Do EEG Foundation Models Actually Beat Conventional Deep Learning?
October 30, 2026 | Research Notes & Vlogs
In the rapidly evolving world of NeuroAI, pretrained foundation models have become the default talk of the town. Following massive successes in natural language processing and computer vision, architectures like LaBraM and REVE promise to revolutionize biomedical time-series analysis, particularly for electroencephalography (EEG).
The core hypothesis sounds bulletproof: use massive self-supervised pretraining on diverse multi-terabyte EEG corpuses, fine-tune downstream, and watch performance skyrocket.
But do they actually outperform tried-and-true conventional architectures when evaluated on rigorous, blind-split clinical benchmarks?
To find out, I put state-of-the-art EEG foundation models head-to-head against classical models (Vanilla RNNs and CNN-BiLSTMs) using the Temple University Hospital EEG Seizure Corpus (TUSZ), featuring strict official subject-independent (leave-one-subject-out) validation and test splits to prevent data leakage.
The code is available at https://github.com/asifprotick10/EffectivityOfEEGModels.
The Experimental Setup
To maintain strict parity, all models were evaluated on the same preprocessed TUSZ subsets:
- Sampling Rate: 200 Hz
- Filtering: Bandpass filtered (BPF) from 0.5 to 99.5 Hz, with 60 Hz notch filtering (NF).
- Segmentation: 5-second non-overlapping windows.
- Class Imbalance Mitigation: 20% of background (non-seizure) data retained in training to maintain clinical realism without crashing models into majority-class triviality.
Test Performance At-a-Glance
| Model |
Training Strategy |
Test Accuracy (%) |
Test Macro-F1 (%) |
| Vanilla RNN |
Scratch |
88.79 |
66.26 |
| CNN-BiLSTM |
Scratch |
93.12 |
72.84 |
| LaBraM |
Fine-Tuning (FT) |
94.56 |
80.39 |
| REVE |
Linear Probing (LP) |
84.50 |
61.90 |
| REVE |
Fine-Tuning (FT) |
95.60 |
77.35 |
Key Insights & Takeaways
1. Foundation Models Aren't "Plug-and-Play" Magic
While REVE (FT) demonstrated strong generalization on test sets (achieving a solid 77.35% Macro-F1), LaBraM (FT) exhibited aggressive overfitting early in training. Pretrained models require careful regularization, tailored learning rate warmups, and layer-wise decay.
2. Fine-Tuning Beats Linear Probing Hands Down
For REVE, shifting from linear probing to full fine-tuning yielded a massive jump in test Macro-F1 (from 61.90% to 77.35%). Biomedical signals possess unique spectral-temporal signatures that rigid pretrained feature extractors fail to capture without adaptation.
3. Classical Architectures Still Punch Above Their Weight
The CNN-BiLSTM model proved remarkably resilient. By pairing convolutional feature extractors with bidirectional LSTMs, it bypassed the vanishing gradient issues plaguing the Vanilla RNN and handled full 5-second windows natively.