Using a Machine-Learning Algorithm to Classify Ugandan Preschool Children’s Drawings

Artificial intelligence has revolutionized the ability to process and analyze large-scale data sets, surpassing human capacity in both speed and efficiency. This study presents a machine-learning algorithm based on the ResNet-50 architecture, developed to classify Ugandan preschool children's drawings according to Lowenfeld’s (1947) first three stages of early artistic development: Scribbling, Pre-schematic, and Schematic. Using Scale 10 of the Formal Elements (FE) Scale (Goodman et al., 2022; Tuman, 1998, 1999), the model was trained on a set of 529 randomly selected human-coded drawings and validated on a different set of 228 human-coded drawings. The model achieved an intraclass correlation coefficient (ICC) of 0.85 with human coders, a validation accuracy of 81%, and a test accuracy of 78%. These evaluation metrics indicate strong agreement and reliable performance across all three of Lowenfeld’s early developmental stages. We demonstrated that a deep-learning model can reliably operationalize an established developmental coding system for preschool children's drawings, creating a scalable method for conducting large cross-cultural developmental research. By automating the classification process, this machine-learning algorithm not only reduces the burden on human coders but also enables high-throughput analysis of large data sets, facilitating scalability for similar studies. The findings have implications for education researchers involved in data analysis of preschool children’s drawings, offering valuable insights into their school readiness. Future work aims to expand the model's classification capabilities to include additional FE scales and Content Scale and apply it to other cultural contexts.