New intelligent music therapy method for applications of enhancing social skills of autism children based on TL-GCN and deep learning.

Researchers developed an AI system called EmoMusik-Net that watches children's facial expressions and automatically plays personalized music to help with social skills and emotions. They tested it with 182 children with autism and found significant improvements - younger boys showed nearly 100% improvement in social interest, and older girls improved by 87% in emotional responses. The system was very accurate at recognizing emotions and choosing appropriate music.

This study introduces EmoMusik-Net, a deep learning system that combines emotion recognition technology with personalized music therapy for autism interventions. The system uses advanced AI to analyze facial expressions and automatically recommend tailored music interventions. Testing with 182 children with autism showed high technical accuracy (97% emotion recognition) and significant improvements in social skills. Boys aged 1-6 demonstrated 98.44% improvement in social interest scores, while girls aged 7-12 showed 86.77% improvement in emotional response scores.

The system achieved strong expert validation with 94% matching accuracy. This technology-driven approach offers a novel tool for delivering personalized music therapy interventions for autism.

This AI-driven music therapy system shows promise for personalized autism interventions, particularly for younger children. The technology could support families and clinicians with automated, responsive therapeutic tools. However, further research is needed to establish long-term effectiveness and practical implementation guidelines before widespread clinical adoption.

Study design unclear from abstract. Sample characteristics not fully described. Questionnaire-based outcomes may have measurement limitations. Long-term sustainability of improvements unknown. Generalizability across different autism presentations unclear. Technical complexity may limit real-world implementation accessibility.

To address the long-standing challenges children with autism face in social skills and emotional regulation, this study introduces Emotion-based Music Intelligent Network (EmoMusik-Net)-a deep learning model designed for intelligent music therapy. The model focuses on emotional impairments exhibited during social interactions, integrating Transformer-based temporal modeling with a Transfer Learning-based Graph Convolutional Network (TL-GCN). This combination enables high-precision recognition of facial expression sequences and supports a dynamically adaptive, closed-loop mechanism for personalized music recommendation. EmoMusik-Net was trained and optimized using three publicly available emotional video datasets.

A pre- and post-intervention study, conducted in collaboration with the families of 182 children with autism, employed questionnaire-based assessments to systematically evaluate the model's real-world feasibility and effectiveness. Experimental results demonstrated that EmoMusik-Net achieved an emotion recognition accuracy above 0.970, an F1-score consistently over 0.960, and an Area Under the Curve (AUC) of 0.978. The model also showed outstanding robustness on large-scale datasets, with a stability score of 0.994, indicating strong classification performance and generalizability. In terms of intervention outcomes, boys aged 1-6 showed a marked increase in social interest scores, rising from 1.280 to 2.540-a 98.44% improvement.

Girls aged 7-12 exhibited significant gains in emotional response scores, from 1.670 to 3.120-an 86.77% increase. Further statistical analysis using the Mixed-effects Model for Repeated Measures (MMRM) and bootstrap confidence interval estimation confirmed the intervention's significance both statistically and clinically, with particularly strong effects observed in younger participants. Expert blind evaluations further validated the system's effectiveness, showing high consistency in rhythm and emotion matching. The Intraclass Correlation Coefficient (ICC) ranged from 0.75 to 0.91, with matching accuracy surpassing 94% in certain subgroups.

EmoMusik-Net not only addresses the current research gap in integrating intelligent emotion recognition with music-based interventions but also offers a responsive, technology-driven support tool for parents, educators, and clinicians. This approach holds strong potential to advance autism spectrum disorder interventions toward personalized, data-driven methodologies.