Obstructive Sleep Apnea (OSA) is a disorder that causes breathing to be interrupted repeatedly during sleep. An interruption in breathing results in deoxygenation, leading to a brief arousal (a period of light sleep or wakefulness which is usually not noticed by the individual). Repeated arousals lead to short-term problems, such as daytime sleepiness and poor concentration. OSA is also associated with increased risk of serious chronic conditions, including cardiovascular disease (Monahan and Redline, 2011) and diabetes (Rosenfeld, 2014).

OSA is relatively common, affecting 2–7% of the global adult population; the prevalence is similar in the developed and developing world, with a high variance across ethnic groups (Young et al., 1993; Bearpark et al., 1995; Ip et al., 2001, 2004; Kim et al., 2004; Udwadia et al., 2004; Sharma et al., 2006; Lam et al., 2007). It is usually diagnosed on the basis of an overnight sleep study, where data including photoplethysmography (PPG), respiratory effort, electrocardiography, audio, and activity are typically recorded (Roebuck et al., 2014). Because of the amount of data that are collected and analyzed, the screening process can be expensive, typically costing more than $1000 (Epstein and Dorlac, 1998). In addition, the screening process removes patients from their normal sleeping environment, preventing repeatable unbiased studies. It is estimated that up to 90% of individuals with OSA are undiagnosed and untreated (Young et al., 1997). Screening of OSA is particularly poor in developing countries, where the resources required for conventional screening and diagnosis are often unavailable. While conventional screening of OSA is expensive, some treatment for those diagnosed with the condition can be relatively cheap; oral appliances, which attempt to enlarge the pharyngeal airway during sleep by holding the tongue or jaw forward, are available for as little as £35 (Somno-Guard 3, Tomed, Bensheim, Germany). Lifestyle changes such as losing weight can often resolve mild to moderate OSA, and oral appliances are available for more serious cases. A cheap method of screening for OSA could therefore greatly reduce the burden of OSA on the healthcare system, particularly in developing countries where sleep lab facilities are very limited.

Since audio recordings are particularly easy to collect, and most mobile phones have in-built high quality audio recording capabilities, this article explores whether audio analysis alone is sufficient to diagnose OSA. This is a comparison between two contrasting methods; a standard approach from speech analysis, which has been done before, and a novel approach which captures non-stationary periodicities. There are a number of approaches that are used to determine whether an event is apneic or not, or whether a subject suffers from OSA. Events can be analyzed using LPC (Ng et al., 2008a), mel-frequency cepstrum coefficients (MFCCs) (Cavusoglu et al., 2007), power spectrum (Ng et al., 2008b), energy distribution (Jones et al., 2005, 2006a,b; Cavusoglu et al., 2007), and pitch (Abeyratne et al., 2005). Active devices (Michaelson et al., 2006) and using audio recorded while the subject is awake (Goldshtein et al., 2011) are both methods used to determine whether a subject has apnea or not. Although event detection is more common, it is also more laborious as there needs to be a gold standard for comparison. This involves annotating the entire night’s recording for each subject in the data set. Active devices and using speech signals allow for the subject to be diagnosed rather than identifying individual events. LPC and MFCCs were chosen as the standard approach from speech analysis because they are two of the most common methods used to determine whether an event is apneic or not. From the literature, it can be seen that LPC classifies apneic snores from benign snores with Se = 88% and Sp = 82% for LPC (Ng et al., 2008a); while MFCCs achieve Se = 82% (Cavusoglu et al., 2007). LPC can provide accurate estimates of speech parameters while not being computationally intensive (Rabiner and Schafer, 2007). However, it does assume stationarity, which may not be true over the event duration. MFCCs were used (where the frequency bands are equally spaced on the mel scale) as they approximate the human auditory system’s response more closely than the linearly spaced frequency bands used in the normal cepstrum. This frequency warping can allow for better representation of sound.

As a comparison, Table 4 shows the performance when using the common clinical thresholds on a variety of features. It is worth noting that both the AHI and ODI are associated with multiple thresholds used for classifying subjects into different categories. Subjects are said to be normal or a snorer if below a threshold of 5 and have mild, moderate, or severe OSA otherwise; 10 or 15 is normal/snorer/mild vs. moderate/severe; 20 or 30 is normal/snorer/mild/moderate vs. severe.

The results of the standard speech-based event classification analysis can be found in Table 5, while the results of the MSE analysis using LDA and a RF can be found in Tables 6 and 7 respectively. The best results for each analysis are in bold font in the tables.

Note that the standard speech analysis techniques have an extremely low sensitivity. Adding demographics to the analysis almost always increases the sensitivity, but only by a moderate amount. Increasing the window size has a moderate effect, but the sensitivity remains low and the variance is high. For the classic speech analysis approach, the best results were consistently obtained when using a combination of LPC, MFCCs, and demographics, for a 3-s window size, achieving Ac = 76.9%, Se = 29.2%, Sp = 88.7%, AUC = 0.73.

As can be seen in Tables 6 and 7, the best performance is obtained when MSE + demos are used in the analysis; the RF achieved Ac = 80.5% and AUC = 0.88 during validation. However, using MSE alone achieves comparable results within the variance over the folds. There is a better balance between Se (66.0%) and Sp (88.8%), unlike in the event classification approach, and the AUC is also higher (0.86 compared to 0.84).

As can be seen in Table 4, demographics or questionnaires alone lead to a poor performance for classifying OSA severity; they are either sensitive or specific, never both, and accuracy is only slightly better than random chance (ranging from 53 to 62%). This is not unexpected, particularly when subjects have been asked to note down these figures themselves rather than being measured and recorded by a healthcare professional; studies have shown that people are poor at self-reporting height and weight (Engstrom et al., 2003). It would therefore be prudent to move away from the use of such information and perhaps rely on objective signals only, if they provide a lower error rate. Both the AHI and ODI are good classification features. It is clear from Table 4 that a threshold of 15 provides the best classification. This separates normal/snorer/mild vs. moderate/severe, and is appropriate for this work as this is the classification problem being addressed. Using this threshold gives performance statistics in the high-80% to low-90% range.

It is clear that the data used for the standard speech-based classifier approach is heterogeneous, and that each fold is significantly different, as evidenced by the large SD values for some of the metrics. This is likely due to the lack of annotations, i.e., there are insufficient data for this analysis or there are too many features. It should be noted that in the training data set, there are up to 30 features (when LPC + MFCCs + demos are combined) but only 56 subjects. This causes the problem to be under-specified, and leads to poor results on the test data. In addition, using three or more annotators would make the annotations more robust. Only those events where two or more annotators agree would be used in the analysis, and could improve the performance of the classifier.

The standard speech-based approach was unable to exceed an Ac of 82% when classifying by subject. Table 5 indicates that using LDA results in a very specific classifier (in the high 90 s), but very low sensitivity (less than 40%). In addition, the PPV could not be computed for some of the combinations of features, because the classifier never identified a true positive.

The fact that the combination of LPC and MFCC consistently provided the best accuracy in determining whether the event was either a choke or noise/snore, regardless of window size, indicates that the LPC coefficients and the MFCCs provided complementary information. The addition of demographics also improved performance. It is worth noting that less data were used at 3 s than at 0.5 s (decreased from 391 noise/snore events, 175 chokes at 0.5 s to 326 noise/snore, 82 chokes at 3 s). The ratio between the classes has completely changed at 3 s.

There are a number of limitations to this approach. Annotating the data is labor intensive and, ideally, there should be three annotators to ensure the quality of the annotation. In this work, one clinical research fellow with 2 years of training labeled the data, which is not optimal.

Assuming that the data were recorded in a low noise environment, the performance achieved is insufficient for screening subjects. Even with the performance metrics as good as they are here, the accuracy needs to be much higher (high 80%) for this approach to be clinically acceptable (Collop et al., 2007), although the ability of portable monitors to correctly diagnose subjects with OSA needs further validation.

The results presented here are lower than those reported in the literature [Se = 88% and Sp = 82% for LPC (Ng et al., 2008a); Se = 82% for MFCCs (Cavusoglu et al., 2007)]. However, many approaches consider in-sample classification which leads to a significant over-estimation of out-of-sample performance. In addition, the literature considers a different classification problem, i.e., thresholding on a given feature to differentiate apneic snores from benign snores, whereas, in the analysis above, the first breath after an apnea has been differentiated from benign snores and noise, which will also influence performance. The poorer performance could be due to the lack of annotations in this analysis, resulting in an under-specified system. In addition, none of the snoring events used in this analysis have been graded. The availability of a grading may have improved performance. It is worth noting that the classification performance is on the classes of sounds themselves, and not for a subject overall.

Finally, the need for an event detector, and then downstream classification to turn the classified events into a diagnosis are likely to degrade the performance further.

In contrast to the speech-based approaches and events classification, the MSE approach yielded a much higher Se, PPV, and AUC with marginal drops in Ac and Sp. Moreover, the addition of demographics made little difference to the classification performance, indicating that the technique may be more suitable as a general tool. The use of an RF applied to MSE coefficients requires very little preprocessing and no segmentation or event detection. While the sensitivity does not surpass 70%, this technique may be suitable as a post-screening test after screening with a pulse oximeter, particularly in settings where there is significant noise from external sources.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

This research was supported by the RCUK Digital Economy Programme grant # EP/G036861/1 (Oxford Centre for Doctoral Training in Healthcare Innovation) and the University of Oxford Sleep and Circadian Neuroscience Institute (SCNi), Centre Grant # 098461/Z/12/Z.