# Postmortem: Biased Model Training with EEG Signals 



## 

Persistent model bias observed during training on EEG dataset for wheelchair navigation commands:

- Models consistently best identified `stop/neutral` → moderately performed on `forward/left` → consistently failed on `right`

- Observed across MATLAB Classification Learner models and custom 

- Best outcomes: ~50% accuracy for problematic classes ("right") with XGBoost/Bayesian 



## Root 

1. **Imbalanced Command Representation**  

   Dataset captured identical durations/sessions per command, ignoring natural variance in EEG pattern 



2. **Signal Similarity**  

   Movements `forward`, `left`, and `right` generate similar frontal lobe (FP1/FP2) activation 



3. **Feature Limitation**  

   Frontal electrodes (FP1/FP2) insufficient to capture directional intent 



4. **Data Validation Gap**  

   Lack of EEG-specific verification methods (e.g. ERP visualization, spectral comparison)



## Why This Happens in Real 

- **Physiological Constraints**  

  Motor-imagery EEG signals exhibit high inter-subject variability and low spatial 

- **Sensor Placement Bias**  

  Frontal electrodes prioritize facial/muscle artifacts over motor cortex 

- **Protocol Flaws**  

  Fixed-duration sessions ignore that complex commands require longer differentiation 

- **Preprocessing Blind Spots**  

  Frequency-domain conversion (FFT) loses temporal dependencies critical for motion 



## Real-World 

If deployed, the wheelchair would exhibit:

- Erratic navigation behaviors in 52% of 

- Catastrophic failure modes:

  - Ignoring "stop" commands → collision 

  - Right-turn misinterpretation → navigational 

- User trust erosion after repeated command 



## Example or 

Problematic feature aggregation approach:

% Typical feature extraction pipeline (simplified)

features = [];

for i = 1:

data = readtable(sprintf(‘command%d/file%d.csv’, command_id, i));

signal = data.Channel1;

% Uses only aggregated

features(i,:) = [mean(signal), std(signal), kurtosis(signal), …];

% Result: Directionally similar commands yield near-identical

Proper spatial feature enhancement:

Baseline correction + topographic mapping (hypothetical improvement)

from scipy.signal import

import numpy as

def enhance_features(signal):

corrected = detrend(signal) # Remove DC

Emulate multi-electrode spatial correlation (requires additional sensors)

spatial_corr = np.correlate(corrected, simulated_motor_cortex_template)

return [np.max(spatial_corr), np.argmax(spatial_corr), …]

## How Senior Engineers Fix 

1. **Sensor Expansion**  

   Add electrodes at C3/C4 (motor cortex) and validate with topographical 



2. **Temporal Augmentation**  

   Record directional commands at 2x duration of neutral 



3. **Artifact Subtraction**  

   Implement EMG/EOG rejection algorithms during 



4. **Feature Engineering**  

   Introduce time-lagged features (e.g. cross-correlation peaks)



5. **Class Reweighting**  

   Apply cost-sensitive learning with 5x penalty on "right" 



6. **Domain Validation**  

   Generate ERP plots per command to verify neural differentiation 



## Why Juniors Miss 

- **Data Assumption Gap**  

  Assumes uniform command difficulty → neglects neurological 



- **Toolchain Dependency**  

  Over-relies on MATLAB's automated workflows (Classification Learner App)



- **Feature Myopia**  

  Focuses on textbook statistical metrics (mean, kurtosis) ignoring EEG temporal 



- **Signal Naivety**  

  Treats all EEG channels as equally informative → misses regional 



- **Validation Oversimplification**  

  Prioritizes aggregate accuracy over per-class