تشخیص تنگی نفس موقتی (آپنه) در خواب بر پایه مدل سازی Rician تغییرات ویژگی ها در سیگنال های EEG در چند باند

ABSTRACT Sleep Apnea Detection Based on Rician Modeling of Feature Variation in Multiband EEG Signal

Sleep apnea, a common sleep disorder deteriorating sleep quality of the patients, affects about 5-20% of adult population [1], [2]. According to American Academy of Sleep Medicine (AASM) criteria, apnea is scored where reduction in airflow is ≥90% and it stays like so for more than 10 seconds. Hypopnea criterion requires ≥30% reduction in airflow for more than 10 seconds in association with either ≥3% oxygen desaturation or an arousal [3]. Sleep apnea patients generally experience severe headaches, daytime sleepiness and several cardio-respiratory disorders [4]-[5].

In overnight polysomnography (PSG), the whole night apnea events are manually scored by an expert, which is expensive, tedious, time consuming and prone to human error [6]. As a result, there is a great necessity for an automatic sleep apnea detection algorithm. Different automatic apnea detection schemes using various biomedical signals including EEG are presented in [7]-[13]. For example, in [7], heart rate variability, nasal pressure, EOG, EMG, oronasal temperature, in [8], Oxygen saturation, heart rate variability and the respiratory signals, in [9] EMG signal, in [10] pupil size, in [11] only ECG signal, in [12] oximetric signal, in [13] ECG, EMG, EOG signals are used.

Instead of utilizing several physiological signals, EEG signal alone is getting special attention by the researches because of its successful application in analyzing sleep related problems [14]-[23]. In [14] non-linear behavior of EEG signal is studied. EEG scaling exponents computed by detrended fluctuation analysis (DFA) are used as features to classify apnea and healthy subjects in [14]. In [15], Hermite decomposition algorithm based on particle swarm optimization is proposed and [16], [17] employ wavelet transform of EEG to identify sleep apnea events. Instead of utilizing the full band EEG signal, an effective way is to divide the EEG signal into well known EEG sub-bands, namely delta, theta, alpha, sigma and beta and analyze the band limited signals. In [18], energy and variance computed from each sub-band are used as features for apnea classification. Bispectral characteristics of EEG signal are studied in [19], where in each subband the degree of quadratic phase coupling (QPC) is analyzed. Sleep apnea is detected from the variation of Hilbert spectrum frequency in [20]. Cumulative delta-power ratio of overlapping frames is used for classification in [21] while in [22], multi-band entropy values are used as features to exploit the random characteristics of EEG signal. In [23], statistical features are extracted from the variation of Beta band energy within an EEG frame and used for the purpose of classification. Most of the reported methods consider classification between apnea and healthy subjects and the difficult task of discriminating apnea and non-apnea events is rarely attempted.

In this paper, a sub-frame based model fitting approach is proposed where both these classification scenarios are taken into consideration. First, a multi-band sub-frame based scheme is introduced to extract the feature variation pattern within a frame. Next, the feature variation patterns are processed using statistical analysis and modeled with characteristic probability density function. Resulting model parameters and some statistical measures are used in K nearest neighbor (KNN) classifier to classify apnea and non-apnea frames. Detail experimentations and performance analyses are carried out in three different publicly available databases. The uniqueness of the proposed method lies in modeling the within-frame feature variation pattern and utilizing the fitted model parameters as potential features in the classification scheme, which offers very low feature dimension. Unlike using multiple bio-signals, this paper focuses on automatic detection of sleep apnea using single lead EEG signal which makes the system cost effective and can lead to an auto-diagnostic device favorable for in home care.