Developing a Machine Learning Model for Automated Scoring on the Cube in Cognitive Assessments: A Pilot Study

Abstract

Psychological assessments are often used to help assess cognitive impairments. Inconsistencies in marking these assessments in general, and in cube drawing tests in particular, can lead to misdiagnoses and irregularity in accurate monitoring of the cognitive status; that can be crucial especially in multi-site studies. As a pilot study, a machine learning model using a convolutional neural network was developed to classify drawn cube shapes as ”correct” or ”incorrect” automatically. Techniques such as K-fold cross validation, image augmentation, and early stopping were used to optimize the model using training data. A model with a final validation accuracy of 85.7% was developed as a proof of concept; suggestions for further improvement are presented in this paper. This model will eventually help to ensure similar scoring across different sites when patients are assessed by different assessors.