Bridging minds and limbs: novel hybrid deep learning approach for low-cost EEG-based lower limb motor imagery classification

• Published in: Biomedical signal processing and control Vol. 112; p. 108528
• Main Authors: Sun, Zhengdi, Mu, Anle, Qian, Ye, Sun, Xudong
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2026
• Subjects: Deep learning with attention mechanisms; EEG signal processing; Low-cost openBCI platform; Lower limb motor imagery classification; Stroke rehabilitation BCI; Deep learning with attention mechanisms; Low-cost openBCI platform; Lower limb motor imagery classification; Stroke rehabilitation BCI; EEG signal processing
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.108528

•Low-cost OpenBCI-based lower limb MI detection, validated with healthy subjects for stroke rehab use.•Hybrid CNN-Bi-LSTM with attention mechanism boosts lower limb MI detection performance.•PSO algorithm tunes model parameters, enhancing accuracy and computational efficiency.•Experiments show superior recognition accuracy, efficiency, and low system cost. This study aims to develop a highly cost − effective brain − computer interface (BCI) system specifically for detecting lower limb motor imagery during stroke rehabilitation. Existing BCI systems often suffer from high costs and are easily affected by environmental interference. To tackle these problems, our system is built on a modified OpenBCI platform with wireless communication capabilities, featuring improved signal processing algorithms and hardware design. We collected electroencephalogram (EEG) data from ten healthy subjects using a 16 − channel EEG system. The data were preprocessed and augmented through an overlapping window method. To perform feature extraction and classification, we developed a novel hybrid deep learning approach that combines Convolutional Neural Networks (CNN), Bidirectional Long Short − Term Memory (Bi − LSTM), Particle Swarm Optimization (PSO), and attention mechanisms. The system achieved an average classification accuracy of 72.14 % (±3.60 %) for left and right leg motor imagery, with the attention-enhanced CNN-LSTM model showing a 4.1 % improvement over baseline models. These results were comparable to existing literature, demonstrating good potential despite the small sample size. Future work will focus on expanding the sample size and conduct long-term clinical validation to address current limitations and further improve the system’s efficacy in real-world applications.