INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Snoring is a common symptom seen in 20% of the overall adult population and in even larger proportions of males, elderly subjects, and obese patients of either sex (1). Chronic snoring became a target of research efforts, however, after it was found to be significantly associated with sleep apnea syndrome (SAS).

Overnight polysomnography in a sleep laboratory is the method most commonly used to confirm or to refute a suspected diagnosis of SAS. However, polysomnography is expensive, time-consuming, and labor-intensive. Few centers can afford to perform polysomnography in all patients presenting with snoring. The high prevalence of snoring and the steady increase in the numbers of patients attending sleep clinics to evaluate snoring have resulted in long waiting lists at sleep laboratories.

With the goal of optimizing health resource utilization by identifying those patients most likely to benefit from overnight polysomnography, several studies have examined the value of clinical symptoms as predictors of SAS (2). However, both sensitivity and specificity proved low. Thus the clinical impression of a sleep specialist is not reliable for diagnosing SAS in the individual patient. This seems to be particularly true in snorers without clinical complaints who are referred to the sleep laboratory because they have risk factors for SAS, such as obesity or cardiovascular disease.

The aim of our study was to develop a screening tool for identifying snorers at low risk for SAS who do not require polysomnography. To hold appeal for the clinician, such a tool would have to show a high negative predictive value (NPV) and to be quick and easy to use. In an earlier study, we identified a number of pulmonary function abnormalities associated with SAS (7). Two parameters were highly correlated with SAS severity, i.e., specific respiratory conductance (sGrs) and daytime arterial O₂ saturation (Sa_O2). We hypothesized that these parameters may be useful for predicting the likelihood of SAS. In the present study, we sought to determine whether applying logistic regression to these parameters to obtain a statistical index was effective in separating nonapneic snorers from patients with SAS (7). We validated our index in a prospective group of obese snorers suspected of having SAS. The index proved to be repeatable and to have a high NPV, indicating that it may be useful for identifying obese snorers whose risk of SAS is so low as to make polysomnography unnecessary.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Patients

Our objective was to develop an index for predicting the absence of SAS in snorers. We studied two groups of patients, the first to develop the index and the second to test its validity. The first group (Group 1) was composed of 168 patients (135 males and 33 females) who attended our sleep clinic for snoring and suspected SAS. The characteristics of this group were described in a previous article (7). The second group (Group 2) included 101 patients with similar clinical characteristics. The study population was composed of consecutive patients with a body mass index (BMI) greater than 25 kg/m² but lower than 35 kg/m².

Exclusion criteria were a history of alcoholism; regular use of hypnotic medication; upper respiratory tract disorders; previous treatment for sleep apnea; cardiopulmonary disease; or airway obstruction defined as a ratio of forced expiratory volume in one second to forced vital capacity less than 80%, due to asthma or to chronic obstructive pulmonary disease (COPD). Patients with a history or clinical evidence of neuromuscular disease were also excluded.

Sleep Studies

All patients underwent sleep studies. The methods and equipment used for polysomnography have been described in detail previously (7). Apnea was defined as cessation of airflow lasting 10 s or more, and hypopnea as a 50% or greater reduction in oronasal airflow for at least 10 s. SAS was defined as a combined apnea + hypopnea index (AHI, events per hour of sleep) of 15 or more (8).

Sleepiness was defined as a score of 11 or more on the Epworth Sleepiness Scale (ESS), a questionnaire that asks the subject to rate his or her likelihood of falling asleep in different situations (13, 14).

Pulmonary Function

Spirometry, flow-volume curves, and arterial blood gas analysis were done in all patients using conventional methods, as previously described (7). FRC was measured using the helium dilution method, as the mean of two determinations.

Respiratory impedance was determined using the standard forced oscillation technique, as described elsewhere (7) (Oscilink; Datalink-MSR, Rungis, France). The subjects were equipped with a mouthpiece and a nose clip, and were comfortably seated with the head in the neutral position and the cheeks held firmly. A pseudorandom noise signal mixing integer frequencies between 4 and 32 Hz was generated by a loudspeaker and superimposed on the spontaneous breathing of the subject. Auto- and cross-spectra of mouth flow and pressure were averaged over a 16-s period to yield a mean estimate of impedance (Zrs) and coherence function (²) for each frequency component. A measurement period was considered acceptable if ² was higher than 0.9 for more than 80% of the frequency components. The real component of Zrs (Rrs), which is related to the resistive properties of the respiratory system, was submitted to linear regression analysis over the 4 to 16 Hz frequency range to obtain the intercept (R₀, resistance extrapolated to 0 Hz). At least three acceptable 16-s measurement periods were averaged to yield the final value of these parameters. Respiratory conductance (Grs) was calculated as the reciprocal of Rrs, and sGrs was obtained as the ratio of Grs over FRC.

Statistical Analysis

All analyses were performed using SPSS software. Data were obtained blindly; the physician who performed the pulmonary tests was not aware of the sleep status of the patient and, conversely, the physician who performed the polysomnography was not aware of the lung test results.

Data were expressed as means ± SD. Correlations between variables were evaluated using least-square linear regression techniques.

Subjects were classified based on their AHI. AHI values of 15 or more were considered abnormal.

Logistic regression was used to model the probability of a binary (yes/no) outcome, namely presence of SAS (AHI  15) or absence of SAS (AHI < 15). Explanatory variables were selected by comparing the data from patients with an AHI < 15 to those from the other patients, using Student's t and chi-square tests. Stepwise logistic regression was performed to determine which factors were independently associated with AHI.

The statistical model developed from Group 1 data was used to predict the presence or absence of SAS in each Group 2 patient. The results thus obtained were compared with the polysomnography diagnosis. Sensitivity, specificity, positive predictive value (PPV), and NPV were calculated. Receiver-operator characteristic (ROC) curves were constructed to assess the relationship between sensitivity and specificity of the index. ROC curves were used to determine the true-positive rate (sensitivity) versus the false-positive rate (1  specificity) at various levels of the index and to identify the cutoff yielding to the largest number of correctly classified patients.

For all comparisons, p values less than 0.05 were considered significant.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Patient characteristics in the two groups are shown in Table 1. The two groups had similar proportions of SAS patients (57% versus 51%, chi-square not significant [NS]) and contained both patients with mild SAS and patients with severe SAS (Table 2). No significant difference in smoking history was found between the groups.

Construction of the SAS Prediction Model (Group 1)

sGrs, Sa_O2, Pa_O2, and forced expiratory flow rates at 50% of the vital capacity (FEF₅₀) were significantly different between patients with and without SAS. Accordingly, in the stepwise logistic regression analysis the dependent variable was AHI and the independent variables were sGrs, Sa_O2, Pa_O2, and FEF₅₀. Only two variables were significantly and independently related to AHI, i.e., sGrs and Sa_O2. Logistic regression with these two significant variables as the independent variables and AHI as the dependent variable was used to develop a model for predicting the probability (p) of having a polysomnography positive for SAS:

where sGrs is expressed in cm H₂O¹ · s¹ and Sa_O2 in percent. The estimated value of p was derived from logit (p) = log_e(p/1  p) and was in the 0 to 1 range. Statistical significance was < 10⁶ for both sGrs and Sa_O2.

Logistic regression analysis indicated that the p value cutoff providing the largest number of correctly classified patients was 0.5. Sensitivity and specificity of the model were calculated across the range of predicted probabilities (p) using a ROC curve (Figure 1). The 0.5 cutoff was associated with 98% sensitivity and 86% specificity. PPV and NPV were 90% and 97%, respectively.

View larger version (16K): [in this window] [in a new window]   Figure 1.   Comparison of ROC curves in the 168 Group 1 patients (solid squares) and in the 101 prospective Group 2 patients (open circles) for the model separating nonapneic snorers from SAS patients (AHI  15). Each point indicates the sensitivity and specificity of a given p value derived from the logistic regression model and ranging from 0 to 1. The cutoff value producing the best sensitivity and specificity is 0.5. Note that the two curves are comparable, with a similar cutoff point.

Figure 2 shows the p values predicted by the model for Group 1 patients. Only two of the 96 patients with SAS were misclassified by the model. These two patients had AHIs of 16 and 17, respectively, with a predominance of hypopneic episodes.

View larger version (12K): [in this window] [in a new window]   Figure 2.   Distribution of probabilities predicted by the model in Group 1 (168 obese snorers with suspected SAS) and Group 2 patients (prospective group of 101 obese snorers with suspected SAS), according to the AHI. SAS was defined as an AHI  15 h1. A cutoff of 0.5 was chosen based on the logistic regression model and ROC curve analysis. Among the patients with a p value less than 0.5, all Group 1 patients but two and all Group 2 patients had an AHI < 15, demonstrating the high negative predictive value of the model.

Validation of the Model (Group 2)

We then evaluated the model in the 101 prospective Group 2 patients. Using 0.5 as the cutoff, all 51 patients with a polysomnography AHI of 15 or greater were correctly classified by the model. Eight of the 50 patients whose polysomnography AHI was lower than 15 were misclassified by the model (false-positives) (Figure 2). These numbers yielded a sensitivity of 100%, a specificity of 84%, a PPV of 86%, and an NPV of 100%. ROC curve analysis confirmed that 0.5 was the best cutoff (Figure 1) and that Group 1 and Group 2 were comparable.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

The results of this study demonstrate that a statistical model based on lung function parameters is effective in excluding SAS in individual obese patients suspected of having this condition.

At present, the diagnostic evaluation of patients with suspected SAS almost invariably includes overnight polysomnography, which is both cumbersome and costly. There is a need for a quick and reliable screening test capable of identifying those patients most likely to benefit from overnight polysomnography, i.e., who have a high risk of SAS. Several simplified nocturnal screening tests used alone or in combination with clinical symptoms have been evaluated. Most involved overnight recording of one or multiple cardiorespiratory parameters, including oximetry, airflow, tracheal sounds, respiratory movement, and heart rate (15).

A different and extremely appealing approach consists in using clinical and lung function parameters to predict SAS in snorers suspected of having this condition. Although a number of studies have used such an approach, opinion remains divided on which and how many symptoms, in addition to snoring, are required to suspect SAS. Statistical models using patient age, sex, obesity, neck circumference, snoring, reports of apnea during sleep, and hypertension proved sensitive but lacked specificity (Table 3) (4, 8). It was suggested that these disappointing results might be ascribable to absence from the models of a cardinal symptom, namely daytime sleepiness. However, including daytime sleepiness failed to improve screening performance (2, 6, 19).

The subjective impression of physicians regarding the presence or absence of SAS has also been assessed. Viner and coworkers (6) concluded that subjective impression based on clinical history and physical findings had a diagnostic accuracy of approximately 63%. In a study of 594 patients referred to the sleep clinic because of suspected SAS, Hoffstein and coworkers (20) found that subjective impression had 60% sensitivity and 63% specificity for the diagnosis of SAS (defined as an AHI of 10 or more). Similarly, Flemons and coworkers (5) found that subjective impression was inferior to a model taking into account neck circumference, hypertension, habitual snoring, and bed partner reports. All these results may be ascribable to the fact that self-reported sleepiness is a nonspecific symptom frequently observed in subjects without SAS (10, 20).

From these studies of clinical features, it can be concluded that clinically derived predictive indexes may be useful in the diagnosis of SAS, but only to a limited extent. As seen in Table 3, the PPVs varied from 49 to 77%, and higher PPVs were associated with lower NPVs. Even when symptoms, such as sleepiness and bed partner's reports of apnea, were used in combination with physical features, such as being male, obese, or hypertensive, a number of patients were misclassified, as stated by Douglas in a recent review of the literature (12).

Patients referred to a sleep clinic for suspected SAS usually undergo lateral cephalometry aimed at visualizing upper airway size, soft tissues, and craniofacial structure, to help in the choice of therapy (23). A number of studies (23) have shown differences in cephalometric indices between SAS and non-SAS patients. However, in a recent study, Brander and coworkers (25) showed that inclusion of cephalometric variables explained only part of the observed variance in AHI. Recently, Kushida and coworkers (27) described a new morphometric evaluation method taking into account oral cavity measurements. An index combining these measurements with BMI and neck circumference was found to have high sensitivity, specificity, PPV, and NPV for the diagnosis of SAS (Table 3). Although there is still some disagreement about the apnea and hypopnea frequency that represents clinically significant SAS, most sleep centers use a threshold of 10 or 15 episodes per hour. In a recent report by Littner and Shepard (28) of various severity indexes based on event frequency, mild sleep-disordered breathing was defined as a number of events < 15 per hour. The Task Force (29) also recommended that 15 events per hour be used as the cutoff. In the study by Kushida and coworkers (27), the cutoff was 5 or more, and as a result the prevalence of SAS was high (Table 3), a fact that may explain in part the high post-test probability of the diagnostic test.

In the present study, we found that simple daytime evaluation of lung function abnormalities predicted the presence or absence of SAS rapidly and accurately in the individual patient. To our knowledge, no studies have systematically assessed pulmonary function parameters, including respiratory conductance, in obese snorers with the goal of developing a statistical model for separating subjects with and without SAS. Previous studies of lung function in SAS were restricted to an analysis of flow-volume curve parameter changes (30) denoting extrathoracic upper airway obstruction, such as the saw-tooth sign (presence of flow oscillations) or a midvital capacity flow ratio higher than 1. These changes showed high specificity but low sensitivity (Table 3), a feature that limits their usefulness in screening for SAS.

In a previous study (7), we found that two lung function parameters, namely sGrs measured by forced oscillation and Sa_O2, were strongly correlated with the severity of sleep-disordered breathing in a population of obese snorers. The significant decrease in sGrs with increasing AHI was explained by obstruction of both the upper and the peripheral airways, whereas disturbances in daytime gas exchange were determined by alterations in ventilatory mechanics, and were related to obesity and diffuse airway obstruction.

Based on these results, we hypothesized that these two parameters could be combined to derive a statistical model for predicting SAS. This model was accurate, as shown by its high sensitivity and specificity values, and repeatable. Although the PPVs were as high as 86% and 90% in our two groups of patients, it should be emphasized that the NPVs were 97% and 100%, indicating that the model identifies non-SAS with almost complete certainty. Therefore, our model may be clinically useful for excluding SAS in patients who attend a sleep laboratory for snoring.

The PPV and NPV of a given test provide a direct assessment of the usefulness of this test in practice. However, numerical results for predictive values are highly dependent on the prevalence of the abnormality in the specific population studied, as shown by Bayes' theorem (33). In our study, the prevalence of SAS was approximately 50%, as is usually the case in sleep clinic populations (Table 3). The PPV and NPV of our index would be very different in the general population, where the prevalence of sleep apnea has been estimated at 4 to 8% (22). It should be borne in mind that our results apply to the indications of sleep studies in patients referred to sleep clinics.

Because of the limited value of clinical parameters in predicting SAS, a number of male, obese, habitual snorers are sent to sleep clinics to eliminate SAS. In this situation, our model may help physicians to determine that polysomnography is unnecessary. However, even when our model indicates that SAS is unlikely, daytime sleepiness (13, 14) requires additional investigation. Daytime sleepiness can indicate upper airway resistance syndrome (34) or other diagnoses, including narcolepsy or periodic limb movement disorder, and should therefore be investigated by overnight polysomnography, possibly combined with a daytime multiple sleep latency test. Among our 42 Group 2 patients with a negative prediction of SAS, four had abnormal daytime sleepiness requiring polysomnography. When our predictive index is positive, polysomnography should be performed to assess the severity of SAS, irrespective of the Epworth score, because SAS severity influences treatment decisions. Our screening strategy would have made polysomnography unnecessary in 38% of patients, thus considerably lightening the sleep laboratory workload.

In conclusion, a statistical model based on daytime respiratory function abnormalities is helpful in eliminating SAS in snorers presenting to a sleep-related breathing disorders clinc. The model proved accurate, feasible, quick, and easy to use.