Using normative modeling and machine learning for detecting mild traumatic brain injury from magnetoencephalography...

Area of Science:

• Neuroscience
• Biomarker Discovery
• Medical Diagnostics

Background:

• Diagnosing mild traumatic brain injury (mTBI) is challenging due to nonspecific symptoms.
• Existing EEG/MEG biomarkers for mTBI are limited by significant inter-individual variability.
• Objective biomarkers are crucial for accurate mTBI diagnosis and management.

Purpose of the Study:

• To develop a machine learning approach for detecting mTBI using resting-state MEG data.
• To address inter-individual variability in brain signals using normative modeling.
• To identify reliable MEG-based biomarkers for mTBI diagnosis.

Main Methods:

• Utilized a multivariate machine learning approach with support-vector-machine classifiers.
• Employed normative modeling with a dataset of 621 healthy participants to establish normal variation in MEG power spectra.
• Trained classifiers on quantitative deviation maps comparing 25 mTBI patients to 20 healthy controls.

Main Results:

• The best classifier achieved 79% accuracy in distinguishing mTBI patients from controls.
• Normative modeling of MEG data effectively accounted for inter-individual variability.
• Low-frequency theta band activity (4-8 Hz) was identified as a significant indicator of mTBI.

Conclusions:

• Machine learning combined with normative modeling of MEG data is feasible for mTBI diagnosis.
• This approach can identify patients who may benefit from targeted treatment and rehabilitation.
• The methodology holds potential for developing biomarkers for various brain disorders using MEG/EEG.