As online and hybrid learning environments expand, traditional instructional frameworks often struggle to address cognitive overload, fragmented attention, and learner disengagement. This work introduces PRISM, an engagement-responsive pedagogical framework comprising four interdependent pillars: Personalised Adaptive Learning, Responsive Engagement, Immersive Learning Experiences, and Social Collaboration. PRISM is designed to support real-time instructional adaptation based on the evolving states of learners. To computationally operationalise PRISM’s core logic, a multimodal transformer model was developed using the publicly available PEEK dataset, containing more than 290,000 learner–video interaction records with behavioural engagement labels. The model achieved better predictive performance (F1 = 82.6%), outperforming text-only and metadata-only baselines. Simulation experiments examined how engagement predictions could trigger micro-level pedagogical interventions within an adaptive instructional loop. Although not tested in live classrooms, ablation studies and temporal attention analyses suggested the additive value of combining textual and behavioural features and highlighted patterns consistent with engagement fluctuations across viewing timelines. A simulated comparison of static versus adaptive conditions indicated that timely interventions may help reduce learner drop-off and improve engagement metrics. These findings suggest the technical feasibility of engagement-aware instruction and provide an initial conceptual foundation for future implementations of PRISM in authentic learning environments. The study concludes with design implications, limitations of simulation-based inference, and directions for future work on deployment, inclusivity, and learner perceptions in adaptive learning systems.
Ibitoye et al. (Tue,) studied this question.