Fig. 1: Development of a deep learning model for predicting global cognitive performance.

a Schematic representation of the model development and interpretation process for predicting global cognition performance. The process includes (1) Model Building: selection of a high-performing vision recognition architecture from 47 established deep-learning models, (2) Output characterization: relations with comprehensive motor examinations and neuropathologies, and (3) Key feature identification: development of a pentagon-drawing simulator to identify key drawing characteristics in individuals with lower cognitive function. b Validation performances of the 47 deep learning models. Spearman’s correlation and RMSE between predicted cognition scores from each model and measured values were computed for validation samples. We repeated model training five times, each time using a distinct training set. The median and median absolute deviation of the metrics from the five runs was plotted. Models were ranked based on the average ranking of Spearman’s correlation and RMSE. c Comparison of model’s performances between validation and test sets. The composite ranking was obtained as an average of rankings based on Spearman’s correlation and RMSE. The composite ranking, Spearman’s correlation, and RMSE are compared between validation (x-axis) and test (y-axis) sets. The based on the VGG-19 BN architecture is highlighted as it showed the best permanence with the validation sets.