BCIwiki (bciwiki.com) — Two occipital electrodes on an EEG headband were enough to separate blinks, left saccades, right saccades and a neutral state, reaching 72.37% test accuracy on the 4-class problem. According to Neurotechnology, whose BrainAccess brand published the post, researchers from Wroclaw University of Science and Technology, with colleagues from Kazimierz Wielki University and Adam Mickiewicz University, reported the work in a conference paper published in June 2026 at the 2026 Progress in Applied Electrical Engineering (PAEE) conference, titled "Preliminary Study on Leveraging Ensemble Convolutional Neural Networks for Robust Eye Movement Pattern Recognition via EEG Systems," DOI 10.1109/PAEE71887.2026.11660679.
Four healthy participants wore the BrainAccess HALO while a custom Python script built on the BrainAccess Board API streamed EEG at 250 Hz from O1 and O2. Each class was segmented into fixed 2-second (500-sample) windows, producing a final dataset of 1,284 samples: 1,208 for training and 76 held out for testing. An Isolation Forest stripped outlier windows before training, and 3 augmentation strategies were applied: Gaussian noise injection (μ = 0, σ = 0.1) to simulate sensor noise, amplitude scaling drawn from U(0.8, 1.2) to mimic electrode-impedance variation, and temporal shifting of up to ±30 samples (about 120 ms) with zero-padding to tolerate small timing misalignments.
The team first evaluated 4 stand-alone architectures as baselines: a 1D-CNN, EEGNet, an LSTM and a Transformer encoder. All 4 topped out around 50% weighted accuracy on the same data, with the CNN-family models outperforming the recurrent and attention-based alternatives. They then trained the 1D-CNN as 4 binary one-vs-rest experts — Left, Right, Blink and Neutral-vs-active — and combined their outputs 2 ways: heuristic voting, which applies a confidence threshold (θ = 0.8) to each expert's softmax output, and stacked generalization, which freezes the pretrained experts and concatenates their penultimate-layer features plus logits into a joint embedding fed to an MLP meta-classifier (one hidden layer, 64 units, ReLU) trained to resolve conflicts into a final 4-class call.
The results ran against intuition. Experts trained to full convergence ("strong") scored much better individually than early-stopped ("weak") experts — 84.21% versus 34.21% for Blink — yet once aggregated, the weak experts consistently produced the stronger ensemble. Stacking with the MLP meta-classifier beat both voting schemes in every configuration tested, and dropping a dedicated Neutral expert from the stack improved things further, landing on 72.37% accuracy with weak base learners as the best result overall. That configuration showed balanced performance across all 4 classes, with Blink and Neutral as particular strengths and Left/Right discrimination the hardest to call, consistent with known difficulties in fine-grained oculomotor discrimination from noisy EEG.
Neurotechnology said in the post that it is continuing to explore a HALO-based eye tracker itself as part of an upcoming EOG suite expected in 2027. The work is a preliminary study: the data come from 4 healthy participants and have not been validated in patient populations or on larger datasets.