MOVE (Motor Output vs Visualized Execution)

MOVE classifies 64-channel EEG motor imagery into 4 classes using EEGNet and subject-independent validation, reaching ~42% CV accuracy (vs 25% chance).

Comment

Inspiration

Our project, MOVE, was developed for the 2026 Rice Datathon Neurotech Challenge with a clear mission: to decode complex human intent from noisy brain signals. We were inspired by the potential of Brain-Computer Interfaces (BCIs) to restore autonomy to individuals with motor impairments. While simple "Left vs. Right" classification is common, we wanted to tackle a more realistic and challenging problem: distinguishing between four distinct motor intentions—Left Hand, Right Hand, Both Hands, and Both Feet—using 64-channel EEG data

What it does

MOVE is an end-to-end Deep Learning BCI pipeline that takes raw, 64-channel EEG signals and translates them into actionable motor classifications.

  • Input: Raw EEG recordings from 86 subjects
  • Processing: It cleans the data by removing eye blinks (EOG) and isolating motor-specific brain rhythms (Mu and Beta bands)
  • Output: It predicts the user's specific motor intention (e.g., "Moving Both Hands") with a mean accuracy of ~42.23%, significantly outperforming the random chance baseline of 25%.

How we built it

We built a robust, scientifically grounded pipeline using Python, MNE, and TensorFlow/Keras

  1. Data Engineering: We developed a custom aggregation loader that merges individual subject files into a single unified tensor. To handle the massive dataset size on Google Colab, we implemented Memory Mapping and aggressive RAM optimization techniques
  2. Preprocessing (The Winkler Pipeline): We used MNE-Python to implement a standard artifact removal pipeline
    • High-Pass Filtering (1Hz): To remove drift and skin potential artifacts.
    • ICA (Independent Component Analysis): To automatically detect and remove eye blinks using the Fp1 channel as a proxy
    • Band-Pass Filtering (8-30Hz): To isolate the Mu (8-13Hz) and Beta (13-30Hz) rhythms, which are the neurological signatures of motor imagery
  3. Model Architecture: We implemented EEGNet (Lawhern et al.), a compact Convolutional Neural Network optimized for BCI. It features Depthwise Convolutions that act as learnable "Spatial Filters," allowing the model to automatically learn which electrodes (e.g., C3 vs. C4) are most important for each movement
  4. Validation: We used Group K-Fold Cross-Validation (5 Splits) to ensure strict subject independence, preventing the model from "memorizing" specific users

Challenges we ran into

  • Data Integrity Issues: We discovered that real-world EEG recording often results in dropped packets, leading to mismatched signal and label lengths. We had to write custom logic to detect and trim these inconsistencies automatically
  • The "Both Hands" Confusion: Our error analysis revealed that the model struggled to distinguish "Both Hands" from "Right Hand" or "Left Hand" Neurologically, this makes sense because "Both Hands" activates both motor cortices, creating a signal that overlaps heavily with single-hand movements
  • Resource Constraints: Processing 86 subjects required more RAM than available. We solved this by implementing a "Run-Once" preprocessing workflow that caches cleaned data to disk and reconstructs metadata on the fly without reloading the heavy signal data

Accomplishments that we're proud of

  • Scientific Rigor: We didn't just throw data at a model; we validated our signals using neuroscience principles (Mu/Beta ERD) and standard 10-05 montage mapping
  • Performance: Achieving a 42.23% accuracy on a 4-class problem is a strong result, demonstrating that our model learned generalized features rather than noise
  • Stability: Our F1-Scores tracked closely with Accuracy across all folds, proving that our model wasn't biased toward the majority class

What we learned

  • Subject Independence is Key: We learned the importance of using GroupKFold instead of standard KFold. Without it, data leakage from the same subject would artificially inflate our accuracy[
  • Neuroscience Meets AI: We gained a deeper appreciation for how Spatial Filtering in CNNs mimics the biological reality of the Motor Cortex
  • Efficient Data Pipelines: We mastered technical skills in handling large-scale biomedical datasets using memory mapping and efficient tensor manipulation

What's next for MOVE

  • Hierarchical Classification: To resolve the "Both Hands" confusion, we plan to implement a two-stage model: first distinguishing "Hands vs. Feet," and then classifying the specific hand movement
  • Real-Time Inference: Optimizing the pipeline to run on live data streams for real-time BCI control.

  • Riemannian Geometry: Exploring geometric deep learning methods to better capture the covariance matrices of EEG signals.

  • Citation: Vigário, R. N. (1997). Extraction of ocular artifacts from EEG using independent component analysis. Electroencephalography and Clinical Neurophysiology.

  • Citation: Pfurtscheller G, Lopes da Silva FH. Event-related EEG/MEG synchronization and desynchronization: basic principles. Clin Neurophysiol. 1999 Nov;110(11):1842-57. doi: 10.1016/s1388-2457(99)00141-8. PMID: 10576479.

Citation: Lawhern, V. J., Solon, A. J., Waytowich, N. R., Gordon, S. M., Hung, C. P., & Lance, B. J. (2018). EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces. Journal of Neural Engineering

Built With

Share this project:

Updates