7th Dutch Bio-Medical Engineering Conference
January 24th & 25th 2019, Egmond aan Zee, the Netherlands
10:30   Sleep
Chair: Jurriaan de Groot
15 mins
Cardiorespiratory Interactions in Sleep Apnea and Associated Comorbidities
Carolina Varon, Margot Deviaene, Dries Hendrikx, Sara Van de Putte, Dries Testelmans, Bertien Buyse, Sabine Van Huffel
Abstract: Aims: The severity of sleep apnea is often assessed using the apnea/hypopnea index (AHI), which is known to be inaccurate in the phenotyping of apnea patients. Hence, better approaches are needed to characterize these patients and to allow cardiovascular risk stratification, and treatment selection. In this context, this work studies the cardiorespiratory interactions in patients suffering from both sleep apnea and apnea associated comorbidities by means of graph theory and kernel methods. Methods: Heart rate variability (HRV), respiration (airflow and effort), and SpO2 signals recorded from 100 sleep apnea patients (AHI>15) and 10 controls (AHI<5) were used to construct the cardiorespiratory graph. A subgroup of 50 (comorbidity) patients presented apnea associated comorbidities. From both the HRV and the respiration, the powers in the classical LF and HF bands were computed using a moving-window approach (=60s, 50s overlap). The mean desaturation in the 60s window was used as feature for the SpO2 signal. In total, 5 time series, 2 for HRV, 2 for respiration, and 1 for SpO2, were obtained with a sampling period of 10s. These series correspond to the vertices of the cardiorespiratory graph and their degree and interactions during the full night were analysed using the RBF kernel. The connectivity of the graph for each patient was estimated using the evolution of the graph algebraic connectivity. Results: The total connectivity of the graph is reduced with higher AHI and this reduction is significant (p<0.05) in the comorbidity group. Furthermore, in the comorbidity patients with AHI>35, the vertex degree of the LF band of HRV is significantly lower and the link between the respiration vertices and SpO2 is significantly weaker. Conclusions: These results are in line with studies that report stronger oxygen desaturations in patients with apnea associated comorbidities, and more unstable control systems, which could be used for a better characterization of apnea patients.
15 mins
Unsupervised Artefact Detection in Emfit Sensor with Screening of Sleep Apnea Patients
Huysmans Dorien, Bertien Buyse, Dries Testelmans, Sabine Van Huffel, Carolina Varon
Abstract: Aim: The commercial Emfit QS (Emfit) pressure sensor is prone to large deviations in the signal (artefacts) due to posture changes and movements. The aim is to identify these artefacts while not relying on visual scoring or information from other sensors. Therefore, an unsupervised approach is developed. It was hypothesized that the percentage of artefacts increases with the patient’s Apnea-Hypopnea Index (AHI). Methods: The Emfit sensor and polysomnography recorded data of 37 patients during sleep diagnosis in UZ Leuven. The pressure signal was preprocessed, decomposed by wavelet transformation and analysed with time-domain features. Next, unsupervised feature selection was based on Robust Spectral Learning (RSFS). The parameters of RSFS, number of features and clusters were optimized by a grid search and running the k-means algorithm for two clusters. The optimised parameters clustered the training points in two clusters, i.e. clean and artefact. Every test data point was mapped to the closest centroid of both clusters in Euclidean distance to label it clean or artefact. The percentage of artefacts per patient was grouped according to standard sleep apnea classes based on AHI. To separate severe sleep apnea (AHI≥30) patients from other sleep apnea patients (AHI<30), a receiver operating characteristic curve analysis was performed. Results: Three features highlighting peak variation were optimal for artefact detection. The percentage of artefacts in the data was a potential parameter to target severe sleep apnea patients. A threshold of 14.6% artefacts resulted in a sensitivity of 80% and specificity of 87%. Conclusion: The Emfit pressure sensor was explored in its potential for sleep apnea screening. An unsupervised algorithmic pipeline based on clustering was developed to detect artefacts and relate these artefacts to sleep apnea classes. The percentage of artefacts in the data was shown a potential parameter to target severe sleep apnea patients.
15 mins
An Improved Deep Convolutional Neural Network for Sleep Stage Classification in Preterm and Term Neonates
Amir Hossein Ansari, Kirubin Pillay, Ofelie De Wel, Anneleen Dereymaeker, Katrien Jansen, Alexander Caicedo, Jan Vervisch, Sabine Van Huffel, Gunnar Naulaers, Maarten De Vos
Abstract: Sleep stage analysis provides valuable information about the maturation progress and brain development in preterm and term babies [1], [2]. Accordingly, in neonatal brain monitors, an automated sleep stage classifier can play an important role to improve potentially the quality of monitoring and consequently the treatment. Recently, we introduced a deep neural network for EEG quiet sleep detection in preterm newborns [3]. In the current research, we redesign the architecture of that network in order to be able to classify the two sleep stages of preterm babies with higher accuracy, as well as the four sleep stages of term neonates. Our previously developed deep convolutional neural network has been improved using more filters in the convolutional layers, as well as employing dropout and batch-normalization layers. Then, this deep 12-layer neural network was trained by an advanced back-propagation optimizer, Adadelta [4], with 70 multi-channel EEG recordings of preterm and term neonates. As a result, in addition to the detection of quiet sleep periods in preterm and term babies, this network is able to classify the four term neonates’ sleep stages: 1) trace alternante, 2) high-voltage slow-wave, 3) active sleep type I, and 4) low-voltage irregular. The proposed algorithm has significantly better performance with respect to different metrics including sensitivity, specificity, and kappa compared to our previously developed algorithms, as well as considered state-of-the-art methods, which have been trained and tested on common datasets. The proposed method is also suitable for real-time monitoring.
15 mins
Modelling Sleep State Misperception at Sleep Onset
Lieke Hermans, Tim Leufkens, Merel van Gilst, Tim Weysen, Marco Ross, Peter Anderer, Sebastiaan Overeem, Annemiek Vermeeren
Abstract: Insomnia patients often overestimate their sleep onset latency (SOL). The mechanism underlying this type of sleep state misperception is not fully understood. We hypothesize that the length of uninterrupted sleep fragments after sleep onset influences the perception of the SOL because too short sleep fragments might be overlooked. We attempt to make a model of the minimum length that a sleep fragment should have in order to be perceived as sleep, and we fit the model to subjective data of insomniacs and healthy controls. Standard in-lab polysomnographic recordings were performed in 20 elderly, untreated insomniacs and 21 age-matched self-defined good sleepers. Recordings were visually scored according to R&K criteria1. In the model sleep onset was defined as the first epoch of the first sleep fragment longer than L minutes, with L varying from 0.5 to 40. We selected the length L that resulted in the smallest Mean Square Error (MSE) of the difference between modelled SOL and SOL perceived by the subject. This was done for both subject groups separately. For insomniacs, the lowest MSE was found for a length L of 30 minutes (MSE without model: 7195 vs. L=30: 3927). In the healthy subjects, applying the model only resulted in small improvements of the MSE. The lowest MSE was found for L=10 (MSE without model: 1185 vs. L=10: 969), although the results for all model parameters L below 20 were very similar. The aim of this study was to investigate the mechanisms underlying sleep onset misperception, by modelling the influence of sleep interruption on subjective SOL. The results indicate that in insomnia patients the perception of sleep onset can be influenced if a sleep fragment is interrupted after less than 30 minutes. The different results for the two groups suggest that, for the perception of sleep onset, insomniacs are more sensitive to sleep interruption than healthy subjects. In order to extend our findings to the general population, the analysis should be repeated in different age groups. Additionally, other parameters could be added to the model, for instance sleep depth and the duration of the sleep disruption.
15 mins
Night to Night Variability Analysis of the Saturation Signal in Children with Suspected Sleep Apnea.
Xenia Hoppenbrouwer, Parastoo Dehkordi, Aryannah Rollinson, Dustin Dunsmuir, Mark Ansermino, Guy Dumont, Ainara Garde
Abstract: Background Obstructive Sleep Apnea (OSA), the most common type of sleep disordered breathing, can result in developmental complications in children. Polysomnography (PSG), the gold standard to diagnose OSA, is resource intensive and confined to the hospital. The Phone oximeter (PO), a smartphone-based pulse oximeter that measures blood oxygen saturation (SpO2), can be a promising alternative to screen for OSA. To assess the PO’s reliability detecting OSA at home, we evaluated the night-to-night variability of overnight pulse oximetry recorded with the PO. Material and Methods Dataset: 74 children suspected of OSA and referred for PSG were recruited and pulse oximetry over multiple nights was recorded using the PO, including one night at the hospital simultaneously with the PSG and two nights at home. Analysis: SpO2 values below 70% and above 100% were considered artefacts and removed. Only SpO2 recordings with more than 3 hours of good quality were studied. The SpO2 analysis consisted of investigating the night-to-night variability of the oxygen desaturations index (ODI) and additional time-frequency features. ODI was defined as the amount of desaturations (>3% decrease from baseline) lasting at least one second. In addition, the SpO2 signal was characterized using a 2-minute sliding window (no overlap) from which variability features (e.g. time below 94%, t94%) and modulation features (i.e power in the modulation frequency band, PM) were extracted. The overnight distribution and the night-to-night variability of every feature were studied using linear mixed models. Results We characterized the SpO2 signals of 68 children (42 non-OSA, 26 OSA), with 63 PO recordings at the hospital and 53 and 46 recordings for the first and second night at home. ODI and SpO2 variability features showed significant differences between non-OSA and OSA (ODI: 1.41±1.32, 3.70±3.98 (m±std), p<0.01), but without any significant variability between nights (ODI: 2.03±2.76, 2.71±3.53, 2.19±2.21, combining both OSA and non-OSA groups). In general, no significant night-to-night variability was found in the overnight distribution of the rest of the studied features (i.e. PM). Conclusion Generally, the distribution of all features showed no significant night-to-night variability, which shows that PO could be used at home to screen for OSA.
15 mins
From Free Text to Structured Information in Sleep Description Analysis Using Deep Learning Networks
Heereen Shim, Stijn Luca Luca, Dietwig Lowet, Bart Vanrumste
Abstract: Human activity recognition and vital sign monitoring play a significant role in tailoring personal health. However, some aspects are difficult to be measured by sensors, such as problems, satisfaction, or wellbeing. One possible solution is to use direct feedback from the user via voice or text input. We propose deep neural networks based sleep description analysis to extract structured information relevant for sleep quality from unstructured free-text input data. In this study, we present two types of sequential deep neural network architectures. Word-level sequential neural network final classification layer for multi-label topic classification achieves 97.8% of accuracy on custom sleep description dataset collected via Mturk service. Character-level sequence-to-sequence model consisting of encoder and decoder networks for timestamp extraction shows 92 % of accuracy on generated timestamps dataset. Experimental results suggest that word-level features can be used for extracting semantic information and character-level features can be used for extracting numeric information from the free-text data. Also, the results highlight that the deep neural network models are robust to unusual expressions and misspelled words compared to rule-based model and the performances are highly dependent on embedding layers.