目前，癲癇是最常被診斷出的神經性障礙之一，其特徵在於自發性發作的突然發生。大約三分之一的患者具有耐藥性癲癇，這意味著他們的癲癇發作不能用藥物控制。在這種情況下，癲癇預測系統至關重要，因為它們允許患者在突發發作之前避免危險的活動並使自己處於安全的環境中。
腦電圖用於診斷癲癇，檢測或預測癲癇發作。到目前為止，醫學專家們仍然缺乏可靠的算法，用於識別發作發生機率增加的時期。因此，開發一種依賴和可靠的算法，能夠對癲癇發作之前的一段時期的數據和大腦正常活動時期的數據進行分類至關重要。這是具有挑戰性的，因為腦電圖上的癲癇發作表現在相同患者的不同時期和不同患者之間都相差甚大。
我們的癲癇預測演算法為基於影像之腦電信號特徵表示和遞歸卷積神經網絡。通過同時捕獲頻譜、時間和空間信息，我們的遞歸卷積神經網絡能夠學習到空間不變性的癲癇預發作時期腦電訊號特徵。此外，我們提出了一種基於我們演算法的癲癇預測系統。本論文中討論了相關的系統概述和硬體實現細節。

Epilepsy is one of the most commonly diagnosed neurological disorders, which is characterized by the abrupt occurrence of spontaneously seizure. About 1 of the patients has a drug-resistant epilepsy. In such cases, seizure forecasting systems are vitally important, since they allow patients to avoid dangerous activities and bring themselves in a safe environment before a seizure with sudden onset. EEG is used for diagnosing epilepsy and detecting or predicting seizures. So far, practitioners still lack reliable algorithms for identifying periods of increased probability of seizure occurrence. Thus, developing a reliant and robust algorithm, able to classify data clips of a pre-seizure brain activity covering a period prior to a seizure onset and data clips of inter-seizure activity is crucial. This is challenging because seizure manifestations on EEG are extremely variable both inter- and intra-patient. Our approach to the seizure prediction is based on video-based EEG representation and recurrent convolutional neural network. By simultaneously capturing spectral, temporal and spatial information our recurrent convolutional neural network learns a general spatially invariant representation of a seizure. In addition, we propose a seizure prediction system based on our algorithm. Its system overview and hardware implementation are discussed in this thesis.

口試委員審定書 i
誌謝 ii
中文摘要 iii
ABSTRACT iv
CONTENTS v
LIST OF FIGURES vii
LIST OF TABLES x
Chapter 1 Introduction 1
1.1 Motivation 1
1.2 Organization 3
Chapter 2 Basic Background Information 5
2.1 Deep Learning Models 5
2.1.1 Artificial Neural Networks (ANN) 5
2.1.2 Activation Function 7
2.1.3 Training Process 8
2.1.4 Back-Propagation Algorithm 9
2.1.5 Overfitting 11
2.2 Convolutional Neural Networks (CNN) 13
2.2.1 Convolutional Layers 14
2.2.2 Pooling (Sub-Sampling) Layers 17
2.2.3 Fully-connected layers 18
2.3 Recurrent Neural Networks 19
2.3.1 LSTM (Long-Short Term Memory) 19
2.4 Fundamental of Electroencephalography (EEG) 22
2.5 Previous Works of Seizure Prediction 25
Chapter 3 Seizure Prediction Algorithm and Experimental Results 30
3.1 The EPILEPSIAE Database 30
3.2 Data Labeling and Splitting 31
3.3 Seizure Prediction Algorithm 34
3.3.1 Video-based EEG Representation 34
3.3.2 Data Normalization 41
3.3.3 Recurrent-Convolutional Neural Networks 43
3.3.4 Model Evaluation Using ROC AUC Score 46
3.4 Experimental results 51
Chapter 4 Hardware Implementation of Real-time Seizure Prediction System 62
4.1 System Overview 62
4.2 Fixed-point Conversion 65
4.3 Hardware Implementation of Building Blocks 67
4.3.1 Fast Fourier Transform (FFT) Module 67
4.3.2 CORDIC Module 70
Chapter 5 79
References 80

[1]Deckers, C. L. P., et al. "Current limitations of antiepileptic drug therapy: a conference review." Epilepsy research 53.1 (2003): 1-17.
[2]Fisher, Robert S. "Therapeutic devices for epilepsy." Annals of neurology 71.2 (2012): 157-168.
[3]Bennewitz, Margaret F., and W. Mark Saltzman. "Nanotechnology for delivery of drugs to the brain for epilepsy." Neurotherapeutics 6.2 (2009): 323-336.
[4]Fujii, Masami, et al. "Application of focal cerebral cooling for the treatment of intractable epilepsy." Neurologia medico-chirurgica 50.9 (2010): 839-844.
[5]DeGiorgio, Christopher M., et al. "Randomized controlled trial of trigeminal nerve stimulation for drug-resistant epilepsy." Neurology 80.9 (2013): 786-791.
[6]Springer, Utaka S., et al. "Long-term habituation of the smile response with deep brain stimulation." Neurocase 12.3 (2006): 191-196.
[7]Motamedi, Gholam K., et al. "Optimizing parameters for terminating cortical afterdischarges with pulse stimulation." Epilepsia 43.8 (2002): 836-846.
[8]Fountas, Kostas N., et al. "Implantation of a closed-loop stimulation in the management of medically refractory focal epilepsy." Stereotactic and functional neurosurgery 83.4 (2005): 153-158.
[9]Lehnertz, K. "Seizure anticipation techniques: state of the art and future requirements." Engineering in Medicine and Biology Society, 2001. Proceedings of the 23rd Annual International Conference of the IEEE. Vol. 4. IEEE, 2001.
[10]Aschenbrenner‐Scheibe, R., et al. "How well can epileptic seizures be predicted? An evaluation of a nonlinear method." Brain 126.12 (2003): 2616-2626.
[11]Haut, Sheryl R., et al. "Seizure clustering during epilepsy monitoring." Epilepsia 43.7 (2002): 711-715.
[12]Gemulla, Rainer, et al. "Large-scale matrix factorization with distributed stochastic gradient descent." Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2011.
[13]Dauphin, Yann N., et al. "Identifying and attacking the saddle point problem in high-dimensional non-convex optimization." Advances in neural information processing systems. 2014.
[14]Yu, Kai, Wei Xu, and Yihong Gong. "Deep learning with kernel regularization for visual recognition." Advances in Neural Information Processing Systems. 2009.
[15]Srivastava, Nitish, et al. "Dropout: a simple way to prevent neural networks from overfitting." Journal of Machine Learning Research 15.1 (2014): 1929-1958.
[16]Prechelt, Lutz. "Automatic early stopping using cross validation: quantifying the criteria." Neural Networks 11.4 (1998): 761-767.
[17]LeCun, Yann, and Yoshua Bengio. "Convolutional networks for images, speech, and time series." The handbook of brain theory and neural networks 3361.10 (1995): 1995.
[18]Mikolov, Tomas, et al. "Recurrent neural network based language model." Interspeech. Vol. 2. 2010.
[19]Schmidhuber, Jürgen. "Deep learning in neural networks: An overview." Neural networks 61 (2015): 85-117.
[20]Bengio, Yoshua, Patrice Simard, and Paolo Frasconi. "Learning long-term dependencies with gradient descent is difficult." IEEE transactions on neural networks 5.2 (1994): 157-166.
[21]Hochreiter, Sepp, and Jürgen Schmidhuber. "Long short-term memory." Neural computation 9.8 (1997): 1735-1780.
[22]Gers, Felix A., Jürgen Schmidhuber, and Fred Cummins. "Learning to forget: Continual prediction with LSTM." (1999): 850-855.
[23]Cascino, Gregory D., et al. "Magnetic resonance imaging–based volume studies in temporal lobe epilepsy: pathological correlations." Annals of neurology 30.1 (1991): 31-36.
[24]Shen, Chia-Ping, et al. "A physiology-based seizure detection system for multichannel EEG." PloS one 8.6 (2013): e65862.
[25]Schulze-Bonhage, Andreas, Hinnerk Feldwisch-Drentrup, and Matthias Ihle. "The role of high-quality EEG databases in the improvement and assessment of seizure prediction methods." Epilepsy & Behavior 22 (2011): S88-S93.
[26]Masko, David, and Paulina Hensman. "The impact of imbalanced training data for convolutional neural networks." (2015).
[27]Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks." Advances in neural information processing systems. 2012.
[28]Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016.
[29]Tieleman, Tijmen, and Geoffrey Hinton. "Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude." COURSERA: Neural networks for machine learning 4.2 (2012): 26-31.
[30]J. W. Cooley and J. Tukey, “An algorithm for machine calculation of complex fourier series,” Math. Comput., vol. 19, pp. 297–301, Apr. 1965.
[31]A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-Time Signal Processing, 2nd ed. Englewood Cliffs, NJ: Prentice-Hall, 1998.
[32]P. Duhamel, “Implementation of split-radix FFT algorithms for complex, real, and real-symmetric data,” IEEE Trans. Acoust., Speech, Signal Process., vol. 34, no. 2, pp. 285–295, Apr. 1986.
[33]S. He and M. Torkelson, “A new approach to pipeline FFT processor,” in Proc. of IPPS, 1996, pp. 766–770.
[34]L. R. Rabiner and B. Gold, Theory and Application of Digital Signal Processing. Englewood Cliffs, NJ: Prentice-Hall, 1975.
[35]E. H. Wold and A. M. Despain, “Pipeline and parallel-pipeline FFT processors for VLSI implementation,” IEEE Trans. Comput., vol. C-33, no. 5, pp. 414–426, May 1984.
[36]A. M. Despain, “Fourier transfom using CORDIC iterations,” IEEE Trans. Comput., vol. C-233, no. 10, pp. 993–1001, Oct. 1974.
[37]E. E. Swartzlander, W. K. W. Young, and S. J. Joseph, “A radix-4 delay commutator for fast Fourier transform processor implementation,” IEEE J. Solid-State Circuits, vol. SC-19, no. 5, pp. 702–709, Oct. 1984.
[38]E. E. Swartzlander, V. K. Jain, and H. Hikawa, “A radix-8 wafer scale FFT processor,” J. VLSI Signal Process., vol. 4, no. 2/3, pp. 165–176, May 1992.
[39]G. Bi and E. V. Jones, “A pipelined FFT processor for word-sequential data,” IEEE Trans. Acoust., Speech, Signal Process., vol. 37, no. 12, pp. 1982–1985, Dec. 1989.
[40]J. E. Volder, “The CORDIC trigonometric computing technique,” Trans. Electron. Computers, vol. EC-8, pp. 330–334, Sept. 1959.
[41]J. E. Volder, “The birth of CORDIC,” J. VLSI Signal Process., vol. 25, pp. 101–105, 2000.
[42]J. S. Walther, “A unified algorithm for elementary functions,” in Proc. 38th Spring Joint Computer Conf., Atlantic City, NJ, 1971, pp. 379–385.