Abstract Details

Deep learning has shown utility for calculating “brain age” using magnetic resonance imaging (MRI) as a proxy indicator of brain integrity. The difference between a person’s brain age and their actual biological age (Δage) indicates whether a person’s brain shows age-related changes.

Here, we assess whether bio-signatures of brain age captured from MRI in patients with acute post-traumatic headache (aPTH) attributed to mild traumatic brain injury (mTBI) may serve as surrogate neuroimaging biomarker for predicting PTH persistence.

We trained a ResNet-18 model on a cohort of 7,377 HC (public datasets: IXI, ICBM, ABIDE, OASIS, NACC) to predict biological age using T1-weighted brain MRI scans. Our model outperforms other baseline methods with mean absolute error (MAE)=2.56 on held-out HC test set. We applied this trained model to a different HC cohort (N=40, mean age=38.5) and an aPTH cohort (N=37, mean age=42.4). Following mTBI, aPTH patients underwent imaging within 0-59 days.  At three months post-mTBI, 19 patients (mean age=44.8) had headache recovery and at 6 months post-mTBI 20 patients (mean age=42.8) had headache recovery.

Using baseline imaging, predicted Δage for HC and aPTH patients was -0.6±11.0 and 0.5±9.9 respectively. For aPTH patients who had headache recovery at 3-month follow-up, predicted Δage was 0.5±8.3 and 0.8±11.7 for those without headache recovery. Patients who continued to have headache persistence at 6-month had predicted Δage of 1.3±9.8.

Compared to HC, mean predicted Δage signature was older for patients with aPTH suggesting more structural decline related to TBI and aPTH. Compared to patients with headache recovery, aPTH patients who had headache persistence at 3 and 6 months showed older brain signatures indicating greater brain structural decline.  Results suggest the potential for deep learning to capture an early imaging signature for structural decline using MRI acutely following TBI for detecting patients at risk for developing persistent PTH.