1.

Introduction

Obstructive sleep apnea (OSA) is a condition that involves multiple episodes of airway closure and/or reduction in airflow that affects 2% to 4% of the population.¹ OSA has been associated with adverse medical conditions including congestive heart failure,² stroke,³ pulmonary⁴ and systemic⁵ hypertension, cancer,⁶ and increased mortality.⁷ Traditionally, sleep apnea has been assessed via a nighttime polysomnogram. This diagnostic procedure continues to be the gold standard for the evaluation of sleep apnea.

Numerous studies have looked at the role of pulse oximetry in the assessment of OSA.^8–14 These previous investigations have developed several methods for predicting an apnea–hypopnea index (AHI) from oximetry data; these include: cumulative time below an oxyhemoglobin saturation (${\mathrm{SaO}}_2$) of 90% (CT90),⁸ the number of ${\mathrm{SaO}}_2$ events that drop either 3% or 4% [oxygen desaturation index (ODI)],^9,10,15 a measure of ${\mathrm{SaO}}_2$ variability ($\mathrm{\Delta }$ index),^11,12 a central tendency measure,¹³ and a multilayer perceptron neural network.¹⁶ However, in our review of the literature, we uncovered only two other studies that have attempted to convert oximetry to an AHI.^14,16 In the Magalang et al. study, both the $\mathrm{\Delta }$ index and a composite measure of various oximetry indices were found to have coefficients of determination of 0.60 and 0.70, respectively. The Marcos et al. study produced a more accurate predictive formula than Magalang et al. However, scoring of the polysomnograms did not appear to utilize the updated guidelines recommended by the American Academy of Sleep Medicine (AASM).¹⁷

Therefore, the goal of this study is to develop a new predictive model of AHI using oximetry data by utilizing more recent pulse oximetry technology, a large sample size, and the newer AASM recommended scoring guidelines to evaluate the gold standard attended polysomnograms. We believe that by taking into account these factors, we will be able to develop a new predictive model that outperforms all previous models. Our predictive model will focus on ODI since it appears to be the most sensitive and specific oximetry index,^9,18 and is therefore an ideal target for conversion to AHI.

2.

Methods

3.

Results

The coefficient of determination ($r^2$) between the actual AHI and the predicted AHI (PredAHI) was 0.80 ($r=0.90$), which was significant at $p<0.001$. PredAHI determined a correct AHI $\pm 5/\mathrm{h}$ in 76% of subjects. The intraclass correlation for single measures using an absolute agreement definition was 0.88 (CI 0.87 to 0.90). The subjects that had a PredAHI greater than $\pm 5/\mathrm{h}$ were significantly older, $t(732)=-5.311$, $p<0.001$; had a higher AHI, $t(732)=-16.89$, $p<0.001$; and a lower sleep efficiency, $t(732)=5.12$, $p<0.001$.

The AUC was $0.97\pm 0.005$ (SE), CI 0.96 to 0.98 for an AHI of $\ge 5/\mathrm{h}$, $0.96\pm 0.007$ (SE), CI 0.94 to 0.97 for an AHI $\ge 10/\mathrm{h}$, $0.96\pm 0.007$ (SE), CI 0.95 to 0.98 for an AHI of $\ge 15/\mathrm{h}$, and $0.97\pm 0.008$ (SE), CI 0.96 to 0.99 for an AHI of $\ge 30/\mathrm{h}$ [see Fig. 1 for receiver operating characteristic (ROC) curves and Table 2 for other measures of test precision]. The asymptotic significance level for AUC at all tested thresholds was $p<0.001$.

Fig. 1

ROC curve for predicted AHI at different cutoff points compared to polysomnograph: (a) 5, (b) 10, (c) 15, and (d) 30.

Table 2

Stratified results for predicting sleep apnea using the PredAHI algorithm. PPV, positive predictive value; NPV, negative predictive value.

4.

Discussion

This analysis shows that an accurate prediction of AHI can be made using a regression formula derived from ODI, with areas under the ROC curve ranging from 0.96 for thresholds of $\ge 10$ and $\ge 15/\mathrm{h}$ to 0.97 for thresholds of $\ge 5$ and $\ge 30/\mathrm{h}$. This is better than most previously published comparisons of ODI to AHI.^10,14,20 Only one other study appears to outperform our model;⁹ however, this study does not attempt to convert ODI to a predicted AHI. Moreover, our model was developed and compared to an AHI calculated using the AASM’s currently recommended scoring guidelines for respiratory events,¹⁹ and therefore is more applicable for use today. In comparison to the Magalang et al. models,¹⁴ which showed $r^2\mathrm{s}$ of 0.60 and 0.70 for predicting AHI with the $\mathrm{\Delta }$ index alone and a composite measure of oximetry, respectively, our model outperformed these algorithms with an $r^2$ of 0.80. The Marcos et al.¹⁶ model slightly outperformed our model at an AHI of 15/h (93% versus 91%). However, our model was more accurate at AHIs of 5/h (84% versus 91%) and 10/h (87% versus 90%). In addition, our model was very accurate at an AHI of 30/h (see Table 2).

Due to the fact that ODI cannot differentiate between central and obstructive events, we do not intend for this algorithm to replace traditional polysomnograms. However, this formula can be useful in diagnosing patients who have had either a home or in-laboratory sleep study, but for whom the flow sensor data are unavailable. It is not uncommon in clinical practice to have patients remove the flow sensor because of discomfort; in these cases, a predicted AHI can be calculated from the ODI, and the respiratory effort activity can be viewed to gain an estimate of central versus obstructive apnea. This formula can also be used with pulse oximeters to convert nighttime oximetry data to AHI in order to determine if additional testing or treatment for sleep apnea is warranted.

As mentioned in Sec. 1, since the 1990s, studies have shown the value of using oximetry in assessing apnea.^8–14 However, the use of oximetry indices for clinical decision making in sleep disordered breathing remains uncommon. This may be due, in part, to the fact that the AASM (the main governing body for sleep disorder clinics in the United States) has no diagnostic criteria for defining sleep apnea severity using any pure oximetry index.¹⁷ Moreover, all major insurance providers, including government-run programs such as Medicare, define sleep apnea based on either an AHI or respiratory disturbance index. Therefore, converting oximetry data to an AHI allows a more readily understandable metric that sleep specialists and most nonsleep specialists alike can interpret.

As seen in Fig. 2, PredAHI tends to underestimate AHI in the severe range. This is likely due to the fact that as the actual AHI becomes very high, the individual apneas tend to be shorter, resulting in fewer respiratory events resulting in $\ge 4\%$ drop in blood oxygen saturation (e.g., if an average apnea length is 15 s, the maximum AHI is 120/h versus 90/h for an average apnea length of 20 s). However, the accuracy (see Table 2) of PredAHI at the $\ge 30/\mathrm{h}$ AHI threshold is high at 95%. Differentiating whether a true AHI is 50/h versus 90/h is not as important as determining if an AHI is in the severe range; therefore, we do not believe this discrepancy will significantly affect clinical decisions to treat sleep apnea based on our formula. Another limitation of this study is the use of a single testing site. However, the heterogeneity of our patient population in New York City moderates this concern to some degree. Moreover, the use of a large sample size in both developing and validating our model gives confidence in its accuracy.

Fig. 2

The center line of the plot represents the mean; the other two lines indicate $\pm 2$ SD. It shows data from 734 subjects (the validation set). The gold standard measure of AHI was derived from in-laboratory polysomnography scored according to AASM 2007 guidelines.

In summary, the algorithm described in this paper provides a convenient and accurate way to convert ODI to a predicted AHI. This tool makes it easier for clinicians to understand oximetry data in the context of traditional measures of sleep apnea. Our goal is to pair our formula with a commercially available pulse oximeter in order to provide a low-cost screening tool for sleep apnea to determine the need for additional evaluation and/or treatment of sleep disordered breathing. We believe this instrument will be useful for trades such as the transportation industry, which is in need of a low-cost and accurate way to assess for sleep apnea, or for patients living in rural or impoverished areas where traditional attended polysomnography may not be available or may be too costly.