Using deep learning to detect early signs of cognitive disease

Handwriting and hand drawings have historically been used as a proxy metric to evaluate psychological and cognitive traits, in addition to fine motor skills. Detecting onset of cognitive diseases early is a very challenging task due to the expertise required to evaluate each individual subject, as well as the time that the process entails. In this work, we evaluated the application of state-of-the-art deep learning and transfer learning models to determine if the author of a copied hand drawing of a template displays signs of cognitive disease. Compared to expert cognitive evaluation, our best performing method yielded a mean accuracy of 67.60% and area under ROC of 0.595 in determining if an undiagnosed subject displays signs of cognitive disease. Our results suggest that state of the art techniques in deep learning have the potential to help alleviate the difficulty of screening for early signs of cognitive disease. Results also suggest that transfer learning did not perform as well as a purpose-built network architecture trained from scratch. Lastly, our results indicate that models which consider all drawings made by a subject outperform models that look at drawing-evaluation pairs independently.