小鼠神經動作電位四聯電極訊號分群演算法之研究

Abstract

本研究提供一種同時擷取神經動作電位空間與波形特徵的方法，稱作平行動作電位主成份分析(parallel spike principal component analysis, PSPCA)，並以此為基礎配合affinity propagation (AP)分群演算法進行神經動作電位分群(spike sorting)。此方法依照動作電位隨發射源與四聯電極距離增加而衰減之關係，以主成份分析(principal component analysis, PCA)擷取空間與波形兩種特徵，將這兩種特徵計算相似度矩陣，而後AP分群演算法依此相似度矩陣進行分群。在排除動作電位重疊(overlapping)與連續發射(bursting)的情況下，經由不同訊雜比(1~12dB)之模擬實驗結果得知，PSPCA所得之波形特徵與發射源之原始動作電位模板具有高相關性，可將其視為一種濾波方式。在相同條件下計算峰值、峰值比和SSPC (serial spike principal component)特徵之DBVI (Davies-Bouldin validity index)值進行比較，波形特徵具有較好且穩定的效能。配合空間特徵作為波形特徵之間的相似度權重，經由AP分群演算法得到最後動作電位分群結果，利用其自動決定分群群集數的優點，可減少人為因素之影響。由模擬實驗結果得知，藉由調整AP演算法之優先權(preference)與阻尼因子(damping factor)兩項參數可避免過度分群(over-sorting)的情況發生，將其結果以adjusted Rand index作指標與k-means分群結果做比較，AP在各相同訊雜比條件下之分群正確率皆較k-means為佳，平均約提昇38%，分群群數誤差也較小，平均約降低67%。本研究之動作電位分群演算法最後實際應用至小鼠大腦RT和VPL腦區之動作電位四聯電極訊號，48組四聯電極訊號處理後得到1~10群不等的分群結果。依據模擬實驗與實際應用結果之分析，證實本研究之演算法能夠有效分群動作電位四聯電極訊號，相較k-means演算法也具有更好的效能。

The PSPCA (parallel spike principal component analysis) method developed in this study can efficiently extract both spatial and waveform feature from a spike (action po-tential) simultaneously. Affinity propagation (AP) clustering algorithm with those fea-tures is used for spike sorting of neuronal signals. PSPCA is based on principal compo-nent analysis (PCA) and the signal decay function of the distance between neuronal spike source and tetrode. Spikes are sorted using AP clustering algorithm with similarity matrix computed from those features. According to the simulation results with different signal noise ratios (S/N ratio), waveform feature is highly correlated with original spike pattern and can be regarded as denoised spike. Comparing the Davies-Bouldin validity index (DBVI) value of waveform feature with three other features, peak, peak ratio, and serial spike principal component (SSPC), the performance and stability of waveform feature are better than that of other features. We used spatial feature as weighting value of similarity matrix computed from waveform feature for AP clustering. As a result, AP clustering determined the amount of clusters automatically and gave reasonable results that are not dependent on experimenter’s experience. By tuning the parameters of AP, preference and damping factor, the over-sorting results can be avoided. Comparing ad-justed Rand index of AP with k-means, AP is about 38% higher than k-means method in accuracy under different S/N ratios. Also, clustering number error of AP is about 67% lower than that of the k-means method. Finally, the PSPCA spike sorting algorithm was applied to 48 experimental tetrode signals recorded from mice RT and VPL. There are 1~10 units sorted out from these data. As indicate above, we conclude that the PSPCA algorithm is useful for sorting spikes recorded by tetrode and performs better results than the k-means spike sorting algorithm.