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

Abstract

New biomarkers are urgently needed for many brain disorders; for example, the diagnosis of mild traumatic brain injury (mTBI) is challenging as the clinical symptoms are diverse and nonspecific. EEG and MEG studies have demonstrated several population-level indicators of mTBI that could serve as objective markers of brain injury. However, deriving clinically useful biomarkers for mTBI and other brain disorders from EEG/MEG signals is hampered by the large inter-individual variability even across healthy people. Here, we used a multivariate machine-learning approach to detect mTBI from resting-state MEG measurements. To address the heterogeneity of the condition, we employed a normative modeling approach and modeled MEG signal features of individual mTBI patients as deviations with respect to the normal variation. To this end, a normative dataset comprising 621 healthy participants was used to determine the variation in power spectra across the cortex. In addition, we constructed normative datasets based on age-matched subsets of the full normative data. To discriminate patients from healthy control subjects, we trained support-vector-machine classifiers on the quantitative deviation maps for 25 mTBI patients and 20 controls not included in the normative dataset. The best performing classifier made use of the full normative data across the entire age and frequency ranges. This classifier was able to distinguish patients from controls with an accuracy of 79%. Inspection of the trained model revealed that low-frequency activity in the theta frequency band (4-8 Hz) is a significant indicator of mTBI, consistent with earlier studies. The results demonstrate the feasibility of using normative modeling of MEG data combined with machine learning to advance diagnosis of mTBI and identify patients that would benefit from treatment and rehabilitation. The current approach could be applied to a wide range of brain disorders, thus providing a basis for deriving MEG/EEG-based biomarkers.

Fig 1. Analysis pipeline.

(A) Data preprocessing. Source-level power spectra were first calculated for each cortical region and morphed to a common brain template. (B) Construction of normative models. The mean μ and standard deviation σ were calculated across subjects within the reference dataset to obtain normative data for each location and frequency. Three different types of normative models were constructed: a full normative model containing all participants from the reference dataset across the entire age range (depicted in blue, mean values across all frequencies and cortical locations are shown), age-matched models containing a subset of participants within the same age range as the patient/control (exemplar depicted in green), and for comparison a random model with a subset of reference data of random ages (shown in red). (C) Classification procedure. The normative models were used for converting the power spectra of the patients and controls into deviation scores (z-scores). The deviation scores, binned into 448 cortical parcels and six frequency bands, were then entered into the classification procedure.

Fig 2

(A) Group-average, relative power spectra of mTBI patients and healthy controls. Horizontal axis is frequency (Hz), vertical axis the cortical location (indices of the 448 cortical regions ordered alphabetically), and the color indicates the Z-score with respect to the normative data at that frequency and cortical location. (B) Relative power spectra associated with different classification results. The average Z-score maps for patients (first row) and controls (second row) by classification output, from left to right: correctly classified samples, incorrectly classified samples and the difference between the correctly and incorrectly classified samples.

Fig 3

(A) Feature importance. The feature importance (horizontal axis) is defined as the reduction in accuracy when the feature is randomly permuted. The labels of the features indicate the cortical region, the index of the subarea within the subdivided region, and the hemisphere (L for left, R for right) that the feature corresponds to. Only the 30 features with the largest mean importance are shown. Values are sorted according to mean importance across folds, and the median values are shown in red. (B) Cortical sources contributing to the classification of patients and controls at two frequency bands. The spatial distribution of the average feature importance across folds, shown for the theta and alpha frequency bands. The values were calculated as the permutation feature importance.

Fig 4. Deviation score maps for theta-band power in four patients.

The color indicates the Z-score with respect to the level of 4–8-Hz activity in the full normative data.