Facial Emotion Recognition for Intellectually Disabled Children Using an Explainable CNN-LSTM Framework

Deep learning has made significant strides in the development of Facial Emotion Recognition (FER) systems, but the majority of these systems are trained on datasets of facial expressions from neurotypical individuals and are thus only moderately applicable to children with intellectual disabilities (ID). The child might show emotions with a delayed, subtle or atypical facial dynamics, which makes the traditional static FER models inadequate for inclusive education and assistive care of such children. This paper offers a comprehensive paper of research style on this drawback and presents an explainable CNN-LSTM model for emotion recognition among the intellectually disabled children. The proposed idea is based on ethical video acquisition, face detection, landmark based alignment, frame-level CNN feature extraction, Temporal modelling using LSTM and Interpretability using Grad-CAM. A simulation oriented experimental protocol is developed based on the benchmark FER data and domain-adaptation assumptions in low-resource neurodiverse environments. The accuracy, macro precision, macro recall, macro F1-score, AUC, specificity and the class-wise error analysis are reported against CNN, ResNet-50, EfficientNet-B0, and vision-transformer baselines. The proposed CNN-LSTM model shows an accuracy of 82.6% and a  F1 score of 0.801, which highlights the improvement in the recognition of the emotion that changes gradually. The study highlights that the FER with intellectually disabled children should be assisting, transparent, privacy preserving and human supervised and not diagnostic.