Paediatric obstructive sleep apnoea syndrome (OSAS) is a sleep-related breathing disorder characterised by intermittent and repetitive episodes of partial or complete collapse of the child’s upper airway while sleeping [1
]. Recurrent apnoeic events lead to gas exchange abnormalities and sleep disruption [2
], which may cause major long-term adverse consequences in several body systems, such as neuropsychological and cognitive deficits, cardiovascular and metabolic dysfunction, and growth impairment [1
]. Consequently, this condition severely affects health, development and quality of life of infants and young children [4
]. In addition, untreated OSAS increases healthcare utilization and associated costs [5
]. Therefore, early detection is essential in order to initiate treatment. In this regard, a recent report of the American Academy of Paediatrics re-emphasised the need for OSAS screening in every habitually snoring child [2
The prevalence of OSAS is estimated to range 1% to 5% of children in the general paediatric population [2
]. Despite its major negative consequences, childhood OSAS is considered a relatively under-diagnosed condition [6
]. Overnight polysomnography (PSG) in a supervised sleep laboratory is the gold standard technique for a definitive diagnosis [2
]. One of the most important factors responsible for this under-diagnosis is the limited availability of paediatric sleep units in most countries [4
]. An additional major limitation is the intrusiveness of PSG for children, who showed high aversion to spend the whole night in the sleep unit with several sensors attached [4
]. These drawbacks limit the effectiveness of conventional PSG as a screening technique for OSAS in every symptomatic child as suggested by the international medical community. Therefore, during the last decade, it has emerged a great demand for novel and simplified screening tools for the disease [9
In the context of simplified alternatives to PSG, attended respiratory polygraphy (RP) has become a reliable method for OSAS detection in clinical settings [11
]. In addition, unattended RP at home has been recently proposed as a feasible approach in low resource settings when in-lab PSG is not available [10
]. Nevertheless, RP, which measures airflow (thermistor and/or nasal pressure), respiratory movements (chest and abdominal effort), body position, pulse rate and blood oxygen saturation (SpO2
), also manage several sensors, being still potentially intrusive for infants and young children. In this regard, recording of single-channel SpO2
from overnight oximetry has been also proposed as a highly simple as well as effective screening technique for paediatric OSAS due to its suitability for children [10
]. Moreover, automated processing of oximetric recordings has been proposed to enhance the diagnostic performance of overnight oximetry as a single screening test for childhood OSAS [16
Several automated signal processing methods have been applied during the last years to parameterise changes in the overnight SpO2
profile due to apnoeic events. Previous studies in the framework of paediatric OSAS detection by means of oximetry assessed conventional desaturation indexes [16
], common statistics in the time domain [16
], spectral features in the frequency domain [16
] and nonlinear measures [16
]. Among these complementary approaches, nonlinear methods have been marginally explored. Approximate entropy (ApEn) [22
], sample entropy (SampEn) [23
], central tendency measure [24
], and Lempel–Ziv complexity [25
] have demonstrated their usefulness to characterise desaturations linked to apnoeic events both in adults [26
] and children [16
]. Nevertheless, we hypothesise that different nonlinear metrics could gain insight into the dynamics of oximetry leading to additional and essential information. Furthermore, common apnoeic events in children with OSAS lead to slight fluctuations in SpO2
recordings compared with deeper desaturations commonly present in adult patients. Consequently, screening for paediatric OSAS using only information from nocturnal oximetry is more challenging and thus more powerful methods are needed to thoroughly characterise all the changes linked with the disease. In the present paper, we propose the multiscale entropy (MSE) as a method able to exhaustively inspect nonlinear dynamics of SpO2
MSE is a nonlinear measure of complexity previously applied in different medical frameworks to quantify entropy changes in biomedical recordings over time scales [33
]. In this regard, MSE has demonstrated to be useful to characterise differences in the heart rate dynamics due to age [33
], obesity [34
] and cardiac disease [35
] or to analyse human gait [36
], as well as to quantify changes in the complexity of the electroencephalogram (EEG) background activity in Alzheimer’s disease patients [37
] and EEG changes due to pharmacological intervention in schizophrenia [38
]. Similarly, MSE has been applied to cerebral oxygenation signals from infrared spectroscopy in order to study mortality and brain injury in preterm infants [39
]. In the context of OSAS, MSE has been recently used to analyse heart rate dynamics in adult patients. Particularly, in the study by Pan et al. [40
], MSE was applied to estimate the deterioration in autonomic and vascular regulatory function linked with increasing OSAS severity and the subsequent improvement after continuous positive airway pressure treatment. Similarly, MSE has demonstrated to be useful in the analysis of speech signals in order to quantify disorderliness in vocal patterns indicative of sleep apnoea [41
]. In a previous study by our group [42
], MSE was also applied to characterise the dynamics of heart rate variability time series in order to derive new patterns able to detect adult OSAS.
The aim of this study was two-fold: (i) firstly, to accomplish a comprehensive analysis of oximetry dynamics by means of MSE in order to characterise differences between non-OSAS children and paediatric patients suffering from the disease; (ii) and second, to assess the usefulness of MSE-derived features in order to compose an optimum model from unattended oximetry able to accurately screen for paediatric OSAS at home.
plots individual SampEn values as a function of τ
for every overnight oximetric recording in the population under study, as well as the average for the whole OSAS-positive and OSAS-negative groups. Despite the inherent variance, OSAS-positive patients showed greater averaged entropy, i.e., irregularity, than OSAS-negative children due to desaturations caused by apnoeic events for every time scale. We performed a visual inspection of the averaged MSE profiles to properly parameterise each curve. Regarding the smaller time scales, it is important to note that SampEn values of OSAS-positive patients increased with a substantially higher slope than for the OSAS-negative group from scales τ
= 1 to τ
= 6. Then, SampEn values increased monotonically for both groups until reaching a similar slope. In addition, we can observe that there was a maximum difference between both MSE averaged profiles for time scale τ
= 14. In order to gather this information, the following parameters were derived from the MSE profile of each oximetric recording:
Slope of the MSE curve between scale τ = 1 and scales τ = 2 (Slp1-2), τ = 3 (Slp1-3), τ = 4 (Slp1-4), τ = 5 (Slp1-5) and τ = 6 (Slp1-6). It is estimated as the slope of the straight-line connecting the MSE values of the time scales under study. Higher slope accounts for a larger entropy increase between the original signal (τ = 1) and coarse-grained versions in consecutive short time scales (τ = 2 to 6), i.e., the control mechanisms regulating peripheral blood oxygen saturation on such short time scales are the most affected by recurrent apnoeic events.
Individual SampEn values from scale τ = 1 to scale τ = 6 (SE1 to SE6). Single-scale SampEn is a measure of entropy or disorderliness and thus larger individual values are linked with more complex underlying mechanisms governing the dynamics of the oximetric signal for these time scales.
SampEn single value in the scale reaching the maximum margin between MSE curves of the groups under study, i.e., τ = 14 (SEmax). This feature quantifies the irregularity of the oximetric recording for the time scale where the maximum difference between the classes under study (OSAS-negative vs. OSAS-positive) is expected.
Area enclosed under the MSE curve between scale τ
= 1 and scales τ
= 2 (Ar1-2
= 4 (Ar1-4
) and τ
= 6 (Ar1-6
). MSE curves allow us to compare the relative complexity of time series [33
]. Higher area is achieved when SampEn values are higher for the majority of the time scales, suggesting that the time series is more complex.
Area enclosed under the MSE curve between scale τ = 1 and the scale reaching the maximum margin (τ = 14) between the averaged MSE curves (Ar1-max). After time scale τ = 14, the MSE curves of OSAS-negative and OSAS-positive groups monotonically increase with a similar slope, showing almost equal behaviour. From short time scales to scale τ = 14, the MSE curves of both groups show the greatest differences regarding shape and individual entropy values. Thus, this feature gathers the contribution of the time scales showing the maximum differences in the dynamics of nocturnal oximetry between the groups under study.
Time scale where the maximum SampEn value is reached (τmax). This feature is related to the level of depth of changes in the underlying complexity of the signal, i.e., it shows the time scale up to which entropy increases.
summarises the median and interquartile range (IQR) of all these MSE-derived parameters for the OSAS-negative and the OSAS-positive groups. It is noticeable that almost all features achieved statistically significant differences between groups (p
< 0.05). On average, OSAS-positive patients showed significantly higher slopes (Slp1-2
), higher irregularity (SE1
), and higher area under the MSE curve (Ar1-2
) in the smaller time scales than OSAS-negative children. Similarly, OSAS-positive patients also showed significantly higher area under the MSE curve between time scales 1 and the maximum-margin scale (τ
= 14) as well as higher entropy in such a maximum-margin scale than OSAS-negative children.
summarises the diagnostic performance of every entropy-based parameter derived from MSE analysis. Almost all features under study showed balanced sensitivity and specificity values, as well as moderate diagnostic accuracy. Regarding slope-based MSE features, accuracy ranged 71.4% to 73.4% and both Slp1-3
reached the maximum AUC (0.80). Similarly, the accuracy of area-based MSE features ranged 71.3% to 72.3% and Ar1-2
reached 0.80 AUC. In regard to SampEn values at individual time scales, accuracy ranged 68.1% to 73.4% and both SE1
achieved 0.81 AUC. Finally, τmax
reached poor accuracy (Acc = 55.8%) and poor area under the ROC curve (AUC = 0.60). Table 4
shows the performance of every conventional oximetric index involved in the study. Accuracy ranged 58.8% to 74.5% and all indices showed balanced sensitivity and specificity. It is important to highlight that ODI3 performed notably higher than the remaining conventional features, reaching 74.5% Acc (71.9% Se and 77.6% Sp) and 0.85 AUC.
shows the number of times each variable was selected after FSLR feature selection using the bootstrapping approach proposed to improve model generalization. A total of nine features (five from MSE analysis and four conventional indices) were above the specified threshold (10% of total runs): Slp1-2
. Using these features, an optimum LR model was composed. Figure 4
shows the ROC curves during the training stage of the bootstrapping procedure (average across all bootstrap replicates) for each feature subset under study, i.e., MSE-derived features (MSE), conventional oximetric indices (OX), all features together without feature selection (MSE-OX) and the optimum feature subset from FSLR (OPT). We can observe that the feature subsets using MSE-derived features and oximetric indices jointly achieved notably higher AUC than each single approach individually, which supports our initial hypothesis of complementarity between both approaches.
summarises the diagnostic performance of all approaches involved in the study. A LR model composed of all the MSE-derived features reached 75.2% Acc (75.7% Se, 75.3% Sp) and 0.79 AUC, whereas a LR model built with all the oximetric indices reached 76.0% Acc (74.7% Se, 77.7% Sp) and 0.82 AUC. Similarly, a LR model composed of all the MSE and oximetric variables under study reached comparable performance, achieving 79.0% Acc (79.4% Se, 79.3% Se) and 0.80 AUC. Interestingly, our optimum LR model composed of features automatically selected by FSLR performed significantly better, reaching 83.5% Acc (84.5% Se, 83.0% Sp) and 0.86 AUC.
In this study, we analysed the usefulness of MSE in the context of screening for paediatric OSAS by means of overnight unattended oximetry at home. Nocturnal oximetry has emerged as a reliable technique for simplified OSAS detection in children. The present research gains additional insight into the diagnostic capability of portable oximetry. Our results showed that MSE is able to provide relevant and complementary information to conventional oximetric indices commonly used by physicians. MSE profiles of oximetric recordings showed a consistent behaviour, with significant differences between OSAS-negative and OSAS-positive children, particularly for lower time scales. In addition to single-scale entropy measures from scales τ
= 1 to τ
= 6, where maximum visual differences were found, all the proposed slope- and area-based parameters derived from the MSE curve achieved statistically significant differences between groups. According to these MSE-measures, OSAS-positive patients showed significantly higher average irregularity than non-OSAS children due to desaturations linked with OSAS along all the time scales. Regarding the diagnostic performance, individual MSE-based parameters reached moderate accuracy (maximum Acc of 73.4%), similar to that achieved by conventional ODI3 (74.5% Acc). A multivariate analysis showed that putting all this data together led to a notably overall performance increase, which demonstrated that MSE is able to provide additional as well as complementary information. A LR model composed of all features under study reached 79.0% Acc (79.4% Se, 79.3% Sp) and 0.80 AUC. Furthermore, our novel LR model fed with an optimum feature subset derived from FSLR feature selection reached 83.5% Acc (84.5% Se, 83.0% Sp) and 0.86 AUC, which significantly outperformed all the approaches under study. Comparing the ROC curves from training shown in Figure 4
and the final performance of each LR model in terms of the AUC values shown in Table 5
, we can observe that LROX
showed the smallest decrease in performance. On the other hand, the model composed of all the variables from the original feature space (LRMSE-OX
) showed a significant performance decrease, which supports the convenience of our optimum feature subset derived from FSLR. Therefore, LROPT
reached higher diagnostic accuracy as well as higher generalization ability.
It is essential to highlight that the optimum feature subset was composed of variables from both MSE analysis (five features: Slp1-2
) and conventional indices (four features: ODI3
), which points out the complementarity of the proposed MSE approach. Regarding the physiological interpretation of the optimum features, it is widely known that ODI3
quantifies the number of desaturations per hour of recording, whereas the remaining conventional indices account for the overall (SatAVG
) and maximum (SatMIN
) severity of the desaturations. On the other hand, it is unknown how the proposed MSE-derived measures are related to particular changes in the dynamics of oximetry. As proposed by Costa et al. [33
], we used MSE curves to compare the relative complexity of normalised time series, i.e., oximetric recordings belonging to OSAS-negative and OSAS-positive children. Accordingly, Slp1-2
reflect that the degree of change in the complexity of overnight oximetry (computed as the slope of the MSE curve) due to apnoeic events is more relevant in smaller scales (scales 1-2 and 1-6). The single-scale entropy measure SE1
shows that original (τ
= 1) oximetric recordings from OSAS-positive children have significantly higher irregularity than OSAS-negative children. Moreover, the influence of apnoeic events is still relevant in moderate time scales since SEmax
reflects significantly higher irregularity in oximetric recordings from OSAS-positive patients for time scale τ
= 14. Finally, according to Table 2
did not show statistically significant differences between groups. Nevertheless, it was automatically selected by FSLR, suggesting that the level up to which entropy increases due to the influence of apnoeic events provides complementary information to direct measures of entropy.
In order to gain insight into the performance of our binary classifier, we analysed polysomnographic and polygraphic features of misclassified children. Notice that a bootstrapping technique was carried out to validate our approach and thus all performance metrics were computed as the average across all the bootstrap replicates. Therefore, we analysed those patients misclassified in a significant number of repetitions of the algorithm. Accordingly, there were five false positive children. Two of them were borderline, showing an OAHI from PSG equal to 2.5 and 2.6 events/h, respectively. It is important to highlight that children were diagnosed according to in-lab PSG, which is the gold standard, whereas our screening method is based on the oximetry signal recorded at-home in a different night. Hence, it is essential to consider common night-to-night variability inherent to OSAS when analysing misclassifications. In this regard, three out of five false positive patients showed an at-home OAHI > 3 events/h, ranging 3.6 to 7.7 events/h. Regarding false negative patients, there were five OSAS-positive children incorrectly classified as no-OSAS using our screening oximetry-based tool. It is important to note that they were all moderate-to-severe OSAS patients (in-lab OAHI > 5 events/h), showing a similar behaviour at home (unsupervised OAHI ranging 4.0 to 11.1 events/h). Nevertheless, four out of five false negative children showed an at-home ODI3 significantly lower, ranging 0.5 to 2.3 events/h. This suggests that apnoeic events did not lead to a matching desaturation in the oximetric profile, which is probably the main limitation of oximetry as a single screening tool for the disease.
Previous studies used MSE in the context of automated characterisation of OSAS in adults. In a recent study by Roebuck and Clifford [41
], MSE was applied to characterise irregularity of speech patterns from subjects suspected of suffering from sleep apnoea. MSE coefficients of speech signals for small (τ
= 1, 2, 4, 8) and large (τ
= 16, 32, 65, 130, 180) scales were used to detect moderate-to-severe OSAS (AHI ≥ 15 events/h) using a random forest classification paradigm. An overall accuracy of 79.9% (66.0% Se, 88.8% Sp) was obtained, whereas the performance increased up to 80.5% Acc (69.2% Se, 87.9% Sp) when demographic variables were added to the model. Pan et al. [40
] analysed heart rate variability (HRV) time series of snoring patients with and without OSAS by means of MSE in order to assess changes in autonomic and vascular regulatory function. A significant irregularity decrease in HRV dynamics both for small (τ
< 6) and large (τ
> 6) time scales were found for moderate-to-severe OSAS patients (AHI ≥ 15 events/h), whereas non-OSAS subjects and patients with continuous positive airway pressure therapy showed a similar increase in MSE features in the larger scales. Similarly, Gutiérrez-Tobal et al. [42
] applied MSE to HRV recordings in order to model adult OSAS in independent populations separated by gender. Using together MSE coefficients and spectral entropy measures automatically selected by means of FSLR, a LR model achieved 85.2% Acc (80.8% Se, 89.3% Sp) in the classification of women with OSAS (AHI ≥ 10 events/h). The performance decreased to 77.6% Acc (87.1% Se, 56.1% Sp) when modelling OSAS in men.
To our knowledge, this is the first study assessing MSE in the context of childhood OSAS. Moreover, in the present research, we focused on the diagnostic ability of unattended oximetry at home as a single screening tool for the disease, which is a major novelty in the framework of paediatric sleep apnoea. Previously, few studies assessed the usefulness of automated analysis of overnight portable oximetry [16
]. Kirk et al. [20
] analysed a dataset composed of 57 children suspected of suffering from OSAS. The oxygen desaturation index > 4% (ODI4) from unattended portable oximetry was used to characterise OSAS (AHI ≥ 5 events/h), reaching 66.7% sensitivity and 60.0% specificity when a cut-off of ODI4 ≥ 5 events/h was used. In the study by Garde et al. [16
], time and spectral features from overnight pulse oximetry were used to assist in paediatric OSAS detection (AHI ≥ 5 events/h). The Authors analysed 146 SpO2
recordings acquired by means of a portable device, although all sleep studies were carried out in a supervised hospital setting. Stepwise linear discriminant analysis (LDA) reached 78.5% accuracy (80.0% Se, 83.9% Sp) using only features from SpO2
, whereas the performance increased up to 84.9% accuracy (88.4% Se, 83.6% Sp) when features from SpO2
and pulse rate were used jointly. In the study by Sahadan et al. [49
], pulse rate time series from unattended pulse oximetry was analysed to assist in the management of childhood OSAS (AHI ≥ 1 event/h). The quantification of pulse rate increases of 15 bpm (PRI-15) reached the highest performance, achieving 18.0% sensitivity and 97.0% specificity when a cut-off of PRI-15 > 35/h was used. In a previous study by our group [18
], ODI3 from nocturnal unsupervised SpO2
was combined with spectral measures from at-home airflow recordings to characterise children with suspected OSAS (OAHI ≥ 3 events/h). A LR model from stepwise feature selection reached 86.3% accuracy (85.9% Se, 87.4% Sp). Cohen and De Chazal [17
] analysed a large dataset composed of 288 children showing suspicion of suffering from OSAS. ECG and SpO2
from unattended PSG at home were automatically processed in order to detect every individual apnoeic event. An accuracy of 74.7% (39.6% Se, 76.4% Sp) was reached under an epoch-based classification approach using a LDA model only composed of features from time-frequency analysis of ECG. Conversely, the accuracy decreases up to 66.7% (58.1% Se, 67.0% Sp) when statistics from SpO2
were added to the model. In a recent study by our group [19
], single-scale non-linear measures of entropy, complexity and variability were combined with conventional statistics, spectral features and oximetric indices to parameterise unattended SpO2
recordings. Optimum LR models were composed from stepwise feature selection for different cut-offs for the disease, reaching 83.4% accuracy (82.9%, 84.4%) for a clinical threshold of OAHI ≥ 3 events/h. In the present research, we reached similar diagnostic performance using MSE as unique complement of conventional oximetric indices, suggesting that different nonlinear methods other than SampEn, central tendency measure (CTM) or Lempel-Ziv complexity (LZC) are also able to provide relevant information in the context of paediatric OSAS detection from oximetry.
A trade-off between the reduction of complexity of the diagnostic methodology by means of simplified techniques and the diagnostic accuracy have to be taken into account. Previous studies reported contradictory data regarding the complementarity of information from different signals in the context of paediatric OSAS. Some studies [16
] showed a slight-to-moderate performance increase when using jointly features from oximetry and other cardiorespiratory signals such as pulse rate or airflow, whereas other studies [17
] reported a significant decrease in performance when oximetry and ECG recordings are combined. Similarly, a recent study by Álvarez et al. suggested that automated analysis of unattended oximetry at home might be as accurate as manual scoring of at-home respiratory polygraphy, particularly when using low OAHI cut-off points for a positive diagnosis of the disease [19
]. Therefore, additional robust evidences are still needed to define the best combination of cardiorespiratory signals in order to design an accurate as well as simplified screening tool for paediatric OSAS.
Some limitations should be taken into account to be able to properly generalise our conclusions. Firstly, a larger population would allow for optimal design and assessment of the proposed LR models. Notwithstanding, a bootstrapping approach was conducted both for feature selection and classification in order to overcome this drawback. In the same way, a wider dataset would let us a better characterisation of overnight oximetry dynamics by means of MSE for childhood OSAS. Nevertheless, our results revealed a quite consistent trend of average MSE curves for OSAS-negative and OSAS-positive groups, as well as statistically significant differences between groups for almost all MSE-based parameters. Regarding feature extraction from overnight oximetry, our findings suggest that multiscale processing methods provide relevant information from oximetric recordings. In this regard, additional time-scale techniques, such as the wavelet transform, could provide useful as well as complementary features to MSE and conventional clinical indices in the framework of childhood OSAS detection from oximetry. Finally, the proposed methodology focused on binary classification. Although this is a very useful approach in order to implement automated screening tools for the disease, it would be very interesting to develop a pattern recognition scheme aimed at classifying patients into the four common categories of severity, i.e., non-OSAS, mild, moderate and severe.
LR could be considered the reference classifier in the context of automated pattern recognition to assist in childhood OSAS. Linear discriminant analysis (LDA) [16
] and LR [1
] have been predominantly used for binary classification of children suspected of suffering from the disease. LDA assumes that all the input variables show normal distribution and equal variances, assumptions that are not always consistent in real-world pattern classification tasks. LR provides a more general approach that fits better to the characteristics of the problem under study. Nevertheless, additional automated pattern recognition techniques such as decision trees, artificial neural networks or support vector machines, which have demonstrated its usefulness in the context of adult OSAS [31
], need to be assessed in the context of paediatric OSAS.