Objective To develop a clinically feasible, brain-based tool for sideline concussion classification.

Design Prospective cohort study.

Setting High school athletes.

Participants 35 healthy control male athletes; 23 male athletes with concussion.

Interventions (or Assessment of Risk Factors) N/A

Outcome Measures (1) Clinical assessment of concussion by a physician with expertise in concussion using diagnostic criteria consistent with the Berlin consensus statement. (2) Assessment of symptoms using SCAT-3 or Child SCAT-3. (3) 5 minutes of raw, resting state EEG data.

Main Results All of the concussed participants met the Berlin criteria and exhibited between 4 to 22 SCAT3 symptoms, at the time of testing. Using our previously trained deep learning network on 64 EEG channels, we identified the top three channels that had the highest impact on concussion classification and re-trained the network using these three channels. We found that the re-trained network classified concussion with an accuracy of 93.3% (SD=.005, 95% CI=.926, .940). All three channels were located in the occipital region of the head.

Conclusions This is the first proof of concept showing that a deep learning network can be trained for concussion classification using raw, resting state data from only 3 EEG sensors. This is a critical step in developing a portable, easy to use EEG systems that can be used in a clinical setting and for sideline assessment of concussion.