# Literature Review

[Drive Link](https://drive.google.com/drive/u/0/folders/1-ij16DfMhFp6306TKOydW9v9dJifG89y)

[toc]

## Automated Characterization of Cyclic Alternating Pattern Using Wavelet-Based Features and Ensemble Learning Techniques with EEG Signals
-- AW

* The cyclic alternating pattern (CAP) is a physiological recurring electroencephalogram (EEG) activity occurring in the brain during sleep and captures microstructure of the sleep and can be used to identify sleep instability.
* Conventionally, sleep is analyzed using polysomnogram (PSG) in various sleep laboratories by trained physicians and medical practitioners. However, PSG-based manual sleep analysis by trained medical practitioners is onerous, tedious and unfavourable for patients.
* Sleep consists of periodic repetition of unconsciousness (physical-inactivity) called non rapid eye moments (NREM) followed by high activity called rapid eye moments (REM).
* Sleep is categorized into five stages: wakefulness (W), N1, N2, N3, and REM. N1, N2, N3 forms NREM
* Although, phasic events like K-complexes and delta bursts show characteristics similar to arousal, but they are not considered as arousal if there is no to short-term frequency increase in EEG. To overcome these shortcomings of macrostructure based sleep scoring, a new microstructure based sleep scoring technique named cyclic alternating pattern (CAP) was devised, which includes such phasic events in brain activity as an alternative scheme to describe NREM sleep.
* The NREM sleep stage is observed to have alternating patterns of cerebral activation (phase A) followed by duration of deactivation (phase B). The CAP phase A typically includes events like K-alpha, K-complex sequences, delta bursts, alpha waves, vertex sharp transients and arousals.
* The combination of phase A and phase B is termed as a CAP cycle and this cycle begins with phase A and ends with phase B. Two successive CAP cycles are needed to form a CAP sequence
* If a phase A is not accompanied by phase B, then it is termed as an isolated phase A and is considered as non-CAP (an absence of CAP for > 60 s duration). Thus, a CAP sequence contains minimum three A phases (A–B–A–B–A) followed by non-CAP period.
* CAP sequence always follows a continuous NREM sleep EEG pattern with a minimum duration of 60 s. CAP phase A can be detected using any EEG lead
* Phase A has three subtypes
1. A1: high amplitude slow waves
2. A2: Low amplitude slow waves
3. A3: low amplitude fast rhythms
* to remove the noise artifacts and preserve only the necessary information, EEG signals are bandpass filtered using an infinite impulse response (IIR), butterworth filter of order four.
* Wavelet entropy is used for quantitative analysis of transient fea- tures of non-stationary physiological signals including EEG. It is able to measure the uncertainty involved in a random process and for a wavelet subband it can be defined as
$E_{Wave}=-\sum_{i=1}^nx_ilog(x_i)$ where, x i represents the magnitude of i th wavelet coefficient.
* We observed that, for most of the classification tasks EBagT and EBoosT algorithms attained the best classification performance after extensive simulations, however, for few classification tasks SVM algorithm showed better results.

## A machine learning model for identifying cyclic alternating patterns in the sleeping brain
-- AW

* Using data points, feature engineering is performed to form the feature matrix. The feature matrix acts as an input to the logistic regression model, which will account for the time series analysis of the brain signals and classify the start of phase A in a given CAP cycle.
* After cleaning, data file has three columns, column one represents the absolute time (in seconds) at which a reading was recorded, column two is the actual EEG reading (in microvolts), column three, are binary representations (0 or 1), where 0 indicates the absence of phase A at this given time and 1 indicates presence of phase A at this time
* For pre-processing Differential moving average is used to make the data stationary. The Dickey-Fuller test (Said, 1984) is used to test stationarity. It was found that a differential moving average with window size of 15 made the data the most stationary

## Automated phase classification in cyclic alternating patterns in sleep stages using Wigner–Ville Distribution based features
-- SM

* Sleep consists of 4 stages: 1 REM 3NREM
* Each CAP cycle is characterized by two stages, the phase that includes the phasic transients (phase A) and background activities (phase B).
* Phase A is either slow varying high amplitude signals or fast varying low amplitude signals or a combination of both. Further divided into A1, A2, & A3.
* for CAP phase segregation from EEG, six EEG bands (low delta, high delta, theta, alpha, sigma and beta) for feature extraction have anayzed. The frequency ranges corresponding to the bands are as delta: 0.75–4 Hz, theta: 4–8 Hz, alpha: 8–12 Hz, sigma: 12–15 Hz and beta: 15–25 Hz.

## Automatic A-phase Detection of Cyclic Alternating Patterns in Sleep Using Dynamic Temporal Information
-- SM
[Database](https://physionet.org/content/capslpdb/1.0.0/)
* The A-phase is commonly restricted to sleep stages without rapid eye movement and characterised by slower high-voltage rhythms, faster lower voltage rhythms or by both
* Subtype A1 is dominated by EEG synchrony, i.e. high-voltage slow waves, whereas faster low-amplitude rhythms are more prevalent in subtype A3. Subtype A2 contains a mixture of both waveforms.

## Sleep Stage classification based on noise-reduced fractal property of heart rate variability
-- AP

* Heart beat signals are relatively easy to detect in real life, and so various studies have been conducted to use them for estimating the sleep stages.Here they extracted the DFA alpha1 value in each time interval from the heart beat signal and chained the values in chronological order to construct DFAseq for whole recording time. Then developed a classification model for determining the sleep stage based on the NR-DFAseq, and evaluated the performance of the model.
* Nomenclature:
DFAseq DFA alpha 1 sequence
NR-DFAseq Noise-reduced DFA alpha1 sequence
RRI RR-interval
HRV Heart rate variability
* The ECG data of 13 subjects because the rest are either irregular format data or too noisy.
* 30 seconds of recording time according to the R & K rules.(epoch)
* labeled a sleep stage (wake, REM, S1, S2, S3, or S4) for each epoch. S1 and S2 as light sleep stage, while S3 and S4 as deep sleep stage.
* From each recording of ECG, detected R-points and obtained RR intervals (RRIs). Among these RRIs, excluded RRIs of less than 0.3 seconds or more than 3 seconds. Also, such RRIs were excluded that have less than 70% or greater than 130% of the average of 10 surrounding RRIs.
* For the calculation of DFA alpha 1 value at each R point, used 256 RRIs before the R point.
* Used the DFA calculation method given in Hardstone’s study. Then, the mean of DFA alpha1 values over each epoch was defined as the DFA alpha1 value of the epoch. By chaining the DFA alpha 1 values of all epochs for each subject in a chronological order, we obtained the DFAseq for each subject.
* For noise reduction, applied the empirical mode decomposition (EMD) method 16 to the DFAseq. The EMD method extracts and subtracts such signals satisfying the intrinsic mode functions (IMFs) from the original signal. At next iteration, signal satisfying the IMFs is newly extracted and subtracted from previously subtracted signal, instead of the original signal.s. In this study, this procedure was repeated four times.
* Figure 1 and Figure 2 show an example of DFAseq and NR-DFAseq for a subject, respectively. From Figure 1, it is hard to intuitively recognize any relationship between DFA alpha 1 and sleep stage. After noise reduction, however, we can recognize relatively well the relationship.
* The Pearson correlation coefficient between the sleep stage sequence and DFAseq or NR-DFAseq was measured for each subject. To do this, each sleep stage was assigned a numeric value. Wake stage was assigned to 0, REM stage to -0.5, and S1 to S4 stage to -1 to -4.
* From this figure, its correlation coefficients with the sleep stages showed an increase of 0.23 on the average in the case of NR-DFAseq than DFAseq. In addition, 11 out of 13 recordings showed the correlation coefficients of 0.6 or more. This implies that the sleep stages and NR-DFAseq are highly correlated in many ECG recordings.
* One model is to distinguish between wake stage and sleep stage. The other model is to distinguish between light sleep stage and deep sleep stage. For wake/sleep classification, used 0.9 of noise-reduced DFA alpha 1 value as a threshold, while we used 0.77 as a threshold for light/deep sleep classification.
* Performance of the model was evaluated by k-fold cross-validation with unit of ECG recording (i.e., k=13).
* The NR-DFAseq might be one of good distinctive features for sleep stage classification. By looking into a changing pattern of the noise-reduced fractal property of HRV over time, we could develop better classification models.
* It seems reasonable to consider that a big change of fractal property in a short time can be noise while a big change in a long time can be a good indicator for sleep stage change.
* Applied the EMD method for noise reduction in HRV DFA alpha 1 data and confirmed that the noise-reduced fractal property of HRV can be closely related to the sleep stage in many cases.