Adaptive Learning for STEM Gender Equity: Framework and Protocol with Potential Transfer to Art and Interaction Design

Gender inequities in STEM learning environments are reinforced by uneven instructional support, biased content, and under-recognized well-being risks. This study introduces a single, end-to-end adaptive framework that integrates learning personalization, bias detection, and engagement/affective-state monitoring to foster inclusive STEM learning. Learner-state features are derived from platform logs including quiz/performance scores, response-time and timing features, navigation/clickstream indicators, and submission history, followed by data cleaning, normalization, and dimensionality reduction using Principal Component Analysis (PCA). A reinforcement learning module (Deep Q-Network, DQN) then generates next-step recommendations (content selection and difficulty adjustment), while a refined language model (BERT-based classifier) screens instructional text to flag biased language and representation patterns in educational materials. In parallel, a Residual Attention-Based Radial Movement Bidirectional Network (RA-RM-BNet) models temporal engagement and affect signals to output risk scores and intervention triggers (e.g., threshold-based alerts) for timely support. The workflow was evaluated on 1,800 de-identified learner-interaction records from a publicly available STEM e-learning repository, achieving accuracy of 0.95, precision of 0.92, recall of 0.90, and F1-score of 0.89, while maintaining low execution times compatible with near-real-time deployment. Implementation scripts, configuration files, and example inputs/outputs are provided alongside the protocol (see Data Availability). Beyond STEM education, the cross-disciplinary discussion is positioned as a theoretical transfer to art and interaction design, which would require domain-specific dataset construction and module re-tuning before empirical deployment.