Modular analysis of biosensor data streams, with visualization component, especially designed for EEG signals. Currently only supports OpenBCI hardware.

Primary objective is a modular processing chain, which can analyze incoming data streaming from an EEG device. Signal processing tasks include:

• Mock signal generation for testing (random, sine wave)
• Class label generation (these can also be thought of as "cues" for the BCI user)
• Notch filter
• Bandpass filter (high, low)
• FFT (frequency analysis)
• DWT (frequency-time-phase analyis)
• Signal conversion (from scalar to cartesian coordinates for plotting)
• Fixed length windowing (on a rolling basis, with overlap, for training phase)
• Class segregated windowing (for testing phase, and online use)
• Downsampling (using simple decimation, or advanced LTTB algorithm)
• Machine Learning (realtime data with scikit-learn)

This software is designed with the open-source user in mind. In the neurotech community, this type of use case is often designated "consumer-grade", as distinguished from "research-grade". This distinction serves to set expectations about the resources and expertise of the user. For example, a home user operating a $500 device, with no formal neuroscience background, will be expected to have a different set of requirements than an academic user, in a lab setting, operating a $10,000+ system in the service of funded research.

In general, "consumer grade" means a device offering lower spatial resolution (1-16 electrodes, vs 32-128 in higher end systems), and lower sampling rate (100-250Hz, vs 1000Hz), which means more modest BCI challenges can be targeted. Some examples of "consumer-grade" projects include: motor imagery detection, neurofeedback training, P300 speller, SSVEP controlled applications, and general entertainment projects, like controlling remote control toys and vehicles.

Python 2.7 (note: the excellent Anaconda scientific package for Python is highly recommended to satisfy the most commonly used dependencies quickly)

• numpy (install with conda or pip)
• scikit-learn (install with conda or pip)
• json (install with conda or pip)
• tornado http://www.tornadoweb.org
• cloudbrain https://github.com/marionleborgne/cloudbrain (note: you'll need to install CloudBrain's dependencies too, such as liblo. You can skip the Cassandra install step, unless you're really sure you will be using it.)

Install using normal procedure

Choice A: you plan to modify the code yourself (likely)

python setup.py develop --user

Choice B: you plan to only use this as third-party tool

python setup.py install --user

Sometimes, you might prefer to start/stop the AnalysisService.py and VisualizationServer.py in their own separate terminal window, and this is supported. However, for convenience, most people will want to start/stop both at once. You can start both processes by a script at root dir called "run.py"

python {path-to-bcikit}/run.py  -i octopicorn -d openbci -c conf/viz_demos/eeg.yml

Parameters to run.py are as follows:

In the future, run.py will likely be the preferred method to run bcikit, since visualization and analysis will share certain startup variables and conf file.

• if you plan to use a device connector from CloudBrain, you can start that up first to begin streaming data (mock connector example given)

  python {cloudbrain path}/cloudbrain/publishers/sensor_publisher.py --mock -n openbci -i octopicorn -c localhost -p 99

• Start both Analysis and Visualization by a single convenience script at root dir called "run.py" (should be started with the same device id, name, rabbitmq params as above)

  python {bcikit path}/run.py -i octopicorn -c localhost -n openbci

• Point your web browser to http://localhost:9999/index.html - Currently the eeg/flot is the only demo working
• Ctrl-C to stop analysis process & viz server
• Edit analysis processing chain and parameters by modifying Analysis/conf.yml You can try commenting out module blocks to turn them on or off. Set debug to True to see live output from command line.

In the folder "Analysis", find the conf.yml. This file is used to set a processing chain of analysis modules, defining the order, names of input and output metrics, and any special params used by each module.

For now, the defined analysis modules include:

• SignalGenerator (generate mock data, either random or sine wave)
• TimeWindow (collect raw data into rolling windows/matrices of fixed size, with optional overlap)
• ClassWindow (collect raw data + class labels into rolling windows/matrices of variable size, grouped by class label, no overlap)
• Convert (conversions, example: convert matrix of raw data to coordinate pairs (x,y) - used for plotting)
• Downsample (decrease number of points while still retaining essential features of graph, used for plotting only)
• ModuleTest (used as a template for new modules)

It is up to the user to make sure that, if one module follows another, that module 1 output is compatible with module 2 input. For example, if a module is expecting scalar input, don't have its input connected to output from a module that emits matrices.

Some limitations to be aware of:

• only works with openbci device type for now
• the config vars inputs > data > message_queues and inputs > data > message_queues define metric names to read/write on rabbitmq
• only rabbitmq is supported, not pipes
• there are some random things hardcoded

• start streaming data using cloudbrain with a mock openbci (assumes you're running rabbitmq locally)

python {cloudbrain path}/cloudbrain/publishers/sensor_publisher.py --mock -n openbci -i octopicorn -c localhost -p 99

• start the analysis modules script in a separate terminal window, using same device id and rabbitmq host

python {CloudbrainAnalysis path}/Analysis/AnalysisService.py -i octopicorn -c localhost -n openbci

If debug is on for a given module, it should output to command line.

This is very rough. Under the folder "Viz", there are demos, intended to show a chart visualization per metric/module, per library. For example, starting with raw eeg, there will be an example using libraries flot, chart.js v1.0, chart.js v2.0, rickshaw, and other libraries.

The intent here is to provide for many different impementations to be demoed so we can compare performance and feasibility of different chart libraries for a specific type of visualization.

For example, in general, we prefer chart libraries that support WebGL. However, the main library we've tried that supports this, chart.js, only offers 5 types of chart. So, it makes sense to branch out and use different libraries for different visualizations, rather than try and find one lib that works well for everything.

The visualizations run using a tornado server copying the pattern used in cloudbrain's rt_server and frontend. The server was modified to use the "multiplex" capability, so that we are not limited to one connection per window. This was modelled after a sockjs example found here: https://github.com/mrjoes/sockjs-tornado/tree/master/examples/multiplex That is why there is reference to "ann" and "bob". The server is defined in Viz/VisualizationServer.py.

The tornado server just passes through all get requests to their relative location based on the "www" folder as root. That is why all the visualization code is presently stuffed into the folder "Multiplex".

If you establish an "ann" type connection, it will use a PlotConnection, which is just a proxy for the connection defined in connection_subscriber.py. This is meant for output from cloudbrain to a javascript visualization via websocket.

If you establish a "bob" type connection, it will use a ClassLabelConnection, which, as the type name suggests, is meant for the javascript frontend to actually send data back to cludbrain via websocket. This is not yet implemented. The idea here is that the frontend visualization will require the capability to show some UI to the user meant for training (or calibration) sessions, in which capturing class label tag is critical.

The current limitations are so many it's not worth listing out. Only outputs of type (message_type: "MATRIX", data_type: "RAW_COORDS") are supported in any of the visualization demos, so that means the only outputs that can be charted must come from either Convert or Downsample.

Different libraries perform differently in different browsers. Overall, Canvas.js library has (so far) shown the best performance, as seen here: [http://jsperf.com/amcharts-vs-highcharts-vs-canvasjs/16]. For optimal performance, we recommend the latest Chrome browser + Canvas.js. Although Canvas.js was shown to have excellent performance in static drawing benchmarks in Safari (at the above link), in real-world tests of dynamic drawing (i.e. live streaming), Safari is very choppy and laggy.

Pro-tip: In Chrome you can turn on an FPS clock, which will help to measure real performance and visualize GPU memory usage. Go to chrome://flags and turn on FPS counter

1. assuming you're running a mock connector as specified in step 1 above, you can start your server by

python {CloudbrainAnalysis path}/Viz/VisualizationServer.py

2. open your browser and visit the path you want, relative to the www folder, like this http://localhost:9999/index.html (only working demo for now is the flot eeg)

3. the "eeg" metric should work if you have mock connector streaming. The basic idea is that you pick the metric you want to see and click "connect" to start streaming it. The actual websocket connection is opened when the page loads.

For calculations on EEG data, we will follow the machine learning conventions of "vector" and "matrix". For anyone new to this field, it's helpful to know that a vector is simply an array. And a matrix is an array of arrays.

Specifically, in EEG data, all of the data coming from a single electrode is considered an array of voltage values over time, or a vector containing time-series data for that electrode. Each electrode, or "channel", is represented as a vector of voltage readings.

When you have more than one electrode, you now have multiple vectors, and we will put the vectors together in a matrix.

Suppose that, for an EEG system with output resolution of 250Hz, you have 3 electrodes, and you're analyzing one second of data from this system. Since our system is 250Hz, or 250 readings per second, this means we have 250 readings x 3 channels. By using a matrix, this will be represented as a grid. We have two choices:

The choice of whether to represent electrode channel data vectors as rows (example 1), or as columns (example 2), is debatable. Our convention in CloudbrainAnalysis will be to use a 250 columns and 3 rows (example 1). Each row will represent all the datapoints of a single electrode channel. Each column will represent data across all channels at one point in time.

One justification for this is simply that this is intuitively how we tend to think of time series data, with the horizontal axis of a graph meaning "time". Additionally, this is a similar convention to what is used in scikit-learn and MNE packages.

It has been suggested that there is performance optimization which can be realized by performing calculations on elements which are contiguous in memory. You can read more about that in the section "Note on array order" here: http://scikit-image.org/docs/dev/user_guide/numpy_images.html

Here's a quick primer on numpy's matrix for those who might be new to python's way of handling matrices.

The np.matrix() object can be thought of as a wrapper around the basic np.array() (type='ndarray').
A matrix is just an array of arrays:

a = np.array([[1,2,3,4], [5,6,7,8], [9,10,11,12]])
[[ 1  2  3  4]
 [ 5  6  7  8]
 [ 9 10 11 12]]

By using the matrix object, and not just ndarray, we get access to some useful matrix math functions. However, just keep in mind that the actual data is contained internally in a np.array(). You can always access this by the A property of the matrix. Or, if you just want to print string, just print the matrix directly.

window = np.matrix(np.random.rand(3,6))
foo = window.A
print foo
print window

Once you have a matrix, you can access specific channels and slices of data using matrix notation. This may be familiar to anyone who has used Matlab.

# matrix of random values
# 3 EEG channels (rows)
# 6 timepoints (columns)
window = np.matrix(np.random.rand(3,6))
print window
#
#[[ 0.01252493  0.76089514  0.90413342  0.14933877  0.95271887  0.62169743]
# [ 0.4461164   0.67827997  0.47488861  0.66459204  0.96701774  0.65374514]
# [ 0.22725775  0.35366613  0.04000928  0.362111    0.49086496  0.04899759]]
#
print "vector representing all channels' data at timepoint 5:"
print window[:,4]
#
#[[ 0.95271887]
# [ 0.96701774]
# [ 0.49086496]]
#
print "vector representing only channel 2 data for the entire time series in the window:"
print window[1,:]
#
#[[ 0.4461164   0.67827997  0.47488861  0.66459204  0.96701774  0.65374514]]
#
print "vector representing only channel 2 data between timepoint 3 and 5 (notice that range includes starting point 3, but not endpoint 5, so by saying [between 3 and 5], you're getting [3,4]):"
print window[1,2:4]
#
#[[ 0.47488861  0.66459204]]
#
print "vector representing only channel 2 data from timepoint 3 to the end (including starting point 3):"
print window[1,2:]
#
#[[ 0.47488861  0.66459204  0.96701774  0.65374514]]
#
print "vector representing only channel 2 data from beginning to timepoint 3 (not including endpoint 3):"
print window[1,:2]
#
#[[ 0.4461164   0.67827997]]
#

• establish a convention for modules to specify what kinds of visualization they are compatible with.
• establish a convention whereby, if any module in configuration has specified a visualization component, the visualization server will be auto-started

• Bandpass Filter (High, Low)
• Common Spatial Pattern (CSP)
• Discrete Wavelet Transform (DWT)
• Notch Filter (60 Hz, etc)
• Noise Removal
• Eyeblink and EMG Artifact removal
• Channel Visibility (i.e. 8 channels coming in, 2 channels coming out)