Fuzzy detection of EEG alpha without amplitude thresholding

Fuzzy detection of EEG alpha withoutamplitude thresholding

Eero Huupponena,b,*, Sari-Leena Himanenc, Alpo Varrib,Joel Hasanc, Antti Saastamoinenb, Mikko Lehtokangasa,

Jukka Saarinena

aDigital and Computer Systems Laboratory, Tampere University of Technology,

Hermiankatu 12 C, P.O. Box 553, FIN-33101, Tampere, FinlandbSignal Processing Laboratory, Tampere University of Technology,

Hermiankatu 12 C, P.O. Box 553, FIN-33101, Tampere, FinlandcDepartment of Clinical Neurophysiology, Tampere University Hospital,

P.O. Box 2000, FIN-33521, Tampere, Finland

Received 8 March 2001; received in revised form 25 June 2001; accepted 25 June 2001

Abstract

Intelligent automated systems are needed to assist the tedious visual analysis of polygraphic

recordings. Most systems need detection of different electroencephalogram (EEG) waveforms. The

problem in automated detection of alpha activity is the large inter-individual variability of its

amplitude and duration. In this work, a fuzzy reasoning based method for the detection of alpha

activity was designed and tested using a total of 32 recordings from seven different subjects.

Intelligence of the method was distributed to features extracted and the way they were combined.

The ranges of the fuzzy rules were determined based on feature statistics. The advantage of the

detector is that no alpha amplitude threshold needs to be selected. The performance of the alpha

detector was assessed with receiver operating characteristic (ROC) curves. When the true positive

rate was 94.2%, the false positive rate was 9.2%, which indicates good performance in sleep EEG

analysis. # 2002 Elsevier Science B.V. All rights reserved.

Keywords: EEG alpha; Automatic detection; Fuzzy reasoning

1. Introduction

Analysis of polygraphic sleep recordings is still carried out visually today because the

expert knowledge of neurophysiologists and sleep technicians has not yet been transferred

Artificial Intelligence in Medicine 24 (2002) 133–147

* Corresponding author. Tel.: þ358-3-365-3858; fax: þ358-3-365-3095.

E-mail address: [email protected] (E. Huupponen).

0933-3657/02/$ – see front matter # 2002 Elsevier Science B.V. All rights reserved.

PII: S 0 9 3 3 - 3 6 5 7 ( 0 1 ) 0 0 0 9 8 - 7

well enough to computer analysis due to inter-subject variability in different waveforms

and the complexity of expert knowledge involved. Sleep is visually quantified into the

following stages, wake (W), rapid eye movement (REM), light sleep (S1), deep sleep (S2,

S3, S4) in epochs of 20 or 30 s based on the waveforms in the signals [15]. Visual analysis

of the recordings requires a considerable amount of effort. Therefore, several attempts have

been made to automate this process [4].

Artificial intelligence may be applied at least on two levels in the analysis of polygraphic

recordings. The primary level consists of the analysis (or detection) of waveforms on single

channels of the recording, like the electroencephalogram (EEG), the electromyogram

(EMG) or the electro-oculogram (EOG). The secondary level is the combination of the

outputs from primary level detectors. This study concentrates on the primary level analysis

of the EEG alpha waves.

Alpha activity (8.0–12.5 Hz) in the EEG is an important indicator of wakefulness [18].

In alert wakefulness with eyes open no alpha is observed. After eye closure it appears over

the posterior regions of the scull with runs of varying durations. During drowsiness and

sleep onset alpha becomes more diffusely distributed showing also in central and frontal

areas [2]. Alpha activity amplitudes vary considerably from moment to moment and from

individual to individual [14]. This variability is increased by age-related changes, for

instance, a decrease in alpha amplitude with increasing age [1]. There may even be gender-

dependent changes in alpha amplitude in relation to drowsiness [18]. The variability of

amplitude is a problem for automated alpha detection. Alpha detection is very challenging

also due to the variability of the durations of alpha segments, during which alpha activity is

burst-like, with waxing and waning waves.

Time domain properties can be used to characterise the EEG patterns in terms of

amplitude, time scale and complexity [5]. Parametric methods that fit a mathematical

model to EEG data have been applied to detect transients in EEG, such as spikes and sharp

waves [10]. In a computer-based model different EEG segments were characterised by

components of power spectrum, such as power on delta, alpha and beta bands and also the

corresponding frequencies [23]. Absolute power values were used in that model and that

may cause difficulties in some applications. Non-parametric methods, on the other hand,

assume no model for the EEG. Only stationarity during short segments is assumed. A

simple non-parametric method is period (or interval) analysis but it may be sensitive to

noise and other artefacts and it is mostly being replaced by fast Fourier transform (FFT)

based analysis [10].

Rather few automated methods for alpha detection have been proposed. A model-based

approach [11] and band-pass filtered alpha amplitude combined with period analysis have

been used [20,21]. Clear guidelines as how the threshold should be selected for analysis of

all-night recordings are usually not given. In any case, a fixed amplitude threshold is

undesirable because of the intra and inter-subject variability in alpha amplitudes. Short

reference measurements prior to the long-term recordings were used in another work to

adapt to the subject specific alpha amplitude levels [22]. In spindle studies, alpha amplitude

has been compared to spindle amplitude to avoid false spindle findings [3,19].

In the present work, a fuzzy reasoning based method for the detection of alpha activity

was designed and tested. The initial approach in waveform detection has previously been

applied to spindle detection [7]. The method used ratios of amplitudes on different EEG

134 E. Huupponen et al. / Artificial Intelligence in Medicine 24 (2002) 133–147

bands instead of absolute amplitudes in order to make it robust against inter-subject

variability in EEG amplitudes. The advantage of the presented detector, is that no alpha

amplitude threshold needs to be selected prior to the analysis of EEG traces. The presented

method aims at detecting well runs of alpha of any duration.

2. Methods

2.1. Subjects and recordings

Multiple Sleep Latency Test (MSLT) [8,16] recordings made at Tampere University

Hospital were studied in this work. MSLT is a daytime polygraphic study widely used in

clinic practice to quantify daytime sleepiness. Subjects were instructed to go to sleep at

1000, 1200, 1400 and 1600 O’clock. Test duration was 20 min, but if sleep occurred, the

test was continued for 15 min and thus the maximum duration of one nap was 35 min. Six

EEG channels (C3-A2, C4-A1, Fp1-A2, Fp2-A1, O1-A2, O2-A1), two EOG channels (P8-

A1, P18-A1) [9] and two EMG channels (mental, tib.ant.) were recorded with a 200 Hz

sampling rate using a digital polysomnograph (Embla1, Flaga Inc.). The data were

converted into the European data format (EDF) [12]. Each nap was scored by the rules

of visual adaptive scoring system by an experienced clinical neurophysiologist [8]. In

visual adaptive scoring the changes in electrophysiological characteristics of the recorded

signals are used to determine the segment boundaries instead of a fixed epoch. Stage

changes shorter than 1 s were not scored separately. More stages than in the conventional

scoring method [15] were used. The stages without alpha activity were wake-low, drowsy-

low, S1, S2 and REM. Alpha activity with occipital dominance was scored as wake-alpha

and drowsy-alpha (occipital alpha with slow eye movements). More diffuse alpha was

scored as wake-alpha-F or alpha-SEM-F. Arousals were scored separately.

Recordings from three subjects, two females aged 54 and 63 (with slight apnea) and one

male aged 25, were studied in training the method. The recordings from five healthy

subjects with no sleep complaints, three females aged 43, 49 and 70 and two males aged 25

and 74 served as an independent test set. All four naps from each subject were included in

the analysis.

2.2. Analysis method

The presented method (Fig. 1) was designed to detect both short and long alpha

segments that a given recording might contain. For this purpose, FFT windows of lengths

2.5 and 10.0 s (experimentally selected) were used in parallel in feature extraction.

During method design, the effect of FFT windowing in alpha detection was examined.

Using a short FFT window the method was quick to react to a burst of alpha activity but

was often not able to stay high during waning of longer runs of alpha activity. A long

window was better in this respect but was slow to react to shorter alpha segments. Thus,

combining the information obtained using two FFT windows of different length should

provide good performance in all cases. In the following, the structure of the method is

described in detail.

E. Huupponen et al. / Artificial Intelligence in Medicine 24 (2002) 133–147 135

2.3. Feature extraction

The EEG signal was analysed using 1 s time resolution. The Saramaki [17] FFT window

function (with parameter beta in value three) was used in forming the FFT. All spectra were

formed using zero padding to 2048 samples giving a frequency resolution of 0.098 Hz and

scaled to one-sided amplitude spectra. Three partly overlapping 2.5 and 10.0 s window

functions were used to slide along the EEG signal (Fig. 2). Let us first consider the 2.5 s

FFT windowing. Three overlapping 2.5 s FFT windows were centred on the kth second,

with maxima of the windows appearing 0.25 s apart. Corresponding amplitude spectra

were formed and the spectrum containing the highest peak, alpha amplitude was selected

for further analysis to ensure as good as possible alignment of feature calculus with respect

to the alpha burst.

Fig. 1. Alpha detection method. FFT windowing of length 2.5 and 10.0 s (Fig. 2) were used in parallel to obtain

spectra from two different time-spans of local EEG around the kth second. Spectral features described the ratios

of alpha amplitude to other EEG bands. Spectral fuzzy reasoning provided the initial alpha detections o1,k and

o2,k, which were combined with temporal fuzzy reasoning to method output yk.

Fig. 2. FFT windows used in feature extraction. Three 2.5 s windows are seen in the middle and three 10.0 s

windows appear wider. Windows of both lengths are centred on the kth second and are peaking 0.25 s apart.

Different time-spans of EEG signal around the kth second are included when using the two window lengths; in

case of 2.5 s FFT window only a short segment of EEG is included whereas with 10.0 s window a longer

segment is taken. The two window lengths were used in different parts of the method (Fig. 1) but are shown

together for temporal comparison.


The peak alpha amplitude at kth second, denoted as Aa,k, was determined from the alpha

band of 8.0–12.5 Hz. The three spectral feature values were then obtained as ratios of peak

alpha amplitude to mean amplitude level on other main EEG frequency bands directly as:

m1;k ¼Aa;k

Ad;k(1)

where Ad,k is the mean delta activity of 0.8–3.0 Hz,

m2;k ¼Aa;k

Ay;k(2)

where Ay,k is the mean theta activity of 3.5–7.5 Hz,

m3;k ¼Aa;k

Ab;k(3)

where Ab,k is the mean beta activity of 20–40 Hz.

Using the same procedure with the 10.0 s FFT windowing, spectral feature values ni,k,

i ¼ 1; 2; 3, were determined.

2.4. Spectral fuzzy reasoning

In the first level of fuzzy reasoning, or ‘‘spectral fuzzy reasoning’’ the spectral feature

values mi,k, and ni,k, i ¼ 1; 2; 3, were combined to initial alpha detections (o1,k, o2,k) (Fig. 1).

The fuzzy reasoning implemented a flexible mapping where basically large feature values

gave rise to large initial detections and if, for instance, one feature value was rather small

the other two were able to compensate.

In order to bring the spectral feature value into the fuzzy rule base, the ranges of the input

values needed to be determined. The range of each spectral feature value was limited to a

range more narrow than the total range as follows:

~x ¼rhigh; x > rhigh

x; rlow � x � rhigh

rlow; x < rlow

8><>:

(4)

where x ¼ mi;k, and ni,k, i ¼ 1; 2; 3.

The ranges (rlow–rhigh) used in this study were 1.32–2.90, 1.54–3.12, 4.65–9.12 and

1.75–3.35, 1.91–4.02, 5.89–10.80 for feature values mi,k, and ni,k, i ¼ 1; 2; 3, respectively.

Feature values were then fuzzyfied using three triangular overlapping membership

functions, named ‘‘small’’, ‘‘medium’’, ‘‘large’’. The function ‘‘medium’’ covered the

whole range, peaking in the middle. The other two functions covered half of the range,

peaking at the start and end of the range. Fuzzy rules combining the membership functions

are given in Appendix A. Centroid method was used for defuzzification [13]. The range for

the initial alpha detections was 0–10, with large values indicating initial alpha activity.

The rationale for using the above-mentioned ranges was as follows. The FFT amplitude

spectrum of well-formed alpha activity had one major peak at alpha frequency and only

little other frequency components (Fig. 3). The spectral feature values became very large


and thus this case was easy to classify correctly as alpha. Then again, if there were only

mixed background EEG activities having fairly equal amplitude on all frequencies, the FFT

amplitude spectrum appeared nearly even. Then, the spectral feature values became very

small (�1.0) and these cases were easy to classify correctly as non-alpha. The spectral

feature values typically ranged from 0.7 to 25, with large values observed during alpha

segments. In order to classify correctly not only the easy cases but also those alpha

segments that appeared less distinct, the classification was focused on the most decisive

range of feature values. During method design, different ranges were tested using the

central C4-M1 channel. The initial detections (o1,k, o2,k) were compared to the EEG signal

and alpha scorings and suitable ranges were experimentally searched. Then, using training

data of 12 recordings, statistical distributions of spectral feature values from alpha seconds

were formed (Fig. 4). Using those, 15 and 70% points of cumulated distribution were found

to provide approximately those ranges experimentally selected. This rule was used for all

spectral features to give the exact ranges (rlow–rhigh).

2.5. Temporal fuzzy reasoning

In the second level of fuzzy reasoning, or ‘‘temporal fuzzy reasoning’’, the initial alpha

detections (o1,k, o2,k) were combined to method output yk. The initial alpha detections

Fig. 3. Principle in feature extraction. Amplitude spectrum obtained using 2.5 s FFT window in the middle of a

fairly well-defined alpha segment. The peak alpha amplitude and mean amplitudes on delta, theta and beta bands

are indicated with horizontal lines and labelled as Aa,k, Ad,k, Ay,k and Ab,k, respectively. The spectral feature

values were m1;k ¼ 4:48, m2;k ¼ 5:71, m3;k ¼ 11:70, strongly indicating alpha and they were limited to the upper

end of the ranges (Eq. (4)) as ~m1;k ¼ 2:90, ~m2;k ¼ 2:96 ~m3;k ¼ 9:12.


conveyed contextual information about alpha activity with different time-spans in the EEG

surrounding the kth second. The 2.5 s FFT windowing based detection (o1,k) reacted

quickly to any alpha activity. The 10.0 s FFT windowing based detection (o2,k) reacted

more slowly and provided insight to alpha activity in a longer time-window (before and

after the kth second, Fig. 2). A high value of o2,k indicated somewhat certain alpha activity

at kth second and o2,k was given some more weight than o1,k in temporal fuzzy reasoning.

The other advantage of o2,k was that it tolerated waning of alpha. Triangular overlapping

membership functions were used in temporal reasoning too. The fuzzy rules are given in

Appendix B. The range of method output yk was from 0 to 10, with higher values indicating

higher probability of alpha activity.

3. Results

The alpha detection accuracy of the presented method was compared to visual alpha

scorings. Most of the scorings indicated dominant occipital alpha which was also often

seen somewhat attenuated on the central channels. The method output yk was simply

compared to all alpha using 1 s intervals (Fig. 5). The method performance was determined

Fig. 4. Ranges for spectral features obtained using alpha segments in training data. Distributions of spectral feature

values mi,k, i ¼ 1; 2; 3 are shown. Using cumulative distribution, 15 and 70% points provided the range limits (rlow–

rhigh). The limits 1.32–2.90, 1.77–4.10, 4.65–9.12 for feature values mi,k, i ¼ 1; 2; 3 are drawn with vertical lines.


testing central (C4-A1) and occipital (O2-A1) EEG channels in turn. Mean and median

durations of the visually scored alpha segments in the test data were 6.2 and 3.5 s,

respectively, ranging from 1.0 to 83.3 s. The scorings were considered using 1 s resolution.

The ‘‘transition seconds’’ containing the beginning or the end of the scoring were ignored

because with this time resolution they were partly both alpha and non-alpha EEG segments.

The arousals in some test recordings (a total duration of only 119 s) were also ignored. All

other segments were included as non-alpha EEG segments. The test data in 20 test

recordings treated this way included a total of 13,534 s of visually scored alpha segments in

2557 separate runs and 11,277 s of other EEG segments.

A true positive second was counted when the method output yk indicated alpha

simultaneously, with a scored alpha (1 s intervals were used also within scored alpha

segments). If the method output indicated alpha and there was no alpha scoring, a false

positive finding was counted. The true positive rate was the number of detected alpha

seconds divided by the number of scored alpha seconds. The false positive rate was the

number of false positive findings divided by the number of all findings by the method. The

receiver operating characteristic (ROC) curves [24] provided by the method were deter-

mined by using different threshold values for the method output yk. By selecting two near-

optimal points from the ROC curves, true and false positive rates of (94.2 and 9.2%) and

(86.0 and 4.6%) were reached using an occipital EEG channel and true and false positive

rates of (90.0 and 12.1%) and (86.0 and 8.9%) were reached using a central EEG channel

(Fig. 6). The performance with respect to duration of alpha segments was consistent.

The frequency of alpha activity could be obtained readily from the presented method as

the frequency of peak alpha amplitude. This information was not used but merely observed

in this study. In the test data, 78% of alpha seconds had a frequency of 10 Hz or lower.

Because the subjects were awake or in light sleep during MSLT tests, there was a relatively

high level of muscle tone, affecting also the EEG.

Fig. 5. Example of alpha detector output yk (k ¼ 800 . . . 1600) as compared to visual alpha scorings in one of

the test MSLT recordings. The performance of the detector was good as it was on all the test data (Fig. 6).


Fig. 6. ROC curves provided by alpha detection method on the test data of 20 MSLT recordings. (a) Results

from analysis of a posterior (O2-A1) and central (C4-A1) EEG channel are drawn with solid and dashed lines,

respectively. Two near-optimal points on both ROC curves showing performance obtainable are indicated with

dots and circles. True positive rate (at the performance level indicated by the dots) as a function of the duration

of visually scored alpha segments. (b) The true positive rate is high all the time, increasing slightly as the

duration of alpha segments increases. The curves become more irregular with increasing duration (>15 s) as

there were a smaller number of long alpha segments than short ones.


4. Discussion

Fuzzy reasoning provided a convenient way to design the classification needed in the

two fuzzy stages of the presented method. Expert knowledge (in a simple form) about how

the fuzzy stages should behave could be incorporated. Furthermore, fuzzy reasoning based

classification was deterministic so that for every pair of values (o1,k, o2,k) the method output

was known and could be expressed as a function of input values. For implementation

purposes, some other method could be used instead of the fuzzy reasoning, provided that it

performs the same mapping. In fact, the mapping could be stored into a look-up table. Then

again, it was the small number of features that made the design the fuzzy reasoning

convenient in this work. If there were a larger number of features, artificial neural network

solutions might offer a good alternative [7].

A secondary level sleep analyser could use the method output as such (as a continuous

value) to indicate the likelihood of alpha in the current EEG segment. A complete sleep

analysis system needs, however, first level detectors also for K-complexes, delta and eye

movements, etc. They are topics of further studies (as well as optimal combination of their

outputs). A combination of fuzzy alpha and spindle detectors might provide valuable

information for differentiation between these two waveforms.

The presented method used a common 1 s time resolution instead of trying to indicate the

exact beginning and end of an alpha segment. Testing was done using the same time

resolution, which brought some difficulty. Because runs of alpha did not follow the exact

second borders, a split second of the beginning and end of the scored alpha could not be

included in testing. There were about the same amount of alpha and non-alpha seconds in the

test data, which should be taken into account when the ROC curves are studied (Fig. 6). If

necessary for some application, it might be possible to implement the method to provide an

output with 0.5 s time resolution, which is close to the accuracy achieved in visual scoring.

Visually scored arousals were ignored in the testing because in some of the arousals there

was alpha and in some there was not. The total duration of the arousal scorings was, however,

so short that it would have had an insignificant effect on the results anyway in this study.

The method was designed to be usable without any subject or channel specific adjust-

ment. Our selection of the ranges also contributed to that goal. The central channel (C4-A1)

was used in selection of the ranges (Fig. 4) to ensure as good as possible performance on all

channels. No strict EMG artefact indicator was used in this work because the alpha detector

should still work in presence on considerable EMG tonus. An additional test was made

including an EMG artefact feature [6], but it had no effect on the results. This seems natural

since it is very unlikely that background EEG contaminated with even a severe EMG

artefact would give rise to a spectrum of the type typical for alpha activity (Fig. 3). The fact

that spectral properties of virtually all the main EEG frequencies (0.8–40.0 Hz) were

included in the analysis (Fig. 3) contributed positively to the reliability of the method.

The presented detector had the appealing theoretical property to work consistently well

with both short and long runs of alpha activity, which was the outcome in testing, too

(Fig. 6). Thus, the aim of the study was met. This was achieved by combining the initial

alpha detections o1,k and o2,k conveying contextual information about alpha activity with

different time-spans in the EEG. During testing it was found that positive contribution

of long windowing (o2,k) was better the longer runs of alpha a recording contained.


The positive contribution of short windowing (o1,k) became most obvious with runs of

alpha that were short and separate from other alpha segments.

In some test MSLT recordings of duration of 35 min the subjects fell into S1 or S2 sleep

and also in those cases the performance of the detector was good. This seems to confirm the

method design, since during sleep there was plenty of theta and/or delta activity instead of

dominant alpha, which affected directly the spectral features, as intended. Spectral features

became very small during S1 or S2 sleep and they were limited to lower end of the ranges

(Eq. (4)) and the fuzzy reasoning stages made sure that method output became small in

those cases. Overall, the presented alpha detection method worked very well.

Acknowledgements

This study was financially supported by the Academy of Finland, project no. 44876

(Finnish Centre of Excellence Program 2000–2005), the BIOMED-2 project SIESTA

(BMH4-CT97-2040) and the Research fund of the Tampere University Hospital.

Appendix A. Fuzzy rules employed in spectral reasoning

(Limited) feature values ~mi;k; i ¼ 1; 2; 3 and ~ni;k; i ¼ 1; 2; 3, fed into the fuzzy rule base

(input i, i ¼ 1; 2; 3), provided the initial alpha detections oi,k, i ¼ 1; 2, respectively. The

membership functions and rules were experimentally selected keeping the number of rules

as low as possible.

1. If input1 is large and input2 is large and input3 is large then output1 is very large.

2. If input1 is large and input2 is large and input3 is medium then output1 is very large.

3. If input1 is large and input2 is medium and input3 is large then output1 is very large.

4. If input1 is medium and input2 is large and input3 is large then output1 is very large.

5. If input1 is large and input2 is medium and input3 is medium then output1 is very

large.

6. If input1 is medium and input2 is medium and input3 is large then output1 is very large.

7. If input1 is medium and input2 is large and input3 is medium then output1 is very

large.

8. If input1 is medium and input2 is medium and input3 is medium then output1 is

large.

9. If input1 is large and input2 is large then output1 is large.



12. If input1 is small then output1 is small.



15. If input1 is medium and input2 is medium then output1 is medium.




Example of spectral fuzzy reasoning. Rows 1–17 represent the fuzzy rules. The input

values (spectral features) are seen as the vertical lines. The rules that are activated (rules 1–

11, 15–17 in this case) contribute sub-outputs that are aggregated (maximum method) [13].

The output (initial alpha detection) of the rule base (7.19 in this case) is obtained via

defuzzification (centroid method).


Mapping performed in spectral fuzzy reasoning (input3 is held constant) showing the

effect of membership functions and fuzzy rules.

Appendix B. Fuzzy rules employed in temporal reasoning

Initial alpha detections oi,k, i ¼ 1; 2, fed into the fuzzy rule base (input i, i ¼ 1; 2),

provided the method output yk. The membership functions and rules were experimentally

selected keeping the number of rules as low as possible.

1. If input1 is large then output1 is large.

2. If input1 is medium and input2 is large then output1 is large.


4. If input2 is large then output1 is large.

5. If input2 is medium then output1 is medium.


7. If input1 is small and input2 is small then output1 is small.

8. If input1 is small and input2 is medium then output1 is small.

Example of temporal fuzzy reasoning. Rows 1–8 represent the fuzzy rules. The input

values (initial alpha detections) are seen as the vertical lines. The rules that are activated

(rules 1–5, 7 in this case) contribute sub-outputs that are aggregated (maximum method)

[13]. The output (method output yk) of the rule base (7.26 in this case) is obtained via

defuzzification (centroid method).


Mapping performed in temporal fuzzy reasoning showing the effect of membership

functions and fuzzy rules.

References

[1] Basar E, Yordanova J, Kolev V, Basar-Erogly C. Is the alpha rhythm a control parameter for brain

responses. Biol Cybern 1997;76:471–80.

[2] Broughton R, Hasan J. Quantitative topographic electroencephalographic mapping during drowsiness and

sleep onset. J Clin Neurophysiol 1995;12(4):372–86.

[3] Fish DR, Allen PJ, Blackie JD. A new method for the quantitative analysis of sleep spindles during

continuous overnight EEG recordings. J Sleep Res 1988;70:273–7.

[4] Hasan J. Past and future of computer-assisted sleep analysis and drowsiness assessment. J Clin

Neurophysiol 1996;13(4):295–313.

[5] Hjorth B. EEG analysis based on time domain properties. Electroenceph Clin Neurophysiol 1970;29:306–

10.

[6] Huupponen E, Varri A, Himanen SL, Hasan J, Lehtokangas M, Saarinen J, et al. Optimization of spindle

detection threshold. J Sleep Res 2000;9(4):327–34.

[7] Huupponen E, Varri A, Himanen S-L, Hasan J, Lehtokangas M, Saarinen J, Comparison of fuzzy

reasoning and autoassociative MLP in sleep spindle detection. Proc. EUSIPCO 2000: 55–8.

[8] Himanen S-L, Saastamoinen A, Hasan J. Increasing the temporal resolution and stage specificity by visual

adaptive scoring (VAS)—a preliminary description. Sleep Hypnosis 1999;1(1):22–8.

[9] Hakkinen V, Hirvonen K, Hasan J, Kataja M, Varri A, Loula P. The effect of small differences in electrode

positions on EOG vigilance signals. Electroenceph Clin Neurophysiol 1993;86:294–300.

[10] Isaksson A, Wennberg A, Zetterberg A. Computer analysis of EEG signals with parametric models. Proc

IEEE 1981;69(4):451–61.

[11] Kemp B, Blom HAP. Optimal detection of alpha state in a model of the human electroencephalogram.

Electroenceph Clin Neurophysiol 1981;52:222–5.

[12] Kemp B, Varri A, Rosa A, Nielsen K, Gade J. A simple format for exchange of digitized polygraphic

recordings. Electroenceph Clin Neurophysiol 1992;82:391–3.

[13] Mendel J. Fuzzy Logic systems for engineers: a tutorial. Proc IEEE 1995;83(3):345–77.

[14] E. Niedermeyer FL, Da Silva. Electroencephalography: basic principles, clinical applications, and related

fields. Baltimore: Urban and Schwarzenbeck, 1987 [chapter 9].


[15] Rechtshaffen A, Kales A. A manual of standardized terminology, techniques and scoring system for sleep

stages of human subjects, US Public Health Service. Washington DC: US Government Printing Office,

1968.

[16] Richardson GS, Garskadon MA, Flagg W, van dan Hoed L, Dement WC, Mitler MM. Excessive daytime

sleepiness in man: multiple sleep latency measurements in elderly and young adult subjects. Distribution

Electroenceph Clin Neurophysiol 1978;45:621–7.

[17] T. Saramaki. A class of window functions with nearly minimum sidelobe energy for designing FIR filters.

In: Proceedings of the IEEE International Symposium on Circular Systems. Portland, Oregon, 1989: 359–

62.

[18] Santamaria J, Chiappa KH. The EEG of Drowsiness in Normal Adults. J Clin Neurophysiol

1987;4(4):327–82.

[19] Schimicek P, Zeitlhofer J, Anderer P, Saletu B. Automatic sleep-spindle detection procedure: aspects of

reliability and validity. Clin Electroencephalogr 1994;25(1):26–9.

[20] Smith JR, Karacan I, Yang M. Automated measurement of alpha, beta, sigma and theta burst

charasteristics. Sleep 1979;1(4):435–43.

[21] Smith JR. Automated EEG analysis with microcomputers. Med Instrum 1980;14(6):319–21.

[22] Varri A, Hirvonen K, Hasan J, Loula P, Hakkinen V. A computerized analysis system for vigilance studies.

Comp Meth Progr Biomed 1992;39(1-2):113–24.

[23] Wennberg A, Zetterberg LH. Application of a computer-based model for EEG analysis. Electroenceph

Clin Neurophysiol 1971;31:457–68.

[24] Zweig MH, Campbell G. Receiver-operating charasteristic (ROC) plots: A fundamental evaluation tool in

clinical medicine. Clin Chem 1993;39(4):561–77.


Date post:	17-May-2023
Category:	Documents
Upload:	independent
View:	0 times
Download:	0 times

Fuzzy detection of EEG alpha without amplitude thresholding

Documents