The procedure of CAP detection algorithm can be divided into four main steps: (1) EEG preprocessing, (2) feature extraction, (3) A-phase detector, and (4) context decision. The flow chart of CAP A-phase subtypes has three steps: (1) feature extraction, (2) classifier-based classification, and (3) rule-classifier-based classification. First, EEG raw data is recorded from PSG, then preprocessing for noise removal. Second, some features of the preprocessed EEG signals are extracted. We use seven bands of band power and Hjorth parameter activity to calculate their descriptors and to compute fractal dimension for A-phase detection and subtypes classification. In A-phase detector, we propose an automated method which can adjust parameters and make some extension to detect A-phase. In context decision, we choose some parameters set as same as previous researches. In subtypes classification, we choose k-nearest neighborhood (k-NN) algorithm.
2.3.1 Signal preprocessing
At first, we use the raw EEG signal which is C3-A2 channel, and the sampling fre-quency is 128 Hz. Then, the continuous time EEG signal are segmented with every 0.5-second interval. We choose the segment length of 0.5 0.5-seconds because earlier studies con-sidered about the microstructure and decided a suitable time length. For instance, an 8 hour records will divide into 57600 segments.
in CAP consensus report, we extract the feature sets in the frequency domain. We divide the frequency band into seven sub-bands between 0.5 Hz and 45 Hz. In addition, we also divide the δ band into lower and higher band. The δ band is 0.5-4 Hz. The band is divided into delta (0.5-2.5 Hz) and delta (2.5-4 Hz) ; The θ band is 48 Hz; The α band is 8-12 Hz; The σ band is 12-15 Hz; The β band is 15-25 Hz. Besides, the total band is defined with the range of 0.5-45 Hz. The EEG signal is filtered by using the butterworth filter and the bandwidth of each frequency bands is defined above. These eight frequency bands is shown in Table 2.1.
Table 2.1: Eight frequency bands with its frequency range respectively Frequency band Range (Hz)
ShortAveragei = 1
where i denoted each point of time and the band power is average each 0.5 s. So we compute short average between -2 to 2 that depends on the band power average length.
LongAveragei = 1
where i denoted each point of time and the band power which average each 0.5 s. So we compute long average between -64 to 64 that depends on the band power average length.
In the second part, the descriptors for each band are computed according the following formula :
descriptori = ShortAveragei− LongAveragei
LongAveragei (2.9)
Because the descriptor provides a relatively measured between the short interval signal instantaneous and background signals variation.
Fractal dimension
In the time domain analysis, it is hard to present the signal variability between the different of A-phase or non-A-phase since scoring is based on the analysis of EEG power spectrum. But fractal dimension is an EEG complexity measure which can be used to estimate EEG signal variability. The proposed method computes one fractal dimension
shape). The activity values are used in the proposed method and calculate in 500 ms time window without overlapping. We compute two activity average values each 0.5 s of the short interval with 2 s and long interval with 64 s, then calculate descriptor as formula 2.9.
2.3.3 A-phase detector
We plan to construct an A-phase detector. The A-phase detector determines the pe-riod of time of the A-phase for all night sleep. In the A-phase detector, we combine two methods that are the variable length template and the EXIST/LENGTH thresholds because we plan to inherit the two methods advantages and to avoid their disadvantages.
The variable length template method for CAP detection was proposed in 1999 and the EXIST/LENGTH thresholds method for CAP detection was proposed in 2004. Using the variable length template method to detect A-phase is a more reasonable method than the EXIST/LENGTH thresholds method. Because the EXIST/LENGTH thresholds method determines short-long term ratios of each time point and compares short-long term ra-tios with the EXIST/LENGTH thresholds to decide A-phase or not, the EXIST/LENGTH thresholds method could not make adjustments according to features near the time. The variable length template method could solve the problem. The variable length template can vary template length to cover features near the time to fit template. In the previous A-phase detectors, the EXIST/LENGTH thresholds method used two thresholds (EXIST/LENGTH) is better than the variable length template method used one threshold. In the definition of A-phase, the length of A-phase is decided by all band descriptors. But the exist of A-phase is decided by delta band particularly. The proposed A-phase detector uses two thresholds
Step2. Initial a 2 s sliding unit square pulse to be template.
Step3. Decide what band descriptors we want to use for detecting A-phase.
Step4. Calculate convolution value between template and descriptor values for every band descriptors we use, if one of convolution values is larger than length threshold, then repeating the convolution with the square pulse by increasing pulse length 0.5 s. if not, then sliding the pulse window right 0.5 s.
Step5. stop the convolution process when the convolution value is smaller than threshold.
the feature values convoluted by square pulse are candidate A-phase.
Step6. Compare every feature values of candidate phase. If one of the candidate A-phase descriptors of every bands larger than exist threshold, the candidate A-A-phase is called A-phase. The non-A-phase is candidate B-phase.
Step7. We repeat the A-phase detection procedure until all night EEG signals are checked.
2.3.4 Context decision
In this part, we mention about CAP scoring method used to detect CAP sequence and divide non-A-phase into B-phase and non-CAP. CAP scoring is defined by the context decision method considered by previous studies, so we use the same requirements. First, we examine the CAP sequence epoch by epoch. If the epoch’s stage is NREM, then we
than 60 s, is scored non-CAP. 3) The A-phase in the end of the CAP sequence is scored as non-CAP. When all epoch is calculated completely, we can calculate CAP parameter such as CAP rate, average length of CAP sequences to identify the performance of CAP detection algorithm. We follow these rules to develop our algorithm which is shown in Figure 2.8.
2.3.5 Classification of A-phase subtypes
In the classification of A-phase subtypes, we plan to combine previous classification method which was proposed in 2002 and k-NN (k-nearest-neighborhood) classifier to im-prove the classification accuracy. In fact, the sub-phase of the A-phase classification method proposed in 2002 was based on their classification criteria of A-phase. The recog-nition of an A-phase is required that either the delta or the theta descriptor overcomes the threshold. In addition, the recognition of a sub-phase of the A-phase is considered about different bands descriptor which overcomes the recognition threshold. The theta, alpha and beta descriptors are considered for the recognition of the sub-phase. The recognition of the sub-phase should be sufficed that one of the three descriptor overcomes the threshold. If no sub-phase is recognized, then the A-phase is an A1-phase. If a sub-phase is recognized, then the A-phase is an A2-phase or A3-phase. Then the classification of A2-phase and A3-phase is based on tail. The tail is the portion of the A-phase with low values of the delta descriptor. If the tail is shorter than two-fifths of the A-phase length, it is an A2-phase.
Otherwise it is an A3-phase. The classification procedure is show in Figure 2.9.
0.5 second and then average all fractal dimension values of entire A-phase. Second, the k-NN classifier is used to classify A1-phase and non-A1-phase by the features of fractal dimension values and time dimension values. Third, delta band descriptor should be cal-culated for the proposed method. Finally, it is used in the proposed method of the criteria to classify A2-phase and A3-phase with the tail. If the tail is shorter than two-fifths of the A-phase length, it is an A2-phase. Otherwise it is an A3-phase.
Figure 2.6: The top curve represents the F4-C4 channel with the duration of 20 seconds.
The other rows show the descriptors of seven frequency bands. These bands from top to bottom are delta, lower delta, higher delta, theta, alpha, sigma, and beta. These descriptors are sampled every 0.5 s. (Figure source: Sleep Medicine 5 Original article, Umberto Bar-caro et al.: A general automatic method for the analysis of NREM sleep microstructure, p 567-576, 2004)
Figure 2.7: The A-phase detector of the proposed method
Figure 2.8: The context decision procedure of the proposed method.
Figure 2.9: The A-phase subtypes classification method used in the previous study.
Figure 2.10: The proposed A-phase subtypes classification method which is combined with the k-NN classification method and A-phase classifier.