mTBI patients exhibit acute symptoms including headache, blurry vision and memory problems, though CT/MRI scans often appear normal^1,2. While these symptoms resolve within few weeks, a second impact during this vulnerable phase can lead to persistent damage³. For sports-related injuries, the decision for return-to-play is the most difficult responsibility facing the physician, and currently this decision is largely based on self-reported symptoms. rs-fMRI may be an additional, more objective tool to assess the severity of the injury. The purpose of this study was to identify functional biomarkers of mTBI during the 3 months following mTBI, with the ultimate goal of predicting patient’s symptom severity and recovery profile

After obtaining informed consent, rs-fMRI was recorded from 78 patients at four time points (3 days, 7 days, 1 month and 3 months) after mTBI and 26 healthy controls (2 sessions, 1 week apart). After eliminating missing and noisy data, we analyzed 184 time points in total. Using GE 3T MRI scanner, multi-band (acceleration factor 3) 2D-EPI, TR/TE = 900/30 ms was acquired for 6 minutes (395 volumes), with 1.875mm2 in-plane resolution and 3mm slice thickness to cover the whole brain. T1-weighted scan (1mm resolution) was acquired at each time point. All participants filled a neuropsychological self-assessment questionnaire that was used to calculate symptom severity scores⁴ (SSS). rs-fMRI data were motion corrected, rigid registered to T1-weighted image, non-rigid registered to MNI atlas, nuisance removed using aCompCor⁵, spatial smoothed using gaussian filter (FWHM 4mm) and temporally band-pass filtered (0.01 -0.1 Hz) using custom-built software. Mean time courses of 90 functional ROIs⁶ were extracted. Correlation coefficients were computed between every pair of time courses, producing a 90x90 matrix. Correlation values were transformed to z-scores using Fisher transform. Average correlation between ROIs belonging to same network (14 networks, within network connectivity, WNC) and between networks (between network connectivity, BNC) were computed. This yielded a total of 105 features: 14 WNC features and 91 BNC features (₁₄C₂).

To automatically identify features associated with SSS, recursive feature elimination (RFE)⁷ method using decision trees as prediction models were used (Figure 1). Initially, a machine learning (ML) model was built using all 105 features and was used to rank the importance of features based on contribution to increase in node purity at various splits of the decision tree. Smaller subsets of top-ranked features were used to build efficient ML models and the cross-validation (CV) performance was evaluated for every subset.

Average connectivity across 90 ROIs was negatively correlated with SSS (Figure 2A). On binning the patients with low (SSS<10) and high (SSS>30) symptoms, there was a visible difference between the two patient groups (Figure 2D, note BNC changes in off-diagonal elements). To identify if the changes in connectivity with SSS were contributed by interactions within or between networks, we performed feature sub-selection using a ML framework (Figure 1). The top-ranked features identified by RFE included BNC between higher and primary visual, executive control (ECN) and precuneus networks (Figure 3). Using these features, we achieved a best performance of ~23 CV RMSE for predicting SSS. 9 of the top 10 relevant features were BNC measures (Figure 3A). BNC between L-ECN and higher visual networks was more strongly anti-correlated in high-score mTBI patients (Figure 3B). In contrast, BNC between primary and higher visual networks increased in mTBI patients (Figure 3B).

We observed an overall decreased connectivity in mTBI patients with high symptom scores. We used RFE based ML framework to sub-select modular network-based connectivity features that showed correlation with SSS. Typically in neuroimaging studies, number of features are much larger than the number of observations. Here, we introduce a novel ML framework for feature selection to improve interpretability and reduce over fitting.

The changes with SSS were more strongly reflected in BNC compared to WNC features, indicating that brain injury disrupts association between functional brain network modules. The observed decrease in connectivity between L-ECN (including superior and middle frontal gyri) and higher visual areas in mTBI is in agreement with a previous study that showed similar altered patterns of activation in fMRI during working memory tasks⁸. The observed increase in visual region connectivity has been shown when subjects become aware of stimuli⁹ and top-down attention also has an influence in this connectivity¹⁰; implying mTBI patients with higher scores are perhaps explicitly paying more attention to external stimuli. To understand the neural mechanisms underlying these changes, we plan to investigate specific sub-scores related to visual, motor and attention with the corresponding networks.