07 — People
Team
Supervisors
EEG-based AI detection of Epilepsy, Alzheimer's & Parkinson's Disease using TEECNet-inspired Error Correction Architecture
Neurological disorders such as Epilepsy, Alzheimer's Disease (AD), and Parkinson's Disease (PD) affect hundreds of millions globally, yet their diagnosis relies heavily on specialist-dependent, resource-intensive clinical assessment. Electroencephalography (EEG) offers a non-invasive, affordable window into brain activity—but automated analysis remains computationally expensive and clinically inaccessible.
This project proposes a unified family of lightweight EEG-based machine learning models for the detection of all three neurological disorders. Drawing inspiration from the TEECNet architecture originally designed for physics simulations, we introduce a two-stage training paradigm: a lightweight Base Network learns dominant EEG patterns, followed by a small Error-Correction Network (ECN) that refines predictions by targeting the systematic errors the base model makes.
Across all three disorders, the ECN consistently improves accuracy and reduces misclassifications—demonstrating that error-aware residual correction is a powerful, generalizable strategy for EEG classification with constrained model budgets.
Final output:
Y = Base + ECN Correction
A lightweight EEGNet-inspired encoder extracts primary temporal–spatial features from the raw multi-channel EEG input. Key components:
A small gated corrector network, TEECNet-inspired, that refines the base prediction:
Down-weights easy, correctly classified examples so the model focuses training capacity on hard borderline cases—critical for highly imbalanced EEG datasets.
Adjusts loss contribution by inverse class frequency, compensating for the natural scarcity of seizure windows relative to interictal EEG.
ECN minimises mean-squared error on the prediction residuals from the frozen base—directly optimising the correction term rather than the full classification loss.
| Component | Parameters | Role |
|---|---|---|
| Base EEGNet | 94,333 | Primary classifier |
| FCN / ECN | 132,804 | Error corrector |
| Total | 227,137 | Full pipeline |
Base: EEGNet — learns dominant temporal–spatial EEG patterns, outputs class logits (seizure / non-seizure).
ECN: Lightweight MLP / small CNN operating on logits + uncertainty estimates. Outputs correction or refined seizure probability.
A Modified EEGNet-inspired base with Spectral Band Power branch fused at the embedding level. On top, a Feature Correction Network (FCN) refines predictions through:
Depthwise-separable CNN, lowest parameter count, focal loss with PD class boost.
Separates temporal rhythm and spatial channel learning. Noise + masking + channel dropout augmentation.
CNN spatial front-end with dilated causal TCN backbone for long-range EEG dependencies. Mixup augmentation.
This section consolidates the strongest findings across Epilepsy, Alzheimer's disease, and Parkinson's disease, then drills into task-specific metrics, cross-dataset behaviour, and model-to-model comparisons.
Each task is evaluated with the metric profile most meaningful to its setting, while the shared question remains the same: can a compact ECN improve lightweight EEG classifiers consistently across distinct neurological conditions?
Strong seizure discrimination with fewer false alarms under cross-dataset evaluation.
EEGNet-FCN achieves the strongest overall classification performance among the lightweight baselines shown.
ECN refinement improves all three lightweight Parkinson's models, with the best final performance from CNN-TCN.
| Model | Accuracy | F1 | AUC |
|---|---|---|---|
| HeavyTeacher-ResNet | 0.97 | 0.90 | 0.96 |
| LightBase-EEGNet | 0.79 | 0.34 | 0.62 |
| EEGNet + ECN (Distilled) | 0.86 | 0.56 | 0.72 |
| Model | Accuracy | Bal. Acc. | Macro F1 |
|---|---|---|---|
| EEGNet-FCN (Ours) | 0.9326 | 0.9312 | 0.9321 |
| DSCNN | 0.8448 | 0.8491 | 0.8448 |
| TinyResNet | 0.8302 | 0.8491 | 0.8296 |
| TCNLite | 0.8506 | 0.8447 | 0.8474 |
| MobileNet1D | 0.8245 | 0.8182 | 0.8212 |
| ShuffleNet1D | 0.8009 | 0.8063 | 0.8189 |
| SqueezeNet1D | 0.7830 | 0.7926 | 0.8025 |
Training measured on NVIDIA RTX 6000 Ada Generation GPU with CUDA AMP. EEGNet-FCN trains in 581s.
| Model | Base Acc | Base F1 | +ECN Acc | +ECN F1 | Acc Gain | Error Reduction |
|---|---|---|---|---|---|---|
| Modified EEGNet | 88.32% | 90.65% | 95.35% | 96.35% | +7.03% | 265 fewer |
| LW Temporal-Spatial CNN | 89.91% | 91.97% | 95.89% | 96.77% | +5.97% | 225 fewer |
| CNN-TCN | 90.71% | 92.61% | 96.42% | 97.19% | +5.71% | 215 fewer |
This project demonstrates that a TEECNet-inspired Error Correction Network (ECN) is a powerful, general-purpose enhancement for lightweight EEG classifiers. By training a small corrector network exclusively on the systematic prediction residuals of a frozen base model, we achieve consistent, meaningful accuracy gains across three neurologically distinct conditions—without increasing the base model's inference-time parameter count.
The results validate a key hypothesis: lightweight models contain systematic, learnable blind spots, and a targeted corrector is a more efficient route to accuracy recovery than simply making the base model larger.
For Alzheimer's detection, our EEGNet-FCN achieves 93.26% accuracy and 93.21% Macro F1—outperforming all evaluated lightweight baselines (DSCNN, TinyResNet, TCNLite, MobileNet1D, ShuffleNet1D, SqueezeNet1D) by a significant margin, while remaining within practical parameter budgets.
Consistent improvement across Epilepsy, Alzheimer's, and Parkinson's with the same base+corrector design philosophy.
Total parameter count under 230K supports real-time inference on wearable and edge devices.
The base never degrades from ECN application; corrections are bounded and stabilised by clipping.