Intelligent Tutoring Systems (ITS) promise personalized instruction at scale, yet most remain affect-blind, using static, rule-based logic that cannot detect a learner's real-time emotional state, contributing to disengagement and suboptimal outcomes. This gap is acute in low-resource settings, where affect-aware ITS research remains scarce and existing deep learning solutions are often too costly for constrained infrastructure. This study designs, implements, and validates a hybrid 1D CNN-LSTM model for classifying learner engagement. A 1D CNN extracts localized behavioral and affective features from interaction telemetry, a stacked LSTM (50 and 120 hidden units) models their temporal evolution, and a modality-level attention layer fuses emotional, behavioral, and assessment-related streams before a softmax layer produces the engagement decision. Hyperparameters were tuned via grid search over a small, bounded space for reproducibility. The model was evaluated on a DAiSEE subset and two supplementary datasets via five-fold cross validation, modality ablation, temporal robustness testing, and attention weight analysis. It achieved 0.91 accuracy, 0.92 precision, 0.94 recall, 0.92 F1 score, and 0.70 AUC-ROC on a binary engagement decision task, with emotional features the strongest predictor. Temporal validation showed marginal degradation over six months. Results suggest lightweight, attention-based affective computing can improve engagement detection in resource-constrained ITS deployments.

Moluno I. E., C. G. Igiri, and V. T. Emmah, "A Hybrid CNN-LSTM Affective State Model for Improving Student Engagement in Intelligent Tutoring Systems" International Journal of Computer Science and Engineering Development, V1(4): Page(68-76) September - October 2026. ISSN: 3139-0862. www.ijcsed.com. Published by Scientific and Academic Research Publishing.