Deep Learning-Based Matrix of Mind Multi-Diagnostic Model for Identifying Dementia, Autism Spectrum Disorder, and Learning Disabilities

Early diagnosis of dementia along with autism spectrum disorder (ASD) and learning disabilities (LD) requires immediate intervention to enable better patient results. The current method of diagnosing these conditions separately creates a broken analysis with decreased diagnostic effectiveness. Prioritizing the development of the multi-diagnostic deep learning model stands critical because it allows simultaneous clinical detection and separation of conditions using delicate behavioral, cognitive, and clinical data patterns. The currently available models face three major problems: overfitting alongside both limited generalization capabilities across disorders, inadequate feature selection, and inefficient optimization approaches, resulting in suboptimal accuracy and reduced interpretability. This paper develops a novel integrated system featuring three fundamental elements, which include the Fisher–Entropy–Laplacian Technique (FELT) for feature selection, followed by Matrix of Mind Network (MOM-Net) for multi-diagnostic classification together with Hiking and Shark Smell Optimization (HiSSO) technique for adaptive weight decay estimation. The FELT technique uses the Fisher score with statistical discrimination analysis and redundancy elimination through entropy and locality preservation from the Laplacian to extract relevant nonredundant features from high-dimensional datasets. MOM-Net represents a deep neural architecture that develops disorder-specific data patterns through shared information between tasks to deliver a simultaneous accurate diagnosis of dementia, ASD, and LD. The novel approach merges hand-generated statistical approaches with deep learning and bio-inspired optimization methods to create an interpretive yet efficient diagnostic model. Research conducted using this model resulted in 99.0% accuracy together with 98.9% recall and 98.7% F1-score performance and decreased training loss to 0.18 with low computational time. The FELT-MOM-Net-HiSSO framework demonstrates superior performance than other traditional convolutional neural networks, recurrent neural networks, Transformers, and recent hybrid techniques based on all evaluation metrics, making it ready for clinical deployment in diagnostic support systems for neurocognitive and developmental testing.