Abstract

A concern in the field of autism electroencephalography (EEG) biomarker discovery is their lack of reproducibility. In the present study, we considered the problem of learning reproducible associations between multiple features of resting state (RS) neural activity and autism, using EEG data collected during a RS paradigm from 36 to 96 month-old children diagnosed with autism (N = 224) and neurotypical children (N = 69). Specifically, EEG spectral power and functional connectivity features were used as inputs to a regularized generalized linear model trained to predict diagnostic group (autism versus neurotypical). To evaluate our model, we proposed a procedure that quantified both the predictive generalization and reproducibility of learned associations produced by the model. When prioritizing both model predictive performance and reproducibility of associations, a highly reproducible profile of associations emerged. This profile revealed a distinct pattern of increased gamma power and connectivity in occipital and posterior midline regions associated with an autism diagnosis. Conversely, model selection based on predictive performance alone resulted in non-robust associations. Finally, we built a custom machine learning model that further empirically improved robustness of learned associations. Our results highlight the need for model selection criteria that maximize the scientific utility provided by reproducibility instead of predictive performance.

Details

Title
Model selection to achieve reproducible associations between resting state EEG features and autism
Author
Carson, William E. 1 ; Major, Samantha 2 ; Akkineni, Harshitha 2 ; Fung, Hannah 2 ; Peters, Elias 2 ; Carpenter, Kimberly L. H. 3 ; Dawson, Geraldine 3 ; Carlson, David E. 4 

 Duke University, Department of Biomedical Engineering, Durham, USA (GRID:grid.26009.3d) (ISNI:0000 0004 1936 7961) 
 Duke University, Duke Center for Autism and Brain Development, Durham, USA (GRID:grid.26009.3d) (ISNI:0000 0004 1936 7961); Duke University, Department of Psychiatry and Behavioral Sciences, Durham, USA (GRID:grid.26009.3d) (ISNI:0000 0004 1936 7961) 
 Duke University, Duke Center for Autism and Brain Development, Durham, USA (GRID:grid.26009.3d) (ISNI:0000 0004 1936 7961); Duke University, Department of Psychiatry and Behavioral Sciences, Durham, USA (GRID:grid.26009.3d) (ISNI:0000 0004 1936 7961); Duke University, Duke Institute for Brain Sciences, Durham, USA (GRID:grid.26009.3d) (ISNI:0000 0004 1936 7961) 
 Duke University, Department of Civil and Environmental Engineering, Durham, USA (GRID:grid.26009.3d) (ISNI:0000 0004 1936 7961); Duke University, Department of Biostatistics and Bioinformatics, Durham, USA (GRID:grid.26009.3d) (ISNI:0000 0004 1936 7961); Duke University, Department of Electrical and Computer Engineering, Durham, USA (GRID:grid.26009.3d) (ISNI:0000 0004 1936 7961) 
Pages
25301
Publication year
2024
Publication date
2024
Publisher
Nature Publishing Group
e-ISSN
20452322
Source type
Scholarly Journal
Language of publication
English
ProQuest document ID
3120698964
Copyright
© The Author(s) 2024. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.