Recurrent neural network-based acute concussion classifier using raw resting state EEG data

Concussion is a global health concern. Despite its high prevalence, a sound understanding of the mechanisms underlying this type of diffuse brain injury remains elusive. It is, however, well established that concussions cause significant functional deficits; that children and youths are disproportionately affected and have longer recovery time than adults; and that individuals suffering from a concussion are more prone to experience additional concussions, with each successive injury increasing the risk of long term neurological and mental health complications. Currently, the most significant challenge in concussion management is the lack of objective, clinically- accepted, brain-based approaches for determining whether an athlete has suffered a concussion. Here, we report on our efforts to address this challenge. Specifically, we introduce a deep learning long short-term memory (LSTM)-based recurrent neural network that is able to distinguish between non-concussed and acute post-concussed adolescent athletes using only short (i.e. 90 s long) samples of resting state EEG data as input. The athletes were neither required to perform a specific task nor expected to respond to a stimulus during data collection. The acquired EEG data were neither filtered, cleaned of artefacts, nor subjected to explicit feature extraction. The LSTM network was trained and validated using data from 27 male, adolescent athletes with sports related concussion, benchmarked against 35 non-concussed adolescent athletes. During rigorous testing, the classifier consistently identified concussions with an accuracy of > 90% and achieved an ensemble median Area Under the Receiver Operating Characteristic Curve (ROC/AUC) equal to 0.971. This is the first instance of a high-performing classifier that relies only on easy-to-acquire resting state, raw EEG data. Our concussion classifier represents a promising first step towards the development of an easy-to-use, objective, brain-based, automatic classification of concussion at an individual level.

. Sensor layout for EGI 64 electrode HydroCel Geodesic Sensor Net. The red ovals show the six electrode clusters used for power spectra analysis. The sensor numbers are listed in Table S1. (The electrode map is used with permission from Magstim EGI, MINNEAPOLIS, MN, and Eugene, OR, USA).

Power Spectrum Comparisons
We computed the power spectral density (PSD) of each participant's 64 raw EEG time series using Matlab [1] function, pwelch, and determined the median PSD for each of the six electrode clusters. The results, for the frequency range [0.1 Hz, 100 Hz], are shown as light orange (concussed) and light blue (control) curves in respective panels in Figure S2. Each row of the matrix plots shows spectral densities for the same cluster, with the PSDs of the concussed participants depicted in the left panel and the PSDs for the controls in the right panel. We also show the median PSD evaluated over all the participants for each class

S2/S6
(control vs. concussed) and electrode cluster. These are also plotted in each panel (in red for the concussed and in dark blue for the control participants).
A visual inspection indicates that the PSDs for the concussed and the control participants are on the whole similar. A detailed row-by-row examination of the plots reveals small differences between the median curves at specific frequencies or in specific frequency bands. Here, we will consider the three most readily apparent of these differences: 1. Alpha Peak Frequency (APF): APF is the frequency of a local PSD maximum in the alpha band [8][9][10][11][12][13]. This peak is thought to be a correlate of cognitive function, and a shift towards lower frequencies has been associated with mild cognitive impairment. In the context of mTBI, empirical results are mixed. Balkan et al. [2] did not find any APF differences between concussed and control participants. On the other hand, Dunkley et al. [3] found a trend towards lower peak frequency in the mTBI group but noted that the differences were not significant, even at the group-level. Our results are similar. We find that the APF of the median PSD of the concussed group, across all six electrode clusters, is lower than that of the control group. However, the distributions of APFs exhibit significant overlap. Considering the frontal region as an example, we note that the APFs of 9 correctly classified controls (i.e. 26% of the sample) are lower than the median APF of the concussed group. 2. Power in the high-gamma band : The amplitude of the median PSD of the concussed group has a slightly lower amplitude over the frequency range [65 − 90 Hz]. The shift is most apparent in the two temporal-parietal clusters and the occipital cluster. Gamma band (30 − 100 Hz) neuronal activity has been associated with a number of cognitive and executive functions. There is some evidence that gamma band activity may be altered by mTBI [4,5]. On the other hand, it is also well known that gamma band can be strongly contaminated by electromyogenic (EMG) activity of the scalp, face and neck muscles [6]. These muscles are very sensitive to emotional state and stress levels, and can experience prolonged contraction throughout an EEG recording that manifests as elevated PSD across the gamma band (see, for example, Figure 1b of Shackman et al. [7]). The temporal and frontal clusters are especially susceptible to EMG contamination. We are presently unable to argue in favor of one or the other origin (i.e., mTBI-based or EMG-based) for the differences between the two groups. We do, however, find that the distributions of broadband gamma power for both the concussed and control groups are wide and exhibit significant overlap; this is the case even for the two temporal-parietal clusters which exhibit the most pronounced differences in the amplitude of the median gamma band PSDs between the two groups. For example, seven of the correctly classified concussed participants have broadband gamma power greater than the control group's median value. 3. 70 Hz peak: The amplitude of the 70 Hz peak (with respect to the local neighbourhood) of the median PSDs differs between the concussed and the control samples. However, as with the two metrics discussed above, the peak amplitude distributions are also broad and overlap considerably. For the temporal-parietal right cluster, for example, seven correctly classified concussed participants have 70 Hz peak amplitude that exceeds the control median value.
The fact that there are differences in median EEG power spectral density of the concussed and the control groups is not surprising. As noted in the Introduction of the main paper, there is considerable literature reporting on group-level differences (see references in main paper). However, given the significant scatter in the PSDs within each class, nobody has yet been able to use these differences, singly or collectively, to statistically differentiate between non-concussed and concussed states at the individual level. Even attempts to apply Machine Learning algorithms (e.g. Support Vector Machine) to a collection of summary spectral features extracted from EEG datasets have not succeeded in identifying concussed individuals with a high degree of confidence. In fact, it is possible that the differences between the two groups are not solely determined from two-point statistical features (such as those embodied in the power spectral density) but rather from differences in higher-order correlations.
In summary, the power spectral densities of the concussed and control participants are comparable and exhibit similar spectral features. Even though small differences in the median curves at certain frequencies in each region may be seen, there is also significant scatter clearly evident amongst the participants in each class. Using only the differences in these few features would consequently lead to significant scatter in the predicted classification scores, and thus in the performance of the network. Since ConcNet achieves an accuracy of >90%, it may be inferred that the classifier does not merely latch on one or more of the differences in the features, but actually uses the information in the full rs-EEG dataset of each participant.