Related Experiment Video
Updated: Jan 17, 2026
08:51
Statistical Modelling of Cortical Connectivity Using Non-invasive Electroencephalograms
Published on: November 1, 2019
6.0K
Quantity versus diversity: Influence of data on detecting EEG pathology with advanced ML models
Martyna Poziomska1, Marian Dovgialo2, Przemysław Olbratowski3
1Faculty of Physics, University of Warsaw, Pasteura 5, Warsaw, 02-093, Poland; Łukasiewicz Research Network - Automotive Industry Institute, Jagiellońska 55, Warsaw, 03-301, Poland.
Summary
Large, diverse EEG datasets improve machine learning model performance for detecting brain pathology. More data can overcome diversity issues, especially with advanced neural networks and meta-models.
Area of Science:
- Neuroscience
- Artificial Intelligence
- Biomedical Engineering
Background:
- Machine learning models are increasingly used for analyzing electroencephalogram (EEG) data.
- The performance of these models is highly dependent on the quantity and diversity of the training data.
- Understanding these dependencies is crucial for developing robust and reliable EEG analysis tools.
Purpose of the Study:
- To investigate how data quantity and diversity affect the performance of machine learning models in detecting general EEG pathology.
- To introduce and characterize the Elmiko dataset, a large and diverse corpus of EEG recordings.
- To identify optimal strategies for training machine learning models on EEG data.
Main Methods:
- Utilized two EEG datasets: one with 2993 recordings (Temple University Hospital) and the Elmiko dataset with 55,787 recordings from 39 hospitals.
- Evaluated various machine learning models, including attention-based and transformer architectures, and a meta-model combining neural networks with gradient-boosting.
- Assessed model performance based on accuracy and robustness across datasets with varying characteristics.
Main Results:
- Small, consistent datasets allow many models to achieve high accuracy.
- Data variations (pathology, protocols, labeling) significantly degrade model performance.
- Increasing data quantity improves predictive accuracy, potentially compensating for diversity, especially in transformer-based networks.
- A meta-model integrating neural networks and gradient-boosting with handcrafted features showed superior performance.
Conclusions:
- Data quantity and diversity are critical factors in machine learning for EEG pathology detection.
- Larger, more diverse datasets, like the Elmiko corpus, are essential for generalizable model performance.
- Advanced architectures (transformers) and ensemble methods (meta-models) show promise in overcoming data limitations.
