Machine Learning in EEG

Machine learning entered clinical EEG to address a genuine bottleneck: continuous recordings now generate days of multichannel data per patient, far more than any expert can review minute by minute. Algorithms promise to flag seizures, triage normal from abnormal, stage sleep, and remove artifact automatically. Some of these promises are partly fulfilled and embedded in regulator-cleared products; many remain oversold. The purpose of this module is to give you the working vocabulary to read a machine-learning EEG paper or product claim critically - to understand what the model does, how the field has shifted toward large self-supervised models in 2026, and, far more importantly, to understand the specific and recurring ways such models mislead. In clinical EEG the failure modes are not an appendix to the methods; they are the heart of the matter, because a model that is wrong in a plausible, confident way is more dangerous than no model at all.

Seizure detection: the canonical pipeline

Automated seizure detection is the oldest and most mature application, and it remains the clearest illustration of the classical approach. The pipeline has four stages. First, the multichannel signal is segmented into short overlapping epochs, typically one to a few seconds. Second, features are extracted from each epoch - spectral band powers, line length (the cumulative absolute change in the signal, which rises sharply during rhythmic spiking), signal energy, entropy measures, and the rate of sharp transients. Third, a classifier - historically a support vector machine, a random forest, or gradient-boosted trees - maps the feature vector to a seizure-or-not decision. Fourth, a temporal smoothing or post-processing stage suppresses isolated false alarms by requiring evidence to persist over time and to involve a plausible spatial pattern across electrodes. The morphology these detectors hunt for - rhythmic, evolving, spatially organized discharge - is exactly what defines an electrographic seizure to the human reader, which is why hand-crafted features built around rhythmicity and amplitude work as well as they do.

It is worth grounding this in reality rather than hype. Commercial seizure-detection software exists and is in clinical use; the field is small and dominated by a few validated systems. A widely used detector built on a deep neural network trained on large volumes of expertly annotated EEG was reported, in a peer-reviewed study, to perform comparably to expert human reviewers on adult continuous inpatient EEG - the first time an algorithm reached that bar. Such systems are positioned as review aids that mark candidate segments and compute seizure probability for a human to confirm, not as autonomous diagnosticians, and their performance is characterized by sensitivity and false alarms per hour rather than by a single accuracy figure. This is the responsible deployment pattern: the algorithm triages the data flood so the expert spends attention where it is most likely to matter, and the expert remains accountable for the diagnosis.

The modern alternative is end-to-end deep learning, in which a neural network - a convolutional network, a recurrent network, or increasingly a transformer - learns its own features directly from the raw or minimally processed multichannel signal. Convolutional networks excel at finding local spatial-temporal patterns; recurrent networks model how the signal evolves over time; transformers, through their self-attention mechanism, relate every part of a sequence to every other part and have become the dominant architecture for sequence data across machine learning, EEG included. Deep models can capture spatial-temporal structure that hand-designed features miss and have produced strong benchmark results for seizure detection and sleep staging. Their cost is opacity, a large appetite for labeled data, and - critically - a tendency to learn whatever spurious cue correlates with the label in the training set, including artifacts, electrode types, and site-specific quirks rather than neurophysiology. Whether to engineer features or to learn them is a genuine trade-off: feature-based models are interpretable, data-efficient, and easy to audit; deep models are more flexible but more opaque and more prone to silently exploiting confounds, a hazard the validation section returns to in detail.

Foundation models, transformers, and self-supervised learning in 2026

The most significant shift in EEG machine learning over the past few years mirrors the broader revolution in artificial intelligence: the rise of self-supervised foundation models. The central problem these models confront is that labeled clinical EEG is scarce and expensive - expert annotation is slow and inconsistent - while unlabeled EEG exists in enormous quantity. Self-supervised learning exploits the unlabeled abundance by inventing a pretext task that requires no human labels. Two strategies dominate. In masked prediction, portions of the signal (or of a tokenized representation of it) are hidden and the model is trained to reconstruct them, forcing it to learn the structure of normal EEG. In contrastive learning, the model is trained to recognize that two augmented views of the same segment belong together while different segments do not, learning representations invariant to nuisance variation. A model pretrained this way on thousands of hours of heterogeneous EEG learns a general-purpose representation that can then be fine-tuned on a small labeled dataset for a specific task - seizure detection, sleep staging, abnormality screening - often with far less labeled data than training from scratch would require.

These EEG foundation models - a research-stage family that includes architectures pretrained with masked or contrastive objectives, frequently using a transformer backbone that operates on patches of signal converted into discrete tokens, sometimes by a learned spectral tokenizer - have shown meaningful gains in the low-label regime on public benchmarks and now routinely win community challenges. They are, however, firmly investigational for clinical use, and the field is candid about this in its own critical reviews. Several limitations are well documented as of 2026. Generalization across datasets, recording montages, and clinical sites remains the central unsolved problem rather than a solved one; truly general-purpose zero-shot models that work out of the box on a new task or site have not been demonstrated, and most reported successes still involve task-specific fine-tuning. Benchmarks are not yet standardized - differences in preprocessing, data splits, and metrics mean published results frequently cannot be compared head to head - so the literature is harder to appraise than its confident numbers suggest. And the models are large and opaque, inheriting every interpretability and confound problem of smaller deep models while adding the difficulty of auditing what an enormous pretrained representation has actually learned. The honest summary is that foundation models are a genuinely promising direction that has improved label efficiency and benchmark performance, not a clinically validated breakthrough, and any claim that one is ready for autonomous clinical decisions should be treated as unproven.

Two adjacent developments deserve mention because they target real deployment obstacles. Federated and privacy-preserving learning trains a shared model across multiple hospitals without the raw EEG ever leaving each institution: only model updates are exchanged and aggregated. This is attractive for EEG because patient recordings are sensitive and legally protected, and because no single center holds enough diverse data to train a robust model, so pooling knowledge without pooling data could in principle improve generalization while respecting privacy. It remains an active research area with its own challenges - heterogeneous data across sites, communication cost, and the fact that model updates can themselves leak information unless additional safeguards are applied. Separately, parameter-efficient adaptation methods, such as low-rank adapters, allow a large pretrained model to be specialized to a new site or task by training only a small number of additional parameters, which makes per-site calibration practical without retraining the whole model. Both ideas address the deployment gap directly, and both are investigational rather than standard clinical practice today.

Artifact removal with algorithms

A large share of machine learning in EEG targets artifact removal rather than diagnosis. The dominant classical tool is Independent Component Analysis (ICA), which decomposes the multichannel recording into maximally statistically independent components. Because eye blinks, lateral eye movements, the electrocardiogram, and many muscle artifacts have stereotyped spatial topographies and statistical signatures distinct from brain rhythms, they often separate into identifiable components that can be inspected and removed before the signal is reconstructed without them. ICA is now routine in research preprocessing, and automated component classifiers can label the artifactual components without manual review, speeding the workflow. Its limits are real and must be respected: ICA assumes the sources are stationary and linearly mixed and that there are at least as many channels as sources, so it performs poorly on low-density montages and on artifacts that overlap brain activity in both space and statistics. Removing a component that contains both artifact and genuine brain signal silently discards real data, so even this mature method requires judgment rather than blind application.

Newer deep denoising methods train neural networks - autoencoders and related architectures - to map a contaminated signal toward a clean target. They can suppress artifacts that ICA struggles with, can operate on few channels, and can run in real time. But they carry a distinctive and underappreciated danger: a denoiser trained to make EEG look clean can hallucinate plausible-looking brain activity, smoothing away or inventing waveforms in ways that change clinical meaning. A network that has learned what normal EEG usually looks like may quietly normalize a genuine abnormality, because reconstructing toward the training distribution is exactly what it was optimized to do. For any denoiser used clinically, the burden of proof is that it preserves diagnostic features - spikes, focal slowing, electrodecrement - and does not fabricate or erase them, and that burden is met by adversarial testing on signals containing real pathology, not by demonstrations that the output looks tidy.

Why reported performance is usually optimistic

The single most important thing a clinician can know about machine learning in EEG is that headline accuracy figures are systematically inflated, and the reasons are specific, recurring, and largely independent of how fashionable the architecture is. The first and most pervasive is data leakage from improper splitting. EEG epochs from one patient are highly correlated with one another. If epochs from the same patient land in both the training and test sets - a common error when data are split by epoch rather than by patient - the model can effectively memorize the patient and report near-perfect accuracy that collapses on truly new patients. Valid evaluation requires subject-wise (patient-independent) splits; any study that does not state how it split, or that splits by epoch, should be read with deep suspicion, and this single error probably accounts for more overstated EEG results than any other.

The second is dataset shift. A model trained at one hospital learns the idiosyncrasies of its amplifiers, electrode types, montages, filter settings, sedation practices, and patient mix. Deployed at another site with different equipment and populations, performance frequently drops sharply because the input distribution has changed - the model was never tested on the world it now inhabits. This is why external validation on independent datasets, not merely a held-out slice of the same dataset, is the real test of a clinical algorithm, and why comparatively few published models survive it. Foundation models were supposed to ease this problem by learning more general representations, and they help in the low-label regime, but as noted above robust cross-site generalization remains unproven even for them. The third is label noise. Expert annotators disagree substantially about seizure onset times, spike identification, and the boundaries of abnormality; inter-rater agreement for many EEG tasks is only moderate. A model trained on one expert's labels inherits that subjectivity, and accuracy measured against the same noisy labels can be close to meaningless - the model may simply have learned to imitate one rater rather than to detect any underlying truth, and a higher number against noisy labels is not necessarily a better model.

Compounding all of this is class imbalance. Seizures occupy a tiny fraction of any continuous recording, so a model that simply predicts no seizure can be over 99 percent accurate while detecting nothing whatsoever. Accuracy is therefore the wrong metric; the meaningful figures are sensitivity, the false-alarm rate per hour (a detector that fires constantly is useless even if it never misses, because alarm fatigue will lead clinicians to ignore it), and, for detection of events in time, the latency to detection. A related and often neglected issue is calibration: a model may rank cases correctly yet output probabilities that do not correspond to real-world frequencies, so that a reported ninety percent confidence does not mean the event occurs nine times in ten. Calibration matters acutely when a probability is used to triage human review, and it commonly degrades under dataset shift, so it must be checked at each deployment site rather than assumed from the original study.

Finally, the rise of opaque deep and foundation models has made explainability a clinical requirement, but the explainability tools themselves must be treated with the same skepticism as the models. Saliency maps, attention visualizations, and attribution methods such as class-activation mapping are offered to show which parts of the signal drove a decision, yet different methods can produce conflicting explanations for the same prediction, and some popular attribution techniques have failed basic sanity checks - performing no better than random when subjected to perturbation and deletion tests. An explanation that highlights a plausible-looking region is not evidence that the model used that region for the right reason; it may equally have latched onto an artifact or a site-specific confound that happens to co-occur. Explainability is necessary for trust and for catching confounds, but a saliency map is a hypothesis about the model's reasoning, not a proof of it, and the explanation method should itself be validated before its output is believed.