A clinical risk matrix for obstructive sleep apnea using Bayesian network approaches

The substantial medical, social, and economic consequences of untreated obstructive sleep apnea (OSA), the overwhelming number of patients who have escaped clinical detection, and the likelihood of successful treatment strongly justify screening [6]. In this clinical outcome, diagnostic models need to have a high sensitivity, as false negatives should be avoided, to prevent excluding patients with moderate or severe OSA from performing polysomnography (PSG) [24, 34], the standard test for OSA final diagnosis.

This study aimed to define auxiliary diagnostic methods that can support the decision to perform PSG, based on risk and diagnostic factors by means of interactive models or risk matrix. The secondary objectives were to describe the population; develop and validate a Bayesian network-based risk matrix in the study population; optimize the need to perform PSG and produce a Bayesian network model for daily use in the primary care setting.

This paper is organized as follows. The background section exposes the related work on the theme. Following section presents the research methodology. Section 4 gives an overview of the achieved results, and Sects. 5 and 6 interpret the results of the work and provide the main findings and recommendations for the work.

Apnea is defined as the complete cessation of airflow for at least 10 s, while a hypopnea is a reduction in airflow (30–50%) that is followed by an awakening or a decrease in oxyhemoglobin saturation (3–4%) [6, 22]. There are 3 types of apneas: central, mixed and obstructive. Central sleep apnea is a reduction in the respiratory effort resulting in reduced or absent ventilation, while mixed apnea begins with central apneas that leads to obstructive events [5, 22]. In OSA, respiratory effort is maintained but ventilation decreases or disappears because of partial/total occlusion of the upper airway [6, 22, 23, 28, 32].

OSA severity is assessed with apnea–hypopnea index (AHI), obtained through PSG, which is the number of apneas and hypopneas per hour of sleep [22]. Recommendations from the American Academy of Sleep Medicine state that OSA is present when AHI \(\ge \) 5. It can be classified as mild (AHI: 5–15), moderate (AHI: 15–30), or severe (AHI \(\ge \) 30) [6, 7, 13, 22]. Approximately 30% of the general public is affected by a significant sleep problem, often of long standing [39]. OSA affects about 4% of men and at least 2% of women worldwide [5, 7, 13, 18, 39]. The signs, symptoms, and consequences of OSA are a direct result of upper-airway repetitive collapse. This leads to sleep fragmentation, hypoxemia, hypercapnia, marked swings in intrathoracic pressure, and increased sympathetic activity.

Reported risk factors for developing OSA include different groups of variables, such as demographic, clinical history, comorbidities and even factors collected during the consultation, for example, male gender [1, 10, 16, 20‐22, 27, 30, 36, 42‐45], aging [1, 16, 20‐22, 27, 30, 36, 42‐45], obesity [1, 10, 12, 16, 20‐22, 30, 36, 42, 43, 45], history of smoking [1, 10, 12, 22, 30, 45], increased neck [1, 10, 16, 20‐22, 27, 30, 36, 43, 45] and abdominal [30, 45] circumferences, arterial [1, 12, 16, 20, 22, 30, 36, 43‐45] and pulmonary [12, 30] hypertension, atrial fibrillation [30, 36], stroke [1, 22, 30, 45], myocardial infarction [1, 30, 45], and high-risk driving populations (such as truck drivers) [10, 12, 30, 42].

Diagnostic criteria for OSA are based on clinical signs and symptoms determined during a sleep consultation, which includes a sleep oriented history, physical examination, and findings identified in an objective exam [13, 28]. It should also include an evaluation, for example, of snoring, witnessed apneas, gasping/choking episodes, daytime sleepiness severity with the Epworth Sleepiness Scale (ESS), nocturia, and morning headaches [22]. The diagnostic methods available are PSG and home testing with portable monitors (tends to underestimate the severity) [3, 6, 7, 13, 18, 22, 39]. PSG is time consuming, labor intensive, limited to urban areas, costly and faces long waiting lists [18, 22], so many studies have been trying to tackle the problem that comes with it. Rodsutti et al. [35] conducted a study to develop and validate a decision rule (based on risk factors) that would allow prioritization on the waiting list, using univariate analysis and multiple logistic regression, achieving a scoring scheme or color-coded tables for easy clinical application. Sun et al. [40] used three questionnaires to improve sensitivity and specificity for discrimination of moderate to severe OSA, based on a genetic algorithm. Montoya et al. [37] based their work on several epidemiological and clinical variables, sought to find alternatives to PSG, using logistic regression analysis and multivariate logistic regression to determine the best model for distinguishing OSA patients from the healthy ones. In the end, the work produced a algorithm to calculate the prediction of AHI in a new patient.

All the previous have focused on traditional simpler methods for decision support. Nowadays, prediction models are generated by artificial intelligence, using decision trees, neural networks, support vector machines, and Bayesian networks [9]. All should have good performance, good ability to handle data entry errors or omissions, transparency of diagnostic knowledge, ability to explain decisions, and the algorithms should be able to reduce the number of tests needed for making a reliable diagnosis [26]. While searching, we found studies that attempted to apply these new techniques in OSA. One study [14] tackled the tedious and time-consuming task of analyzing PSG records, automatizing both the detection and classification of sleep apneas, through analysis of wavelets and Bayesian neural networks. The other [19] classified patients with possible diagnosis of OSA into groups according to the severity of the disease using a decision tree producing algorithm based on nonlinear analysis of three respiratory signals instead of full PSG. However, none of the found approaches addressed only clinical and demographic variables that could be used earlier in the healthcare process flow, as they require diagnostic data from PSG.

In fact, technologies exist to tackle the problem at later stages, e.g., disease management [31], but there is a clear lack of solutions for the early diagnosis of OSA. Bayesian networks have been used in several clinical domains, especially given their balance between accurate predictions and their specific interpretability in the clinical domain, resembling the human reasoning in a probabilistic way, along with their ability to produce predictions with missing values, presenting high performance in areas like pneumonia and breast cancer [26].

This study was designed according to the Standard for Reporting Diagnostic accuracy studies (STARD) list [4], updated in 2015. Its guiding principle was to select items that, when reported, would help readers to judge the potential for bias in the study, to appraise the applicability of the study findings and the validity of conclusions and recommendations. STARD guidelines have been generally used for adequately validating diagnostic tools.

To our knowledge, no one has attempted to analyze risk and diagnostic factors for OSA the way this study does. Our study is one of the first to build and validate risk models for OSA based solely on clinical and demographic variables, which have the key advantage of being easily available and quickly acquired. We focused on the most important risk and diagnostic factors, being aware of the clinical definition of OSA. We obtained a proportion of normal results in 66 patients (34%), revealing a large number of unnecessary exams that are performed every day in Vila Nova de Gaia/Espinho hospital center, with the possibility of this result being generalized to the different hospital centers in the country.

Male gender was more prevalent (63%), agreeing with the literature [1, 10, 16, 20‐22, 27, 30, 36, 42‐45]. Possible explanations are the higher prevalence of craniofacial and upper-airway abnormalities (21%), and also snoring (90%). Also, higher age (> 55 years old) was linking to a higher OSA prevalence. The strata of 55–69 and more than 70 years old had a total of 94 patients (73%), followed by 29 patients (23%) in the 40–54 years old. In previous work, age was not included due to the fact that it was dominating the remaining dependencies, so we have been working in a new approach for dealing with this variable. We performed a new analysis with consequent alterations in the preprocessing levels, making the model more robust and clinically accurate, as in the literature this variable is described as one of the most important risk factors for OSA.

During the physical examination, body mass index, neck, and abdominal circumferences were collected. Regarding obesity, described as an important risk factor for OSA, we found that the percentage of patients, in the pathology group, having a normal BMI is equal to the percentage of the patients with obesity (50%). This could explain why in our study obesity is not included in the selected variables. Analyzing neck and abdominal circumferences, unadjusted for gender, we saw higher percentages of increased level, 60 vs. 46% in NC and 94 vs. 91% in AC. In some studies, neck circumference is described as one of the most important risk factors, but not always considered by several medical doctors, creating the need to perform more studies with this variable. Craniofacial and upper-airway abnormalities are important in the pathogenesis of OSA, particularly in non-obese patients, and our study demonstrated it. In patients with normal weight, we found that 26 (96%) patients had craniofacial and upper-airway abnormalities. The differences in craniofacial morphology may explain some of the variation in risk of OSA. Only 3% of OSA patients reported not snoring during the night, while 100% of the patients in the group without OSA report snoring. This might explain why snoring cannot continue to be an important risk factor. Depression or anxiety is highly related to sleep problems, becoming nowadays one of the most studied factors, but only 53 (27%) patients have the diagnosis.

The clinical definition of OSA includes several diagnostic factors, such as daytime sleepiness. One way to confirm its presence is with ESS. Even though this questionnaire is not specific for OSA, it has been often used in Vila Nova de Gaia/Espinho hospital center. When we analyzed the OSA group, we discovered a median of 10 (3–13), demonstrating a normal result for this group (cutoff in 10 points). However, when we analyzed the median in patients without OSA we found a higher value (13 (7–16)). This highlights the possibility that ESS is not adjusted for the pathology. Another common diagnostic factor is witnessed apneas. Men have a percentage of 60% against 43% in women. This can be explained by female bed partners having a lower threshold for symptom perceptions and reporting it less than male bed partners.

For visual inference Bayesian network we chose TAN6. Even though NB6 presents a higher value of sensitivity (94.1% [92.9–95.4%]), in clinical settings, the variables presented in our study are not independent, so we should assume the relation among variables, especially when not all variables are available to the physician. The TAN6 has a sensitivity of 90.2% [88.0–92.4%], meaning that 90% of OSA patients would perform PSG, rejecting 10% of patients who would be referred to follow-up consultation. Further studies are needed to optimize the model and also to externally validate it preferable with a prospective validation cohort.

Visually, the network model TAN6 is very intuitive. It is a simple and friendly model that can be easily accessed and filled in, helping the diagnosis of this pathology in the primary care centers or other facilities that have a gap in its diagnosis. The substructures that raised from this model were also interesting. Gender is related to witnessed apneas, aspect described by physicians and specialists; likewise, witnessed apneas are associated with age. Additionally, age is associated with nocturia (aging increases the need to use the bathroom more often) and with craniofacial and upper-airway abnormalities (possibility of accidents in the adult life and also obesity development). Another common relation described in care is the association between craniofacial and upper-airway abnormalities and increased neck circumference, which was also present in our model.

The final risk results were arranged into a color-coded and user-friendly matrix that constitutes a preliminary but useful tool that can be used by primary care physicians and others in the diagnostic decision-making process. In the case of female gender, the rate started at 8% (under than 40 years, normal neck circumference and without witnessed apneas) to 86% (higher than 70 years old, neck circumference increased and witnessed apneas). In the case of male gender, the initial rate is almost 3 times higher than for female gender. It started at 20% (under than 40 years, normal neck circumference and without witnessed apneas) to 95% (more than 70 years, with neck circumference increased and witnessed apneas). Low risk of having OSA is related with the female gender under 40 years. In contrast, high risk is prevalent in the male gender above 50 years old.

One limitation to fully use the set of predictive variables was the lack of representativeness of some factors, such as vehicle crashes, humor alterations, decreased libido, pulmonary hypertension, congestive heart failure, renal failure, hypothyroidism, gastroesophageal reflux and genetics, which might have led to a bias in the 38 variables models. Also, we acknowledge that the retrospective nature of the derivation cohort is a limitation to the study and, most of all, that the low specificity of the resulting models make them somewhat limited. However, we manage to provide a sensitive tool (sensitivity \(>90\%\)) which nonetheless prevents 1 out of 5 healthy individuals from unnecessarily performing PSG exam (specificity \(>20\%\)), improving from current clinical practice. This improvement could also result in financial benefits for the healthcare system. Using a simple back of the envelope calculation, the relatively small district hospital where our cohort was recruited could potentially perform 466 PSG (yearly extrapolation of 194 PSG in 5 months). Given a 20% specificity of our model, from the 158 normal PSG (yearly extrapolation of 66 normal PSG in 5 months), 32 would not have been referred to perform PSG. Considering Portuguese mandamus number 207/2017, each PSG is priced at €939,14 leading to a total saving of €30.052,48. Moreover, 31 patients with OSA (yearly extrapolation of 10% of 128 OSA patients in 5 months) would also have their PSG delayed by our approach, representing additional savings of €29.113,34, for a grand total of €59.165,82. Certainly, these 31 patients should have performed PSG, thus diminishing the clinical benefit of our proposal. However, limited inspection on the data supports our belief that these would have been mild or moderate OSA patients, who could perhaps be rescheduled for follow-up consultation and PSG in the subsequent months without serious harm. Either way, given the current wait list situation, the 63 PSG vacancies would most likely be filled with patients, with the expected savings corresponding to one and a half months worth of PSG work. Nevertheless, we defer this discussion to future work, where we will perform a cost-effectiveness analysis (using Monte Carlo simulations) to assess the expected overall impact of our approach.

Our study added important knowledge to the state of art regarding OSA. Moreover, this knowledge is delivered in the form of an intuitive and user-friendly model (TAN6), which can be used by any physician, and in the form of a risk matrix that physicians can use to quantify the probability for having OSA. We present as main risk factors: gender, age, neck circumference, and craniofacial and upper-airway abnormalities, and as diagnostic factors: witnessed apneas and nocturia. Naïve Bayes was considered a better classifier, while TAN showed advantages for visual inference, proving the great advantages of Bayesian network models (dealing with missing information and simple graphical representation, showing not only the probabilities given the patient characteristics but also the relationship between variables).

According to our cross-validated evaluation, we expected around 30% of false positives which, although unwanted, is nonetheless an improvement if we compare with all patients at risk being referred to sleep consultation and polysomnography. Nonetheless, we were able to rule out 25% of healthy patients, which, in our understanding, would alleviate health services by reducing the burden of unneeded consultations and the wait lists for polysomnography, while identifying more than 90% of patients with OSA.

Portugal, like many other countries, does not have a validated method to screen patients with suspicion of OSA, before performing polysomnography, so we think that our models (TAN6 and risk matrix) consist in valid methods. Current work is focused on bringing the model toward primary care facilities.