Predicting motion based on electroencephalographic recordings (EEG).

The electroencephalograph, or EEG, is a common medical device first developed to diagnose and monitor neurological disorders such as epilepsy. To record brain activity, up to 128 electrodes are affixed to the patient's scalp at standardized points. When large numbers of neurons below an electrode fire in unison, the difference in electric potential (measured in microvolts) is registered between that electrode and a reference electrode, which is usually attached to one or both earlobes. The EEG can be administered while the patient performs simple tasks -- moving arms or legs, observing images on a screen, listening to music -- or even while he or she is asleep. The EEG's notable advantages, those being its low cost and minimally invasive nature, lend it to applications beyond the hospital setting. The nascent field of brain-computer interfaces (BCI) is one such application.

EEG Motor Movement/Imagery Dataset, or EEGMMIDB, from PhysioNet:

Subjects performed different motor/imagery tasks while 64-channel EEG were recorded using the BCI2000 system. Each subject performed 14 experimental runs: two one-minute baseline runs (one with eyes open, one with eyes closed), and three two-minute runs of each of the four following tasks: A target appears on either the left or the right side of the screen. The subject opens and closes the corresponding fist until the target disappears. Then the subject relaxes. A target appears on either the left or the right side of the screen. The subject imagines opening and closing the corresponding fist until the target disappears. Then the subject relaxes. A target appears on either the top or the bottom of the screen. The subject opens and closes either both fists (if the target is on top) or both feet (if the target is on the bottom) until the target disappears. Then the subject relaxes. A target appears on either the top or the bottom of the screen. The subject imagines opening and closing either both fists (if the target is on top) or both feet (if the target is on the bottom) until the target disappears. Then the subject relaxes.

Easy and cheap to obtain though they may be, EEG recordings do have inherent drawbacks. For one, the noise-to-signal ratio is uncomfortably high, and artifacts are difficult to avoid: electrical interference, both endogenous (heartbeat, eye movement, changes in skin conductivity due to sweating) and exogenous (AC power mains), can make interpretation of recordings difficult or impossible.

Fortunately, these problems can be substantially mitigated with the aid of certain mathematical techniques. Low-pass and high-pass filters were applied to the raw signal at 30 Hz and 7 Hz, respectively, to remove high-frequency noise (including that produced by mains electricity at 60 Hz) as well as low-frequency artifacts caused by heartbeat and eye movement. Conveniently, the band from 7-30 Hz is likely to carry all the information we seek: normal, waking neural oscillations are dominated by alpha and beta waves. More to the point, motor activity is typically marked by activity within the mu band, from 7.5-12.5 Hz. While one is at rest, neurons in the motor cortex tend to fire in unison. At the onset of motion, however, these neurons' activity becomes desynchronized, and therein lies the key to successful prediction.

Independent component analysis (the FastICA implementation, to be specific) was then performed to achieve signal separation. Why is this necessary? Imagine you have placed three microphones in a room where a lively cocktail party is being held. With multiple conversations occurring simultaneously, each microphone is bound to pick up at least some of the sound captured by the other two. If one can safely assume that the component signals (the voices of individual people at the cocktail party) are statistically independent and that the component signals' values are not normally distributed, then ICA can reliably reproduce those signals.

In their initial continuous form, EEG recordings are not suitable ground on which to build a predictive algorithm. The stream must be sectioned into epochs, time-locked series which contain the event of interest. Like-kind epochs can then be averaged across multiple recordings and compared between and among experimental subjects. These averages are known in the literature as event-related potentials.

At this early stage, I have limited the scope of my project to the first version of the experiments found in EEGMMIDB, in which each subject was instructed to clench the corresponding fist when an image appeared on either the left or right side of a screen.

From the recordings of 109 subjects who participated in the experiment, four sets were removed from the dataset because of formatting discrepancies. Those remaining were divided into epochs (200 milliseconds before the event and 500 milliseconds after). Only the fourteen electrodes located above the primary motor cortex were used in this context.

After packaging the data in a compatible form using MNE and standardizing each feature, I divided them into training and testing sets (80% and 20%, respectively) and then employed scikit-learn's SGDClassifier, specifying one of two applicable loss functions, and RandomizedSearchCV to tune parameters for the classifier.

Here are the results produced by SGDClassifier with 'hinge' loss function, which gives a linear support vector machine:

RandomizedSearchCV took 12.69 seconds for 20 candidate parameter settings.
Model with rank: 1
Mean validation score: 0.499 (std: 0.002)
Parameters: {'n_iter': 9, 'alpha': 17.884975133166026, 'shuffle': True}

Model with rank: 2
Mean validation score: 0.499 (std: 0.002)
Parameters: {'n_iter': 10, 'alpha': 0.12765806765435153, 'shuffle': False}

Model with rank: 3
Mean validation score: 0.499 (std: 0.002)
Parameters: {'n_iter': 18, 'alpha': 0.1495637472206022, 'shuffle': False}

It's plain to see that the classifier is performing no better than chance (a score of approximately 50%). I suspect this stems from the format of the input data. MNE provides an EpochsVectorizer class that, as the name suggests, reduces the dimensions of each epoch. Thus far I have not been able to make it work as expected.

$ git clone https://github.com/omega-thirteen/scanners.git

or, using your SSH key:

$ git clone git@github.com:omega-thirteen/scanners.git