Enhanced EEG decoding with weight freezing in parallel deep neuro-fuzzy networks (PDNFN) for multi-neurological disorder diagnosis

Abstract

This study introduces an improved EEG decoding model. It uses parameter freezing within a Parallel Deep Neuro-Fuzzy Network (PDNFN) to improve the accuracy and efficiency of diagnosing neurological disorders. The method outperforms other evaluated approaches, achieving an accuracy of 91.84%, precision of 92.15% and recall of 91.42%. Statistical validation-through p-values < 0.05 and confidence intervals-confirms the robustness of the results. Unlike traditional deep learning models that update all weights during training, this method freezes less important weights based on a mask matrix and input threshold. This technique encourages model sparsity, alleviates overfeeting, and reduces computational load without compromising key discriminative features. Compared to dropout, parameter freezing offers a more structures from of regularization, improving generalization in EEG analysis. The PDNFN architecture combines fuzzy logic with deep learning to handle the uncertainty in EEG signals and enhance feature learning. Experiments conducted across multiple publicly available EEG datasets demonstrate notable improvements in both classification performance and computational efficiency. Comparative evaluation further highlight its superiority over traditional fully connected networks and existing regularization methods. Overall, it provides a practical and efficient solution for real-time clinical diagnosis using EEG data. © The Author(s), under exclusive licence to Springer Nature Switzerland AG 2025.