Adaptive learning rate optimized hybrid model for early identification of children with autism

Autism spectrum disorder (ASD) in children often leads to academic and social challenges due to difficulties in communication, social engagement, and behavioral adaptation. Delayed diagnosis exacerbates these issues, hindering timely interventions. Recent machine learning (ML) algorithms have shown promise in predicting ASD, enabling early diagnosis and intervention. However, existing datasets suffer from limitations such as small sample sizes, class imbalance, and privacy concerns. The main objective of this study is to design a federated learning (FL)-based hybrid deep learning framework that can accurately and securely detect ASD in children, overcoming data imbalance and privacy challenges. Specifically, the study addresses the following research questions: (i) How can federated learning be employed to securely collect and train ASD-related data across multiple schools while ensuring privacy? (ii) Can a hybrid deep learning model combining convolutional neural network (CNN) and long short-term memory (LSTM) improve ASD prediction accuracy compared to existing models? (iii) To what extent can Dragonfly Optimization (DFO) enhance hyperparameter tuning for robust performance under imbalanced data conditions? To this end, we propose the Adaptive Learning Rate Optimized Hybrid Model (ALROH), which integrates a modified CNN for feature extraction, LSTM for sequential pattern recognition, and DFO for hyperparameter optimization. Evaluated on toddler and children datasets, ALROH achieves accuracies of 99.36% and 99.23%, respectively, outperforming state-of-the-art methods by up to 5%. By enabling privacy-preserving and scalable ASD screening in educational settings, ALROH facilitates early interventions, potentially improving academic outcomes for students with ASD.