HT
← All projects

University project — Completed

EEG and eye state

Eye-state classification from EEG signals and causal analysis, with chronological evaluation.

Period2026
My roleFull analysis — modeling, causal inference, interpretation
TechnologiesPython · Pandas · Scikit-learn · PyTorch · Statsmodels · EconML · Matplotlib
01

The starting point

Can EEG signals predict whether eyes are open or closed? Does predictive ability establish a causal effect on future brain activity? Two questions investigated separately.

Analysis of 14,980 observations from 14 EEG channels, segmented into 117 windows of 128 observations. Comparison of predictive models and study of the effect of eye closure on future EEG activity. No statistically significant causal effect was detected.

14 980Initial EEG samples
117Complete windows
24Test windows
EEG and eye state
Project visual · open to view full size
02

How it works

Simplified functional workflow
  1. 114,980 samples · 14 channels
  2. 2128-sample windows
  3. 3Chronological split
  4. 4Models & evaluation
  5. 5Separate causal analysis
Architecture & implementation
  • Signal preparation and segmentation into 128-observation windows.
  • Chronological split: 78 training, 15 validation and 24 test windows.
  • Comparison of logistic regression, Random Forest and a Transformer Encoder.
  • ROC-AUC, macro F1 and balanced accuracy; five Transformer initializations.
  • Causal estimation using regression with HAC errors and CausalForestDML.
Technical decisions
  • Preserve temporal order and fit preprocessing on training data.
  • Compare simpler models with a Transformer.
  • Report variability and limitations of the small test set.
Lessons learned
  • Predictive performance is not evidence of causation.
  • Class imbalance and a small test set limit what metrics can establish.
Possible next steps
  • Extend the analysis to multiple EEG recordings to improve generalization.
  • Test more complex deep learning architectures (GRU, 1D CNN).
  • Explore additional causal methods (DiD, IV).
03

Results & practical value

Predictive results on the test set

Balanced accuracy: Random Forest 0.8409; Transformer mean 0.7318. Full values in the following table.
Scores range from 0 to 1. Transformer values are means across five initializations.
Recorded evaluation metrics
ModelROC-AUCBalanced accuracyF1 macro
Random Forest0.93180.84090.5872
Transformer (5-run mean)0.85450.73180.5189

Transformer standard deviations: ROC-AUC ±0.2207; balanced accuracy ±0.1741; macro F1 ±0.2146. Only 24 test windows, including 2 closed-eye windows. These are descriptive results, not proof of clinical reliability.

A chronological evaluation

78 training windows → 15 validation windows → 24 test windows. Temporal order is preserved.

  • Compare a deep model with tabular models rather than assume it will be better: Random Forest achieves higher balanced accuracy here than the Transformer’s average.
  • Document a useful negative result: this analysis found no statistically significant causal effect.
  • Transferable skills: time-series preparation, leakage prevention and communication of uncertainty.

Delivered so far

  • Analysis and report completed in 2026.
  • Transformer: mean ROC-AUC 0.8545 and mean balanced accuracy 0.7318 across five initializations.
  • Test set limited to 24 windows, including only two closed-eye windows: results require cautious interpretation.
  • No statistically significant causal effect detected in this analysis.
O1 and O2 channels: the first 1,000 samples. Original figure from the EEG notebook (2026); labels in French.
O1 and O2 channels: the first 1,000 samples. Original figure from the EEG notebook (2026); labels in French.
Proposed DAG: assumptions linking past EEG, eye states, time and future EEG. This diagram does not prove causal relationships; labels in French.
Proposed DAG: assumptions linking past EEG, eye states, time and future EEG. This diagram does not prove causal relationships; labels in French.
Distribution of estimated conditional effects. All CATE confidence intervals include zero: this spread does not establish causal heterogeneity. Original French figure.
Distribution of estimated conditional effects. All CATE confidence intervals include zero: this spread does not establish causal heterogeneity. Original French figure.
04

Scope & limitations

  • The test contains only 24 windows, including 2 with closed eyes. Metrics are fragile and do not establish clinical generalization.
  • Transformer: five initializations; standard deviations reflect their variability, not a population confidence interval.
  • No medical application or business impact has been validated.

Source: 2026 EEG project report and notebook. Values reproduced from recorded results; models were not retrained for this portfolio.