以閉迴路架構實現即時處理之神經元動作電位分類器

Abstract

即時處理之神經元動作電位分類器是神經科學中的重要研究之一，動作電位的分類可以運用在確認於電極附近的神經元個數以及其放電行為，動作預測或是腦機介面上。過往的研究往往將即時處理的動作電位之分類器建立在理想的神經訊號上。然而，在實際活體實驗的神經訊號上有著大量的干擾雜訊以及不斷變動的背景雜訊，如果動作電位偵測無法有效抑制突發的雜訊或是無法適應變動的背景雜訊，將會造成分類的正確率下降以及增加傳輸資料上的負擔。 本研究以閉迴路架構實現即時處理之動作電位分類器，並提出一個兩階層的神經動作電位偵測方法，由於不同的特徵擷取方法會影響分類的準確性，因此我們分別使用兩種不同的特徵擷取方法來評估對系統分類的影響，一種為低運算量的波形特徵擷取，另一種為較高運算量的以迭代法實現的主軸成份分析法。為了評估在不同程度的神經訊號上的效能，我們採用藉由侵入式電極紀錄而得的大鼠神經訊號來模擬與評估。 本系統經過模擬以及實驗的驗證，從結果可以發現兩階層的神經動作電位偵測與迴授控制合作可以減少錯誤偵測的產生，而且波形特徵擷取比以迭代法實現的主軸成份分析法得到更低的錯誤率。在分類的效能上，在一般的神經訊號中，波形特徵擷取的類數極限在三種以下的神經元以及訊雜比在八以上；以迭代法實現的主軸成份分析法的類數極限在四種以下的神經元以及訊雜比在六以上。在具有干擾雜訊的神經訊號中，波形特徵擷取的類數極限在三種以下的神經元以及訊雜比在八以上；以迭代法實現的主軸成份分析法的類數極限在四種以下的神經元以及訊雜比在八以上。因此，本系統具有抗雜訊能力於實際活體實驗中運行動作電位分類，此外系統中的神經訊號處理器也可實現在晶片上，而低資料量的優勢也有利於架設在無線傳輸的系統上。

On-line spike sorting is a relevant research in the neuroscience. Spike sorting can be used to provide the application on confirmation of number neuron and observation of isolated neural activities, movement prediction and brain machine interface (BMI). The previous research of on-line spike sorting is constructed based on the ideal neural signal condition. However, the neural signal of living experiment presents a lot of disturbance of noise and continuously varying background noise. If the spike detection can not resist the unexpectedly noise, the poor performance appear on the final sorting result and the huge burden response on the data transmission. This study proposes the adaptable on-line spike sorter (AOSS) based on the close-loop framework. The two-stage detection used for lower the false detection and adjusts the discriminant criterion according to the background noise level. Since the feature extraction method could influence the performance of spike sorting, we implement two methods including waveform parameter extraction which required less computational complexity and distill PCA extraction which required higher computational complexity, and discuss the performance on different neural signal quality by using simulated neural signal from in vivo rat recording. The proposed adaptable on-line spike sorter were verified and evaluated by simulations and experiments. From the results, the false detection could reduce by cooperating with feedback control. The detection performance of waveform parameter extraction is higher than distill PCA extraction. In the normal condition neural signal, the limitation of class for the waveform parameter extraction on AOSS was under 3 classes of SNR 8 while distill PCA was under 4 classes of SNR 6. In the disturbance condition, the limitation of class for the waveform parameter extraction was 3 classes of SNR 8 while distill PCA was under 4 classes of SNR 8. The proposed adaptable on-line spike sorter has potential to resist noise on living experiment sorting. Furthermore, the neural signal processor on AOSS is on-chip architecture, and the low data rate has the benefit to build on wireless telemetry system.