A Machine Learning Approach to Detect Intellectual Disability: The Bioneurofusionnet Multimodal Perspective
Main Article Content
Abstract
Background: Intellectual Disability (ID) constitutes a highly heterogeneous neurodevelopmental disorder characterized by profound, early-onset deficits in both intellectual functioning and adaptive behavioral domains. Traditional diagnostic pipelines rely heavily on subjective clinical assessments, which are inherently susceptible to observer bias, cultural disparities, and critical delays in intervention.
Objective: The primary objective of this research is to conceptualize, develop, and validate a high-performance, multimodal artificial intelligence framework, designated as "BioNeuroFusionNet" (BNFN). This framework is engineered to automate the identification of ID and achieve precise multi-class severity stratification (mild, moderate, severe, and profound) aligned with clinical taxonomies.
Methods: The proposed synergistic architecture integrates electroencephalography (EEG) signals, facial topography, and unstructured behavioral/clinical metrics. The BNFN pipeline employs a novel Harmony-ReliefF Optimization (HRO) algorithm for high-dimensional feature selection. At its core, the framework utilizes a hybrid Convolutional Neural Network-Bidirectional Long Short-Term Memory (CNN-BiLSTM) architecture utilizing Independently Recurrent Neural Networks (IndRNN), combined with a Swin Transformer and a clinical BERT encoder. These distinct modal embeddings are unified via a Meta-Guided Cross-Attention (MGCA) fusion mechanism.
Results: Extensive empirical benchmarking on diverse clinical datasets—including TORGO for voice-acoustic analysis, ABIDE for neurophysiological baselines, and MIMIC-IV for clinical textual data—demonstrates that the BNFN model achieves state-of-the-art diagnostic accuracy. The system records an accuracy of 99.63% for behavioral data integration and 98.94% for voice-acoustic analysis, significantly outperforming traditional unimodal architectures such as isolated Random Forests and Support Vector Machines. Furthermore, biological validation indicates a statistically significant enrichment for pathogenic de novo variants in the AI-identified high-risk subgroups ( ). Interpretability is preserved through SHAP and LIME integrations, yielding transparent clinical decision support.
Conclusion: The sophisticated integration of multimodal biological, neurophysiological, and behavioral signals via the BNFN architecture provides a robust, objective, and highly interpretable diagnostic adjunct. This framework effectively mitigates the limitations of unimodal systems, thereby enabling proactive intervention, equitable clinical triage, and personalized therapeutic strategies for individuals presenting with intellectual disabilities.