Subject-level dyslexia detection via optimized EEG channel selection and compact CNN analysis of multi-window spectrograms

Dyslexia is a common learning disability where early detection is critical for timely intervention. While EEG-based approaches show promise in revealing neurophysiological biomarkers; existing methods are often limited by computational inefficiency and their reliance on segment-level classification across all channels, which hinders clinical applicability. In this study, we propose a compact and clinically oriented DL framework that addresses these limitations through four key contributions. First, we introduce a subject-level evaluation strategy that aggregates segment predictions, enabling clinically relevant individual-level decisions. Second, we design an optimized channel selection process that combines wrapper-based Random Forest evaluation with variance criteria, identifying the 10 most discriminative channels from standard 19-channel EEG configurations. This reduces complexity while preserving diagnostic accuracy. Third, we conduct multi-window temporal analysis, demonstrating that 4-second segmentation more effectively captures dyslexia-related alpha and beta oscillations compared to shorter time windows. Fourth, we develop a lightweight CNN tailored for EEG spectrograms, which systematically outperforms more complex pre-trained models in ablation studies. Experimental evaluation on 54 subjects (29 dyslexic, 25 control) shows that our approach achieves 81.7% subject-level accuracy, 83.6% F1-score and 86.1% AUC, while reducing computational cost by 32% relative to full-channel baselines. Notably, the selected channels map to neurologically relevant regions, offering biological validation of the method. These results establish a practical framework for developing clinically reliable, computationally efficient EEG-based diagnostic tools for early dyslexia detection.