Well-established sleep apnea is characterized by loud snoring, witnessed apneic episodes, disturbed nocturnal sleep, daytime sleepiness, and impaired cognition and is typically associated with obesity and (in men) a large neck size. Given this profile, it is not surprising that clinical prediction models would arise in an effort to diagnose OSA in larger populations. Virtually all of these studies have been done in sleep clinic populations rather than in the general population. One of the earliest studies showed that witnessed apneic episodes combined with loud snoring predicted an apnea-hypopnea index (AHI) > 10 with a sensitivity of 78% and specificity of 67% (1l). Crocker et al. (12) used an alternative approach and developed a statistical model using clinical data to predict disturbance of sleep-disordered breathing in 114 consecutive patients. Witnessed apneic episodes, hypertension, body mass index (BMI), and age provided a sensitivity of 92% but a specificity of only 51% for an AHI > 15. Using 410 clinic patients, Viner et al. (13) developed a model incorporating snoring, BMI, age, and sex and came up with sensitivity and specificity of 94% and 28%, respectively. The higher the pretest probability of sleep apnea, the better the positive predictive value of their model. Maislin et al. (14) added to this and developed the multivariable apnea prediction index, which includes questions about frequency of symptoms of apnea as well as measurements of BMI, age, and gender. Using this tool, at a BMI > 40, the likelihood of apnea is very high with or without symptoms, while at lower values of BMI, the likelihood of sleep apnea is much more dependent on whether or not the individual has symptoms. Predictive abilities assessed using receiver operating characteristic (ROC) curves noted that for BMI alone the ROC value was 0.73, and for an index measuring a self-report of apnea symptoms it was 0.7. Many other papers have demonstrated varying degrees of variability and specificity with their clinical prediction models (15-17).
Netzer et al. (18) assessed the utility of the Berlin Questionnaire to diagnose sleep apnea in a primary care setting. This questionnaire asks about risk factors for sleep apnea, namely snoring behavior, wake time sleepiness or fatigue, and the presence of obesity or hypertension. A subset of patients underwent overnight portable recording using a six-channel recorder (EdenTrace® Recording System, vide infra). This approach resulted in a sensitivity of 86%, a specificity of 77%, and a positive predictive value of 89% for OSA. This questionnaire appears to be a useful tool but needs to be tested in other populations with neck circumference, age, and race added to the predictive model.
Rodsutti et al. (19) derived and validated a clinical decision rule to assess risk of sleep apnea and prioritized those for polysomnography. Five variables—age, sex, BMI, snoring, and stopped breathing during sleep—were significantly associated with sleep apnea. ROC analysis for both derivation and validation sets gave area under the curve (AUC) values of 0.81 and 0.79, respectively.
The uses of neural networks for predicting or excluding sleep apnea have been demonstrated in a few studies. Artificial neural networks (ANNs) are computer programs modeled after the nervous system and are capable of recognizing complex patterns in data. ANNs are "trained" by presenting a set of data together with the outcomes that the trainer wishes the network to learn. The trained ANN can then be evaluated by inputting similar but previously unseen data. This approach for outcome prediction has been used successfully in medical applications, including the prediction of acute myocardial infarction in patients presenting to an emergency room physician (20), the diagnosis of pulmonary embolism (21,22), and the predicted length of stay of patients in an intensive care unit (23).
El-Solh et al. (24) utilized a back-propagation ANN algorithm on 189 patients as a training set and validated it prospectively on 80 additional patients. Predictive accuracy at different AHI thresholds was assessed by the c-index, which is equivalent to the area under the ROC curve. The c-index for predicting OSA in the validation set was 0.96, 0.95, and 0.935 using thresholds of > 10, > 15, and > 20/hour, respectively. They suggested that ANN may be useful as a predictive tool for OSA. Using a backward error propagation ANN with 23 clinical variables and a leave-k-out strategy, Kirby et al. (25) found the positive predictive value that a patient would not have sleep apnea to be 98%, with a negative predictive value the patient would have sleep apnea (AHI > 10) to be 89%. In that study the use of the ANN would have reduced the number of PSGs performed by 22%.
Additional approaches have also been taken. Kushida et al. (26) incorporated oral cavity measures into a morphometric model of OSA, using a degree of maxillary overjet, intermolar distance, and maxillary mandibular planes and palatal height, combined with neck circumference and BMI. This model had a sensitivity of 98%, specificity of 100% and a positive predictive value of 100% for an AHI > 5. Despite these impressive results, this technique has not been replicated in other centers, possibly because it is somewhat labor intensive. More recently, Tsai et al. (27) noted the three main reliable clinical symptoms of sleep apnea (snoring, witnessed apneas, and hypertension) and three signs of sleep apnea (thyro-mental space less than 1.5 cm, pharyngeal grade > 2, and the presence of an overbite) provided a positive predictive value of 95%. A thyro-mental space of > 1.5 cm excluded sleep apnea with a negative predictive value of 100%.
Rowley et al. (28) prospectively studied the utility of four clinical prediction models for predicting the presence of sleep apnea or prioritizing patients for a split-night protocol. They took four clinical prediction formulae of Crocker, Viner, Flemons, and Maislin to calculate the probability of sleep apnea for each model in each of 370 clinic patients. For an AHI > 10, their sensitivity ranged using these models from 76% to 96%, specificity of only 13% to 54%, and a positive predictive value ranging between 69% and 77%. They concluded that clinical prediction models are not sufficiently accurate to discriminate between patients with or without sleep apnea but could be useful in prioritizing patients for split-night polysomnography.
Was this article helpful?