Neuro-Symbolic Integration Using Knowledge Attention Graphs with Advanced Deep Learning Techniques for Detecting Brain Disorders

Abstract

Recent advances in artificial intelligence have demonstrated remarkable progress in medical image analysis; however, most deep learning models remain black-box in nature and lack interpretability—an essential requirement in clinical decision support systems. This study proposes a Neuro-Symbolic Integration framework based on Knowledge Attention Graphs (KAGs) combined with advanced deep neural architectures for accurate and explainable detection of brain disorders. The framework integrates multimodal neuroimaging data—including MRI, EEG, and patient metadata—through a hybrid CNN–LSTM–Transformer embedding model that captures both spatial and temporal brain patterns. Extracted embeddings are then mapped to a Knowledge Attention Graph, where graph neural networks (GNNs) perform relational reasoning guided by domain-specific neuro-symbolic rules. A symbolic reasoning layer ensures consistency with established medical knowledge, enhancing interpretability and transparency. Experimental evaluation on benchmark datasets (ADNI, OASIS-3, and TUH EEG) demonstrates that the proposed NS-KAG model achieves a classification accuracy of 96.87%, an AUC of 0.98, and an Explainability Index (EI) of 82.4%, outperforming conventional CNN, Transformer, and GNN-based approaches. Statistical analysis (p < 0.01) validates the significance of performance improvements. The integration of symbolic reasoning with neural attention not only improves diagnostic accuracy but also bridges the gap between data-driven learning and human-understandable reasoning, thereby enabling trustworthy and transparent AI-driven neurodiagnostic systems.