Redefining parameter-efficiency in ADHD diagnosis: A lightweight attention-driven kolmogorov-arnold network with...

Area of Science:

• Artificial Intelligence
• Neuroscience
• Medical Diagnostics

Background:

• Deep learning models, especially Convolutional Neural Networks (CNNs), face challenges in interpretability and computational cost for medical analysis.
• These limitations are critical in Attention Deficit Hyperactivity Disorder (ADHD) diagnosis, where efficiency and understanding model decisions are paramount.

Purpose of the Study:

• To propose a novel, parameter-efficient framework for ADHD diagnosis using Kolmogorov-Arnold Networks (KANs).
• To address the limitations of existing complex deep learning models in medical applications.

Main Methods:

• Utilized a Kolmogorov-Arnold Network (KAN) for restructured feature transformations, reducing parameters while maintaining accuracy.
• Incorporated an attention-driven feature selection mechanism for prioritizing significant features.
• Introduced a novel activation function with learnable coefficients for adaptive transformations.
• Employed a sliding window-based data augmentation technique to enhance model generalization.

Main Results:

• Achieved 79.25% accuracy, 78.75% F1-score, and 78.23% precision on the ADHD-200 dataset.
• Demonstrated superior performance compared to many state-of-the-art ADHD studies.
• Required only a few thousand parameters, significantly fewer than millions used by existing approaches.

Conclusions:

• The proposed KAN-based framework offers substantial parameter reduction with enhanced performance and interpretability.
• This lightweight architecture is highly promising for ADHD diagnosis and other resource-constrained medical applications.