本研究目的在於利用復健前測量腦電圖的資料以及機器學習技術預測中風病患於接受復健治療後是否會達到「恢復良好」，並探討不同建模參數與臨床因子使用之下對於準確率的影響。研究收集53位中風患者的做肩關節屈曲/伸直(shoulder flexion/extension)時的腦電圖資料，每位患者都經過總時數24小時的復健治療並且經由三種臨床量表(FM、TEMPA、WMFT)來評估中風患者的動作功能恢復情況。其中37人的資料建立預測模型，而另外16人則用來做為驗證使用。 本研究採用機器學習的方法，將患者分為復健良好(good)以及復健一般(general)兩種情況，本研究根據臨床量表(FM、TEMPA、WMFT) 定義兩種情況：一種是後測時三種臨床量表任一個進步分數是否達到總分的10% (如 ∆FM> =6.6 ; ∆WMFT >=8.5; ∆TEMPA >=13.8 )，稱之為TypeI，結果有20位屬於復原良好；另一種以該患者後測時的三種臨床量表分數減去前測時的分數為基準，與前測的分數比較，看進步的幅度是否有達到10%，稱之為TypeII，結果有27位中風患者屬於復原良好。 在前測時，患者做80次肩關節動作同時收取腦電圖，而後將腦電圖資料經過前處理(濾波、切段)後，在周圍刺激時間點上利用廣義逆電場矩陣將前處理過的腦電圖資料推估出對側初級運動皮質區(CM1)、同側初級運動皮質區(IM1)、對側前運動區(CPM)、同側前運動區(IPM)以及輔助運動區(SMA)等五種運動區在大腦上的近似位置，再來把五種運動區所計算出的光譜密度圖利用莫萊小波(wavelet number: 7)轉換成時間頻譜圖。每次收取的腦電圖經過上述的處理後所得到的時頻圖絕對值平均起來而後進入動態因果模型之誘發響應當作該模型觀察依據並且試圖解釋。經由動態因果模型所得到的參數以及五種運動區所得到光譜密度圖擔任資料的特徵，而後將這些特徵在二分類法下利用包裝法選取特徵，分別使用四種不同的分類器:支持向量機、邏輯回歸法、貝氏分類器、J48。 分類結果在TypeI以及動態因果模型特徵下，在β+γ頻帶組合使用邏輯回歸法最高準確率92.95%，這個預測結果明顯優於使用臨床因子做預測的結果。並使用新的資料驗證後準確率亦達81.25%。此外在TypeI以及動態因果模型特徵下，邏輯回歸法與β頻帶都存在其重要性。且預測的準確率與三大因素:患者年齡、中風發生位置、中風後的時間相關。相信本研究所發現的結果可對於發展個人化最佳復健策略做出相當貢獻。 ;This study is aiming at accurate prediction of the rehabilitation outcome after stroke by using the pre-rehab electroencephalogram (EEG) and machine learning techniques and using new data to validate. 53 stroke patients were recruited for this study. 37 of them are used to build the prediction model and other 16 are used to validation. All patients underwent 24-hour rehab program and rehabilitation outcome was measured with FM, TEMP and WMFT. For supervised machine learning methods, we first divided the data into two groups : good and general recovery, according to two criteria - (1) Type I : a level of 10% improvement of any above mentioned measures (i.e ∆FM> =6.6 ; ∆WMFT >=8.5; ∆TEMPA >=13.8 ) after rehab was considered as good recovery as suggested in several studies, resulting in 20 good recovery patients (out of 37); and (2) Type II: the scale after rehab up to 10 % improvement from the before rehab scale of any above mentioned measures was labeled as good recovery, resulting in 27 good recovery patients (out of 37). The EEG data were acquired during the shoulder flexion/extension for eighty trials before the rehabilitation and were pre-processed offline for filtering and epoching. The spectral density from 4-48Hz at contralesional primary motor cortex (CM1), ipsilateral primary motor cortex (IM1), contralesional premotor area (CPM), ipsilateral premotor area (IPM) and the supplementary motor area (SMA) were obtained by projecting the EEG data to the chosen sources using the generalized inverse of the lead-field matrix over peri-stimulus time and then used a time-frequency Morlet wavelet transform (wavelet number: 7). The absolute value of the resulting time-frequency responses were averaged over trials and entered dynamic causal modeling for induced responses (DCM_IR) as the observations that the model is trying to explain. Both the spectral density at all sources and the parameters given by DCM were served as the data features. These features entered Wrapper method to select features, and the selected features went into the four different classifiers：SVM, Logistic Regression, NaiveBayes, J48 for two-class classification. The classification result suggests that, the best accuracy rate was 0.9295 when using DCM features of β+γ frequencies of Type I data partition and Logistic Regression. And the result of construct validity is 0.8125. The accuracy of prediction with EEG data is much higher than prediction with other clinical factors. Furthermore, beta rhythm within the motor network and Logistic Regression have significant roles in motor recovery prediction. There are three factors related to the predictive accuracy. They are age, time post-stroke and lesion area of stroke. We believe that our findings in this study have great benefits on developing a knowledge-based and individual rehab program.
