Real-world evaluation of deep learning decoders for motor imagery EEG-based BCIs

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Online Brain-Computer Interfaces (BCIs) using Motor Imagery (MI) Electroencephalography (EEG) demand rapid decisions within narrow timeframes.
• Most current Deep Learning (DL) EEG decoders are optimized and validated offline using extended window lengths, limiting insight into their real-time effectiveness.

Purpose of the Study:

• To bridge the gap between offline and online performance assessment of DL EEG decoders.
• To evaluate the real-time applicability of 10 representative DL decoders under a soft real-time protocol with 2-second windows.

Main Methods:

• Assessed 10 DL decoders (CNNs, filter-bank CNNs, TCNs, attention/Transformer hybrids) using a soft real-time protocol (2-s windows).
• Quantified performance via accuracy, sensitivity, precision, miss-as-neutral rate (MANR), false-alarm rate (FAR), information-transfer rate (ITR), and workload.
• Examined lateralization indices, mu-band power, and topographical contrasts to link decoder behavior with physiological markers.

Main Results:

• Performance rankings shifted significantly between offline and online settings, with increased inter-subject variability.
• Top-performing online decoders included FBLight ConvNet (71.7%) and EEG-TCNet (70.0%); attention/Transformer models showed less stability.
• Errors primarily involved Left-Right swaps, while neutral conditions remained relatively stable. Subject-specific patterns correlated with performance metrics.

Conclusions:

• Compact spectro-temporal CNNs with lightweight temporal context excel in stable short-time window performance for online EEG decoding.
• Deeper attention and Transformer architectures are more prone to inter-subject and inter-session variability.
• This study provides a benchmark and guidance for developing robust online-first EEG decoders under real-world constraints.