Example real-time classification

The simplest is to try and classify the tutorial MEG dataset which is available from the ftp server. More information is on the dataset is available here. That dataset contains the stimulus classes FC, IC and FIC, corresponding to trigger values 9, 5 and 3.

cfg = [];
cfg.dataset  = 'Subject01.ds';
cfg.trialfun = 'trialfun_twoclass_classification';
cfg.trialdef.numtrain    = 20;
cfg.trialdef.eventtype   = 'backpanel trigger';
cfg.trialdef.eventvalue1 = 9; % FC
cfg.trialdef.eventvalue2 = 3; % FIC
cfg.trialdef.prestim     = 0.3;
cfg.trialdef.poststim    = 0.7;

The trial definition function trialfun_twoclass_classification that is being used is included in the fieldtrip/trialfun directory. Based on the code above you can already do

dummy = ft_definetrial(cfg);

to see how the configuration and especially the trial definition looks like:

>> dummy.trl
ans =
         211         510          90         NaN           0
        1111        1410          90           2           0
        2011        2310          90         NaN           0
        2911        3210          90           2           0
        3811        4110          90           1           0
        4711        5010          90         NaN           0
        5611        5910          90         NaN           0
        6511        6810          90           2           0
        7411        7710          90           1           0

The first column is the beginsample, the second the endsample, the third column the offset of each segment. The fourth column indicates the class of each data segment (NaN means unknown, which happens for the third trigger type in this dataset) and the fifth column whether it should be used for training (0) or testing (1).

However, here we are not interested in the trial definition for offline processing, but instead for online classification. So based on the cfg structure above, you can run


The ft_realtime_classification function will print the classification result on screen and will open a figure in which the timing is displayed. It being an offline application here, the timing is measured relative to the amount of data that is processed. An acceleration factor larger than 1 means that data is processed faster than realtime, whereas smaller than 1 would indicate that it cannot keep up with the realtime speed. Note that there is quite some time spent on plotting the timing figure. Furthermore note that the timing is relative to the processed data, whereas there is also time between the trials for which the data does not have to be processed.

function ft_realtime_classification(cfg)
% FT_REALTIME_CLASSIFICATION is an example realtime application for online
% classification of the data. It should work both for EEG and MEG.
% Use as
%   ft_realtime_classification(cfg)
% with the following configuration options
%   cfg.channel    = cell-array, see FT_CHANNELSELECTION (default = 'all')
%   cfg.trialfun   = string with the trial function
% The source of the data is configured as
%   cfg.dataset       = string
% or alternatively to obtain more low-level control as
%   cfg.datafile      = string
%   cfg.headerfile    = string
%   cfg.eventfile     = string
%   cfg.dataformat    = string, default is determined automatic
%   cfg.headerformat  = string, default is determined automatic
%   cfg.eventformat   = string, default is determined automatic
% This function works with two-class data that is timelocked to a trigger.
% Data selection is based on events that should be present in the
% datastream or datafile. The user should specify a trial function that
% selects pieces of data to be classified, or pieces of data on which the
% classifier has to be trained.The trialfun should return segments in a
% trial definition (see FT_DEFINETRIAL). The 4th column of the trl matrix
% should contain the class label (number 1 or 2). The 5th colum of the trl
% matrix should contain a flag indicating whether it belongs to the test or
% to the training set (0 or 1 respectively).
% Example useage:
%   cfg = [];
%   cfg.dataset  = 'Subject01.ds';
%   cfg.trialfun = 'trialfun_Subject01';
%   ft_realtime_classification(cfg);
% To stop the realtime function, you have to press Ctrl-C
% Copyright (C) 2009, Robert Oostenveld
% Subversion does not use the Log keyword, use 'svn log <filename>' or 'svn -v log | less' to get detailled information
% this makes use of an external classification toolbox
hastoolbox('prtools', 1);
% set the default configuration options
if ~isfield(cfg, 'dataformat'),     cfg.dataformat = [];      end % default is detected automatically
if ~isfield(cfg, 'headerformat'),   cfg.headerformat = [];    end % default is detected automatically
if ~isfield(cfg, 'eventformat'),    cfg.eventformat = [];     end % default is detected automatically
if ~isfield(cfg, 'channel'),        cfg.channel = 'all';      end
if ~isfield(cfg, 'bufferdata'),     cfg.bufferdata = 'last';  end % first or last
% translate dataset into datafile+headerfile
cfg = ft_checkconfig(cfg, 'dataset2files', 'yes');
cfg = ft_checkconfig(cfg, 'required', {'datafile' 'headerfile'});
% ensure that the persistent variables related to caching are cleared
clear read_header
% start by reading the header from the realtime buffer
hdr = ft_read_header(cfg.headerfile, 'cache', true);
% define a subset of channels for reading
cfg.channel = channelselection(cfg.channel, hdr.label);
chanindx    = match_str(hdr.label, cfg.channel);
nchan       = length(chanindx);
if nchan==0
  error('no channels were selected');
% these are for the data handling
prevSample = 0;
count      = 0;
% measure the timeing
t(1) = toc;
s(1) = 0;
% these are for the classification
W           = [];
correct     = [];
train_class = [];
train_dat   = [];
% this is the general BCI loop where realtime incoming data is handled
while true
  % determine latest header and event information
  event     = read_event(cfg.dataset, 'minsample', prevSample+1);  % only consider events that are later than the data processed sofar
  hdr       = read_header(cfg.dataset, 'cache', true);             % the trialfun might want to use this, but it is not required
  cfg.event = event;                                               % store it in the configuration, so that it can be passed on to the trialfun
  cfg.hdr   = hdr;                                                 % store it in the configuration, so that it can be passed on to the trialfun
  % evaluate the trialfun, note that the trialfun should not re-read the events and header
  fprintf('evaluating ''%s'' based on %d events\n', cfg.trialfun, length(event));
  trl = feval(cfg.trialfun, cfg);
  % the code below assumes that the 4th column of the trl matrix contains
  % the class label and the 5th column a boolean indicating whether it is a
  % training set item or test set item
  if size(trl,2)<4
    trl(:,4) = nan; % don't asign a default class
  if size(trl,2)<5
    trl(:,5) = 0; % assume that it is a test set item
  fprintf('processing %d trials\n', size(trl,1));
  for trllop=1:size(trl,1)
    begsample = trl(trllop,1);
    endsample = trl(trllop,2);
    class     = trl(trllop,4);
    train     = trl(trllop,5)==1;
    test      = trl(trllop,5)==0;
    % remember up to where the data was read
    prevSample  = endsample;
    count       = count + 1;
    fprintf('processing segment %d from sample %d to %d, class = %d, train = %d\n', count, begsample, endsample, class, train);
    % read data segment from buffer
    dat = read_data(cfg.datafile, 'header', hdr, 'begsample', begsample, 'endsample', endsample, 'chanindx', chanindx, 'checkboundary', false);
    % keep track of the timing
    t(end+1) = toc;
    s(end+1) = endsample;
    % compute the cummulative and instantaneous number of samples per second
    % compare these to the sampling frequency to get the relative acceleration factor
    instantaneous = [nan diff(s) ./ diff(t)];
    cumulative    = (s-s(1)) ./ (t-t(1));
    semilogy([instantaneous(:) cumulative(:)]/hdr.Fs, '.');
    title('acceleration factor');
    legend({'instantaneous', 'cumulative'});
    % force Matlab to update the figure
    % from here onward it is specific to the processing of the data
    % apply some preprocessing options
    dat = preproc_baselinecorrect(dat);
    if test
      % retrain the classifier based on the accumulated training data
      if isempty(W) && numel(unique(train_class))==2
        % only if the classifier needs to be retrained and can be retrained
        fprintf('retraining the classifier based on %d examples\n', length(train_class));
        A = dataset(train_dat, train_class);
        W = svc(A);
      % classify this trial
      if ~isempty(W)
        [nchan, nsmp] = size(dat);
        dat = reshape(dat, [1, nchan*nsmp]);
        B   = dataset(dat, class);
        Bc  = B*W;
        estimate = labeld(Bc);          % this is the estimated class
        warning('classifier has not yet been trained');
        estimate = nan;
      % keep track of the classification performance
      fprintf('estimated class = %d, real class = %d\n', estimate, class);
      if ~isnan(class)
        % this can only be done if the true class is known
        correct(end+1) = (estimate==class);
        fprintf('classification rate = %d%%\n', round(mean(correct)*100));
    end % if test
    if train
      % delete the previously trained classifier
      W = [];
      % add the current trial to the training data
      fprintf('adding one example to the training dataset\n');
      [nchan, nsmp] = size(dat);
      dat = reshape(dat, [1, nchan*nsmp]);
      if isempty(train_dat)
        train_dat   = dat;
        train_class = class;
        train_dat   = cat(1, train_dat,   dat);
        train_class = cat(1, train_class, class);
    end % if train
  end % looping over new trials
end % while true