We enrolled 98 patients suspected with OSA, and conducted a full-night PSG for these patients at the Samsung Medical Center in Seoul, Korea as per standard clinical guidelines (1). The exclusion criteria in this study were as follows: patients with central sleep apnea, mixed sleep apnea, and cardiovascular disorders. A total of 45 subjects suffering from OSA were included in this study (Table 1). All patients were subjected to a full-night, attended diagnostic PSG using an Embla N7000 (Embla Systems Inc., Broomfield, CO, USA) device. The PSG studies were scored according to the standard criteria defined by the sleep specialists at American Academy of Sleep Medicine (AASM). All subjects provided written informed consents, and the study protocol was approved by The Institutional Review Board of the Samsung Medical Center. The piezo-electric sensor (REF 1420610; Embla Systems Inc.) was attached to the neck of a patient to measure vital signs during nocturnal PSG. Signals were recorded at a sampling rate of 200 Hz, which is the minimum acceptable sample rate for snoring according to the AASM guideline (1). Snoring episodes were traced via a sleep-specific program, RemLogic (Embla Systems Inc.) and footnoted by an experienced sleep specialist. Signal processing and analysis were performed using MATLAB (Mathworks Inc., Natick, MA, USA).

In this study, snoring episodes were detected via a piezo-electric sensor as shown in Fig. 2. We used the method proposed previously by our research team for automatic snoring detection (13). This method was based on a short-time Fourier transform analysis with a 100 ms hamming window. The snoring episodes so detected were interpreted using SIs, where SI represents the frequency of snoring episodes. The SI, which can be determined heuristically and demonstrates the incidence of snoring within a predetermined interval, was calculated as the number of snoring events occurring within 10 seconds. Snoring induces vibrations of the skin surface in the neck area. A piezo-electric sensor can measure not only snoring vibrations, but also non-snoring vibrations related with breathing, body movements, and coughing during sleep. All non-snoring events were considered as normal breathing and thus, rejected in the elimination process.

Features used for classification of OSA were extracted from the PRV analysis. The PRV signal was identified with a piezo-electric sensor to extract temporal and spectral domain features. The pulse rate was obtained by detecting the pulse waves originating from the carotid artery using the piezo-electric sensor attached to the neck. For recognizing the PRV signals, we used a self-developed, sensor-customized algorithm based on an autocorrelation method to detect the heartbeat peaks (Fig. 3). The pulse rate was calculated from the pulse-to-pulse interval (PP), and an unequal time interval signal was adjusted to an equal interval signal via linear interpolation. The adjusted pulse rate signal was resampled at 4 Hz to analyze the extracted features in both the time and frequency domains. These features were extracted from the resampled pulse rate signal using a length of 2 minutes and a hamming window with 25% overlapping. In the time domain, we calculated the PP, its standard deviation (SDPP) and the root mean square of successive differences (rMSSD). In the frequency domain, we extracted the low-frequency power (LF) range (0.04–0.15 Hz), high-frequency (HF) power range (0.15–0.40 Hz), and LF/HF ratio. These features reflect the short-term HRV and parasympathetic nerve activity, and are known to be associated with sleep-disordered breathing. The values of PP, SDPP, and rMSSD increased during OSA; however, there was no significant difference between normal breathing and OSA. LF, HF, and LF/HF ratio decreased during OSA, also with no significant differences between normal and OSA regions (Table 2). To classify an OSA event, an SVM was used as a classifier in this study.

Before classification, the SI was used to distinguish between normal and OSA suspected regions. Next, the OSA suspected region was analyzed via SVM using temporal and spectral features (Fig. 1). Finally, OSA was classified using an SVM. The aim of SVM is to determine an optimal separating hyperplane that shows the maximum margin between the apneic and non-apneic segments. First, the input data was transformed into a higher dimensional space by employing a kernel function, and then a linear optimal hyperplane was constructed between normal and OSA classes in the transformed space. These data vectors nearest to the constructed line in the transformed space are called support vectors. In this study, we applied a single binary SVM classifier with a radial basis function employed as the kernel function. The multiplier coefficient α and regularization parameter C were determined empirically (C = 1; α = 0.5). The extracted parameters of the PRV time series were used as input features in the SVM, and the output types were represented as follows: −1 = Normal and +1 = OSA. Then, the number of automatically detected apneic events was counted (per hour of recording) and compared with standard cutoffs (5, 15, and 30 events/hr) to designate the recording as a mild, moderate, or severe group. All SVMs were trained and tested on the SVM toolbox of MATLAB (Mathworks Inc.).

To test the performance of the proposed method, we used the piezo-electric signals from the sleep recordings of the 45 patients suffering from OSA (Table 1). These patients were divided into 3 groups based on the severity of the regular apnea to validate the classification ability of the proposed algorithm. We used OSA and snoring footnotes of the PSG, while the pulse rates from the piezo-electric sensor were confirmed using reference ECGs in the PSG recordings. A cross-validation was performed to evaluate the proposed method for each patient group.

Five characteristics, namely, accuracy, sensitivity, specificity, positive predictive value, and kappa value were calculated to assess the performance of the proposed method. Accuracy indicates the overall detection accuracy, sensitivity is interpreted as the ability of the algorithm to accurately classify OSA, and specificity indicates its ability to not generate a false negative (normal breathing). The kappa coefficient (κ) (17) was used to evaluate the agreement between the estimated apnea-hypopnea index (AHI) level of the proposed method and annotated AHI of the clinical diagnosis of PSG. The Kruskal-Wallis test was performed for comparing patient groups. P < 0.01 was accepted as significant.

The protocol of this study was reviewed and approved by the Institutional Review Board of Samsung Medical Center (IRB No. 2012-01-063). All patients provided written informed consent for participating.