Related Experiment Video
Updated: Jun 9, 2025
08:31
Conscious and Non-conscious Representations of Emotional Faces in Asperger's Syndrome
Published on: July 31, 2016
13.1K
Model selection to achieve reproducible associations between resting state EEG features and autism
William E Carson1, Samantha Major2,3, Harshitha Akkineni2,3
1Department of Biomedical Engineering, Duke University, Durham, NC, 27708, USA.
Scientific Reports
|October 25, 2024
Summary
Reproducible autism biomarkers were identified using electroencephalography (EEG) and machine learning. Prioritizing reproducibility over prediction accuracy revealed distinct neural patterns associated with autism in children.
Area of Science:
- Neuroscience
- Biomarker Discovery
- Computational Psychiatry
Background:
- Autism biomarker discovery using electroencephalography (EEG) faces challenges with reproducibility.
- Resting-state (RS) neural activity patterns may hold potential biomarkers for autism.
Purpose of the Study:
- To develop a method for learning reproducible associations between RS-EEG features and autism.
- To compare the impact of prioritizing predictive performance versus reproducibility in model selection.
Main Methods:
- Utilized EEG spectral power and functional connectivity features from autistic and neurotypical children (ages 36-96 months).
- Employed a regularized generalized linear model to predict diagnostic group.
- Developed an evaluation procedure to quantify predictive generalization and reproducibility.
Main Results:
- A reproducible profile of associations emerged when prioritizing both performance and reproducibility.
- This profile indicated increased gamma power and connectivity in occipital/posterior midline regions in autistic children.
- Model selection based solely on predictive performance yielded non-robust associations.
Conclusions:
- Reproducibility is crucial for identifying reliable autism biomarkers from EEG data.
- Prioritizing reproducibility enhances the scientific utility of machine learning models in autism research.
- A custom machine learning model further improved the robustness of learned associations.
