Cognitive AI · 3 min read
AI Models of Cognitive States from EEG: What Works, What Does Not
Machine learning on EEG has been used for years to classify brain states, and deep learning has extended it [1] [2]. But results depend on how models are evaluated, on the amount of data, and on differences between people [3]; a correlation found by a model is not a cause.
What a cognitive state model does
A cognitive state model takes features of brain activity, and often of heart and behaviour, and estimates a state: high or low load, attention holding or slipping, fatigue building. It is trained on recordings where the state is known from the task or from other measures.
Brain-computer interfaces have long used such classifiers; a review of classification algorithms for EEG-based interfaces compares the main families and their conditions of use [3].
What deep learning brings
Reviews of deep learning on EEG describe a fast-growing field, applied to tasks such as mental workload, emotion, sleep or motor imagery [1] [2]. Deep models can learn features directly from the signal rather than relying only on hand-crafted ones.
They also need more data, and their results are harder to compare from one study to another. A systematic review notes, among other points, that many studies do not share their data or code, which limits reproducibility [2].
What does not work, or not yet
- Small datasets. With little training data per person, complex models do not automatically beat simpler methods [3].
- Differences between people. A model trained on some people can perform worse on others; validation must use people the model has never seen.
- Leakage. If data from the same session are used for training and testing, scores look better than they are.
- Causes. A model that links a brain pattern to a drop in performance shows a relation, not a cause.
Good practice
- One model per question: states, load, fatigue, performance, rather than one general model for everything.
- Validation on independent sessions and people, reported as such.
- Measured signals kept separate from inferred indicators, so every result stays traceable to the raw recording.
- Consent and pseudonymisation for every recording used to train models.
How SPARK approaches it
SPARK develops a family of specialised models, trained on synchronised cognitive data and validated on people they have never seen, with the raw recording kept alongside every result.
Questions
Can AI read thoughts from EEG?
No. Models estimate broad states, such as load or fatigue, from features of brain activity, within the limits of their validation.
Is deep learning always better on EEG?
Not automatically. With small datasets, simpler methods can do as well; results depend on the data and on how models are evaluated.
Why validate on people the model has never seen?
Because brain signals differ between people. Testing on new people shows whether a model generalises, rather than memorising its training data.
Sources
Numbers in brackets in the text refer to these sources.
- Craik, He, Contreras-Vidal (2019). Deep learning for electroencephalogram (EEG) classification tasks: a review. Journal of Neural Engineering.doi.org
- Roy, Banville, Albuquerque et al. (2019). Deep learning-based electroencephalography analysis: a systematic review. Journal of Neural Engineering.doi.org
- Lotte, Bougrain, Cichocki et al. (2018). A review of classification algorithms for EEG-based brain-computer interfaces: a 10 year update. Journal of Neural Engineering.doi.org