Analyzing children’s interests aids parents and educators in understanding their needs and preferences, facilitating more personalized educational guidance and support. Traditional methods like observation, interviews, and questionnaires are widely used but are subjective, time-consuming, and lack precision. Emotion computing methods based on single physiological features are insufficient to describe complex interest states comprehensively. This study proposes a multimodal feature fusion method for analyzing children’s interests, integrating facial expressions, eye openness, expression change intensity, head posture, and EEG signals. A Multi-scale Hybrid Attention Weight Concatenation Network (MHAWCN) is employed to predict interest states. It uses multi-scale convolutional operations and hybrid attention mechanisms to comprehensively extract physiological features. The Attention Weight Concatenation Module (AWCM) adjusts focus on different features, enhancing accuracy and robustness. The algorithm achieved 89.87% accuracy on a self-built dataset and 93.4% on the PAMAP2 dataset, demonstrating robustness and generalization capability.