本文關鍵詞：植入式腦機接口神經元鋒電位的時變特征分析與解碼研究　出處：《浙江大學》2014年博士論文　論文類型：學位論文

  更多相關文章： 腦機接口 時變性 初級運動皮層 廣義回歸神經網絡 蒙特卡羅點過程濾波器

【摘要】：腦機接口系統(tǒng)在大腦與外部機械裝置之間建立了一條直接交互的渠道,為殘障病人修復運動功能提供新的方式。其中,解碼算法是腦機接口系統(tǒng)的核心部分,承擔著將神經信號準確翻譯為運動指令的關鍵使命。以往的解碼算法假設神經元活動與運動表達之間的聯(lián)系是靜態(tài)不變的,然而研究發(fā)現(xiàn)神經元的鋒電位發(fā)放規(guī)律可在短期實驗中觀察到明顯的變化,并導致解碼效果逐漸下降。本文在基于大鼠和非人靈長類動物的植入式腦機接口平臺上,分析運動皮層神經元編碼特征的時變規(guī)律,并在此基礎上設計能跟蹤時變性神經活動的解碼算法,用于提高解碼準確性,延長模型的使用時間。 本文搭建了基于大鼠壓桿實驗和猴子二維手臂運動的實驗平臺,同步采集了初級運動皮層(M1)的神經電信號及多種運動參數(shù)。以往研究定性地觀察到神經元鋒電位的發(fā)放模式會隨著時間變化,在此基礎上,本文提出了基于黑盒模型的時變廣義回歸神經網絡算法。該方法能不斷吸收新出現(xiàn)的發(fā)放模式,·忘記不再出現(xiàn)的舊模式,從而動態(tài)實現(xiàn)對神經元時變活動的跟蹤。本文進一步研究了單個神經元鋒電位的編碼模態(tài),設計了具有生理基礎的灰盒模型時變解碼算法。首先建立了神經元編碼函數(shù)時變分析的定量方法,發(fā)現(xiàn)神經元存在多種編碼形式；神經元重要子集的成員和信息量都存在明顯的時變現(xiàn)象,并建立了編碼函數(shù)時變規(guī)律的預測方法。本文將神經元編碼的時變性質融入解碼算法中,提出了雙重蒙特卡羅點過程濾波器。這種基于灰盒模型的算法能跟蹤神經元編碼特征的時變規(guī)律,在仿真數(shù)據(jù)和真實數(shù)據(jù)上實驗都表現(xiàn)出更好的解碼效果。 本研究工作實現(xiàn)了大鼠及猴子運動皮層神經元編碼特征時變規(guī)律的定量分析和解碼研究,主要創(chuàng)新點在于,(1)設計模式層動態(tài)增長的廣義回歸神經網絡算法,降低了大鼠壓桿系統(tǒng)中解碼壓力信號的平均誤差；(2)建立了基于線性-非線性-泊松編碼模型的神經元時變規(guī)律的預測方法,能夠更好地適應捕捉神經元編碼的多樣性和時變性；(3)提出融入神經元編碼特性的雙重蒙特卡羅點過程濾波方法,用于動態(tài)解析神經元集群的時變活動,將猴子二維搖桿的軌跡預測誤差降低5%以上。本研究探索了一條定量描述和解析神經元時變規(guī)律的新思路,為提高解碼效果,設計能更穩(wěn)定工作的腦機接口系統(tǒng)奠定了基礎。
[Abstract]:The BCI system establishes a direct channel of interaction between the brain and external mechanical devices, which provides a new way for disabled patients to repair motor function, in which decoding algorithm is the core part of BCI system. The former decoding algorithms assume that the relationship between neuronal activity and motion expression is static and invariant. However, the study found that the regulation of spikes in neurons can be observed in short-term experiments. This paper analyzes the time-varying characteristics of motor cortical neurons on the implanted brain-computer interface platform based on rats and non-human primates. On this basis, a decoding algorithm which can track time-varying neural activity is designed to improve the accuracy of decoding and prolong the usage time of the model. In this paper, the experimental platform based on rat compression bar experiment and monkey two-dimensional arm movement was built. The neuroelectric signals and various motion parameters of primary motor cortex (M1) were collected simultaneously. Previous studies have qualitatively observed that the mode of neuronal spike release will change with time, and on this basis. In this paper, a time-varying generalized regression neural network algorithm based on black box model is proposed, which can absorb the new distribution mode and forget the old one. In order to dynamically track the time-varying activities of neurons, the coding mode of single neuron spike potential is further studied in this paper. The time-varying decoding algorithm of grey box model with physiological basis is designed. Firstly, a quantitative method of time-varying analysis of neuron coding function is established, and it is found that there are many coding forms in neurons. There is obvious time-varying phenomenon in the members and information of important subset of neuron, and a prediction method of time-varying law of coding function is established. In this paper, the time-varying property of neuron coding is incorporated into decoding algorithm. A double Monte Carlo point process filter is proposed, which is based on grey box model to track the time-varying rule of neural coding features, and performs better decoding performance in both simulation data and real data. In this study, quantitative analysis and decoding of the time-varying characteristics of motor cortical neurons in rats and monkeys have been carried out. The main innovation lies in. 1) the generalized regression neural network algorithm for dynamic growth of mode layer is designed to reduce the average error of decoded pressure signal in rat pressure-bar system. (2) the prediction method of neuron time-varying law based on linear-nonlinear Poisson coding model is established, which can better adapt to capture the diversity and time-varying of neuron coding. A dual Monte Carlo point process filtering method is proposed to dynamically analyze the time-varying activities of neuron clusters. The prediction error of monkey's two-dimensional rocker trajectory is reduced by more than 5%. In this study, a new way of quantificationally describing and analyzing the time-varying rule of neurons is explored to improve the decoding effect. The design of brain-computer interface system, which can work more stably, has laid the foundation.
【學位授予單位】：浙江大學
【學位級別】：博士
【學位授予年份】：2014
【分類號】：TN911.7

【參考文獻】

相關期刊論文 前4條

1 ;Development of an invasive brain machine interface with a monkey model[J];Chinese Science Bulletin;2012年16期

2 明東;萬柏坤;;功能性電刺激技術在截癱行走中的應用研究進展[J];生物醫(yī)學工程學雜志;2007年04期

3 ;Neural decoding based on probabilistic neural network[J];Journal of Zhejiang University-Science B(Biomedicine & Biotechnology);2010年04期

4 王勇;槐瑞托;王敏;楊斌;;基于腦微刺激的智能動物的研究[J];中國生物醫(yī)學工程學報;2006年04期

相關博士學位論文 前1條

1 張巧生;基于猴子M1區(qū)的腕部解碼系統(tǒng)研究[D];浙江大學;2012年

，

本文編號：1429813