EEG Classifier

archivedPyTorch, MNE, NumPy

Reads raw brainwaves and detects schizophrenia at 94.3% accuracy. A deep learning architecture that listens to the brain's geometry and its rhythm at the same time. Peer reviewed, IEEE published.

94.3%
Accuracy
CNN+LSTM
Architecture
IEEE
Published

The problem and the build

Diagnosing schizophrenia today still leans on clinical interviews. EEG sees the disorder in the signal, but the signal is noisy, high dimensional, and never the same twice. The architecture I designed treats it the way the brain actually behaves: a CNN front end learns the spatial fingerprint across electrodes, an LSTM back end learns how that fingerprint evolves through time, and the two streams fuse before classification. Months of preprocessing, hyperparameter ablation, and cross subject validation later, the model held 94.3% on a benchmark where prior work plateaued in the high 70s. Published in IEEE Xplore. The takeaway that stayed with me: applying ML to physics, biology, and medicine is where the work is genuinely hard and genuinely matters.

Key insight

Most teams threw bigger models at noisier signals. The win was structural. Match the architecture to the physics. Space and time are different problems. Solve them with different layers, then fuse.

Stack

  • PyTorch
  • MNE
  • NumPy
  • scikit-learn

More case studies

Need something like this in production?

Fixed price, 4 to 6 weeks. I reply to every inquiry within 24 hours.

Reply within 24 hours