EEG Brain-State Classification

Team

Donggyu Park (Computer Science), Seong Hun Lee (Computer Engineering), Hyoungju Lim (Physics), Hwanhee Cho (Electrical Engineering).

Problem

Patients with severe brain damage often appear unresponsive, yet studies find that roughly a quarter retain cognitive awareness. A brain-computer interface can give them a channel to communicate — but fMRI and clinical-grade EEG systems are far too expensive to put in front of an individual patient. We set out to find how much of that capability survives on cheap hardware.

What we built

• A two-channel analog front end: an instrumentation amplifier at a gain of about 90, a 60 Hz notch filter for mains interference, a 7 Hz high-pass to reject galvanic skin response, and a 31 Hz low-pass to isolate the beta band (13–30 Hz).
• A double-sided PCB with short traces, dedicated ground planes, and the analog and digital sections kept apart so switching noise stays out of the signal.
• Real-time processing on an STM32 microcontroller: buffered ADC capture, preprocessing and feature extraction using the CMSIS DSP library.
• A support vector machine classifier — chosen for its small footprint on embedded hardware — trained on the open PhysioNet mental-arithmetic EEG dataset.

Results

76.9% accuracy on training data and 66.4% on held-out test data. The honest read: because the dataset is 75% relaxed and 25% focused, a distribution-matched random guess already scores about 62.5%, so the pipeline works end to end but the current features do not yet separate the focused state well — most focused samples were misread as relaxed.

Next steps

Frequency-domain and time-frequency features, better artifact rejection, and a lightweight deep model sized to run on the microcontroller.

Presented

MRC 2025 and Engineering Open House 2025, University of Illinois Urbana-Champaign.