基於小波混沌分析法之癲癇預測及電路實現

Abstract

在大腦生理訊號分析的研究中，如何從長時間的腦電波訊號找出可靠的特徵來進行癲癇疾病的預測是目前熱門的研究議題，有鑑於傳統的統計分析方法對於非穩定、非線性動態系統的訊號，容易因錯誤的結論而影響預測準確性。 本論文提出一個基於小波及混沌理論分析的架構包含離散小波轉換、相關維度、及相關係數。因小波具有多解析度、時頻分析的特性，相對於傅立葉轉換更適合用在非穩定訊號。而混沌理論對於非穩定、非線性動態系統的基本推論，比傳統的統計學方法能更有效地重建腦波的特徵，有助於腦波訊號的分析。因此，結合小波與混沌理論將可有效提高癲癇預測的準確性及降低誤判率。 本論文提出的作法是先將訊號分解成若干子頻帶，並針對子頻帶，利用腦電波在癲癇發作前後，其相關維度收歛速度的不同來做為預測的依據。經實驗結果，本論文所提出的演算法在11位具癲癇病患的測試中，可達到87%的預測準確率，誤判率為0.24次/小時，其平均預測時間約在發病前27分鐘。 為了能讓本論文提出的方法能應用於可攜式生理監控設備，我們將其設計成癲癇分析的處理電路，並針對演算法特性，使用提升式小波轉換、改良記憶體定址、及其他算數化簡來降低電路面積及功耗，也可於將來進一步整合到生醫的訊號處理器中。

The Epilepsy and epileptic seizure prediction algorithm by extracting useful features from Electroencephalography (EEG) is a hot topic in the current research of physiological signals. In view of the erroneous conclusions from the traditional statistical analysis methods for non-stationary and non-linear dynamics system of signals may affect the accuracy of forecasts. This thesis presents a novel architecture based on wavelet and chaos theory, including Discrete Wavelet Transform (DWT), correlation dimension, and correlation coefficient. The wavelet transform is more suitable for non-stationary signals than Fast Fourier Transform (FFT) due to the ability of multi-resolution and time- frequency analysis. The fundamentals of Chaos theory for non-stationary and non-linear dynamics systems are more in line with the characteristics of brain waves than statistics. Therefore, combining DWT and Chaos analysis can achieve a high prediction rate. In this thesis, first EEG signals are decomposed into several subbands. We predict seizures by the difference of convergent radius between the correlation dimension of EEG before a seizure and the one during a seizure for each subband. The proposed algorithm is evaluated with intracranial EEG recordings from a set of eleven patients with refractory temporal lobe epilepsy. In the experimental results, the algorithm with global settings for all patients predicted 87% of seizures with a false prediction rate of 0.24/h. Seizure warnings occur about 27 min ahead the ictal on average. To apply the algorithm proposed to a portable physiological monitoring device, a seizure analysis circuit is also designed. Some techniques, such as lifting wavelet transform, enhanced memory addressing, and arithmetic reduction etc., are used to reduce circuit area and power consumption of circuit. In the future, the seizure analysis circuit can be further integrated into a digital signal processor for biomedical applications.