| 摘要: | 癲癇為一常見的腦部疾病，起因於腦部神經細胞過度活躍，而引起的異常放電現象，腦電圖 (Electroencephalography, EEG) 是目前診斷癲癇主要依據之一，用以檢測大腦的功能性是否正常。已有許多論文針對癲癇患者與一般人之腦波差異進行研究，並已整理出許多癲癇樣波的波型。然而，對於中風病人的癲癇診斷，醫師並不建議用已知的癲癇樣波來分辨，原因是中風會造成部分腦區域受損，使腦電圖在中風急性期也可能出現癲癇樣波，造成判斷是否為癲癇的效果不準確。因此，我們設計了針對中風病患EEG之訊號處理及特徵抽取方法，並使用機器學習模型進行偵測。
研究資料來源為台北榮總2012年至2017年針對中風病人所蒐集的腦電波資料，在量測腦電波訊號方面，電極採用10-20 System，實驗中使用不同光照赫茲作為誘發癲癇的外在刺激，受測者在時間軸上依序處於「休息、光照刺激、休息」狀態，總共量測831位中風病患，收集1323筆EEG。腦電波前處理先透過Resampling、Notch Filtering、Bandpass Filtering、Epochs、CSP等步驟去除雜訊、強化訊號特徵，再利用小波轉換 (Wavelet Transform) 將不同頻帶腦電波分離出來以獲得更多資訊，最後透過統計學對訊號提取出富含資訊性且不冗餘的特徵值，並以機器學習方式進行腦波分析。透過大量實驗探討在中風癲癇上適合的前處理、特徵與模型，其中採用Coherence、Entropy、Kurtosis、Skewness四種特徵的Logistic Regression模型，在1分54秒的EEG訊號分析，可以獲得0.7192的F1-score、0.4479的Sensitivity及0.8313的Specificity，與當前的癲癇偵測機器學習方法與EEG深度學習模型比較，有顯著的效果提升。此外，我們開發了一個視窗程式介面，可以選擇癲癇偵測模型，觀察EEG腦波時間序列分析，用以輔助醫師決策，冀望於此次的初步探索，可以開啟腦波分析於中風癲癇的研究濫觴。
;Epilepsy is a common brain disease that is caused by abnormal discharging from overactive nerve cells in the brain. Electroencephalography (EEG) is one of the main diagnostic measurements for epilepsy detection. EEG based signal analysis in epilepsy seizures have been studied to find out brain wave patterns of patients being diagnosed with/without epilepsy diseases. However, doctors usually do not recommend using epileptiform patterns as epilepsy detection features for stroke patients, whose abnormal brain signals originating from certain brain injuries. Therefore, we design a series of brain signal processing to extract discriminative features for a machine-learning-driven epilepsy detection on stroke patients.
The experimental data came from Taipei Veterans General Hospital (TVGH), Taiwan. A total of 831 stroke patients from 2012 to 2017 had been measured their EEGs using the 10-20 system. Different light hertz had been used as external stimulations. A total of 1,323 EEGs were collected in a state sequence of “resting, light stimulation, resting”. We firstly removed signal noises and strengthened features through brainwave preprocessing such as notch filtering, bandpass filtering, epochs, and CSP. We then used a wavelet transform method to separate brainwaves in terms of different frequency bands. Several statistical methods were extracted to obtain informative and non-redundant features. Finally, we exploited well-known machine learning models to detect epilepsy seizures on stroke patients. Based on a series of experiments and their result analysis, we find that the Logistic Regression model with four features (coherence, entropy, kurtosis, and skewness) significantly outperforms machine-learning-based epilepsy detection methods and EEG-based deep learning models, achieving the best F1 score of 0.7192, sensitivity of 0.4479 and specificity of 0.8313 from EEG data (each duration is 1 minutes and 54 seconds). Besides, we implemented a windows based programming interface, with which doctors can select epilepsy detection models and observe EEG distribution with the time change for clinical decision supports. This is our pilot study of this research issue. Hope our explorations can benefit the studies for EEG based epilepsy detection on stroke patients. |
