以軟硬體協同設計架構進行即時自動化之多通道神經動作電位辨識

Abstract

神經動作電位辨識是神經科學領域用於理解大腦功能不可或缺的重要程序，動作電位的分類對於更進一步的應用，例如運動軌跡預測或是腦機介面(Brain machine interface)上，提供了一個腦神經活動以及動物外在行為兩者關係上的連結。如果神經動作電位辨識無法正確有效地進行動作電位分類，當運用於上述的應用時，將會因錯誤的判斷而造成嚴重的影響。因此，神經動作電位辨識的精確性跟穩定度必須盡可能地提高。而隨著植入式裝置製程的進步以及不同的應用範疇，多通道神經訊號紀錄已成為電生理領域之研究上的基本工具。 本研究軟硬體協同設計架構實現即時自動化處理之多通道神經動作電位辨識系統。本系統所應用的兩階層神經動作電位偵測方法，結合了閾值偵測及非線性能量運算法(Non-linear energy operator, NEO)之優點。在特徵擷取上，應用離散微分方法(Discrete derivative)來提高動作電位在特徵空間的可辨識性，並使用主成分分析法(Principal component analysis, PCA)降低資料維度。單一連結法(Single linkage method, SLM)結合馬氏距離(Mahalanobis distance)被應用於動作電位分群。此外，跨電極驗證方法(Cross electrode validation)被提出用以確認是否有單一神經元被兩個以上的電極所記錄。 多通道神經動作電位辨識系統所使用的演算法經過模擬以及實驗的驗證，由結果可以發現兩階層神經動作電位偵測結合回授控制能有效減少錯誤動作電位偵測的發生。而動作電位於特徵空間的分佈，使用離散微分方法之後，將獲得更高的可分辨性，因而能提高動作電位辨識的準確度。在完成多通道神經動作電位辨識後，跨電極驗證方法能降低多餘的神經資訊被系統紀錄，從而能夠減少往後資訊分析的複雜度。因此，本研究所提出的即時自動化之多通道神經動作電位辨識架構能適用於神經科學家解開動物大腦功能的第一步

Spike sorting is a primary and essential procedure for the realization of brain in the neuroscience and it provides a connection between the neural behavior and external behavior of animal for further application such as movement prediction and brain machine interface (BMI). With different objectives and improvement in implantable device technology, multichannel recording has become a standard tool for the research on neurophysiology. Besides, the accuracy of the spike sorting has crucial relations with the stability of the advanced application, and it would result in fatal influence for the application related to humans if the spike sorting was not reliable. A real-time and automatic spike sorting system for 16-channel neural recording based on hardware-software co-design is proposed in this study. The two-stage spike detection, combining the benefit of threshold method and nonlinear energy operator (NEO), is presented as the initial step of spike sorting process. The feature extraction in this study utilizes the discrete derivative method to improve the spike separation and chooses the principal component analysis to select few dominant features for reduction of indistinctive data. The single linkage method, with Mahalanobis distance as the distance metric, is used for spike clustering. The cross electrode validation is presented for the purpose that validates whether there is a single neuron recorded by two or more electrodes. The algorithms for this multichannel spike sorting system were verified and evaluated through simulations and experiments. The two-stage spike detection cooperating with feedback rule could decrease the probability of false detection. There is a significant improvement for spike separation on feature space with the help of the discrete derivative method, and, thus, the accuracy of the spike sorting is enhanced on indistinguishable data set from the result. After spike sorting, the cross electrode validation could lower the redundant neuronal information to be recorded. The proposed framework for real-time and automatic spike sorting of multichannel neuronal activity is feasible as the first step for neuroscientist to figure out the brain function of animals.