Using deep learning predictions to study the development of drawing behaviour in children

• Deep learning models classify children’s drawings across 10 distinct age categories. • Accuracy of classification models improves as children’s drawing skills develop. • Bayesian optimization enhances CNN models for predicting drawing age categories. • Grayscale models generally outperform RGB in classifying younger children’s drawings. • Grad-CAM and Captum provide insights into model interpretability of drawing features. Drawing behaviour in children provides a unique window into their cognitive development. This study uses Convolutional Neural Networks (CNNs) to examine cognitive development in children’s drawing behaviour by analysing 386 drawings from 193 participants, comprising 150 children aged 2–10 years and 43 adults from France. CNN models, enhanced by Bayesian optimization, were trained to categorize drawings into ten age groups and to compare children’s drawings with adults’ ones. Results showed that model accuracy increases with the child’s age, reflecting improvement in drawing skills. Techniques like Grad-CAM and Captum offered insights into key features recognized by CNNs, illustrating the potential of deep learning in evaluating developmental milestones, with significant implications for educational psychology and developmental diagnostics.