Energy-Efficient EEG-Based Autism Spectrum Disorder Detection Using a Hyperbolic Attention Neural Network

Long-term physiological monitoring using wearable wireless systems represents a paradigm change in next-generation e-health applications. Specifically, electroencephalography (EEG) represents a noninvasive and trustworthy way of recording brain activity and is a likely candidate for the early diagnosis of autism spectrum disorder (ASD). Yet, conventional methods involving the streaming of raw EEG signals to outside servers for classification consume significant energy and drastically shorten the operational life of wearable sensors. In response to these gaps, this research introduced an energy-aware, sensor-based scheme for ASD detection during early childhood from EEG signals. The system exploits on-node signal denoising via chaotic signal models, feature extraction by dual tree discrete wavelet transform (DT-DWT), and lightweight feature selection by parrot optimization (PO). The core detection is executed via a new Hyperbolic Cross-Head Attention-Based Neural Network (HyperCrossNet) that proposes deep reversible learning in conjunction with spatial and channel-oriented attention mechanisms. The network weights are then optimized by the Pied Kingfisher Optimization Algorithm (PKO) for improved accuracy. Experimental outcomes indicate 99.92% classification, 99.91% recall, and a 99.90% F1-score not mentioning that it has lowered considerably the amount of energy used to transmit the raw data. This effective design enables real-time wearable detection useful and applicable to long-term monitoring.