Electroencephalogram (EEG) offers insights into brain function for conditions like Attention Deficit Hyperactivity Disorder (ADHD). Accurate diagnosis is critical, yet existing EEG methods struggle with capturing multi-faceted frequency-band characteristics, complex long-range spatio-temporal dynamics, and avoiding redundant feature learning. We propose Multi-level Expert Disentangled Network (MEDNet), a novel architecture addressing these challenges. MEDNet integrates dedicated multi-expert layers: one for adaptive frequency-band feature extraction, and another utilizing Mamba for efficient temporal modeling and a learnable graph for spatial relationships. An expert disentanglement loss is introduced to ensure learned features are diverse and complementary. Extensive experiments on the DEEG-ADHD dataset demonstrate MEDNet achieves state-of-the-art performance, outperforming existing methods. Ablation studies confirm the contribution of each proposed component.