Background: With the advance of technology, our
capacity to assess patients with dementia is also developing. It is possible
to administer cognitive examinations using technology, such as the
iPad-based Toronto Cognitive Assessment, but hitherto difficult to
autonomously administer them. Many of the ’inputs’ from patients could be
easily scored with software, but highly variable inputs such as the clock
drawing are extremely difficult to score, precluding automated
administration and scoring. This work focuses on the development of a neural
network designed to assess cube drawings, infinity drawings, and clock
drawings. Methods: 3200 drawings, evenly split
between clocks, cubes and infinities were generated, with half being correct
and half incorrect. A SqueezeNet was trained on 2000 images, validated on
800 drawings, and then tested on 400 drawings.
Results: The SqueezeNet was able to achieve 97%
accuracy on 400 images it had never seen before in categorizing images as
“Cube”, “Clock”,
“Infinity”, or “Other” (incorrectly drawn).
Conclusions: This neural network can successfully
determine the difference between correctly and incorrectly drawn images
commonly used in cognitive examinations, overcoming the final barrier to
autonomously administering and scoring cognitive examinations. Next steps
are to clinically validate an autonomous examination program which has been
modeled after the Addenbrooke Cognitive Examination-3.