AFF-BPL: An adaptive feature fusion technique for the diagnosis of autism spectrum disorder using Bat-PSO-LSTM based framework

Autism spectrum disorder (ASD) is a neurological condition revealed by deficiencies in physical well-being, social communication, hyperactive behavior, and increased sensitivity. The delayed diagnosis of ASD showcases a significant obstacle in mitigating the severity of its impact. Individuals with ASD often exhibit restricted and repetitive behavioral patterns. In this context, we proposed a novel adaptive feature fusion technique with a BAT-PSO-LSTM-based network for the diagnosis of autism spectrum disorder. Our focus is on three distinct autism screening datasets namely, Toddlers, Children, and Adults for comprehensive analysis of techniques. Bat and PSO concurrently select features and the selected features will go through an adaptive feature fusion algorithm and an LSTM-based classifier. This research addresses various challenges encountered in the existing techniques including concerns related to overfitting, faster training, interpretability, generalization capability, and reduced computation time. The work incorporates baseline techniques like a neural network, CNN, and LSTM with evaluations based on key parameters like precision, specificity, accuracy, sensitivity, and f1-score. The experimental simulations reveal that AFF-BPL outperforms considered baseline techniques achieving remarkable accuracy on all three datasets. Specifically, the model attains the accuracy of 0.992, 0.989, and 0.986 on toddler, children, and adult datasets respectively. Additionally, the exploration of functional and structural images will provide deeper insights into the underlying mechanism of ASD.