Related Experiment Video
Updated: Apr 25, 2026
14:27
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
15.2K
On the generalizability of resting-state fMRI machine learning classifiers
Wolfgang Huf1, Klaudius Kalcher1, Roland N Boubela1
1Center for Medical Physics and Biomedical Engineering, Medical University of Vienna Vienna, Austria ; MR Centre of Excellence, Medical University of Vienna Vienna, Austria ; Department of Statistics and Probability Theory, Vienna University of Technology Vienna, Austria.
Frontiers in Human Neuroscience
|August 15, 2014
Summary
Machine learning classifiers show promise for single-subject fMRI analysis. However, their generalizability across different study samples is limited, necessitating multi-site data for robust applications.
Area of Science:
- Neuroimaging
- Machine Learning
- Computational Neuroscience
Background:
- Machine learning classifiers are increasingly used for single-subject functional Magnetic Resonance Imaging (fMRI) inferences.
- This represents a shift from traditional group-level analyses.
- Limited information exists on the generalizability of these classifiers to new subject samples.
Purpose of the Study:
- To investigate the generalizability of machine learning classifiers trained on resting-state fMRI data.
- To assess classifier performance on independent datasets from different studies.
Main Methods:
- A simulation study was conducted using resting-state fMRI data from the 1000 Functional Connectomes and COBRE projects.
- Classifiers were based on regional homogeneity of resting-state time series.
- Generalizability was tested on datasets from different studies.
Main Results:
- Classification accuracies up to 0.8 were achieved when training and testing on data from the same study (using sex as the target variable).
- Generalizability to classifiers applied to different study samples was limited, though accuracy remained above chance.
- A degree of generalizability can be expected, but it should not be overestimated.
Conclusions:
- Classifiers trained on fMRI data exhibit some generalizability across studies, but performance significantly decreases.
- Overestimation of classifier generalizability is a concern.
- Future studies aiming for generalizable machine learning classifiers in fMRI should incorporate multi-site data.
