基于贝叶斯联合模型的皮层脑机接口实现: 动作电位的实时检测、分类和解码

Accurate decoding is critical to the clinical use of cortical brain-machine interfaces (BMI). The signal processing in typical cortical BMI systems consists of three steps: the detection of action potentials via a voltage threshold, the identification of the neuron which emitted the action potential (spike sorting), and the decoding of the action potentials from a population of neurons into prosthesis control signals. In prior work, these steps were conducted in serial order, precluding the sharing of uncertainty information between steps. For example, in the second step, a spike is labeled as coming from either neuron A or B, without the ability to express the notion of 70% probability from A and 30% probability from B. This completely neglects uncertainty information and prevents it from being used to improve decoding. The proposed project aims to use the uncertainty information in each step to better handle the noise in the neural signal. To achieve this goal, we will unify spike detection, spike sorting, and decoding into one joint probabilistic inference problem and perform joint inference with a Bayesian algorithm. This will enable the sharing of uncertainty information between each step, allowing greater individual-step and overall decoding accuracy and robustness to noise. We will test our methods with experiments on Rhesus macaques. We hope that the development of such a Bayesian unified brain machine interface (BNU-BMI) will bring a step closer the clinical use of cortical BMIs.

精确的皮层脑机接口解码对于临床应用至关重要。传统皮层脑机接口的信号处理分三阶段：用阈值电压来检测动作电位；将动作电位归属于某一个神经元(分类)；最后对神经元群的动作电位序列进行解码，输出假肢控制信号。前人工作中，上述步骤均依次进行，导致各个阶段的不确定性信息无法传递给其他阶段。比如第二阶段中，只能将一个动作电位归属于A细胞或B细胞，不可输出70%可能属于A、30%可能属于B的信息。这彻底忽视并阻碍了利用不确定性信息来促进解码的可能性。本项目旨在利用各阶段的不确定性信息来更好地应对信号中的噪音。为了实现这目标，前人研究中独立进行的检测、分类和解码将被整合到统一的联合概率推论框架之中，通过一个贝叶斯算法联合推论，实现各个阶段的不确定性信息共享，提高各阶段的精度，最终增强解码效果和噪音容忍度，促进皮层脑机接口的临床应用。我们将用猕猴实验来检验该算法。

皮层脑机接口的信号处理过程传统来说有三步：检测动作电位，分类动作电位，然后解码动作电位。本项目的目标是提出并验证一个新的信号处理范式，把三步统一用概率推理来解决，经过减少不确定性信息的丢失来提高解码精确度。本项目设定四个具体目标。第一，合并动作电位的检测与分类，用基于隐马尔科夫模型（hidden Markov model, HMM）的算法同时举行这两步，验证该算法提高检测与分类的准确性。第二，验证该算法能提高解码的精确度。第三，把三步合并，用分层贝叶斯算法举行，并验证该算法比传统的范式更精确。第四，探讨怎么降低该算法的计算量需求，确保实时运行能力。..我们实现了本项目提出的各算法:HMM联合动作电位检测与分类算法,分层贝叶斯的联合动作电位检测、分类与解码算法,以及改进文献里的同时分类与解码算法。该同时分类与解码算法计算动作电位波形的特征的统计量，然后用该统计量作为线性解码器的输入。我们记录猕猴的脑信号来验证这些方法。结果表明联合动作电位检测与分类的确优于传统方法。联合方法在动作电位分类精确度、综合检测与分类精确度、解码精确度、编码模型拟合精确度和处理后脑信号与运动意图的互信息等指标上有提升。不过，联合检测、分类与解码的算法没有比联合检测与分类然后解码算法更好，而且这算法计算量大。因此我们探讨了另外一个思路：用非概率方法同时分类与解码。同时分类与解码方法比传统的分类然后解码方法在解码精确度上表现更优。我们将探讨在同时分类与解码基础上搭建的同时检测、分类与解码这个替代思路。本项目研制的算法提高了植入电极脑机接口的精确度，从而提高脑机接口的实用性，让脑控制的假肢和康复系统更近于临床应用。