Modeling motor task activation from resting-state fMRI using machine learning in individual subjects

Area of Science:

• Neuroimaging
• Brain Physiology
• Machine Learning

Background:

• Resting-state functional MRI (rs-fMRI) offers insights into brain function and is a growing alternative to task-based fMRI for presurgical mapping.
• A standardized method for identifying eloquent brain areas using rs-fMRI in clinical settings is lacking.

Purpose of the Study:

• To evaluate a general linear model-based machine learning (GLM-ML) approach for predicting individual motor task activation from rs-fMRI data.
• To compare the accuracy of the GLM-ML approach against conventional independent component analysis (ICA).

Main Methods:

• A GLM-ML model was trained using rs-fMRI network maps and hand movement task fMRI data from 47 healthy subjects.
• The trained model predicted task activation maps from rs-fMRI data alone for unseen subjects.
• A low-resolution fMRI protocol was employed for resting-state, active, and passive motor tasks.

Main Results:

• The GLM-ML approach accurately predicted individual task activation differences using rs-fMRI.
• GLM-ML outperformed ICA in detecting task activation within the primary sensorimotor region.
• Predicted activation maps from GLM-ML showed strong agreement with passive hand movement fMRI results on an individual level.

Conclusions:

• The GLM-ML approach robustly predicts individual task activation from low-resolution rs-fMRI data.
• This method holds significant promise for future clinical applications in presurgical brain mapping.