Please refer to xAAEnet to have the updated version

This repository provides the open-source codes and supplementary materials related to the publications:

• La Fisca et al, "Explainable AI for EEG Biomarkers Identification in Obstructive Sleep Apnea Severity Scoring Task", NER 2023.
• La Fisca et al, "Enhancing OSA Assessment with Explainable AI", EMBC 2023.

The pending patent application 23168189.1 describes a general method for scoring the severity of sleep-related disorders from polysomnographic (PSG) signals. Due to non-disclosure considerations, we are only able to make the codes for the xVAEnet model available as open-source data here.

Each trial in the dataset is composed of 23 channels and 3001 timestamps, as shown on Figure 1 Fig. 1. Data overview. Example of a preprocessed 60-seconds trial with OSA event. Channels: 1) nasal airflow, 2-3) abdominal-thoracic respiratory motions, 4) oxygen saturation, 5-6) electrooculograms, 7) pulse rate variability, 8) abdominal-thoracic motions phase shift, 9-23) EEG signal of the 3 selected electrodes at different frequency ranges.

The EEG signals have been preprocessed following the COBIDAS MEEG recommendations from the Organization for Human Brain Mapping (OHBM) [1]. Trials significantly affected by ocular artifacts have been excluded from the database, based on the correlation between the EOG and the FP1 signals. Trials with non-physiological amplitudes are also excluded, based on their peak-to-peak voltage (VPP): VP-P < 10^-7V and VP-P > 6 ∗ 10^-4V are excluded. A baseline correction was applied using a segment of 10 seconds preceding each trial as the baseline. The EEG delta band powe7 being the most varying frequency band during sleep apneahypopnea occurrence [2], we focused our analysis on low frequency EEG components by filtering the signals into 2Hz narrow bands: 0-2Hz, 2-4Hz, 4-6Hz, 6-8Hz, and 8-10Hz. We also rejected trials based on physiological fixed range criteria on VP-P for EOG and SAO2 signals, moreover trials with VAB, VTH and NAF2P statistical outliers in amplitude are rejected. Two additional signals have been computed from the aforementioned recorded signals: 1) the Pulse Rate Variability (PRV) being the difference between a PR sample and the next one, and 2) the belts phase shift (Pshift), computed as the sample by sample phase difference between VAB and VTH phase signals, as suggested by Varady et al. [3]. The normalization has been performed by channel independently as a z-score normalization with clamping in the [-3; 3] range. After the exclusion and preprocessing phases, the final dataset is composed of 6992 OSA trials from 60 patients divided into a training set of 4660 trials from 48 patients, namely the trainset, and a validation set of 2332 trials from the 12 remaining patients, namely the testset.

[1] C. Pernet, M. I. Garrido, A. Gramfort, N. Maurits, C. M. Michel, E. Pang, R. Salmelin, J. M. Schoffelen, P. A. Valdes-Sosa, and A. Puce, “Issues and recommendations from the OHBM COBIDAS MEEG committee for reproducible EEG and MEG research,” Nature Neuroscience, vol. 23, no. 12, pp. 1473–1483, Dec. 2020, number: 12 Publisher: Nature Publishing Group.

[2] C. Shahnaz, A. T. Minhaz, and S. T. Ahamed, “Sub-frame based apnea detection exploiting delta band power ratio extracted from EEG signals,” in 2016 IEEE Region 10 Conference (TENCON), Nov. 2016, pp. 190– 193, iSSN: 2159-3450.

[3] P. Varady, S. Bongar, and Z. Benyo, “Detection of airway obstructions and sleep apnea by analyzing the phase relation of respiration movement signals,” IEEE Transactions on Instrumentation and Measurement, vol. 52, no. 1, pp. 2–6, Feb. 2003, conference Name: IEEE Transactions on Instrumentation and Measurement.

Fig2. Detailed xVAEnet architecture

The entire training phase has been performed on an NVIDIA GeForce GTX 1080Ti 12Go RAM on 12 workers. The first module to be trained is the VAE module. The training process has been performed on the trainset on 100 epochs with a batch size of 16, a learning rate of 5 · 10^-3 and a ranger optimizer, while the validation has been done on the testset with a batch size of 32. A gradient accumulation of 64 samples and an early stopping option based on the validation loss with a patience of 10 epochs have also been used. Then, the GAN module has been trained by initializing the generator with the best weights of the encoder obtained during the VAE training phase and the discriminator is randomly initialized. At each batch, the discriminator is first trained by freezing the generator and using the loss function of the discriminator described in Section III, then the generator is trained by freezing the discriminator and using the correspond- ing loss function. This training phase is performed on the trainset on 100 epochs with a batch size of 16, a learning rate of 2 · 10^-3 and a root mean square propagation (RMSprop) optimizer, while the validation process is done with a batch size of 32. A gradient accumulation of 64 samples and an early stopping option with a patience of 30 are also used. Every 15 epochs, the updated network is used in inference to compute a new Zd vector given as real input for the 15 following epochs in order to avoid the deterioration of the “real” space to be responsible for the increase of the GAN performance. Finally, the classifier module is initialized with the weights of the best generator previously obtained and the single- layer perceptron is randomly initialized. In the philosophy of curriculum learning, the classifier is trained on each severity feature sequentially, starting with the low vs. high severity classification on the hypoxic burden, then on the arousal event and finally on the event duration. All the training processes for the classification task have been performed on the trainset on 100 epochs with a batch size of 16 and a ranger optimizer, and have been validate on the testset with a batch size of 32 only the learning rate and the combination of loss functions differ. A gradient accumulation of 64 samples and an early stopping option based on the validation loss with a patience of 30 epochs have also been used. On the hypoxic burden, the learning rate was set to 10^-3 and, once every 5 epochs, the whole model was trained using a global loss combining the three modules:

On the arousal event, the learning rate was set to 5 · 10^-4, the global training was performed every 5 epochs and, once every 2 epochs, the classification has been performed on both the hypoxic burden and the arousal event using a weighted sum of both losses:

On the respiratory event duration, the learning rate was set to 2·10^-4, the global training was performed every 5 epochs and, twice every 3 epochs, the classification has been performed on all the severity features using a weighted sum of all classification losses:

Once trained, the VAE module is able to properly reconstruct the input PSG signals, with a final root-mean-square error (RMSE) of 0.196 on the trainset and 0.247 on the testset. The model including the GAN module still performs well in reconstruction with a RMSE of 0.214 on the trainset and 0.273 on the testset. The Gaussian distribution and generative ability are evaluated through the closeness between Ze and Zd defined by the discrimination efficiency of the discriminator. When giving samples from Zd (real input) to the discriminator, its mean accuracy reaches 81.6% (trainset) and 76.4% (testset) meaning that it performs well in recognizing real inputs. However, this mean accuracy decreases to 52.3% and 49.8% (testset) when samples from Ze (fake input) is given as input, meaning the generator generates Ze samples sufficiently similar to Zd samples to fool the discriminator.

The xAAEnet architecture is a variation of the xVAEnet, with several key differences. While both models have an encoder-decoder structure with a latent space in between, the xAAEnet's latent block has a simpler design, consisting of only one dense layer and one batch normalization layer. This block does not use any reparameterization technique, in contrast to the xVAEnet, which employs a variational autoencoder approach. Additionally, the xAAEnet's final block is a regressor block that includes a single-layer perceptron with one output, and no activation function. These changes were made to adapt the model for the specific task of severity scoring in obstructive sleep apnea, and to simplify the training process.

The entire training phase has been performed on an NVIDIA GeForce GTX 1080Ti 12Go RAM on 12 workers. The first module to be trained is the autoencoder (AE) module. The training process has been performed on the trainset on 50 epochs with a batch size of 16, a learning rate of 1 · 10^-3 and a ranger optimizer, while the validation has been done on the testset with a batch size of 32. A gradient accumulation of 64 samples and an early stopping option based on the validation loss with a patience of 15 epochs have also been used.

Then, the scoring unit is initialized with the weights of the best encoder previously obtained and the single- layer perceptron (SLP) is randomly initialized. The training process for the scoring task, based on the hand-crafted score S_h described in the paper, has been performed on the trainset on 50 epochs with a batch size of 16, a learning rate of 5 · 10^-4 and a ranger optimizer, and have been validate on the testset with a batch size of 32. A gradient accumulation of 64 samples and an early stopping option based on the validation loss with a patience of 25 epochs have also been used.

Finally, the GAN module has been trained by initializing the generator with the best weights of the encoder obtained during the training phase of the scoring unit and the discriminator is randomly initialized. At each batch, the discriminator is first trained by freezing the generator and using the loss function of the discriminator described in Section III C, then the generator is trained by freezing the discriminator and using the correspond- ing loss function. This training phase is performed on the trainset on 50 epochs with a batch size of 16, a learning rate of 2 · 10^-4 and a root mean square propagation (RMSprop) optimizer, while the validation process is done with a batch size of 32. A gradient accumulation of 64 samples and an early stopping option with a patience of 25 are also used.

[1] A. S. Jordan, D. G. McSharry, and A. Malhotra, “Adult obstructive sleep apnoea,” The Lancet, vol. 383, no. 9918, pp. 736–747, Feb. 2014.
[2] R. B. Berry, R. Budhiraja, D. J. Gottlieb, D. Gozal, C. Iber, V. K. Kapur, C. L. Marcus, R. Mehra, S. Parthasarathy, S. F. Quan, S. Redline, K. P. Strohl, S. L. D. Ward, and M. M. Tangredi, “Rules for Scoring Respiratory Events in Sleep: Update of the 2007 AASM Manual for the Scoring of Sleep and Associated Events,” Journal of Clinical Sleep Medicine, vol. 08, no. 05, pp. 597–619, 2012, publisher: American Academy of Sleep Medicine.
[3] D. A. Pevernagie, B. Gnidovec-Strazisar, L. Grote, R. Heinzer, W. T. McNicholas, T. Penzel, W. Randerath, S. Schiza, J. Verbraecken, and E. S. Arnardottir, “On the rise and fall of the apneahypopnea index: A historical review and critical appraisal,” Journal of Sleep Research, vol. 29, no. 4, p. e13066, 2020.
[4] A. Malhotra, I. Ayappa, N. Ayas, N. Collop, D. Kirsch, N. Mcardle, R. Mehra, A. I. Pack, N. Punjabi, D. P. White, and D. J. Gottlieb, “Metrics of sleep apnea severity: beyond the apnea-hypopnea index,” Sleep, vol. 44, no. 7, p. zsab030, Jul. 2021.
[5] D.-H. Park, C.-J. Shin, S.-C. Hong, J. Yu, S.-H. Ryu, E.-J. Kim, H.-B. Shin, and B.-H. Shin, “Correlation between the Severity of Obstructive Sleep Apnea and Heart Rate Variability Indices,” Journal of Korean Medical Science, vol. 23, no. 2, p. 226, 2008.
[6] G. Bachar, B. Nageris, R. Feinmesser, T. Hadar, E. Yaniv, T. Shpitzer, and L. Eidelman, “Novel grading system for quantifying upper-airway obstruction on sleep endoscopy,” Lung, vol. 190, no. 3, pp. 313–318, Jun. 2012.
[7] A. Kulkas, P. Tiihonen, P. Julkunen, E. Mervaala, and J. T ̈oyr ̈as, “Novel parameters indicate significant differences in severity of obstructive sleep apnea with patients having similar apnea–hypopnea index,” Medical & Biological Engineering & Computing, vol. 51, no. 6, pp. 697–708, Jun. 2013.
[8] A. Muraja-Murro, A. Kulkas, M. Hiltunen, S. Kupari, T. Hukkanen, P. Tiihonen, E. Mervaala, and J. T ̈oyr ̈as, “Adjustment of apnea-hypopnea index with severity of obstruction events enhances detection of sleep apnea patients with the highest risk of severe health consequences,” Sleep and Breathing, vol. 18, no. 3, pp. 641–647, Sep. 2014.
[9] H. Korkalainen, J. T ̈oyr ̈as, S. Nikkonen, and T. Lepp ̈anen, “Mortality- risk-based apnea–hypopnea index thresholds for diagnostics of obstruc- tive sleep apnea,” Journal of Sleep Research, vol. 28, no. 6, p. e12855, 2019, eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/jsr.12855.
[10] W. Cao, J. Luo, and Y. Xiao, “A Review of Current Tools Used for Evaluating the Severity of Obstructive Sleep Apnea,” Nature and Science of Sleep, vol. 12, pp. 1023–1031, Nov. 2020, publisher: Dove Press.
[11] A. Zinchuk and H. K. Yaggi, “Phenotypic Subtypes of OSA: A Chal- lenge and Opportunity for Precision Medicine,” CHEST, vol. 157, no. 2, pp. 403–420, Feb. 2020, publisher: Elsevier.
[12] G. Labarca, J. Gower, L. Lamperti, J. Dreyse, and J. Jorquera, “Chronic intermittent hypoxia in obstructive sleep apnea: a narrative review from pathophysiological pathways to a precision clinical approach,” Sleep and Breathing, vol. 24, no. 2, pp. 751–760, Jun. 2020.
[13] S. Pusk ́as, N. Koz ́ak, D. Sulina, L. Csiba, and M. T. Magyar, “Quantita- tive EEG in obstructive sleep apnea syndrome: a review of the literature,” Reviews in the Neurosciences, vol. 28, no. 3, pp. 265–270, Apr. 2017.
[14] E. Sforza, S. Grandin, C. Jouny, T. Rochat, and V. Ibanez, “Is waking electroencephalographic activity a predictor of daytime sleepiness in sleep-related breathing disorders?” The European Respiratory Journal, vol. 19, no. 4, pp. 645–652, Apr. 2002.
[15] K. Dingli, T. Assimakopoulos, I. Fietze, C. Witt, P. K. Wraith, and N. J. Douglas, “Electroencephalographic spectral analysis: detection of cortical activity changes in sleep apnoea patients,” The European Respiratory Journal, vol. 20, no. 5, pp. 1246–1253, Nov. 2002.
[16] M. Younes, A. Azarbarzin, M. Reid, D. R. Mazzotti, and S. Redline, “Characteristics and reproducibility of novel sleep EEG biomarkers and their variation with sleep apnea and insomnia in a large community- based cohort,” Sleep, vol. 44, no. 10, p. zsab145, Oct. 2021.
[17] S. Nikkonen, H. Korkalainen, S. Kainulainen, S. Myllymaa, A. Leino, L. Kalevo, A. Oksenberg, T. Lepp ̈anen, and J. T ̈oyr ̈as, “Estimating daytime sleepiness with previous night electroencephalography, elec- trooculography, and electromyography spectrograms in patients with suspected sleep apnea using a convolutional neural network,” Sleep, vol. 43, no. 12, p. zsaa106, Dec. 2020.
[18] G. C. Guti ́errez-Tobal, D. ́Alvarez, A. Crespo, F. del Campo, and R. Hornero, “Evaluation of Machine-Learning Approaches to Estimate Sleep Apnea Severity From At-Home Oximetry Recordings,” IEEE Journal of Biomedical and Health Informatics, vol. 23, no. 2, pp. 882– 892, Mar. 2019, conference Name: IEEE Journal of Biomedical and Health Informatics.
[19] M. Ghassemi, L. Oakden-Rayner, and A. L. Beam, “The false hope of current approaches to explainable artificial intelligence in health care,” The Lancet Digital Health, vol. 3, no. 11, pp. e745–e750, Nov. 2021.
[20] P. Angelov and E. Soares, “Towards explainable deep neural networks (xDNN),” Neural Networks, vol. 130, pp. 185–194, Oct. 2020.
[21] A. Gramfort, M. Luessi, E. Larson, D. Engemann, D. Strohmeier, C. Brodbeck, R. Goj, M. Jas, T. Brooks, L. Parkkonen, and M. H ̈am ̈al ̈ainen, “MEG and EEG data analysis with MNE-Python,” Frontiers in Neuroscience, vol. 7, 2013.
[22] D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,” May 2014, arXiv:1312.6114 [cs, stat].
[23] X. Zhang, L. Yao, and F. Yuan, “Adversarial Variational Embedding for Robust Semi-supervised Learning,” in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, ser. KDD ’19. New York, NY, USA: Association for Computing Machinery, Jul. 2019, pp. 139–147.
[24] N. A. Eiseman, M. B. Westover, J. M. Ellenbogen, and M. T. Bianchi, “The Impact of Body Posture and Sleep Stages on Sleep Apnea Severity in Adults,” Journal of Clinical Sleep Medicine, vol. 08, no. 06, pp. 655– 666, 2012, publisher: American Academy of Sleep Medicine.
[25] S. G. Jones, B. A. Riedner, R. F. Smith, F. Ferrarelli, G. Tononi, R. J. Davidson, and R. M. Benca, “Regional Reductions in Sleep Electroencephalography Power in Obstructive Sleep Apnea: A High- Density EEG Study,” Sleep, vol. 37, no. 2, pp. 399–407, Feb. 2014.