Extracting last (non NaN) 200 columns from matrix with varying number of NaNs ending the rows

I have pupil size data from an eye tracking experiment. In the experiment, to start each trial the participant looks in the centre of the screen for 1000 ms (prefixation period), then the trial begins. If they look away or blink, the 1000 ms period restarts. So, each trial has a different length prefixation period. When I create a matrix (with each row being a different trial, and each column being a pupil size sample over time) it creates the matrix with the number of columns based on the trial with the longest prefixation period, and then adds varying numbers of NaNs to the end of each other row. I need to extract the last 200 samples (columns) of each trial, but these are not the last 200 columns of the matrix because of the additional NaNs that are added.

At the moment I have this:

Row1 = PreFixBase(1,:); % extract the first row

Row1(isnan(Row1)) = []; %get rid of the NaNs

Row1Base = Row1(end-200+1:end); %extract the last 200 samples / columns

which I do for each row separately and then paste them back together. It works, but is really inefficient (I have 324 rows / trials) and I'm sure there must be a more concise way of doing this, but haven't been able to find the answer.

Any help appreciated.

Answer 1

[Walter Roberson]

%first trick: compute row lengths by cumsum of logical tests
temp = cumsum(~isnan(PreFixBase), 2);
used_length = temp(:,end);

%second trick: use code equivalent to sub2ind() to compute starts of rows
nrow = size(PrefixBase, 1);
row_starts = nrow * (used_length - 199) + (1:nrow).';

%third trick: row entries are "nrow" apart in linear memory. Build a
%complete table of linear indices corresponding to the desired entries
linear_indices = bsxfun(@plus, row_starts, nrow * (0:199));

%fourth trick: a 2D array of linear indices gives a 2D array of output
RowBases = PreFixBase(linear_indices);

[Amy Maddock]

Thanks for your response. I get the error:

Index exceeds matrix dimensions

on the last line of code, even though PreFixBase and linear_indices both have 324 rows. PreFixBase has 24352 columns and linear_indices has 200.

[Walter Roberson]

I should have said

row_starts = nrow * (used_length - 200) + (1:nrow).';

[Amy Maddock]

Thank you so much! Works perfectly.