A novel deep learning approach for privacy-preserving encoded EEG-based brain-computer interfaces with clinical LLM applications.

Purpose: The rise of large language models (LLMs) such as GPT-4 and DeepSeek has transformed healthcare information processing by enabling natural language-based clinical reasoning. However, the integration of LLMs with privacy-sensitive biomedical signals, particularly electroencephalogram (EEG) data used in brain-computer interface (BCI) systems, remains underexplored. EEG signals, especially during motor imagery (MI) tasks, are critical for assistive neurotechnologies but pose significant privacy risks due to their capacity to reveal cognitive and medical information. Traditional encryption techniques often distort signal structure or require decryption with additional noise, compromising classification performance and real-time usability. Methods: To address this gap, we propose a deep denoising structure-preserving neural encoding network (DSNet) that enables accurate classification of privacy-preserving encoded EEG representations without requiring decryption. EEG features were extracted using common spatial pattern (CSP) and transformed into privacy-preserving encoded representations while preserving their statistical structure. Here, encoding refers to a non-reversible neural transformation designed for privacy preservation rather than a formal cryptographic guarantee. Two deep learning architectures, a feedforward neural network (NN) and a recurrent neural network (RNN), were evaluated for classification in the encoded feature space. Furthermore, we integrated an LLM (GPT-4) to generate clinical-style summaries based on model outputs, enhancing interpretability for clinician review and potential clinical support use. Results and conclusion: Using publicly available datasets, DSNet-NN achieved over 87% accuracy for every subject, outperforming both the RNN variant and baseline models. It also demonstrated resilience to simulated privacy attacks. LLM-generated reports provided clinician-friendly interpretations of MI predictions, supporting potential real-world applicability. This study introduces an AI framework that bridges privacy-preserving EEG decoding with LLM-based clinical reasoning, offering a practical solution for privacy-preserving neurorehabilitation and digital health systems.