Interdisciplinary Approaches to Automated Obstructive Sleep Apnea Diagnosis Through High-Dimensional Multiple Scaled Data Analysis

Abstract

Obstructive sleep apnea (OSA) is a wide-spread condition that results in debilitating consequences including death. Diagnosis is a lengthy and expensive process because OSA is a multifactorial disorder, making it necessary to study many different types of data, including DNA sequences, multiple time series, metabolites, airflow in airway, and shape analysis of airway and patients’ faces. OSA data are an example of complex and multi-dimensional data for which analysis and interpretation can be challenging, requiring sophisticated analytic techniques. It may be no longer effective to independently apply methods from a specific discipline such as statistics, mathematics, or computing science. In this article, combining the analyses of three datasets from independent OSA studies, we illustrate the complementary nature of the techniques. Specifically, we apply techniques in statistics, machine learning, geometry, and computational topology to derive automated analytic tools for each data type. Taken together, these techniques provide a sophisticated diagnostic tool. A novel geometric OSA severity index (GSI) is developed using methods from computational geometry. This index measures the volume of the airway obstruction in OSA patients. The lower the GSI value is, the more severe the airway obstruction is. Persistent homology is employed to extract the importance information from 28-dimensional polysomnography (PSG) data. Random forests and principal component analysis are used and compared to identify important variables in the PSG, while logistic regression and random forest are used and compared to verify the prediction power of the identified variables. The results indicate that persistent homology can accurately extract importance information from PSG, and the identified important variables are meaningful for predicting obstructive apnea–hypopnea index (ahi). Cluster analysis is used to identify the pattern of the survey information, and the importance of responses to individual questions in survey questionnaires is also identified by random forest. The results from all three independent studies are very meaningful in clinical studies and can be used as guidance for clinical practitioners.

Appendix

The appendix is a brief description about polysomnography and its signals. Polysomnography is a multi-parametric test used in the study of sleep and as a diagnostic tool in sleep medicine. The test result is called a polysomnogram also abbreviated PSG. PSG is a comprehensive recording of the biophysiological changes that occur during sleep. It is usually performed at night, and in some special cases, it could also be done during the day time. The PSG monitors many body functions including brain (electroencephalography or EEG), eye movements (electrooculography or EOG), muscle activity or skeletal muscle activation (electromyography or EMG), and heart rhythm (electrocardiography or ECG) during sleep. In the 1970s, the sleep efficiency and duration, sleep stages, apnea–hypopnea index, oxygen saturation, carbon dioxide level, sleep stage changes, spontaneous arousal index breathing functions respiratory airflow, and respiratory effort indicators were added to PSG records together with peripheral pulse oximetry [32]. Basically, polysomnography records a lot of time series associated with human sleep and provides rich information about the quality of sleep. Each channel is a time series. Figure 14 shows how typical PSG data looks like. There are several channels in the PSG and each channel is a time series recorded by the units of 10 s. During the whole sleeping period (9.5–10 h often), there are millions of time points recorded and for each participant, their PSG data are multivariate time series with millions of time points. Figure 14 is from the NSRR website.

Fig. 14

The figure displays the different channels in PSG. There are several channels in the PSG and each channel is a time series recorded by the units of 10 s. During the whole sleeping period (9.5–10 h often), there are millions of time points recorded for each participant. Figure is taken from the NSRR website

Particularly, in our study, each of the 100 participants with their PSG recorded has 28 signals in their PSG, namely, electroencephalography (EEG, which has 4 channels of signals, namely, C3, C4, A1, and A2), left outer canthus (LOC), right outer canthus (ROC), electrocardiogram (which has two signals, namely, ECG1 and ECG2), LEFT LEG1, LEFT LEG2, RIGHT LEG1, RIGHT LEG2, electromyogram (which has three signals, namely, EMG1, EMG2, and EMG3), airflow via thin catheters placed in front of nostrils and mouth (AIRFLOW), absence in the effort in the thoratic (THOR EFFORT), absence of effort in the abdominal (ABDO EFFORT), snoring (SNORE), sum channels (SUM), body position (POSITION), oxygen saturation (OX STATUS), pulse oximetry (PULSE), oxygen level (SpO2), light, heart rate (HRate), plethysmography (Pleth WV), and nasal pressure (NASAL PRES).

We write the definition of respiratory events. Respiratory events were scored if they were at least 8 s long, which represents at least 2 missed respiratory cycles at this stage. Obstructive apneas were scored when chest and abdominal efforts were asynchronous and estimated tidal volume was < 25% of baseline, irrespective of associated desaturation. Hypopneas were scored when respiratory efforts were accompanied by a 25–50% reduction in estimated tidal volume and accompanied by at least 3% oxyhemoglobin desaturation or when clearly discernible decreases in estimated tidal volume were associated with similar desaturation. Central apneas (absent effort in both channels) were excluded from sleep-disordered breathing indexes.