睡眠呼吸中止症是一種呼吸的慢性疾病，患者在睡眠時會反覆的呼吸暫時停止，在臨床上，會造成病人甦醒、嗜睡，而長時間下來則會引發高血壓等心血管疾病，嚴重甚至猝死。睡眠呼吸中止症大致上可以分成兩種：阻塞型睡眠呼吸中止症(OSA)及中樞型睡眠呼吸中止症(CSA)。另外，一樣有著嚴重影響程度的睡眠呼吸徵狀：睡眠淺呼吸(Hypopnea)，一併有以上三種睡眠呼吸中止事件。呼吸中止在醫學上定義為呼吸氣流下降超過90%以上並且持續10秒鐘，而淺呼吸則是呼吸氣流下降超過30%持續10秒鐘並伴隨3%以上血氧濃度下降或大腦甦醒。醫學上定義睡眠呼吸及淺呼吸指標(AHI)來辦別患者的嚴重程度。由於病發的時間處在睡眠，常容易導致患者忽略甚至未察覺。當今最標準檢測呼吸中止的方法是透過睡眠多項生理檢查(Polysomnography, PSG)，經由睡眠技師逐一觀察睡眠期間的各項生理訊號，來診斷整晚睡眠呼吸中止的情況。由於整個檢測流程昂貴，並且須配戴多條感測器使患者不舒服並且受環境的限制，本研究使用低成本的兩顆整合型的感測器，包含三軸加速器及心電描技術，分別黏在左胸及左腹來取得胸、腹起伏訊號以及心電圖，再搭配PSG中的血氧濃度訊號，總共三種訊號透過一個人工智慧演算法，來自動感測患者整晚睡眠的呼吸中止現象。其中，心電圖(ECG)是透過人工智慧的演發法來偵測醒睡，目的是減少AHI預估的誤差。

本研究中，兩個透過三軸加速器的訊號先經過低通濾波器過濾心跳雜訊，再經由三軸選擇器(TAA selection method)將三軸整合成單一軸的起伏變化。接下來每0.5秒鐘萃取胸腹起伏訊號的振幅比例、頻率比例和延遲後血氧濃度訊號取每段20秒中的最小值、最大值、平均值、變異數，以上8個特徵加上當下血氧濃度基準值總共9個特徵值透過長短期記憶模型類別的遞歸神經網路分類出當下不同種類的病症。最後，將此結果結合血氧濃度3%下降萃取機制(SpO2 desaturation detector)以及醒睡判別機制(Sleep-Wake classification)來提升準確度並得到最終結果。

透過記憶性模型來偵測呼吸中止症，整體的敏感度及精密度比起已往使用SVM的結果有很大的提升，最終病人嚴重程度的準確度可達89%，AHI差值的平均為5.0，而每20秒的整體準確度為92%。

Sleep Apnea-Hypopnea Syndrome (SAHS) is a respiratory chronic disease that harms body health and worsens sleep quality of patients. The disease causes complete or partial cessation of breathing while sleeping, which are known as Apnea and Hypopnea event. On clinic, it abrupt awakenings accompanied by choking which cause fluctuation of blood pressure and heart rate. Furthermore, SAHS leads to severe disease such as hypertension and cardiovascular failure. Polysomnography (PSG) is the gold standard for diagnosing SAHS. However, PSG is limited environment, uncomfortable, and time consuming. To improve these shortcomings, this thesis uses two low-cost sensor, tri-axial accelerometer and pulse oximeter, as the only biophysiological parameters for automatically screening SAHS by a self-learn artificial intelligence system.

This thesis proposed an algorithm for detecting apnea and hypopnea events by thoracic movement (THO), abdominal movement (ABD), and blood oxygen saturation levels (SpO2) signal during sleep. Besides, to obtain the apnea-hypopnea index (AHI), a neural network model is proposed for classifying sleep or wake with electrocardiography (ECG) signals and SpO2 signal as features.

First, the algorithm reconstruct two three-dimensional signals from two tri-axial accelerometers (TAA) into two one-dimensional signals which represent the thoracic (THO) and abdominal (ABD) movement, which are called TAA-THO and TAA-ABD. Second, segment TAA-THO and TAA-ABD in 10-second to extract both two features, fundamental frequency ratio and 95% quantile amplitude ratio, in every 0.5 second. On the other side, the SpO2 signal is segmented into 20-second after a 20-second delay to extract four features, minimum, maximum, mean, and variance of the first derivative. With above eight parameters and the original SpO2 index in every 0.5 second as features, the proposed algorithm use a LSTM-RNN model to classify every segment into 4 events types such as obstructive sleep apnea events (OSA), central sleep apnea event (CSA), hypopnea event (HYP), and normal event (NOR). Further, a corrective algorithm which detects SpO2 desaturation is added to the classification result. Finally, combining with the sleep-wake detector, the algorithm reports an AHI and sleep status of whole night second by second.

With the memorable and sequential characteristic in RNN model, the proposed algorithm is able to confront pathological characteristics difference between patients. According to our database of 115 subjects in CGUMH, the final AHI classification accuracy is 89.6%, and AHI difference between PSG is 5.0. The overall accuracy of detecting apnea events second by second is 92.1%, and the accuracy of classify 4 classes in every second is 83%.

The results indicates that LSTM-RNN has great potential to classify SAHS due to the similar memorable function to PSG labeling procedure.

1 Introduction 1
1.1 Sleep Apnea Syndrome . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2
1.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Study Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Organization of This Thesis . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Medical Diagnosis of Sleep Apnea Syndrome 9
2.1 Polysomnography Examination . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Wireless Sensing Architecture . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Sleep Apnea Syndrome Severity and Sleep Apnea Event Types . . . . . . . . 13

3 Proposed Respiration Signal Processing Algorithm 23
3.1 Reconstructed Movement Signal from Tri-Axial Accelerometer . . . . . . . . .25
3.1.1 Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.2 Trial-Axial Accelerometer Selection Algorithm . . . . . . . . . . . . . . 25
3.2 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . .28
3.2.1 Features of THO and ABD . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.2 Feature of SpO 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .32
3.3 Neural Network Model Classifier . . . . . . . . . . . . . . . . . . . . . . .34
3.3.1 Neural Network Model for SAHS . . . . . . . . . . . . . . . . . . . . . . 34
3.3.2 LSTM-RNN Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .35
3.3.3 Proposed LSTM-RNN Architecture . . . . . . . . . . . . . . . . . . . . . .37
3.4 Signal Adjustment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4.1 Oxygen Desaturation Detection Algorithm . . . . . . . . . . . . . . . . . 42
3.4.2 Sleep-Wake Classification Algorithm . . . . . . . . . . . . . . . . . . . .44
3.5 Assessment of Event Detection Accuracy . . . . . . . . . . . . . . . . . . .45

4 Simulation Result 49
4.1 Thoracic and Abdominal Waveform . . . . . . . . . . . . . . . . . . . . . . 49
4.1.1 Normal Event . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .50
4.1.2 OSA Event . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1.3 CSA Event . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.1.4 HYP Event . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2 Database and Study Design . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3 Support Vector Machine Classification Performance . . . . . . . . . . . . . .54
4.3.1 Original Support Vector Machine Performance . . . . . . . . . . . . . . . 56
4.3.2 Phenotype-base Support Vector Machine Performance . . . . . . . . . . . . 59
4.3.3 Phenotype-base Support Vector Machine with comorbidity Performance . . . .61
4.4 Sleep-Wake Classification Performance . . . . . . . . . . . . . . . . . . . .64
4.5 LSTM-RNN Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . .65
4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .72

5 Conclusion and Future Work 73
5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .73
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

[1] N. S. F. of USA. Why do we need sleep. USA. [Online]. Available: https://sleepfoundation.org/excessivesleepiness/content/why-do-we-need-sleep

[2] ——. How much sleep do we really need? USA. [Online]. Available: http://sleepfoundation.org/how-sleep-works/how-much-sleep-do-we-really-need

[3] K. D. FNP. Sleep apnea: Causes, symptoms and treatments. [Online]. Available: http://www.medicalnewstoday.com/articles/178633.php

[4] HELPGUIDE.ORG. Sleep apnea: Symptoms, causes, treatments, and cures for sleep apnea. [Online]. Available: http://www.helpguide.org/articles/sleep/sleep-apnea.htm

[5] J. M. Andrew Bates, Martin J. Ling and D. K. Arvind. (2010, July) Respiratory rate and flow waveform estimation from tri-axial accelerometer data. [Online]. Available: https://sleepfoundation.org/excessivesleepiness/content/why-do-we-need-sleep

[6] G. M. Anmin Jin, Bin Yin, H. Duric, and R. M. Aarts. (2009, Nov) Performance evaluation of a tri-axial accelerometry-based respiration monitoring for ambient assisted living. [Online]. Available: https://sleepfoundation.org/excessivesleepiness/content/why-do-we-need-sleep

[7] M. H. Bonnet and D. L. Arand, “Heart rate variability: Sleep stage, time of night, and arousal inuences,” Electroencephalogr. Clin. Neurophysiol., pp. 390–396, May 1997.

[8] F. Chouchou and M. Desseilles, “Heart rate variability: A tool to explore the sleeping brain?” Front. Neurosci., pp. 1–9, 2014.

[9] D. Zemaityte and G. Varoneckas, “Heart rhythm control during sleep,” Psychophysiology, vol. 21, pp. 279–290, May 1984.

[10] Eiseman, N. A., M. B. Westover, J. E. Mietus, R. J. Thomas, and M. T. Bianchi,“Classification algorithms for predicting sleepiness and sleep apnea severity,” ”Journal of sleep research 21, no. 1, pp. 101–112, 2012.

[11] de Chazal, Philip, and N. Sadr, “Sleep apnoea classification using heart rate variability, ecg derived respiration and cardiopulmonary coupling parameters,” In Engineering in Medicine and Biology Society (EMBC), IEEE, pp. 3203–3206, 2016.

[12] Eiseman, N. A., M. B. Westover, R. J. T. Joseph E. Mietus, and M. T. Bianchi, ”Classification algorithms for predicting sleepiness and sleep apnea severity,” Journal of sleep research 21, no. 1, pp. 101–112, 2012.

[13] L. Almazaydeh, M. Faezipour, and K. Elleithy, “A neural network system for detection of obstructive sleep apnea through spo2 signal features,” International Journal of Advanced Computer Science and Applications (IJACSA), vol. 3, no. 5, pp. 7–11, May 2012.

[14] A. Quiceno-Manrique, J. Alonso-Hernandez, C. Travieso-Gonzalez, M. Ferrer-Ballester, and G. Castellanos-Dominguez, “Detection of obstructive sleep apnea in ecg recordings using time-frequency distributions and dynamic features,” 31 st IEEE International Conference on Engineering in Medicine and Biology Society (EMBS), pp. 5559–5562, Sep 2009.

[15] M. Mendez, D. Ruini, O. Villantieri, M. Matteucci, T. Penzel, and A. Bianchi,“Detection of sleep apnea from surface ecg based on features extracted by an autoregressive model,” IEEE International Conference on Engineering in Medicine and Biology Society (EMBS), pp. 6105–6108, 2007.

[16] Novk, D. Mucha, K., and A.-A. T, “Long short-term memory for apnea detection based on heart rate variability,” IEEE Engineering in Medicine and Biology Society (EMBS), pp. 5234–5237, August 2008.

[17] A. Ng, J. Chung, M. Gohel, W. Yu, K. Fan, and T. Wong, “Evaluation of the performance of using mean absolute amplitude analysis of thoracic and abdominal signals for immediate indication of sleep apnoea events,” Journal of Clinical Nursing, vol. 17, no. 17, pp. 2360–2366, September 2008.

[18] Avc, Cafer, and A. Akba, “Sleep apnea classification based on respiration signals by using ensemble methods,” Bio-Medical Materials and Engineering 26, no. sl, pp.S1703–S1710, 2015.

[19] Garde, Ainara, P. Dekhordi, J. M. Ansermino, and G. A. Dumont, “Identifying individual sleep apnea/hypoapnea epochs using smartphone-based pulse oximetry,” Engineering in Medicine and Biology Society (EMBC), pp. 3195–3198, 2016.

[20] M. F. Laiali Almazaydeh and K. Elleithy, “A neural network system for detection of obstructive sleep apnea through spo2 signal features,” International Journal of Advanced Computer Science and Applications(IJACSA), vol. 3, no. 5, 2012.

[21] C-Y.Chen, “An ultra low-power wearable sensor for physiological signal monitoring,” Master’s thesis, National Tsing Hua University, August 2017.

[22] P. Carney, J. Geyer, and R. Berry, Clinical Sleep Disorders. Philadelphia: Lippincott Williams & Wilkins, 2012.

[23] Y.-Y. .Lin, “Sleep apnea syndrome severity abd sleep apnea event types,” Master’s thesis, National Tsing Hua University, May 2014.

[24] H.-T. W. I. Daubechies, J. Lu, Synchrosqueezed Wavelet Transforms: an empirical mode decomposition-like tool. America: Appl. Comput. Harmon. Anal., 2011.

[25] R. K. Pathinarupothi, E. S. Rangan, and G. E. A, “Single sensor techniques for sleep apnea diagnosis using deep learning,” IEEE International Conference on Healthcare Informatics, no. 23-26, September 2017.

[26] F. M.-D. Sheikh Shanawaz Mostafa, Fbio Mendona and A. Ravelo-Garca, “Spo2 based sleep apnea detection using deep learning,” International Conference on Intelligent Engineering Systems(INES), vol. 21, 2017.

[27] S. Hochreiter and J. Schmidhuber, “Long short-term memory.” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997.

[28] a. G. M. a. Q.-C. Auli, Michael and G. Zweig, “Joint language and translation modeling with recurrent neural networks.” EMNLP, vol. 3, p. 0, 2013.

[29] I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to sequence learning with neural networks.” Advances in Neural Information Processing Systems, pp. 3104–3112, 2014.

[30] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan, “Show and tell: A neural image caption generator.” arXiv preprint arXiv, no. :1411.4555, 2014.

[31] Z. C. Lipton, D. C. Kale, C. Elkan, and R. Wetzel, “Learning to diagnose with lstm recurrent neural networks,” International Conference on Learning Representations, November 2015.

[32] J.-C. .Wu, “Sleep apnea syndrome screening by tri-axial accelerometer, oximeter and phenotype information,” Master’s thesis, National Tsing Hua University, August 2017.

[33] A. Lewicke, E. Sazonov, M. Corwin, and S. Schuckers, “Reliable determnination of sleep versus wake from heart rate variability using neural networks,” IEEE International Joint Conference on Neural Networks, vol. 4, pp. 2394–2399, 2005.

[34] J. Malik, Y.-L. Lo, and H. tieng Wu, “Sleep-wake classication via quantifying heart rate variability by convolutional neural network,” IEEE International Joint Conference on Neural Networks, vol. 39, no. 8, pp. 2394–2399, August 2018.

[35] F. Snyder, J. A. Hobson, D. F. Morrison, and F. Goldfrank, “Changes in respiration, heart rate, and systolic blood pressure in human sleep,” J. Appl. Physiol, no. 19, pp. 417–422, August 1964.

[36] A. C. J. C. Iber, S. Ancoli-Isreal and S. Quan, The AASM Manual for Scoring of Sleep and Associated Events-Rules: Technical Specification. America: American Academy of Sleep Medicine, 2007.

[37] “Physionet,” www.physionet.org.