Decoding Field Potentials
KeywordsLinear Discriminant Analysis Support Vector Regression Spike Activity Local Field Potential Gamma Band
The local field potential (LFP) is believed to represent the aggregated neuronal activity of populations of neurons. LFP activity encodes information related to the control of movements as well as information about sensory and other cognitive processes. LFP decoding refers to the process of extracting features of these processes from the measured LFP signals. LFP decoding is typically performed for a brain-machine interface application to allow control of an external device such as a prosthetic limb or a cursor on a screen.
Neuromotor prosthesis is a class of brain-machine interfaces (BMIs) that aims to restore lost motor function by decoding control information from neural signals. The local field potential (LFP) is one such neural signal that can be utilized. The LFP is operationally defined as the low-frequency components (<~300 Hz) of neural activity recorded extracellularly using microelectrodes and is generated via the summation of transmembrane currents (Buzsáki et al. 2012). There is great interest in decoding LFP activity due to the relative ease of recording and long-term stability of the signal in comparison with the spiking activity of individual neurons (Andersen et al. 2004; Hwang and Andersen 2013). Further, the LFP has been shown to be information rich and contains temporally structured information that reflects behavioral processes, such as movement planning and execution (Pesaran et al. 2002; Scherberger et al. 2005). There is mounting evidence that specific frequency bands of the LFP may reflect neuronal processes at different spatial scales (Siegel et al. 2012). Gamma band (30–90 Hz) and high gamma band (90–300 Hz) activity is hypothesized to reflect local encoding due to firing of neurons near the recording electrode (Schomburg et al. 2012). Activity in lower-frequency bands, <30 Hz, is believed to reflect distributed processing across larger-scale circuits (Donner and Siegel 2011; Fries 2005; Pesaran et al. 2008).
Previous work has shown that the LFP can be decoded off-line to reconstruct high-dimensional upper limb kinematics (Bansal et al. 2011; Zhuang et al. 2010), 2-dimensional movement trajectories (Mehring et al. 2003), and the end goal of movements (Hwang and Andersen 2013; Markowitz et al. 2011). The LFP can also be used in conjunction with spiking activity to decode behavioral states (Aggarwal et al. 2013); however, the information content of the LFP is not necessarily dependent on spiking activity being present in the vicinity of the electrode (Flint et al. 2012a). In fact, the information conveyed via spiking and LFPs can differ in how they vary with cortical depth (Markowitz et al. 2011).
Work is now being undertaken to move LFP decoding online, with subjects controlling a cursor on a screen (Flint et al. 2012b).
Applications of LFP decoding include extracting information regarding skeletal joint angles, bone segment kinematics, and movement end goal states. This information can then be used to control external actuators such as an upper limb prosthetic, motorized wheelchairs, speech decoders, a computer cursor on a screen, or function electrical stimulators to restore movements or paralyzed muscles.
Methods of Decoding
The LFP is a continuous time varying voltage signal and comprises a multivariate set of observations when many electrodes are recorded simultaneously. Decoding can be performed by using the amplitude of the time varying signal or by transforming the signal into a set of features. Commonly used features involve applying time frequency spectral techniques and decoding in the frequency domain (Mitra and Pesaran 1999), but other features include those based on wavelet transformations and other sparse, over-complete basis sets. To estimate the frequency spectrum, the LFP is generally assumed to be stationary over a short time window, typically 100–500 ms in duration. The window is then stepped in time allowing decoding across time. To minimize bias in the spectral estimates, the LFP can be first multiplied by one or many orthogonal data tapers before being Fourier transformed.
Decoding in the frequency domain allows for extraction of information carried at each frequency band; however, this greatly increases the number of inputs to the decoder. When decomposing N channels into M frequency bands, the number of potential inputs to the decoder increases to N*M. Due to the challenge of fitting decoding algorithms in many dimensional feature spaces, the real-time demands of applications, and limitations in computing power, dimensionality reduction such as principal component analysis can be applied to the frequency transformed LFP to extract a set of modes (Markowitz et al. 2011). Feature selection can then be performed on these modes to obtain a smaller number of features that are then used for decoding.
Once the LFP has been preprocessed to give features, the features are decoded to obtain control signals. Discrete control signals, such as selecting letters on a keyboard, can be decoded using classification algorithms such as linear discriminant analysis (Pesaran et al. 2002), the naïve Bayes classifier (Bai et al. 2007), and support vector classifiers. Continuous control signals, such as the trajectory of an arm movement, can be decoded using regression algorithms, such as linear regression (Wessberg et al. 2000), penalized linear regression (Mulliken et al. 2008), Kalman filters (Wu et al. 2006), and support vector regression (Shpigelman et al. 2008).
Advantages of the Approach
Local field potential decoding offers the advantage of improved long-term stability and ease of recording compared to decoding of action potentials generated by individual neurons. Recordings of action potentials can be difficult to acquire as neurons need to be within ~100 um from the tip of the recording electrode (Buzsáki and Draguhn 2004). Further, action potential recordings are highly susceptible to changes in electrode characteristics due to local inflammatory responses such as the buildup of scar tissue (Polikov et al. 2005). In comparison, the LFP has proven to be easier to acquire and the information content more stable to tissue responses (Andersen et al. 2004). The low-frequency, <~300 Hz, nature of the LFP also minimizes the sampling frequency required to acquire the signal compared to spiking activity. This greatly reduces the power consumption and data transfer requirements of the end application BMI wireless recording systems.
The main limitation of decoding the LFP is that the integrative nature of the signal results in reduced spatial and temporal information. Currently, it is unclear whether enough information can be decoded from the LFP to control higher-dimensional devices such as an upper limb prosthetic. Finally, challenges presented by performing spectral analysis in real time via a fully implanted device also remain to be resolved.
- Flint RD, Wright ZA, Slutzky MW (2012b) Control of a biomimetic brain machine interface with local field potentials: performance and stability of a static decoder over 200 days. Proc Ann Int Conf IEEE Eng Med Biol Soc 2012:6719–6722Google Scholar
- Shpigelman L, Lalazar H, Vaadia E (2009) Kernel-arma for hand tracking and brain-machine interfacing during 3D motor control. Adv Neural Info Process Syst 21:1489–1496Google Scholar