<![CDATA[Autism spectrum disorder is a neurodevelopmental condition that affects social communication and behaviour, and diagnosis still relies on subjective behavioural assessment. Electroencephalography provides a noninvasive view of brain activity but is noisy and often analysed with handcrafted features or evaluation schemes that risk data leakage. This study proposes a deep learning pipeline that combines wavelet denoising, EEG-to-image encoding, and heavy-light decision fusion for autism detection from EEG. Sixteen-channel EEG from children and adolescents with autism and typically developing peers in the KAU dataset is denoised using discrete wavelet transform shrinkage, segmented into fixed 4 second windows, and rendered as pseudo colour heatmaps. These images are used to fine-tune five ImageNet pretrained architectures under a unified training protocol with 5-fold cross-validation. Heavy-light fusion combines one heavyweight backbone and one lightweight backbone through weighted soft voting on class posterior probabilities. The strongest single model, ConvNeXt Tiny, attains about 97.25 percent accuracy and 97.10 percent F1 score at the window level. The best heavy light pair, ConvNeXt plus ShuffleNet, reaches about 99.56 percent accuracy and 99.53 percent F1, with sensitivity and specificity in the 99 percent range. Fusion mainly reduces missed ASD windows without increasing false alarms, indicating complementary error patterns between heavy and light models. These findings show that the proposed denoise encode classify pipeline with heavy light fusion yields more robust autism EEG classification than individual backbones and can support EEG-based decision support in autism screening.]]>
Copyrights © 2026
