Automated sleep stage scoring using hybrid rule- and case-based reasoning.
ABSTRACT We propose an automated method for sleep stage scoring using hybrid rule- and case-based reasoning. The system first performs rule-based sleep stage scoring, according to the Rechtschaffen and Kale's sleep-scoring rule (1968), and then supplements the scoring with case-based reasoning. This method comprises signal processing unit, rule-based scoring unit, and case-based scoring unit. We applied this methodology to three recordings of normal sleep and three recordings of obstructive sleep apnea (OSA). Average agreement rate in normal recordings was 87.5% and case-based scoring enhanced the agreement rate by 5.6%. This architecture showed several advantages over the other analytical approaches in sleep scoring: high performance on sleep disordered recordings, the explanation facility, and the learning ability. The results suggest that combination of rule-based reasoning and case-based reasoning is promising for an automated sleep scoring and it is also considered to be a good model of the cognitive scoring process.
- SourceAvailable from: R. Lengelle[show abstract] [hide abstract]
ABSTRACT: We describe an approach to automatic all-night sleep analysis based on neural network models and simulated on a digital computer. First, automatic sleep stage scoring was performed using a multilayer feedforward network. Second, supervision of the automatic decision was achieved using ambiguity rejection and artifact rejection. Then, numerical analysis of sleep was carried out using all-night spectral analysis for the background activity of the EEG and sleep pattern detectors for the transient activity. Computerized analysis of sleep recordings may be considered as an essential tool to describe the sleep process and to reflect the dynamical organization of human sleep.Computers and Biomedical Research 05/1993; 26(2):157-71.
- [show abstract] [hide abstract]
ABSTRACT: Manual analysis of sleep, breathing, and oxygenation records is the "gold standard" for diagnosing sleep abnormalities but is time consuming and cumbersome. The accuracy and cost of a computerised sleep analysis system have therefore been investigated. Manual and computerised (CNS Sleep Lab) scores from 43 consecutive clinical sleep studies were prospectively compared for accuracy and the time and costs were recorded. There were good correlations and no systematic differences between manual and computer scoring for total sleep time, sleep onset latency, and duration of REM sleep. There was a small but clinically insignificant systematic difference in breathing pattern analysis, the number of hypopnoeas/hour being lower with manual than with computer scoring (13 (SE 3) v 15 (SE 3)/hour). There was no difference between computer and manual scoring of the frequency of apnoeas, so the frequency of apnoeas + hypopnoeas was clinically insignificantly higher with computer scoring with a highly significant correlation between the two techniques. The time taken to perform the analyses was not different between the two methods (manual 83 (SE 8) v computer 86 (SE 8) minutes). The computer system was six times more expensive than the manual system and annual running costs, including full maintenance contract and 15% depreciation, were twice as great. The CNS Sleep Lab is sufficiently accurate for use in clinical sleep studies but is significantly more expensive and does not save technician time.Thorax 04/1993; 48(3):280-3. · 8.38 Impact Factor
- [show abstract] [hide abstract]
ABSTRACT: Since its discovery some 50 years ago, the electroencephalogram (EEG) has formed the basis for the classification of sleep into several stages. Such classification has been laboriously performed by visual examination of the EEG and related signals, or, more recently, by automated techniques. Both visual scoring and most automated analyses rely on the application of a predefined set of rules. The authors have proposed a method of sleep analysis using a Kohonen self-organising feature map, that requires no such application of rules and aims to give some indication of the dynamics of sleep in humansRadar and Signal Processing, IEE Proceedings F 01/1993;
Computers and Biomedical Research 33, 330–349 (2000)
Article ID cbmr.2000.1549, available online at http://www.idealibrary.com on
Automated Sleep Stage Scoring Using Hybrid Rule- and
Hae-Jeong Park,*,† Jung-Su Oh,* Do-Un Jeong,‡ and Kwang-Suk Park§
*Interdisciplinary Program of Medical and Biological Engineering Major,
†Institute of Biomedical Engineering, ‡Department of Psychiatry, and
§Department of Biomedical Engineering, College of Medicine,
Seoul National University, 28 Youngon-Dong Chongno-Gu, 110-799, Seoul, Korea
Received December 1, 1999
We propose an automated method for sleep stage scoring using hybrid rule- and case-
based reasoning. The system first performs rule-based sleep stage scoring, according to the
Rechtschaffen and Kale’s sleep-scoring rule (1968), and then supplements the scoring with
case-based reasoning. This method comprises signal processing unit, rule-based scoring unit,
and case-based scoring unit. We applied this methodology to three recordings of normal sleep
and three recordings of obstructive sleep apnea (OSA). Average agreement rate in normal
recordings was 87.5% and case-based scoring enhanced the agreement rate by 5.6%. This
architecture showed several advantages over the other analytical approaches in sleep scoring:
high performance on sleep disordered recordings, the explanation facility, and the learning
ability. The results suggest that combination of rule-based reasoning and case-based reasoning
is promising for an automated sleep scoring and it is also considered to be a good model of
the cognitive scoring process.
? 2000 Academic Press
Sleep is not a uniform biological state. It consists of rapid eye movement (REM)
sleep and nonrapid eye movement (NREM) sleep. NREM sleep can be further
classified into stages 1, 2, 3, and 4 according to the current sleep scoring standard
proposed by Rechtschaffen and Kales (1). Sleep stages are scored to all 30-s
epochs during the sleep based on the characteristics of electrophysiological signals
including electroencephalogram (EEG), electrooculogram (EOG), and chin electro-
Stage W (wakefulness) is characterized by EEG alpha activities that appear in
more than 50% of the epoch. Stage 1 is characterized by the low voltage mixed
frequency EEG without rapid eye movements. Stage 2 is defined by the presence
(a high-amplitude biphasic wave of at least 0.5 s duration, an initial sharp positive
voltage followed by a negative deflection slow wave). Stages 3 and 4 are character-
ized by slow delta waves of EEG. Slow delta wave is defined as EEG activity
slower than 2 Hz with peak-to-peak amplitude greater than 75 ?V. If the portion
Copyright ? 2000 by Academic Press
All rights of reproduction in any form reserved.
AUTOMATED SLEEP STAGE SCORING
of slow delta waves in an epoch is 20–50%, the epoch is scored as Stage 3, and
if it is higher than 50%, it is scored as Stage 4. Stage REM is defined by the
relatively low voltage, mixed frequency EEG in conjunction with episodic rapid
definitions of each sleep stage, Rechtschaffen and Kales’ standard proposes several
smoothing rules in order to score obscure epochs that cannot be dealt with by the
above definitions. Smoothing rules are based on the assumption that sleep state has
the tendency to persist through epochs. These rules refer to the electrophysiological
context between epochs prior and posterior to the target epoch. For example, if
there is no spindle, no K-complex, and no other events in the target epoch, then
smoothing rules propose to look for events at prior and posterior epochs and if
spindles are found, then the epoch is scored as Stage 2. This kind of contextual
information is essential in practical scoring. An exemplary display of one-epoch
polysomnographic signals is shown in Fig. 1. One epoch of 30 s, including EEG
(C3/A2 and O2/A1), EOG (left and right), and chin EMG is scored as one of six
sleep stages (Stage REM, Stage W, Stage 1, Stage 2, Stage 3, and Stage 4). In
this figure, sleep spindles and K-complexes are found and this epoch is scored as
Numerous attempts have been made to automate sleep stage scoring. They are
largely based on analytical procedures extracting certain features from EEG, EOG,
and EMG and classifying these features into one of the sleep stages. These ap-
proaches include thresholding the spectral power of the frequency bands (2,3),
(4) and artificial neural networks (5–7), and waveform detection by various pattern
recognition algorithms (8–14). Some groups applied rule-based reasoning methods
in order to include the contextual information (15,16).
According to their results, the agreement rate between automated scoring and
manual scoring is 75–85% in recordings of normal young adults (17–21). In sleep-
disordered cases, where sleep signals are more irregular with various artifacts and
physiological signals including EEG (C3/A2 and O2/A1), EOG (left and right), and chin EMG is
scored one of six sleep stages (Stage REM, Stage W, Stage 1, Stage 2, Stage 3, and Stage 4). In this
figure, sleep spindles (oscillating wave with the frequency between 12–15 Hz) and K-complexes
(high voltage, sharp rising, and sharp falling wave) are found and this epoch is scored as Stage 2.
An exemplary display of one epoch polysomnographic signals. One epoch of 30s electro-
PARK ET AL.
innate complexity of the disturbed sleep, the agreement rate is reported somewhat
lower to the level of 65–75% (22–24). We considered that the unsatisfactory
agreement rates of these algorithms, especially in the sleep-disordered cases, pri-
marily resulted from limited usage of the contextual information.
To overcome this limitation, we analyzed the human approach to sleep scoring.
they often use heuristics and memories from past cases. These two complementary
aspects motivated us to integrate rule-based reasoning (RBR) and case-based rea-
Rule-based reasoning and case-based reasoning have been developed on the
different concept and architecture (25–27). Rule-based expert system is the most
common form of the knowledge-based expert system that consists of knowledge
base, working memory, and inference engine (25). The knowledge base contains
the domain knowledge represented in the form of IF THEN type rules. Inference
engine makes inferences by investigating rules, according to the priorities of the
rules in the knowledge base that are satisfied by facts in working memory. Rule-
based reasoning system solves problems by applying previously established rules
to the given problem. The advantage of rule-based reasoning is easy to modify
and add rules in knowledge base due to the separation of knowledge base and
inference engine. However, this system has difficulties in representing human
knowledge and does not benefit from experience with their use nor learns from
Case-based reasoning is a relatively new field of artificial intelligence (26–29).
Instead of applying rules to the problem, case-based reasoning system solves
problems by making use of solutions to the previous problems. Case is generally
represented as a problem and solution pair and case-base contains previous prob-
lem–solution pair cases. The reasoning procedure is categorized into the following
four cycles: case retrieve, reuse, revise, and retain. When a new problem is inserted,
thereasoning systemretrievesasimilar caseinthe case-baseandreuses thesolution
of the retrieved case. If the retrieved case does not match enough, the solution is
revised to adapt to the given problem situation by repair rules. If the solution is
still not adequate, human intervention may be required to modify the solution and
make a new case and retain to the case-base for future use. Case-based reasoning
system can solve the problem without abstraction of rules from various situations
and without complete knowledge of the application domain. The learning facility
and explanation facility by previous cases are important advantages of this system.
However, the performance is dependent on the representation of cases, similarity
measure, and searching methods.
The application areas of RBR are closed, narrow, and well understood enough
to be covered by rules, while CBR is appropriate for areas in which knowledge
of the domain has difficulty in being represented in the form of rules, but easily
accumulated by experiencing more solved problems (26,29,30). Because of the
above different but complementary characteristics, integration of CBR with RBR
has been researched in different tasks such as classification (31) and diagnosis
(32,33), using different architectures. The usual integration is to use RBR to deal
AUTOMATED SLEEP STAGE SCORING
with knowledge in a standard situation and to use CBR to deal with the problems
that are not covered by rules and which require past problem solving experiences
(29,31,32). In these cases, RBR is usually applied first. When RBR fails to provide
a reliable solution, CBR looks for past cases that best match the current problem.
The area of sleep stage scoring is thought to require hybrid reasoning because
human experts use both rule-based knowledge and experiences. Therefore, our
system adopted the hybrid RBR-first-CBR-last approach to analyze sleep stages.
We decided to use rules to generate a strict sleep stage scoring and cases to deal
with details and exceptions to the rules caused by sleep complexity. To implement
this approach, we designed a three-unit architecture, including a signal processing
unit (SPU), a rule-based scoring unit (RBSU), and a case-based scoring unit
(CBSU). Figure 2 illustrates this architecture. From the 30-s polysomnographic
signals, SPU eliminates artifacts, extracts features from background signal activi-
ties, and detects special events. RBSU applies basic standard rules to the facts that
are composed of features and events derived from SPU. If RBSU fails to reason
out above the confidence threshold, then CBSU executes additional scoring.
DATA PREPARATION FOR SLEEP STAGE SCORING
The purpose of SPU is to extract features from the electrophysiological signals
that will be used as materials for the next reasoning processes. We applied the
following signalprocessing algorithms forartifact rejection, featureextraction from
background signal activity, and event detection.
Polysomnographic recordings have various artifacts such as ECG interference in
EEG, harmonic noises, and drifts caused by sweat. These artifacts are automatically
measured and rejected if the levels of artifacts are above rejection thresholds.
Removing ECG interference to EEG. QRS of ECG regions are detected, and
correlation coefficients between EEG and ECG at QRS region are calculated to
decide whether to apply ECG rejection algorithm or not (34).
functional units. From the 30-s polysomnographic signals, SPU rejects artifacts, extracts features from
background signal activities, and detects special events. RBSU applies basic standard rules to the
facts (features and events) derived from SPU. If RBSU fails to reason out acceptably, i.e., the overall
reliabilityvalue (whichwillbe discussedinthe followingsection)of reasoningislower thanpredefined
threshold and then CBSU executes additional scoring.
Architecture of the automated sleep scoring engine: SPU, RBSU, and CBSU are major
PARK ET AL.
Removing harmonic noises. Adaptive notch filtering is performed when 20
or 60 Hz harmonic noises are continuously detected on the power spectrum of
Removing drift caused by sweat. Low frequency band filter is applied to check
drifts and then high pass filtering is applied if drifts exist.
Feature Extraction from Background Signal Activities
Among various features selected, the most representative features to characterize
the background activities are (1) wave segment of EEG; (2) LPC (linear predictor
coefficients) of EEG; (3) state of EOG; and (4) tone of chin EMG.
Wave segment of EEG. The overall band powers during one epoch 30 s have
been used as main features in several papers (3,5). However, these features give
no precise information on the temporal distribution of the waves of different
frequency bands during an epoch. This temporal distribution of the waves is
important in sleep stage scoring, because many standard rules are defined by the
waveform duration and distribution. For example, if the total duration of alpha
wave is prominent by more than 50%, then the epoch is Stage W, according to
Rechtschaffen and Kales’ rule. In order to include the temporal wave distribution,
one epoch is segmented by the duration of 1 s and Hamming windowed FFT is
calculated to assign indexes that correspond to the dominant frequency band of
the segment. Assigned indexes are denoted as delta (0.5–2.5 Hz), theta (3–7 Hz),
alpha (7.5–12 Hz), beta1 (12–20 Hz), beta2 (20–50 Hz), and spindle (11.5–15 Hz).
LPC of EEG. Total power spectrum of one epoch EEG is estimated using MEM
(maximum entropy method) by calculating LPC.
State of EOG. The state of EOG is classified into one of five classes: SEM
(slow eye movement), DRIFT, DELTA, QUIET, and NORMAL. To classify EOG
state, three types of parameters are calculated: total spectral power between fre-
quency 0.15 and 0.45 Hz, total spectral power between frequency 0.5 and 1.2 Hz,
and correlation coefficient between left and right EOG.
Stage REM from Stage W. By referencing the tone level during biocalibration and
the lowest tone level of whole recording, the threshold of EMG tone is determined.
The tone of each epoch is the average value of every half-second EMG variance
between interquartile ranges. The interquartile range is adopted to reject the burst-
type high activities.
Current sleep stage scoring is largely dependent on the typical waveform events
such as spindle, K-complex, and rapid eye movement. Arousal is also an important
event that shows transition of stages. To detect EEG events, a band pass filter
bank (the frequency band is 0.5–2.5, 3–7, 7.5–12, 12–15, and 16–30 Hz,) is
Sleep spindle. The wave of which the power of 12–15 Hz band is relatively
high is regarded as spindle.