Introduction
Dependent BCI - 不用normal pathways to carry the message, 但需要带有信号的pathways去generate brain activity(比如EEG) Independent BCI - 不依赖normal output pathways - Message is not carried in the brain’s normal pathways
Synchronous BCI(Cue guided) - 从一个mental task切换到下一个时, 必须遵守a fixed repetitive scheme - All commands are connected to the stimulus - BCI depends on proper timing of stimulus onset in EEG recordings
- Asynchronous BCI(self-paced)
- 原理: subject 当mental task stop或者下一个mental task的开始时做出自愿的self-paced decisions.
- 不依赖于外部信号.
- BCI decodes and evaluates the brain signals generated and decided by the subject.
- Approaches to BCI control
- 刚开始: 用neurofeedback and operant learning principle(subject learning)
- 后来: Machine Learning, 允许推断specific brain的statistical signature
- 现在: subject and machine training同步.
- BCI指标
- Accuracy
- speed
- Information transfer rate
Signal Processing
STFT(Short time Fourier transform)
一小段时间内信号的傅里叶变换
- 采集到的信号通常是连续的, 需要合适的采样频率去最小化loss of information
- Spectral filter: FIR and IIR
- 理论上需要很高阶的滤波器, 但实际上做不到.
Machine Learning
subject to subject variability session to session variability trail to trail variability
- feature selection
- 原因:
- curse of dimensionality
- improve performance of classifier
- save computation
- better understanding of the undrelying processing that generated the data.
- 方法:
- filter: rank features based on their usability for classification, 与分类器类型没有关系.
- student t-test(filter)
- $\r^{2}$ matrix of difference(filter)
- wrapper: access subset of features, 根据他们的对于指定分类器的usefulness.
- K-fold cross-validation
- 步骤
-
- start from some subset of features
-
- filter: rank features based on their usability for classification, 与分类器类型没有关系.
- 原因:
Invasive Recording
Classical approach: Wire electrodes MEMS: Utah array, Michigan probes
- 常用Signal processing 方法:
- Spike detection
- Window discriminator
- 缺点:
- need human supervision
- time consuming using multiple electrodes
- statistical properties is not well understood.
- 缺点:
- power method
- principal components
- 优点: use of f-test statistic, 可以用阈值去控制false positive
- 缺点: 需要监督, 需要user提取spike去建立pc.
- matched filtering: 需要监督
- NEO(Non-linear energy operator)
- wavelets
- unsupervised
- Window discriminator
- Spike sorting
- Feature extracted from each spike:
- PCA components
- Wavelet coefficients
- Spike串中的Spikes可能来自很多个neuron, 需要sorting来确定: 1.有多少neurons被捕捉到. 2.把spikes分配到不同的neuron上
- mixture of t-distribution
- 缺点是当neuron fires很快, spike的幅度会衰减.
- 一个neuron可能会产生几个cluster
- Feature extracted from each spike:
- Spike detection
- Recording 方法
- multi-electrode arrays
- Utah array
- Michigan Probes
- NeuroProbes
- Electrode selection
- 因为每个shaft上有上百个electrodes, 不是每一个electrode都可以同步读出来的, 所以为了保证同步, contact pad的数量是有限的.
- 根据一些标准来选择电极~~~
- SNR-based(unsupervised)
- 步骤:
- spike detection
- 计算SNR, (平均N个spikes的root mean square value)
- 因为电极之间靠的太近, 同一个neuron可能会被选择多次
- 步骤:
- Penalized SNR(unsupervised)
- 之前的有问题, 所以要考虑之前选择过的电极, 对已经选择过的进行惩罚, 避免重复选择.
- Expert scores(supervised)
- 人为确定signal quality score
- multi-electrode arrays
Invasive BCI base on action potential and local field potential(LFP)
- Action potential(0.3-250HZ)
- 特征:
- the number of spikes in 20 bins with length 25ms from all 96 electrodes(1920 features in total)
- 特征:
- Local field potential(>500HZ)
- 对于慢速颅内记录(chronic intracranial recording), action potential常常丢失. 原因包括: cell expiration, inflammation, reactive gliosis and scarring
- LFP对于信号degradation更加robust, 更容易记录更长的周期
- LFP可以比action potential表现更好.
- Decoding:
- LFP有相似的neural plasticity. 经过几天训练, 可以区分target和stimuli
- LFP包含了移动的动机(intention), 可以被解码出来.
- LFP的特征:
- Wavelet coefficients
- Phase synchrony相位同步性
- Wave propagation
Noninvasive methods for studying brain activity
Noninvasive BCI的四种
- VEP
- dependent BCI, 因为它依赖末梢神经或肌肉控制.
- SSVEP(steady state VEP, 稳态视觉诱发电位): 当visual stimulus频率高于6HZ时, 上一次的stimulus还没散去, 下一次的stimulus就会出现, 这种周期性的响应称为SSVEP.
- SSVEP因为信噪比很高, 适合用来做short-term identification of evoked responses.
- SCP
- 持续500ms至几秒之间
- 负的SCP对应的是与cortical activation相关的movement或者其它功能
- 正的SCP对应的是reduced cortical activation.
- P300
- 正电位, 在不经常性的stimulus, 或强烈分散的stimulus发生300ms之后产生.
- 可以是visual, auditory or somatosensory产生.
- P300会随时间而衰减(所以要不经常那种刺激.)
- 依赖external stimulation, 这是evoked potential的通病!!!
- mu and beta rhythm
- ERD和ERS
- 人醒着的时候主要的sensory和motor cortical areas会有8-12HZ的EEG信号, mu波. 同时会有18-26HZ的beta波.
- movements or preparation of movements 会导致Mu波beta波下降. 这是ERD.
- 与ERD相反的上升是ERS, 是在movements结束后, 放松时产生的.
EEG
a measure of the brain’s voltage fluctuations as detected from scalp electrodes. 是neurons的累积电位活动的近似. 需要通过胶体附着在scalp上.
EEG是两个或多个electrodes的电位差. 所以有几种方法: Single reference montage.(耳垂, 乳头) Bipolar design Common average reference(CAR)
- 电极分布
- 标准10-20系统
- spaced apart 10-20%
- EEG的频率分布
- Delta <4HZ. During sleep, coma
- Theta 4-7HZ. 与emotional stress相关, 挫折, 失望…
- Alpha 8-12HZ. reduce amplitude with sensory stimulation or mental imagery
- Beta 12-36HZ. 在intense mental activity时会上升
- Mu 9-11HZ. 随着移动而减少,
- Lambda(sharp, jagged)
- Vertex(sharp wave)
- 放松的时候Alpha波高, 兴奋的时候Beta波高.
-
Mu波, 身体运动或者意图去运动的时候高. 频率跟alpha波重叠, 但Mu波是从其它brain activity中区别出来的.
- EEG的神经基础
- EEG是电位活动在肌肉中传输, 从electrical generator到recording electrode. 有很大的噪声.
- Action potential太快, 不产生dipole
- Post-synaptic potentials可以满足记录EEG的需求.
- 当Action Potential到达axon terminal时, neurotransmitter 释放.
- neurotransmitter绑在receptor上
- postsynaptic neuron gets depolarized(EPSP) or hyperpolarized(IPSP).
- EPSP and IPSP 时间空间求和.
- 如果postsynaptic neuron到达了depolarization阈值, action potential会在后一个细胞上产生出来.
- 当EPSP在dendrites中生成时, 细胞外电极会检测到一个负电位, 因为NA离子会回流到细胞内…
- pyramidal neurons空间对齐, 并垂直于皮质表面. EEG主要是pyramidal neurons的postsynaptic potentials. 从更深处的generators传来的electrical activity被volume conduction effects分散及加强.
Artifacts
EEG中有一些噪声, 是由其它因素造成的, 而非我们想要的那部分.
MEG(Magnetoencephalography)
一个电偶极子周围有相应的磁场. 场的极性取决于电流方向. 磁场的累加与voltage累加方式相同. 磁场方向与电流方向垂直. 当电流方向与scalp方向平行时, 磁场冲dipole的一边离开头部, 再从另一边进入, 这样就可以测量了. 当电流与scalp方向垂直, 磁场没法离开scalp, 无法测量. MEG对pyramidal cells更加sensitive EEG可以从sulcus(脑回)或者gyrus(脑沟)获得. MEG包含的brain activity更少, 因为他没有radially aligned axons的信息. MEG不用与头皮直接接触, 可以非接触获得.
-
fMRI MRI注重于脑的解剖结构, fMRI关注于脑的作用. 通过neural activity导致血液含氧量上升, fMRI信号也会随之上升.
-
Near-infrared spectroscopy
- 近红外光可以穿透皮肤, 骨头.
- 对含氧血红蛋白和去氧血红蛋白和线粒体的色素, 有特征吸收带. 含氧血红蛋白的大小可以表征脑活动的大小.
- 测量含氧血红蛋白, 还原血红蛋白, 总血红蛋白 和 细胞色素氧化酶的浓度差别.
- 如果用tracer(oxygen or indocyanine green), 就可以测量大脑的血流量和大脑的总血量.
- 好处:可以测量大脑含氧量
- 缺点:难用, 不能连续记录, 对人为因素很敏感, 很难测大脑血流量.
小结
- Spikes太快, EEG主要用post-synaptic potential
- EEG的原理就是cortial上的Pyramidal neurons受到spikes, 产生postsynaptic potential
- MEG
- f-MRI
- NIRS
13 Hybrid BCI
- 几种: ERD是motor imagery, SSVEP是visual attention, Heart rate是emotion, Optical是mental arithmetic
- ERD->SSVEP
- ERD+Heart rate
- ERD+SSVEP
- Heart rate->SSVEP
- Optical->SSVEP
- SSVEP->ERD
- Eye gaze->ERD
- ERD+SSVEP: 证明可以同时工作, 且使系统更加通用
- ERD+SSVEP: 证明可以同时国祚, 但并不比单用SSVEP更好
- 2D cursor control
- P300+SSVEP: 传统的P300容易在同row或同col中发生错误, 用SSVEP来加强它们的区别
- Simultaneous P300/SSVEP: hybrid的accuracy更低, 但是ITR更高