Deep Wavelet Self-Attention Non-negative Tensor Factorization for non-linear analysis and classification of fMRI data

Objective: This study presents Deep Wavelet Self-Attention Non-negative Tensor Factorization (Deep WSANTF), an innovative framework tailored to address the challenges of dimensionality reduction and classification in multidimensional, highly non-linear fMRI data for Autism Spectrum Disorder (ASD) and Attention-Deficit/Hyperactivity Disorder (ADHD). Methods: The proposed framework integrates wavelet self-attention mechanisms to focus on intrinsic time-frequency features, combines Non-negative Tensor Factorization (NTF) for interpretability with deep learning for non-linear data modeling, incorporates a multi-branch CNN for robust classification of neuropsychiatric disorders, and includes a formal proof of stability theory to ensure the reliability of the model under varying data distributions and perturbations. Results: Deep WSANTF achieves a classification accuracy improvement of up to 15% over state-of-the-art methods, maintains signal-to-noise ratio (SNR) under up to 4.3% noise perturbation, and provides superior feature reconstruction quality, as demonstrated in evaluations on fMRI datasets for ASD and ADHD. Conclusions: The framework effectively integrates interpretability and advanced modeling capabilities, making it an effective tool for analyzing complex fMRI data with the potential to facilitate early diagnosis of neurodevelopmental and neuropsychiatric disorders.