Classification algorithms for predicting sleepiness and sleep apnea severity

Abstract

Identifying predictors of subjective sleepiness and severity of sleep apnea are important yet challenging goals in sleep medicine. Classification algorithms may provide insights, especially when large data sets are available. We analyzed polysomnography and clinical features available from the Sleep Heart Health Study. The Epworth Sleepiness Scale and the apnea-hypopnea index were the targets of three classifiers: k-nearest neighbor, naive Bayes and support vector machine algorithms. Classification was based on up to 26 features including demographics, polysomnogram, and electrocardiogram (spectrogram). Naive Bayes was best for predicting abnormal Epworth class (0-10 versus 11-24), although prediction was weak: polysomnogram features had 16.7% sensitivity and 88.8% specificity; spectrogram features had 5.3% sensitivity and 96.5% specificity. The support vector machine performed similarly to naive Bayes for predicting sleep apnea class (0-5 versus >5): 59.0% sensitivity and 74.5% specificity using clinical features and 43.4% sensitivity and 83.5% specificity using spectrographic features compared with the naive Bayes classifier, which had 57.5% sensitivity and 73.7% specificity (clinical), and 39.0% sensitivity and 82.7% specificity (spectrogram). Mutual information analysis confirmed the minimal dependency of the Epworth score on any feature, while the apnea-hypopnea index showed modest dependency on body mass index, arousal index, oxygenation and spectrogram features. Apnea classification was modestly accurate, using either clinical or spectrogram features, and showed lower sensitivity and higher specificity than common sleep apnea screening tools. Thus, clinical prediction of sleep apnea may be feasible with easily obtained demographic and electrocardiographic analysis, but the utility of the Epworth is questioned by its minimal relation to clinical, electrocardiographic, or polysomnographic features.

Figure 1

Distribution of clinical features in the Sleep Heart Health Study (SHHS) subjects. Histograms of various clinical features are shown in overlapped bars, where dark grey represents subjects with at least one missing data point, and light grey representing subjects with complete data. We studied 27 features in total. Twenty-five are shown in this figure; arousal index (AI) in non-rapid eye movement (NREM) and REM are not shown, but demonstrated similar complete versus missing distributions. See Table 1 for description of features.

Figure 2

Spearman’s correlation of clinical features with Epworth Sleepiness Scale (ESS) and apnea–hypopnea index (AHI). R-values for correlation between ESS (a) and AHI (b) are shown for the listed clinical features. Scatterplots are shown for representative pairs: ESS versus AHI (c) and AHI versus body mass index (BMI) (d).

Figure 3

Performance of the naive Bayes classifier in predicting Epworth Sleepiness Scale (ESS) and apnea–hypopnea index (AHI) classes. The prediction power of the naive Bayes classifier algorithm is shown for ESS based on polysomnogram (PSG) features (a), electrocardiogram (ECG) features (b) or a combination of all 26 available features (c). The prediction power of the naive Bayes classifier algorithm is shown for AHI based on clinical features (d), ECG features (e) or a combination of features (f). For (f), the combination included clinical, ECG and only those PSG values not related to apnea index [excluded arousal, respiratory disturbance index (RDI) and oxygen metrics].

Figure 4

Performance of the support vector machine (SVM) classifier in predicting apnea–hypopnea index (AHI) class. The prediction power of the SVM algorithm is shown for AHI based on clinical features (a), electroencephalogram (ECG) features (b) or a combination of these and non-respiratory polysomnogram (PSG) features (n = 19 features) (c). The parameters gamma, C and epsilon were first searched manually across log-units and then more narrow choices until the shown values were obtained. The parameters used in the data shown were: gamma 0.1, C 0.2, epsilon 0.2. Normalized weights from the SVM are shown in (d), where all values are normalized to the largest weight [elevated low-frequency coupling (e-lfc)]. The absolute value of the weights reflect the strength of the relationship between that feature and the AHI class. The dotted lines mark the range of weighting values obtained when the AHI values were scrambled randomly; feature weights within this range are taken to be non-significant.

Figure 5

Mutual information between various features and Epworth Sleepiness Scale (ESS) or apnea–hypopnea index (AHI). The normalized mutual information is shown for various discrete/continuous features compared with ESS (a) and AHI (b). Categorical features were not computed. Note that the vertical axis range is 0–0.05 in (a), while that in (b) is sixfold larger, 0–0.3, emphasizing the striking difference in information between the features and ESS versus AHI. For both plots, the features are stratified from most to least mutual information.