Machine Learning for Neural Signal Classification

Why EEG Needs Specialized Architectures

If you take a ResNet-50 or a standard transformer and throw EEG data at it, the results will be mediocre at best. The reasons are fundamental, not just inconvenient.

The biggest constraint is data volume. A typical EEG experiment produces 100 to 500 usable trials per subject. Even large multi-site datasets rarely exceed a few thousand subjects. Compare that to ImageNet's 14 million images or the billions of tokens used in language modeling. A model with millions of parameters will memorize your training set before it learns anything generalizable.

Subject variability compounds the problem. Brain anatomy differs across individuals. Skull thickness, cortical folding patterns, and electrode-scalp contact impedance all shape the signal before it reaches your amplifier. A model trained on one group of subjects will see a measurable drop in accuracy on new subjects, sometimes 10-15 percentage points, because the spatial distribution of features shifts substantially.

Normalization Matters More Than You Think

Batch normalization assumes that statistics are stable across samples within a mini-batch. EEG violates this assumption. Signal amplitude can drift within a single recording session due to electrode gel drying, muscle tension changes, or alertness fluctuations. Instance normalization, which computes statistics per sample, handles this nonstationarity far better. Several recent architectures have confirmed this empirically: switching from BatchNorm to InstanceNorm often improves cross-session accuracy by 3-5% without any other changes.

EEGNet and Compact CNN Architectures

EEGNet, introduced by Lawhern et al. in 2018, remains one of the most influential architectures in the field. Its design is elegant because each layer has a clear neurophysiological interpretation.

The first layer is a set of temporal convolutions applied independently to each channel. These learn frequency-selective filters. When you inspect the learned weights, they closely resemble classical bandpass filters tuned to delta, theta, alpha, and beta bands. The network is literally rediscovering neuroscience from raw data.

The second layer applies depthwise convolutions across channels at each time point. This learns spatial filters, effectively determining which channel combinations are informative. It is the learned equivalent of a Common Spatial Pattern (CSP) filter, but end-to-end differentiable.

The third layer uses separable convolutions to combine the temporal and spatial features into a unified representation. By factoring the convolution into depthwise and pointwise components, the parameter count stays extremely low.

The compactness is the point. With so few parameters, EEGNet generalizes well even on small datasets and runs comfortably in real-time on embedded hardware.

Successors and Variants

Since EEGNet, several architectures have extended its ideas:

• EEGNeX: multi-scale temporal convolutions for richer frequency resolution
• TSception: multi-scale temporal and spatial convolutions with inception-style branching
• FBCNet: filter-bank approach that explicitly separates frequency bands before spatial filtering
• CSP-Net: embeds classical CSP computation as differentiable layers within a CNN

Temporal Models: LSTM and Transformers

LSTM Networks

Long Short-Term Memory networks address a specific problem: learning dependencies across long time sequences without gradient degradation. The architecture uses three gating mechanisms, forget gate, input gate, and output gate, that selectively control what information persists, what gets added, and what gets exposed at each time step.

For EEG, this matters because clinically relevant patterns can unfold over hundreds of time steps. A seizure onset might build over 2-3 seconds. Pre-ictal slowing might develop over minutes. LSTMs can track these evolving states without the training instabilities that plague vanilla recurrent networks.

In practice, CNN-LSTM hybrids tend to outperform either architecture alone. The CNN front-end extracts spatial features from the multi-channel EEG montage, compressing the channel dimension into a learned representation. The LSTM back-end then models how those spatial patterns evolve over time. This division of labor plays to each architecture's strengths.

Transformers for EEG

The self-attention mechanism computes pairwise relationships between all time points and channels simultaneously. Unlike LSTMs, which process sequentially and can struggle with very long-range dependencies despite their gating, transformers have direct access to every position in the sequence.

EEG-Conformer combines convolutional layers with self-attention and achieves above 90% accuracy on standard motor imagery benchmarks. What makes this interesting is not just the accuracy, it is the attention maps. When you visualize where the model attends, the patterns align with known neurophysiology: motor imagery attention concentrates over sensorimotor cortex (C3/C4 electrodes), visual tasks highlight occipital channels.

The most compelling advantage of transformers for EEG is in transfer learning. A transformer pre-trained on a large corpus of EEG recordings builds general representations of neural dynamics. Fine-tuned on just 50 sessions of task-specific data, it can outperform a compact CNN trained from scratch on 500 sessions. The pre-trained representations capture cross-subject structure that smaller models cannot learn from limited data.

Transfer Learning & Subject Adaptation

Individual differences are the central obstacle in EEG-based BCI and clinical diagnostics. A classifier that works well for the subjects in your training set will degrade on new individuals. This is not a bug; it reflects genuine physiological variability in how brains produce and propagate electrical signals.

Domain Adaptation

Domain adaptation techniques align the feature distributions of source subjects (training data) with a target subject (new user). Methods like Maximum Mean Discrepancy or adversarial domain discriminators learn representations that are informative for the classification task but invariant to subject identity. The model keeps what is diagnostically relevant and discards what is idiosyncratic to individual anatomy.

Subject-Adaptive Fine-Tuning

A simpler and often equally effective approach: collect 2-5 minutes of labeled calibration data from the new subject, freeze the early layers of a pre-trained model, and fine-tune only the final classification layers. The early layers have already learned general EEG feature extraction. The final layers adapt to the specific signal characteristics of this individual. This works because the early representations (frequency filters, spatial patterns) are more universal than the decision boundaries.

Large-Scale Pre-Training

Datasets like the Temple University EEG Corpus, with over 10,000 clinical recordings, enable meaningful pre-training. Self-supervised objectives, predicting masked time segments, contrastive learning across augmentations, learn representations without requiring task labels. These pre-trained encoders serve as foundation models that can be fine-tuned for diverse downstream tasks: seizure detection, sleep staging, cognitive state classification.

Edge Deployment & Embedded Inference

Running a neural network on a cloud server is straightforward. Running one on a microcontroller with 1MB of flash and 256KB of SRAM is a different engineering problem entirely. For wearable EEG devices and implantable neurostimulators, on-device inference is not optional; latency, power consumption, and connectivity constraints make cloud offloading impractical.

Quantization

8-bit post-training quantization converts 32-bit floating-point weights to 8-bit integers. This provides 3-4x reduction in model size with minimal accuracy loss (typically less than 1% degradation). For EEGNet-sized models, this brings storage requirements well under 50KB.

Pruning

Structured pruning removes entire filters or attention heads that contribute least to output accuracy. Combined with quantization, pruning achieves the best accuracy-to-size ratio. The order matters: prune first, fine-tune to recover accuracy, then quantize.

Deployment Results

Real-time seizure detection has been demonstrated on ARM Cortex-M4 and Cortex-M7 platforms using quantized EEGNet variants. The latency budget for seizure detection is typically on the order of seconds, so even the worst-case inference time leaves comfortable margin for preprocessing and post-processing logic.

Explainability & Regulatory Compliance

In consumer applications, a model's internal reasoning is a nice-to-have. In medicine, it is a clinical necessity. A neurologist will not trust a seizure detector that says "positive" without showing why. And a regulator will not approve it.

Explainability Methods

SHAP (SHapley Additive exPlanations) provides both local and global explanations. Applied to EEG classifiers, SHAP consistently identifies spectral features in the delta (0.5-4 Hz) and theta (4-8 Hz) bands as primary drivers for pathological classifications, which aligns with decades of clinical EEG knowledge.

Grad-CAM highlights which spatial and temporal regions of the input most influence the CNN's decision. For motor imagery classification, Grad-CAM activation maps localize to contralateral sensorimotor regions, providing visual confirmation that the model uses physiologically plausible features.

Transformer-based models offer a built-in advantage: the attention weights themselves serve as interpretable visualizations. You can directly inspect which time points and channels the model considers most informative for a given prediction.

FDA Regulatory Landscape

As of late 2025, the FDA has authorized 1,451 AI/ML-enabled medical devices. Neurology is the second most common specialty category. The regulatory framework is maturing:

• Good Machine Learning Practice (GMLP): 10 guiding principles developed jointly by the FDA, Health Canada, and the UK MHRA covering data management, training practices, and performance monitoring
• Predetermined Change Control Plans (PCCP): allow manufacturers to define pre-approved model update strategies, enabling continuous learning without requiring a new submission for each retrained model

EU Regulatory Framework

The dual regulatory burden of MDR and the AI Act means that EEG classification software deployed in Europe faces some of the strictest oversight globally. Documentation requirements include algorithmic transparency, bias testing across demographic groups, and continuous post-market performance monitoring.