04.03.2026

New Publication on Data Requirements for AI-Based EEG Analysis

New Publication on Data Requirements for AI-Based EEG Analysis

Marc Steffen Seibel
Learning curves and extrapolation. Left: Observed model accuracy as a function of the number of training subjects. Right: Mathematical extrapolation of learning curves beyond the observed data range. The results illustrate that adding new participants continues to improve performance, although the gains gradually become smaller. Bottom: Concept of learning curve estimation and extrapolation from training data subsets.

How much EEG data does artificial intelligence actually need to learn reliably?

This seemingly simple question is crucial in practice: Is it worth recruiting more participants? Or is it sufficient to record longer sessions from the same individuals?

In a new study, we systematically investigated how the accuracy of deep learning models for EEG classification scales with increasing dataset size. We analyzed several classification tasks (e.g., distinguishing normal from abnormal EEG or eyes-open from eyes-closed states) and compared different neural network architectures.

Key findings include:

  • Adding more participants improves performance much more than extending recordings from the same individuals.
  • Accuracy increases with more data, but with diminishing returns.
  • Different neural network architectures showed remarkably similar learning behavior — dataset size was the dominant factor.
  • Using mathematical learning curve models allows us to estimate in advance whether collecting additional data is likely to be beneficial.

These results provide practical guidance for planning future EEG studies. Rather than focusing solely on increasingly complex AI models, investing in larger and more diverse cohorts may often be the more effective strategy.

The study was published in the Journal of Neural Engineering:

Marc S. Seibel, Jens Haueisen, Thomas Jochmann

How much EEG is needed for deep learning with convolutional neural networks? Predicting the benefit from additional data

Journal of Neural Engineering (2026)

doi.org/10.1088/1741-2552/ae453d

 

Contact: Thomas Jochmann