Well this line of thinking is a bit simplisitc compaired to tracking shapes and you may have already incorperate all this.
So you have speed and direction, and maybe rough absolute position from the pulleys right?

So often if you were using a mouse or something you would want to turn the absolute position into speed because speed is generally more emotionally connected. Location gets all tied up with response to physical constraints.
But extrapolate that to acceleration and it gets more interesting and dynamic.

If you are needing pen down and up, maybe put the medium in a holder with a pressure sensor. the pressure might be interesting, it would follow mood and line weight, but might also give a leading indicator of line ending.
maybe try a microphone in that holder rather than on the board.

I did a little gesture recognition code that was shockingly simple. I just broke up the screen into a grid, 3×3 was enough for me which was what was surprising.
i just numbered each region and kept a list of the last 7 locations. (the list was updated each time a region barrier was crossed, and on a no cross timeout.)
i then saved lists from various gestures and used them as a library to compare too.
It’s overly simplistic, but I did not have pen down or up like a block letter recognition system does.
however in looking into block letter recognition, the useful simplifying principle i picked up was to first scale any line to a standard scale before looking for shape patterns.

anyway that just occurred to me because just a total time value of how long in the last xx seconds the pen was in a region could make a nice data map, which would very coarsely relate to how dark that area was getting.