Abstract
Autistic Adolescents usually experience difficulties in the management of emotions, routine transitions and social cue interpretation. Many existing tools that aim to fill in the gap are often non-personalise, static or lack real-time responsiveness in handling these challenges. This study conceptualises and empirically validates a prototype of an emotionally adaptive AI companion that focuses on reducing stress due to routine transition, emotional regulation and social cue interpretation while increasing personalised management by providing contextual support. A quasi-experimental, mixed methods design is adopted. The core of this system conducts facial multimodal emotion recognition through facial expression and simulated voice tone using transfer learning across three CNN architectures (ResNet-18, MobileNetV2, and EfficientNet-B0) as comparison tests. The resulting emotion output is feeds into a contextual engine for real-time personalised interventions which can also be continuously improved through critical feedback-in-the-loop control architecture based on caregiver logs. The key model trade-offs are validated, the findings established that ResNet-18 possesses the highest accuracy of 48%, EfficientNet-B0 with a superior F1 Score of 0.31 and MobileNetV2 proves to be efficient but slightly lower performance compared to other architectures. Simulated user feedback validation resulted in high preliminary acceptability, as high as 87.5% acceptability for an intervention like ”Reassurance”. This validated the utility of this responsive system. This transfer-learning based, multi-modal pipeline is robust. The results of the comparative analysis uncovered a very profound and instructive trade-off between the complexity of models, their efficiency, and performance metrics relating to accuracy versus the F1-score