BCIBench: A BCI Benchmarking Suite

Introduction

Increased demands for applications of *brain computer interface* (BCI) have led to growing attention towards their low-power embedded processing architecture design. Most clinical, wellness, and entertainment applications of BCI require wearable and portable devices. Better understanding of application characteristics in terms of computational complexity, memory usage, and power consumption can lead to more effective system designs for future wearable BCIs. For this purpose, we introduce **BCIBench**, a benchmarking suite which includes a wide range of algorithms used for pre-processing, feature extraction and classification in BCI applications, and in the related publications we analyzed the architectural characteristics of these algorithms such as performance, data-intensiveness and memory behavior. We provide insights into architectural components that can enhance the performance and reduce the power consumption of BCI embedded systems using these applications.

- Roozbeh Jafari, email: rjafari@utdallas.edu

- Omid Dehzangi, email: omid.dehzangi@utdallas.edu

- Chengzhi Zong, email: cxz121430@utdallas.edu

- Viswam Nathan, email: vxn078000@utdallas.edu

** Download all source code, data & scripts for BCI Benchmark suite:** BCIBench.rar

** Download all source code, data & scripts for Preprocessing:** Pre_Processing.rar

In the CAR method the mean of all EEG channels is removed from all channels. This method is effective in reducing the noise which is common to all channels such as 60 Hz power source.

**Download all source code, data & scripts for CAR:** Common_Average_Referencing.rar

FIR filters are very popular because of their simple architecture for hardware and software implementation.

**Download all source code, data & scripts for FIR:** FIR_Filter.rar

Laplacian reference adjusts the signal of each channel by removing the average of neighbor channels. This technique is useful for reducing the effect of noise, when it is focused on a specific region.

**Download all source code, data & scripts for Lap:** Laplacian.rar

** Download all source code, data & scripts for Feature Extration:** Feature_Extraction.rar

Autoregressive (AR) algorithm is a method that estimates the model of current value based on previous samples.
The primary goal behind AR modeling is to extract statistics such as variance of the data to model time series.
Therefore, the coefficients of the AR model can be used as a feature of the brain signals.

Related paper about Autoregressive for BCI: Autoregressive.pdf

**Download all source code, data & scripts for AR:** Autoregressive.rar

An EEG signal consists of several frequency bands which are called delta, theta , alpha , beta and gamma bands. There are no strict frequency ranges for these different bands. Typical range for these bandsare as follows: delta (0-4), theta (4-7), alpha (7-14), beta (15-30) and gamma (30-100+). Therefore, analysis of the EEG signals in these frequency bands is very popular. For instance, alpha waves can be detected during wakeful relaxation with closed eyes.These waves are reduced with open eyes, drowsiness and sleep. Band-power algorithm extracts power of the signal in those frequency bands. In order to compute the band-power features, we use wavelet to decompose the signal to different frequency bands. Then power of signal calculated in time domain. Typically several level of wavelet according to the sampling frequency applies to the input signal to decompose it to these bands.

Related paper about bandpass fitler for BCI: bandpower.pdf

**Download all source code, data & scripts for BP:** Band_Power.rar

Fast Fourier Transform (FFT) is an algorithm that rapidly computes the discrete Fourier transform (DFT) and its inverse.

**Download all source code, data & scripts for FFT:** FFT.rar

**Download all source code, data & scripts for Higuchi:** Higuchi.rar

**Download all source code, data & scripts for Hjorth:** Hjorth.rar

Wavelet Packet Decomposition (WPD) is a wavelet transform where the discrete-time (sampled) signal is passed through more filters than the discrete wavelet transform (DWT).

**Download all source code, data & scripts for WPD:** Wavelet.rar

Related paper about Classification methods for BCI: Classification.pdf

** Download all source code, data & scripts for Classification:** Classification.rar

k-Nearest Neighbors algorithm, a non-parametric classification method

**Download all source code, data & scripts for KNN:** KNN.rar

Linear discriminant analysis (LDA) and the related Fisher's linear discriminant are methods used to find a linear combination of features which characterizes or separates two or more classes of objects or events.

**Download all source code, data & scripts for LDA:** LDA.rar

Multilayer Perceptron (MLP) is a feedforward artificial neural network model that maps sets of input data onto a set of appropriate outputs.

**Download all source code, data & scripts for MLP:** MLP.rar

An SVM model is a representation of the examples as points in space, mapped so that the examples of the separate categories are divided by a clear gap that is as wide as possible. New examples are then mapped into that same space and predicted to belong to a category based on which side of the gap they fall on.

**Download all source code, data & scripts for SVM:** SVM.rar

** Download all source code, data & scripts for End-to-end Applications:** Ene_to_end_applications.rar

CAR+Band-Power+LDA this is a motor imagery BCI.

**Download all source code, data & scripts for Motor Imagery 1:** BCI1.rar

FIR+FFT+LDA this is a SSVEP BCI.

**Download all source code, data & scripts for VEP:** BCI2.rar

Laplacian+Autoregressive+MLP this is a motor imagery BCI.

**Download all source code, data & scripts for Motor Imagery 2:** BCI3.rar

FIR+P300 process+LDA this is a P300 based BCI.

**Download all source code, data & scripts for P300:** BCI4.rar

Roozbeh Jafari, Omid Dehzangi, Chengzhi Zong, Viswam Nathan, **'BCIBench: A Benchmarking Suite for EEG-based Brain Computer Interface'**, *ODES-11 2014 Workshop on Optimizations for DSP and Embedded Systems*, February 15, Orlando, FL. (Link)

This work was supported in part by the National Science Foundation, under grant CNS-1150079, the Semiconductor Research Corporation, task # 1836.103 through the Texas Analog Center of Excellence (TxACE). Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the funding organizations