Keywords

5.1 Diagnostic Procedures

5.1.1 History and Physical Examination

The diagnosis of OSA starts with a sleep history, which commonly gathers data from three complementary approaches: (1) evaluation of OSA-related symptoms, (2) sleep habit evaluation, and (3) comprehensive evaluation of high-risk patients due to potential negative consequences. OSA is a complex disease and thus patients can manifest a large spectrum of different symptoms. Commonly, adult patients with OSA report loud snoring, witnessed apneas, awakenings with choking sensation, and excessive daytime sleepiness, the latter usually assessed by the Epworth Sleepiness Scale (ESS) [1]. Other common symptoms are nonrefreshing sleep, tiredness, morning headaches, and neurobehavioral problems [2]. On the other hand, a less common phenotype includes insomnia and fatigue as predominant symptoms, which is more particularly encountered in women. It is important to highlight that, due to subjectivity, there is a great variability among patients concerning the prevalence of symptoms, particularly sleepiness [3]. In children, the emergence of symptoms is much more subtle, even if younger patients are more susceptible to develop morbid consequences. As such, snoring may be the only symptom that parents may report, or snoring may be apparent along with snorting or gasping episodes, restless sleep, bedtime wetting, frequent nightmares or night terrors, bruxism, morning headaches, grumpy awakenings, tiredness, falling asleep in school or in the car, and problematic progression in academic performance at school. A particular feature of children who are at risk for sleep-disordered breathing is the presence of inattention with or without hyperactivity, the latter being much more frequent among non-obese children. As such, the concept of different phenotypes based on clinical presentation has been proposed to include type I (non-obese, hyperactive) and type II (obesity and excessive daytime sleepiness with inattention). Furthermore, social withdrawal and aggressive or oppositional defiant behaviors are also common.

Besides symptoms, major indicators of OSA include hereditary factors (family history of the disease) and known anthropometric factors linked with increased risk for upper airway obstruction, such as oropharyngeal and nasal abnormalities, large neck circumference, increased body mass index (BMI), or additional markers of obesity [2, 4]. In addition, cardiorespiratory, cerebrovascular, and metabolic comorbidities potentially linked with OSA should be investigated [5]. Table 5.1 summarizes sleep assessment procedure, whereas Table 5.2 shows criteria for OSA diagnosis according to the American Academy of Sleep Medicine (AASM) [6].

Table 5.1 Sleep evaluation in the diagnosis of OSA [4]
Full size table
Table 5.2 Criteria for OSA diagnosis [6]
Full size table

For children, it is important to follow the guidelines as formulated by various professional societies [7,8,9]. More generally, adoption of the concept of habitual snoring (defined as snoring recognized by caretakers at least 3 nights per weak, particularly if loud) that also exhibits an additional problematic symptom is an excellent rule of the thumb to make the decision to evaluate further with some objective diagnostic testing.

5.1.2 OSA Screening Questionnaires

Major advantages of questionnaires are readiness and negligible cost. However, clinical questionnaires are not recommended as single tools for OSA diagnosis due to their limited accuracy. Nevertheless, despite their low specificity (large number of false-positive cases) [10, 11], standardized questionnaires show overall high sensitivity, and thus they can be useful for screening purposes. Accordingly, questionnaires can be easily performed in clinical practice to identify patients with higher risk for OSA both in primary care and in specialized care. Some authors suggest their usefulness in low-income countries where PSG is not available [12].

There are several standardized questionnaires aimed at assisting physicians in the detection of OSA. The Berlin questionnaire (BQ) [13], OSA50 [14], and STOP-BANG [15] are probably the most popular, while No-apnea, GOAL, NoSAS, and Obstructive Airway Adult Test (OAAT) [16] have been also widely assessed [17,18,19,20,21,22,23,24,25,26]. There are few studies comparing the diagnostic ability of these questionnaires with each other, and some inconsistencies have been reported [12, 27]. Table 5.3 shows the most widely used questionnaires.

Table 5.3 Main questionnaires used for identifying patients with a high pretest probability of OSA [10]
Full size table

In children, a large number of questionnaires have also been developed over the years with overall remarkable similarity in their performance even if some are better than others. These questionnaires for the most part exhibit excellent sensitivity but relatively lower specificity, making them robust tools for screening purposes but not accurate enough to be used as an alternative to a diagnostic test [28,29,30,31,32,33].

More complex multivariate predictive models for OSA have been proposed, including a variety of additional symptoms (infections, high blood pressure, attention deficits) and measures (palatal height, maxillary, and mandibular intermolar distances). Despite these models demonstrated higher ability to discriminate OSA, particularly for higher diagnostic thresholds for the disease, they still show a great imbalance between sensitivity and specificity, leading to a marginal use in clinical practice [23].

5.1.3 Comprehensive Sleep Studies

History and questionnaires provide physicians with relevant information on the presence or absence of the disease. Nevertheless, definitive diagnosis of OSA must be confirmed using appropriate devices for comprehensive sleep testing [10]. Table 5.4 summarizes the four traditional categories of devices for sleep analysis according to the complexity (number of physiological signals involved) and the setting (supervised vs. unattended) [34]. These approaches are applicable to both adults and children with some caveats that will be addressed below.

Table 5.4 Main characteristics of equipment used in the different approaches for adult OSA diagnosis according to AASM [10, 34]
Full size table

5.1.3.1 Polysomnography

In-laboratory polysomnography (PSG) is the gold standard method for OSA diagnosis. PSG provides essential information on the duration and quality of sleep, sleep stages, and transient events (electroencephalographic, muscular, cardiac, and respiratory events) during sleep. Current AASM guidelines [35] delineate scoring and interpretation rules as well as technical aspects concerning PSG performance and analysis based on cumulative published evidence.

Electroencephalographic channels and associated recordings (electrooculogram and electromyogram) are an essential part of a PSG. EEG, EOG, and EMG signals are used to analyze the macro- (sleep stages) and micro- (electroencephalographic events) structure of sleep. Table 5.5 summarizes the main characteristics of sleep stages and common electroencephalographic events [35].

Table 5.5 Main characteristics of states and electroencephalographic events during sleep [35]
Full size table

Cardiorespiratory recordings involving airflow, respiratory movements, electrocardiogram, pulse oximetry, and body position compose the second main part of a PSG, aimed at characterizing and quantifying respiratory events. Monitoring snoring and hypoventilation are optional in adults but is required in children, particularly assessment of alveolar hypoventilation since many children may manifest sleep-disordered breathing almost exclusively by developing extended periods of elevated carbon dioxide levels during sleep. Table 5.6 shows the types and definition of polysomnographic respiratory events involved in the diagnosis of sleep-disordered breathing in adults [35]. Additional physiological recordings have been proposed to be included in the PSG, such as the divided nasal cannula [36]. However, no improvement was found in terms of the accuracy of AHI.

Table 5.6 Scoring criteria for respiratory events in adults [35]
Full size table

In order to obtain reliable and confident diagnosis, the AASM recommends a minimum of 4 h of total sleep time (6 h for children) for high-risk patients undergoing PSG, while this limit increases to 6 h (and >16% of total sleep time REM sleep in children) in clinical trials in order to account for a potentially broader spectrum of OSA severity [3].

OSA severity is categorized according to the AHI, which accounts for the combined number of apneas and hypopneas per hour of sleep. Although the most appropriate diagnostic threshold for positive OSA has been under great discussion for years, currently, the following severity degrees are widely accepted in adults [10]: mild OSA, 5 ≤ AHI < 15 events/h; moderate OSA, 15 ≤ AHI < 30 events/h; and severe OSA AHI ≥ 30 events/h. Because the collapsibility of the upper airway is lower in children than in adults, the number of respiratory events in healthy children is <0.5 events/h of sleep. Accordingly, the criteria for OSA severity have traditionally been adopted as follows: >1–5 events/h corresponds to mild OSA; >5 to 10 events/h represent moderate OSA and >10 events/h is viewed as severe OSA in the pediatric age. However, for children older than 14 years of age, application of adult or pediatric OSA severity criteria is at the discretion of the physician.

Despite been the standard index to diagnose OSA, the AHI does not completely reflect the complex heterogeneity of the disease. Its main drawback is related to its inability to differentiate among the clinical phenotypes of OSA [3] or predict all negative consequences of the disease [2]. Similarly, it has been found that different measures of hypoxia are better predictors of outcomes and adverse events than the AHI [37]. However, notwithstanding the limitations of AHI regarding more precise patient phenotyping, severity-dependent associations between AHI and some of the common morbid consequences of OSA have been detected despite the marked heterogeneity of morbid phenotype [38], whereby other morbidities do not exhibit significant associations with AHI [39].

In the same regard, PSG has widely known limitations. It is complex, time-consuming, and intrusive for patients. In addition, regarding equipment and specialized staff, PSG is costly leading to limited accessibility and availability. Therefore, due to the high prevalence of the disease worldwide, it is not a cost-effective approach to evaluate all patients suspected of suffering from the disease. Accordingly, simplified alternatives are being sought, particularly to conduct portable sleep studies at home and under the lead assumption that such transition to home will not curtail the accuracy and reliability of the alternative tests.

5.1.3.2 Home Sleep Apnea Testing

In recent years, abbreviated and portable at-home sleep apnea testing (HSAT) has emerged as an alternative or complementary diagnostic tool for OSA. There is a great number of commercial devices currently available. In 2007, the AASM stablished the characteristics and minimum requirements of HSAT devices [40]. Their use as an alternative to standard PSG was subsequently accepted only in uncomplicated patients showing high pretest probability for the disease [18, 40].

Concerning the physiological signals recorded by abbreviated HSAT devices, three signals are considered relevant and widely recommended: blood oxygen saturation from oximetry, respiration from nasal pressure or alternatively respiratory bands, and heart rate from ECG or from the pulse oximetry signal.

The implementation of HSAT procedures differs among countries. In Europe, simplified portable monitoring has been widely accepted, firstly supervised in the hospital setting and then unattended at home [41]. Similarly, up to 40% of total patients are referred to HSAT in Canada [42]. In contrast, routine HSAT use in the patient’s home in the USA has been slower to be adopted due to some regional limitations related to reimbursement and personnel structure. As no EEG is recorded, abbreviated HSAT devices do not provide information about sleep, nor do these recordings provide sleep time or arousals, which limits their ability for ruling out the disease [10]. Therefore, the numerator of events does not include arousals and the denominator is larger since it denotes total recording time (TRT), while PSG actually includes total sleep time (TST; obviously TRT > TST), such that unless the disease is more severe, there can be clear underestimations or mis-estimations in any given patient. Indeed, a remarkable overall underestimation (high false-negative rate) has been consistently reported [43,44,45]. Accordingly, in-lab PSG or further follow-up should be considered when HSAT results are negative. Therefore, the cost associated with repeated test due to false-negative results (17%) [40] and technical failures (18%) [46] must be taken into account [47]. The main limitations of HSAT are summarized in Tables 5.7 and 5.8. These limitations of HSAT are further magnified in pediatric settings, because the severity criteria are much more stringent and the numerator to calculate the AHI is much smaller [50].

Table 5.7 Limitations of HSAT leading to OSA underestimation [48]
Full size table
Table 5.8 Factors affecting suitability of HSAT [49]
Full size table

Conversely, the main advantages include less cost and less intrusiveness for patients. Particularly, higher cost-effectiveness ratio of HSAT compared to in-lab PSG has been reported in appropriately selected patients [51,52,53,54,55]. In regard to patient’s comfort, contradictory data have been recently reported [56, 57]. “Real-world” studies are needed to prospectively assess preferences and satisfaction of patients with HSAT. Additionally, reduced workload in sleep laboratories and reduced waiting times for diagnosis and treatment are commonly reported benefits for both specialists and patients. Finally, HSAT enables multiple-night assessment at a reasonable cost, minimizing the confounding effect of the widely known night-to-night variability of OSA [3].

Concerning the different types of HSAT approaches (Table 5.4), comparable clinical (BMI, ESS, blood pressure) and treatment (CPAP adherence) outcomes have been reported in OSA patients diagnosed using ambulatory PSG (type II) compared to in-hospital standard diagnosis [58]. Despite more limited cardiopulmonary data, respiratory polygraphy (RP) (type III) is able to expedite diagnosis [53]. Similarly, it has been found useful to increase availability of both diagnosis and treatment resources in patients with mobility issues and major illness [10]. A major meta-analysis on the effectiveness of type III monitors reported good diagnostic performance (AUC ranging from 0.85 to 0.99) for the common OSA severity cutoffs and no significant differences in clinical management parameters between patients diagnosed using either PSG or type III monitoring [53]. In the context of children, the jury is still out and the consensus at this stage pending additional large-scale studies is still in favor of minimizing the use of HSAT and preferably conducting PSG unless the regional circumstances and resources lead to substantial delays in diagnosis [59,60,61].

Accurateness and reliability of more simplified devices (type IV) are still under debate, particularly single-channel blood oxygen saturation (SpO2) from oximetry. Respiratory disturbance index (RDI) from RP and oxygen desaturation index (ODI) from oximetry have been found to provide similar predictions of actual AHI, showing small differences compared to the known night-to-night variability [62]. Nevertheless, the AASM demands further evidence on the efficiency of unattended oximetry and similar single-channel approaches as reliable tools for OSA diagnosis (currently recommended only for screening purposes), particularly in mild or asymptomatic patients and in the presence of other sleep disorders (central sleep apnea) or significant concomitant diseases [10]. Concerning cardiorespiratory comorbidities, several studies have assessed the usefulness of simplified screening tests for OSA in patients with hypertension [63], heart failure [64], and stroke [65]; in the presence of COPD [66, 67] or morbid obesity [68]; in patients remitted for surgery [69]; and in hospitalized elders [70].

The interest on oximetry recently increased due to its ability to characterize intermittent hypoxia. Several studies reported that hypoxia level correlates with mortality, cardiovascular outcomes, and cancer incidence better than the standard AHI [3, 71,72,73]. In this regard, novel oximetric indices have been proposed to parameterize intermittent hypoxia linked with OSA, such as the hypoxic burden [73], the hypoxia load [71], or the desaturation severity parameter [74].

A similar evolutionary trend has occurred in the context of pediatric settings. The Oxygen Desaturation (≥3%) Index (ODI3) and McGill Oximetry Score (MOS) are widely used as predictors of moderate-to-severe OSA (apnea-hypopnea index-AHI >5 episodes/h), an indication for adenotonsillectomy [75, 76]. In a recent study, we have shown that the optimal cutoff values for the ODI3 and MOS were ≥4.3 episodes/h and ≥2, respectively. Nevertheless, the ODI3 emerged as the better performing index for detecting moderate-to-severe OSA in habitually snoring children when PSG is not available [77].

5.1.3.3 Blood Biomarkers

Sleep apnea is known to trigger mechanisms leading to metabolic and endocrine dysfunctions, such as inflammation, hypoxemia, and oxidative stress, which can be readily detected by blood tests. Accordingly, blood biomarkers have been proposed to enhance the diagnosis of OSA [10, 78]. Common biomarkers are glycated hemoglobin (HbA1c), C-reactive protein (CRP), erythropoietin (EPO), interleukin-6 (IL-6), and uric acid [78]. Similar efforts in children have yielded remarkable accuracy under specific circumstances in the detection of disease, in the identification of residual OSA after treatment, or in the evaluation of comorbidities associated with OSA [79,80,81,82,83,84,85].

Some biomarker combinations have been found to correlate with sleep apnea severity, emerging as potential user-friendly and low-cost tools for OSA screening in high-probability patients [86,87,88,89,90]. Nevertheless, further research is encouraged in this topic to obtain successful diagnosis and treatment strategies, mainly by means of big data and machine learning techniques [3, 91].

5.2 Diagnostic Strategy

As mentioned above, attended in-lab overnight PSG is the gold standard tool for OSA diagnosis. Nevertheless, due to the widely known drawbacks of PSG and the increased frequency of visits linked with suspected sleep apnea, the aforementioned simplified methods have been implemented within the diagnostic protocols in order to conduct sleep studies at home, reduce costs and waiting lists, and expedite both diagnosis and treatment. Figure 5.1 shows a flowchart with the diagnostic strategy for OSA detection according to the current evidence and recommendations.

Fig. 5.1
A decision flowchart. It is divided into 3 areas titled primary care, home, and hospital. The components are history and physical examination, significant comorbidities, H S A T, mild, moderate, and severe O S A, and in-lab P S G, negative, mild, moderate, or severe O S A.

Flowchart summarizing the diagnostic strategy for OSA detection involving primary care, HSAT and in-laboratory PSG

Full size image

The main recommendations to conduct standard PSG in a sleep laboratory are:

  • Patients with significant comorbidities (such as severe pulmonary disease, congestive heart failure, or neuromuscular weakness)

  • Patients with risk of central or mixed sleep apnea or other sleep disorders

  • Patients with low probability of OSA

  • Pediatric patients except for resource-constrained settings

Currently, according to the AASM, HSAT is a suitable alternative to in-lab PSG for the diagnosis of OSA in patients showing moderate-to-high pretest probability without certain comorbidities [10]. Positive OSA is diagnosed in patients showing symptoms and AHI/RDI/REI ≥5 events/h from PSG or HSAT or, alternatively, AHI/RDI/REI ≥15 events/h regardless of symptoms (Table 5.2).

The scarcity of sleep specialists and sleep units forces primary care to increase its role in the detection of OSA [92]. In this framework, general practitioners have two possible pathways to manage patients in order to speed up diagnosis and treatment. Patients scoring positive using a validated screening questionnaire (either BQ positive, OSA50 ≥5, or a STOP-BANG cutoff of ≥4) and showing ESS ≥8 are assigned moderate-to-high risk for OSA and referred directly for in-lab PSG (complex patients) or HSAT (uncomplicated patients) [93]. It is important to point out that patients with negative scores in the proposed questionnaires should not be definitively excluded for OSA [93]. In this situation referring patients to a sleep unit will be required.

Alternatively, other authors proposed a two-step screening strategy involving questionnaires and simplified HSAT in primary care. In a first stage, a battery of sleep questionnaires is applied. Then, abbreviated HSAT approaches can be used in the second stage, e.g., single-channel nasal airflow or pulse oximetry. This two-step approach seems to increase specificity becoming very useful to rule out OSA in primary care while not overstretching sleep units [94, 95].

Application of similar strategies with specific caveats that are exclusively applicable to children are being developed and will hopefully become incorporated into consensus guidelines in the few years to come.

5.2.1 Impact of COVID-19 Pandemic on Diagnostic Strategy for OSA

The coronavirus disease 2019 (COVID-19) significantly affected healthcare systems worldwide [96, 97]. Social distancing, confinement of population, and additional countermeasures aimed at minimizing virus spread have changed protocols of all medicine disciplines, including sleep laboratories. Particularly, face-to-face in-hospital visits for PSG involved an increased risk for both patients and sleep-related professionals. Moreover, domiciliary approaches also carry infection risk.

Accordingly, during a pandemic, sleep experts strongly recommended to stop performing all kinds of sleep studies, except to hospitalized patients showing high risk of sleep apnea that could worsen their health status [98]. Table 5.9 summarizes main recommendations when conducting sleep studies in an epidemic context. After the study, the equipment must be disinfected or undergo quarantine if used by infected patients [98]. The epidemic also affects treatment management due to the risk of viral shedding via mask leakage in CPAP devices [99].

Table 5.9 Recommendations for in-laboratory sleep assessment during a pandemic
Full size table

Predicting how this epidemic will finally influence protocols and strategies for management of the disease is difficult. Nevertheless, e-health and HSAT approaches will probably arise as suitable tools to minimize face-to-face diagnosis and follow-up visits [100].

5.3 Decision Support Systems in OSA Diagnosis

Physicians must deal with several data sources to reach a definitive diagnosis of OSA, including symptoms, anthropometric and clinical variables, and data from biomedical recordings. In this regard, PSG is the main source of information, as it allows physicians to score major electrophysiological and cardiorespiratory events in order to quantify the AHI. Nevertheless, many long-term recordings compose each PSG and many and complex rules must be taken into account to score each type of event. In this context, computer-aided diagnosis systems have been developed as valuable tools that are able to assist physicians in the laborious diagnosis of the disease. Particularly, artificial intelligence has been found to expedite interpretation of physiological signals by decreasing heterogeneity and subjectivity in the PSG scoring task, as well as decreasing the high workload for sleep experts. In fact, the AASM Foundation identified artificial intelligence as a strategic area in which research focused on according to its 2020 Strategic Research Award program.

Figure 5.2 shows how artificial intelligence algorithms are usually used to assist in OSA diagnosis. Regarding conventional machine learning methods, artificial neural networks and support vector machines have been predominantly used due to its nonlinear architecture and generalization ability, able to model real-world complex data with fast and effective processing [101]. In the last several years, ensemble learning techniques, based on the combination of multiple classifiers in order to increase prediction ability, outperformed previous approaches in this context [102].

Fig. 5.2
2 vertically arranged blocks are connected by an arrow. 1. Sleep study. It illustrates sleep staging and respiratory events scoring on a polysomnography and graphs for E C G, E E G, A F, and S p O 2. 2. Artificial intellegence. It illustrates machine, ensemble, and deep learning.

Artificial intelligence-based approaches to formulate decision support systems for OSA diagnosis

Full size image

Deep learning is currently replacing conventional machine learning approaches in several fields of industry and medicine, including sleep research. Deep learning minimizes major shortcomings of classical pattern-recognition algorithms based on feature engineering, which could potentially remove relevant information when composing the initial feature space of input variables [103]. Deep learning models use all the data contained in the signal and compose their own internal features, thus maximizing the diagnostic ability of the signal. Deep neural networks, particularly convolutional neural networks (CNNs), are the most widely used deep learning approach. Similarly, auto-encoders, deep generative models, and recurrent neural networks are also deep learning methods able to characterize biomedical signals [104].

The two main tasks focus the efforts of researchers and developers in the context of decision support systems for OSA management, namely, the development of automated algorithms for sleep staging, aimed at analyzing the macro- and micro-structure of sleep, and automated analysis of cardiorespiratory recordings, in order to obtain simplified and reliable diagnostic tests for OSA. Table 5.10 shows the main tasks involving artificial intelligence in the context of OSA diagnosis.

Table 5.10 Tasks commonly addressed by means of machine learning and pattern-recognition techniques
Full size table

5.3.1 Automated Sleep Staging from Polysomnography

The interpretation of a PSG is a cumbersome and time-consuming task even for trained experts. Artificial intelligence provides reliable and accurate algorithms to analyze both the macro-structure (sleep/wake time, sleep stages) and the micro-structure (transient events, such as arousals and spindles) of sleep. Particularly, spectral analysis, nonlinear methods, and patter recognition techniques have been found to provide essential information on electroencephalographic dynamics during sleep and sleep-related breathing disorders [102, 105]. Commercial software by major medical companies in the framework of PSG analysis already implements automated tools for sleep staging. Nevertheless, scoring neuromuscular signals is really challenging for automated algorithms, and accuracy greatly depends on the number of categories (sleep stages) involved in the classification task. Furthermore, performance is also influenced by the type and number of biomedical recordings used (EEG channels and EOG/EMG).

Concerning pattern recognition, according to exhaustive comparisons, a combination of complementary signal-processing approaches including wavelet analysis together with random forest, which is a kind of ensemble learning classifier, is able to reach high efficiency in sleep-staging tasks [106,107,108]. Recently, a number of studies applied deep learning techniques to accomplish sleep-staging both in healthy subjects and patients with suspected sleep disorders, reporting accuracies ranging from 67.4% to 92.2% [109,110,111,112,113,114,115,116,117].

5.3.2 Automated Diagnosis of Sleep Apnea

Decision support systems in the context of OSA diagnosis focus on two main goals: (1) providing a reliable and accurate diagnosis using the least amount of signals, promoting the use of simplified portable monitors able to increase availability and accessibility to sleep-related diagnostic resources, and (2) providing physicians with automated tools able to decrease the high workload linked with scoring respiratory events.

Regarding abbreviated tests for OSA, the scientific community has focused on the analysis of a reduced subset of cardiorespiratory signals, mainly heart rate variability (HRV) from electrocardiogram (ECG), SpO2 from oximetry, and airflow (AF). Two approaches are commonly used to characterize changes in these biomedical recordings: time-domain and frequency-domain analyses [118, 119]. In addition, as biological systems have major nonlinear interactions, nonlinear analysis has demonstrated to provide additional and complementary information to conventional linear and frequency-domain methods [120,121,122,123,124]. Similarly, wavelet transform (time-scale analysis) and bispectrum (high-order spectra) have been proposed as reliable alternatives to conventional Fourier analysis in the context of OSA [118, 119, 121].

In order to implement computer-aided diagnosis systems for OSA, binary and multiclass classifiers as well as regression models have been developed (Table 5.10). Classical statistical methods, such as linear (LDA) and quadratic (QDA) discriminant analysis and logistic regression (LR), reached remarkable diagnostic performance in binary classification tasks, whereas more complex machine learning methods including decision trees, neural networks, and support vector machines have demonstrated its usefulness in multiclass and regression problems [101, 118, 119, 125,126,127,128]. Recently, ensemble learning and deep learning techniques have been found to outperform conventional machine learning techniques in the framework of OSA diagnosis [104, 129,130,131].

5.3.3 Digital Health Technologies in the Management of Sleep Apnea

In the last years, e-Health systems and mobile (m)-health applications through smartphones raised as reliable tools for unattended portable testing and therapy monitoring in the context of OSA management [132, 133]. The increasing recognition of good sleep quality as synonym of health and the need of periodic follow-up in chronic sleep-disordered breathing foster the usefulness of telemedicine in this framework. In this regard, the AASM points out the usefulness of telemedicine for improving accessibility to and availability of sleep experts and sleep-related healthcare resources [134].

Regarding high-quality medical applications for clinical practice, e-health/m-health tools mainly focus on long-term therapy tele-monitoring, while remote PSG or alternative simplified sleep studies at home are less frequent [135,136,137]. On the other hand, the vast majority of smartphone-based applications focus on wellness and lifestyle monitoring for nonspecialized general use, mainly for sleep tracking and sleep quality assessment with a lack of appropriate scientific validation [138,139,140]. Concerning suitable assessment, the FDA recommends validation against PSG since it is currently the gold standard method for OSA diagnosis. In this regard, overall, sleep-related parameters provided by mobile Apps correlate poorly with PSG and further improvement is needed [141].

While recent advances focus on simplifying devices to increase portability, accessibility, and comfort, experts point out the lack of direct measures of brain activity, the need for estimating total amount of sleep, and measuring the duration of respiratory events to obtain a reliable and accurate diagnosis of OSA [3].

5.3.4 Advantages and Limitations of Artificial Intelligence in the Management of OSA

Sleep medicine, particularly for sleep-related breathing disorders, can significantly benefit from artificial intelligence and big data techniques due to the huge volume of data involved in the process of diagnosis, including not only PSG but also data from clinical history, anthropometric variables, biomarkers, or genetics, among others. In the last decade, conventional machine learning techniques demonstrated their usefulness in addressing problems related to automated classification of sleep stages, respiratory events, and, overall, sleep apnea severity [142]. In addition, current increased computational capabilities have led novel deep learning algorithms to outperform stablished limits, reaching correlations above 0.95 in the estimation of the AHI [143] and multi-class kappa above 0.80 in sleep staging [113, 144].

Beyond particular sleep staging and respiratory event scoring tools, artificial intelligence is expected to enhance therapy management and patient outcomes in order to reach the so-called personalized medicine in the context of sleep apnea [142, 145]. In this regard, machine learning has been recently proposed to identify the most appropriate diagnostic pathway for OSA patients. In the work by Stretch et al. [146], a predictive model is trained to decide which patients are referred directly to in-hospital PSG and which ones are more likely to benefit from abbreviated sleep testing at home. Nevertheless, as pointed out in a recent report by the AASM, artificial intelligence is not going to replace sleep experts but to augment efficiency and accurateness of sleep laboratories [147].

On the other hand, common drawbacks of machine learning and artificial intelligence are limited generalizability; output variables and performance metrics difficult to interpret, potentially leading to loss of relevant information; lack of transparency of proprietary algorithms, leading to the so-called black boxes; and subjectivity of standard rules used to label training samples, as automated learning mostly relies on supervised training.

In addition, some major challenges must be still overcome in order to generalize the use of expert systems in clinical practice, such as regulatory laws concerning software certification and logistics linked with computer support, as well as ethical and legal issues [148]. Regarding ethics, it is important to highlight that the clinician is solely responsible for the decisions related to patient diagnosis and treatment.