Subject variability in mild traumatic brain injury (mTBI) data has often been an obstacle in accurate injury-related feature extraction and biomarker identification for successful mTBI diagnosis. To address this issue, we propose an adversarial variational autoencoder model as a novel regularization approach to extract subject-invariant representations for transfer learning in mTBI identification. The proposed method consists of a variational autoencoder with an attached adversarial network. The autoencoder attempts to learn the latent space mappings from neural activity, while the adversary network is used in a discriminative setting to detach the subject individuality from the representations. The trained encoder is then transferred to extract the representations from new subject's data. Several classifiers are utilized to classify the extracted representations into two categories of normal and mTBI. To evaluate the performance of the proposed method, recorded cortical activity of GCaMP6s transgenic calcium reporter mice before and after inducing an injury is used. Experimental results on cross-subject transfer learning exhibit the efficiency of the proposed framework by achieving 89.7% classification accuracy, suggesting the feasibility of the proposed method in learning invariant representations in mTBI data.