CNN-HMM hybrid model for real-time cognitive assessment and adaptive interaction in experiential educational games

Traditional experiential educational games suffer from limited real-time cognitive perception and rigid feedback mechanisms, which fail to deliver personalized learning experiences for users of different age groups. To address this issue, this study proposes an improved convolutional neural network-hidden Markov model (CNN-HMM) hybrid model integrated with channel attention, neural emission network and differentiable forward algorithm. The framework extracts multimodal behavioral features and realizes dynamic cognitive state inference and real-time adaptive teaching regulation. We conduct experiments on 302 participants covering children, adolescents, college students and working adults. The results show that the proposed model achieves an average prediction accuracy of 0.939 for learner cognitive states. It reduces task completion time significantly, with cognitive adaptation rate and feedback adaptation rate both exceeding 0.92. Ablation tests verify the synergistic effect of each core module. Crosscohort generalization experiments and statistical tests confirm the model's strong robustness and universality. This work provides an effective technical solution for intelligent design of educational games and offers a reliable reference for adaptive learning system development.