A Reproducible Classroom-to-Analysis Protocol for AI-Assisted Art Pedagogy to Support Youth Emotional Well-Being: A Secondary Data Worked Example

Art-based pedagogy is widely used to support adolescents’ emotional expression, yet educators and researchers often lack a classroom-ready, reproducible, and ethically explicit workflow for evaluating well-being–related outcomes from art activities. Grounded in self-determination theory and the expressive therapies continuum, this protocol positions AI as a measurement and feedback aid that complements, rather than replaces, teacher-led facilitation. We present the Creative Intelligence Cloud (CIC), a step-by-step classroom-to-analysis protocol that is designed for prospective classroom use, and we demonstrate its full reporting and reproducibility controls using a publicly available youth art and well-being dataset as a secondary-analysis worked example (adolescents aged 14–18; N = 1,753 entries; access and identifiers reported in the Data Availability section). CIC specifies (i) teacher preparation (session goals, autonomy/competence/relatedness-aligned prompts, safeguarding notes, and documentation templates), (ii) in-class student procedures (timing, materials, reflection prompts, and engagement logging), (iii) ethical data governance for youth affect information (de-identification, secure storage, controlled access, and constrained interpretation of model outputs), and (iv) a standardized analysis workflow with fully reported parameters to support replication. The analysis pipeline includes image standardization, augmentation rules designed to preserve emotional semantics, and feature processing prior to training an attention-augmented classifier for affect labeling. Baseline models are trained under matched split indices and comparable training budgets to enable fair comparisons, and reporting follows transparent conventions (software versions, random seeds, hyperparameter search ranges, and repeated-run stability summaries). Where an optional generation/reconstruction baseline is included for methodological completeness, PSNR/SSIM are reported strictly as fidelity indicators for that component; the protocol’s primary evaluative outputs for classroom translation are affect labeling performance (accuracy and F1 with variability across repeated runs) and documentation-linked indicators such as engagement logs and self-reported affect measures. This protocol is implementable on a mid-range workstation and is accompanied by disclosed code and configuration artifacts, together with a submitted Table of Materials, to support reproducibility. Limitations include reliance on a single dataset in the worked example, potential cultural/contextual bias, and the need for prospective classroom validation that links AI-assisted evaluation to sustained learner outcomes. Future work should test CIC across diverse cohorts and compare AI-assisted evaluation against simpler, teacher-administered assessment approaches.