Two-tier nature inspired optimization-driven ensemble of deep learning models for effective autism spectrum disorder diagnosis in disabled persons

Autism spectrum disorder (ASD) includes a varied set of neuropsychiatric illnesses. This disorder is described by a definite grade of loss in social communication, academic functioning, personal contact, and limited and repetitive behaviours. Individuals with ASD might perform, convey, and study in a different way than others. ASDs naturally are apparent before age 3 years, with related impairments affecting manifold regions of a person's lifespan. Deep learning (DL) and machine learning (ML) techniques are used in medical research to diagnose and detect ASD promptly. This study presents a Two-Tier Metaheuristic-Driven Ensemble Deep Learning for Effective Autism Spectrum Disorder Diagnosis in Disabled Persons (T2MEDL-EASDDP) model. The main aim of the presented T2MEDL-EASDDP model is to analyze and diagnose the different stages of ASD in disabled individuals. To accomplish this, the T2MEDL-EASDDP model utilizes min-max normalization for data pre-processing to ensure that the input data is scaled to a uniform range. Furthermore, the improved butterfly optimization algorithm (IBOA)-based feature selection (FS) is utilized to identify the most relevant features and reduce dimensionality efficiently. Additionally, an ensemble of DL holds three approaches, namely autoencoder (AE), long short-term memory (LSTM), and deep belief network (DBN) approach is employed for analyzing and detecting ASD. Finally, the presented T2MEDL-EASDDP model employs brownian motion (BM) and directional mutation scheme-based coati optimizer algorithm (BDCOA) techniques to fine-tune the hyperparameters involved in the three ensemble methods. A wide range of simulation analyses of the T2MEDL-EASDDP technique is accomplished under the ASD-Toddler and ASD-Adult datasets. The performance validation of the T2MEDL-EASDDP method portrayed a superior accuracy value of 97.79% over existing techniques.