Features of Mismatch Negativity for Classification of Schizophrenia Patients

聽覺變異刺激實驗可以在刺激後約100毫秒激發失匹配負波（MMN）特徵波形，同時也在標準或異常聲音刺激後約300毫秒呈現特徵波形。在本研究中，我們使用分類器工具來尋找區分精神分裂症患者與正常人的特徵組合。113位精神分裂症患者與95位正常受試者參與聽覺變異刺激實驗並測量事件相關電位（ERP）。我們先計算在標準或異常聲音刺激後約100毫秒及約300毫秒的平均振幅與波峰時間延遲，也使用離散小波轉換來代表MMN ERP的特徵，結果發現在精神分裂症族群與正常族群中，theta頻段ERP波形的標準差與能量存在顯著差異。我們採用支持向量機來比較精神分裂症患者識別的準確性。分類結果顯示，使用訊雜比來表示MMN波形會得到較好的準確分類率，使用線性核函數會得到比使用高斯核函數好的準確分類率。雖然本研究提出的小波時間序列特徵組合並沒有辦法增加分類率，我們認為這是我們第一次嘗試用聽覺MMN ERP來進行精神分裂症患者的區分。An auditory oddball paradigm can elicit the mismatch negativity (MMN) waveform with characteristic waveform at about 100 ms after the stimulus onset. There was also a characteristic waveform at about 300 ms after either standard or deviant sounds presentation. In this study, we used classification tools to find compositions of features in the hope of devising and imaging biomarkers to differentiate between schizophrenia patients and healthy subjects. Specifically event-related potentials (ERP) were measured from 113 schizophrenia patients and 95 healthy controls using an auditory oddball paradigm. The mean amplitudes and peak latencies of the ERP elicited by standard and deviant sounds at about 100 ms and 300 ms were first calculated. We also used discrete wavelet transform to characterize features the MMN ERP’s. We found that the standard deviation and the energy of the ERP waveform in theta band were significantly different between schizophrenia and control groups. Using support vector machine (SVM) as the classifier, we compared the accuracy of identifying schizophrenics. More accurate classification when MMN waveform is described in the signal-to-noise ratio (SNR) unit. Using a linear kernel in SVM gave higher classification accuracy than using a Gaussian kernel. The proposed wavelet time series feature compositions did not improve the classification accuracy. We consider this study is our first attempt to differentiate between healthy and schizophrenic patients using auditory MMN ERP