Automated ADHD Detection Using Machine Learning, Deep Learning, and Transformer-Based Attention
Loading...
Files
Date
Journal Title
Journal ISSN
Volume Title
Publisher
Unuversity Setif 1 Ferhat Abbas . Faculty of Sciences
Abstract
Automated diagnosis and subtyping of Attention Deficit Hyperactivity Dis-order (ADHD) face critical computational challenges: multi-site heterogeneity, severe class imbalances, and temporal data leakage. This thesis addresses these limitations by developing and evaluating deep learning, multimodal fusion, and decentralized frameworks using two distinct modalities: 3D structural MRI (3D sMRI) and Electroencephalography (EEG).
For 3D sMRI (ADHD-200), a Multimodal Late-Fusion Attention Autoen-coder integrates volumetric scans with 74 tabular phenotypic features to achieve a 62.60% multi-class validation accuracy. Additionally, a hierarchical diag-nostic cascade decouples binary presence screening (62.00% accuracy) from granular subtyping (46.00% accuracy), effectively isolating the complex ADHD-Combined phenotype (0.33 F1-score). To preserve institutional data privacy, a decentralized Federated Learning framework (FedAvg) deployed across 9 clini-cal sites handles non-IID data distribution to yield a robust 58.59% global test accuracy.
For EEG functional streams (Shahed University), a patient-strict validation strategy partitioned by unique Subject IDs eliminates row-wise data leakage, ensuring a valid 80.00% patient-level accuracy. Concurrently, a hybrid 1D-CNN-LSTM network effectively models cross-channel topologies and long-range macro-temporal dependencies, reaching a 79.70% frame-level accuracy across 433,277 streaming segments.
Ultimately, synthesizing spatial-temporal deep architectures, multimodal phe-notypic late-fusion, and privacy-preserving decentralized networks provides a statistically rigorous, highly generalizable foundation for computer-aided neu-rodevelopmental diagnostics.
