) and are based upon visual waveform patterns that are difficult to capture with automated methods (REM, k- complexes). Second, electrophysiological variables comprise, basically, an unconstrained set of signals, so the manual does not cover all possibilities, and the interpretation of the patients EEG/EOG/EMG is left ultimately to the clinician. Therefore, each sleep lab- oratory pursued its own interpretation. Rechtschaffen and Kales' rules are therefore more of a guide that the type of strict definitions found in the exact sciences. In the early 1970s computers were heralded as the technology that could bring quantification to electro- encephalography. Each computer program contains intrinsically the definitions utilized to arrive at the out- 399