癲癇發作預測系統之演算法及硬體架構設計

Abstract

癲癇是世界上最常見的腦部疾病之一，自發性的癲癇發作嚴重影響癲癇患者的日常生活。近年來從腦電信號的觀察來做癲癇預測的方法表現出長足的進步。然而，腦電信號的變化包括患者狀態，癲癇嚴重程度等等，構成了癲癇預測根本性的困難。傳統的離線訓練方法使用固定的訓練資料來求得最佳的模型，並期望模型不管經過多長的時間，表現始終保持穩定，因而預測效果受到腦電信號變化問題的干擾。 在這篇論文中，我們提出一個持續訓練方法，利用最近觀察的數據，逐步擴大訓練資料，並持續訓練找出最佳的模型。此外，一個簡單的後處理步驟降低了誤判的機率。我們所使用的資料庫為德國弗萊堡顱內腦電信號資料庫以及CHB - MIT顱外腦電信號資料庫。我們提出的方法在顱內及顱外的資料庫上分別可達到 74.2％和52.2％的預測率，和傳統離線訓練方式相較分別提高了29.0％和17.4％。實驗結果表明，持續訓練的演算法可大大提高癲癇預測系統的可靠性，且對未來的演算法研究立下值得參考的基礎。 為實現所提出的癲癇預測系統，本系統面臨著三個設計問題，包括吞吐量的要求，功率消耗和系統的可調變性。為了解決設計問題，我們首先提出了一種混合型降維技術，以減少資料維度的大小。資料的降維可以減少系統的時間和空間複雜度，並減少功耗，達成持續訓練的可行性。經過硬體資源的分析和估計，我們再映射系統中的各個功能到各個硬體單元。分析的結論指出，一個高效率的特徵萃取加速器和通用處理器為兩個不可或缺的硬體單元。於是我們設計了一個高效率的多通道小波同步性特徵萃取加速器，和一個通用處理器互相配合，滿足系統的需求，並解決設計的問題。 總之，我們提出了一種新穎的持續學習癲癇預測方法，以及它的系統架構設計，並通過仔細的分析解決了系統在演算法與硬體設計層面所遭遇的挑戰。

Epilepsy is one of the most common brain disorders in the world. The spontaneous seizure onset influences the daily life of epilepsy patients. The studies on feature extraction and feature classification from Electroencephalography(EEG) signal in seizure prediction methods have shown great improvement these years. However, the variation issue of EEG signal (being awake, being asleep, severity of epilepsy, etc.) poses a fundamental difficulty in seizure prediction problem. The traditional off-line training method trains the model using a fixed training set, and expects the performance of the model to remain stable even after a long period of time, and thus suffers from variation issue. In this thesis, we propose an on-line retraining method to leverage the recent input data by gradually enlarging the training set and retraining the model. Also, a simple post-processing scheme is incorporated to reduce false alarms. We develop our method based on the state of the art machine learning based classification of bivariate patterns method. The performance of the method is evaluated on Electrocorticogram(ECoG) recording from Freiburg database as well as long-term scalp EEG recording from CHB-MIT EEG Database and National Taiwan University Hospital. The proposed method achieves 74.2% sensitivity on ECoG database and 52.2% sensitivity on scalp EEG database, while improving the sensitivity of off-line training method by 29.0% and 17.4% in ECoG database and EEG database respectively. The experimental result suggests that on-line retraining can greatly improve the reliability and is promising for future seizure prediction method development. To implement the proposed seizure prediction system, the system faces three design challenges including throughput requirement, power consumption, and system flexibility. To deal with the design challenges, we first propose a hybrid dimension reduction technique to reduce the pattern size from over thousands to only 256 dimensions, while the prediction performance was only dropped by 3%. The reduction of the pattern size improves the system regarding to time, area complexity, and power consumption, and enables the capability of online-retraining method. Through a careful analysis on hardware resources and computation estimation, we then map the function into hardware units. The analysis leads to the conclusion that an efficient feature extraction engine and a general purpose processor are two indispensable units to solve the design challenges. An efficient architecture design of multi-channel wavelet coherence feature extraction engine is analyzed and optimized. A reasonable system architecture design which meets the design challenges is then presented. In sum, a novel online retraining seizure prediction method along with the system architecture design are presented. The hardware implementation results of the core DSP function of the system verifies the proposed solution, and the mentioned design challenges in algorithm and architecture are solved through careful analysis.