MTT: Multi-Tier Transformer for Early Diagnosis of Autism Spectrum Disorder and Language Disorder in Infants

Abstract

Current diagnosis of autism spectrum disorder (ASD) and developmental language/reading disorder (DLD/DD) relies predominantly on subjective behavioral assessments, this underscoring the urgent need for objective biomarkers to enable early intervention. This study proposes a Multi-Tier Transformer (MTT) model for early identification of ASD and DLD/DD using resting-state EEG baseline power values. To address severe class imbalance, we augmented the dataset using the SMOTE method. The MTT architecture integrates a Feature-Embedding layer, a Feature-Attention mechanism that dynamically weights multi-spectral inputs, and a dual-attention encoding block comprising both self-attention and cross-attention to enhance contextual representation learning from limited samples. Transfer learning was further employed to improve robustness by pre-training on augmented data and fine-tuning on original samples. Evaluated on clinical infant EEG data, the proposed MTT achieved an accuracy of 0.91 (95% CI: 0.89–0.93), recall of 0.89, and AUC of 0.97, significantly outperforming the state-of-the-art FT-transformer (p = 0.00091). The results indicate that MTT provides a robust and interpretable deep learning tool for auxiliary diagnosis of neurodevelopmental disorders in infancy.