Predicting outcome in patients with moderate to severe traumatic brain injury using electroencephalography

There is an ongoing need for continuous, bedside-available, preferably non-invasive tools for reliable prediction of neurological outcome in traumatic brain injury (TBI) patients. A prediction model for patient outcome after TBI might provide realistic information to caregivers to inform relatives about expectations. Besides, better outcome prediction could assist in reliable quantification and classification of TBI severity to support clinical decision-making [1].

Currently, the International Mission for Prognosis And Clinical Trial Design (IMPACT) predictor is the best available predictor for neurological outcome after TBI. This online predictor is based on a large multicenter database including both randomized controlled trials and observational studies in TBI [2]. Linear regression models were created on clinical admission parameters to predict mortality (Glasgow Outcome Scale 1) and unfavorable outcome (Glasgow Outcome Scale 1–3) at 6 months after injury [3]. The model is based on age, motor score, pupillary reaction, CT characteristics, signs of hypoxia, hypotension, glucose, and hemoglobin at admission. However, this prediction model, based on parameters at admission, does not take into account evolvement of secondary injury in TBI patients nor effects from treatment [4].

Electroencephalography (EEG) is a non-invasive bedside-available measure of cortical activity and is particularly useful in sedated or comatose patients [5‐7]. Continuous EEG (cEEG) is often part of the multimodal monitoring in TBI patients at the ICU, where it is mainly used for the detection of (non-convulsive or electrographic) seizures [8, 9].

Several attempts have been made to relate EEG features to functional outcome of TBI patients. Worse outcome was associated with lower (regional) EEG power [9], slowing of the EEG [10‐13], decrease in alpha power [9‐11, 14, 15], lower EEG (alpha) variability [9, 16, 17], and increased coherence [18]. However, most studies used single EEG features and were limited to mild TBI in the subacute phase (days, weeks, or even months) after the moment of injury [19]. A combination of multiple EEG features might be even more relevant for outcome prediction in TBI patients, considering the heterogeneity of the primary injury, additional organ dysfunctions, and evolvement of secondary injury in these patients. Furthermore, computer algorithms can assist in EEG interpretation [6, 20, 21], which decreases the labor intensity and subjectivity.

As proof of concept, we report on a cEEG monitoring tool using multiple quantitative EEG (qEEG) features combined with IMPACT predictors for neurologic outcome prediction in patients with moderate to severe TBI.

Fifty-seven patients were included. Patients were divided into a training set (38 patients included between 2013 and 2016) and a validation set (19 patients included between 2017 and 2018). A flow diagram for the inclusion and exclusion of eligible patients is shown in Fig. 1.

Table 1 shows the baseline characteristics. Patients with good neurological outcome were significantly younger in both training set (p = 0.037) and validation set (p = 0.016). Injury Severity Score and duration of EEG monitoring did not differ between the groups in both sets.

Fifteen out of the 38 patients in the training set died during their hospital admission (GOSE 1), and 15 had a good neurological outcome at 1 year with a GOSE between 3 and 6. The remaining eight patients were lost to follow-up, because they could not be contacted by telephone. Patient outcome was therefore estimated based on their medical data records. Two of them were in a vegetative state at discharge and assigned to the poor outcome group (GOSE 1–2), and the other six to the good outcome group (GOSE 3–8). In the validation set, none of the patients was lost to follow-up; 11 out of 19 patients had a good neurological outcome, and the other eight patients had a poor outcome.

We developed a multifactorial model as proof of concept for using a machine learning algorithm including qEEG features and clinically relevant parameters for the prognostication in moderate to severe TBI. Our best models comprised qEEG features and MAP at 72 and 96 h after TBI, combined with clinical (IMPACT) parameters at hospital admission. Our model has an AUC of 0.94 on the training set and 0.81 on the validation set. In both training and validation sets, the sensitivity and specificity of our best model were slightly higher than those of the IMPACT predictor alone, while the AUC of the ROC was similar.

Although several attempts have been made over the last decades to relate individual EEG parameters to patient outcome after TBI [9‐11, 13‐16, 19, 33‐37], no models based on multiple qEEG parameters predicting patient outcome in the acute phase after moderate to severe TBI were reported.

Quantitative EEG features highly contributed to the predictive ability of the model, especially the mean amplitude of the EEG. Although mean amplitude (or total power) of the EEG has been associated with depth of coma [38] and with neurologic outcome in patients after cardiac arrest [6], it has rarely been described in relation to patient outcome after TBI [9]. Spectral edge frequency of 90% and relative alpha power at 72 h were the second and third most relevant qEEG features. This is in line with a recent study from Tolonen et al., who showed that median alpha power was most related to patient outcome after TBI [9].

Age and MAP at 72 h were relevant prognostic features as well, which is in line with earlier literature stating that age is a powerful prognostic factor in recovery after TBI [2]. Prediction was most optimal with features from recordings at 72 and 96 h after TBI. This time interval coincides with being at risk for secondary damage, and this period might therefore have strong impact on the prognosis of TBI patients [39]. Although the model at 72 + 96 h was selected as the best model for this proof of concept, this was however not statistically substantiated and all models performed well at the different (individual) time intervals in the training set (average AUC of 0.86). Therefore, models at other time intervals could be of interest in future studies, too. In addition, temporal evolution might have additional prognostic value. At 72 and 96 h, the number of patients dropped due to patients’ passing or awakening and being transferred to the general ward. Patients with no immediate risk of death and uncertainty about the awakening remain, which is also the category for which caregivers could use better decision support tools.

From the IMPACT parameters, glucose levels at admission contributed most to the model’s predictive ability. Murray et al. already stated that prognostic models in TBI should include laboratory data and especially glucose [2], although they found a smaller effect of glucose compared to other parameters in the IMPACT predictor [24]. Least relevant for our models were pupillary reactivity, hypotension, hypoxia, and the presence of epidural hematoma (EDH) or traumatic subdural hemorrhage at the CT scan. On the contrary, Steyerberg et al. found that most prognostic information was contained by age, motor score, and pupillary reaction and that CT information provided additional information, although this information might be not detailed enough for prognostic purposes [24]. Lee et al. (2019) [40] recently showed that the absence of a posterior dominant rhythm, absence of sleep stage N2 transients, and predominant delta activity and discontinuous background during the first 72 h were individually associated with poor outcome at 3 months and increased the predictive ability of the IMPACT predictor from 0.65 to 0.77. Since we focused on quantitative analysis instead of visual EEG analysis, these EEG characteristics were not incorporated in our proof of concept study; however, they could be added in future prediction modeling after TBI.

EEG measurements are more time consuming than IMPACT prediction. Electrode application takes approximately 45 min, and visual analysis approximately 1–2 h per day. This can be substantially reduced by computer algorithms supporting visual analysis [6] and promising developments in electrode design [41‐44]. If the prognostic ability of a model based on EEG exceeds that of the IMPACT predictor, the benefits may outweigh the extra efforts.

A strength of our study is the recording of 7-day high-quality continuous EEG measurements in patients after moderate to severe TBI. Another strength is that in contrast to the IMPACT predictor that only uses parameters at admission, predictions based on longitudinal qEEG parameters take into account the influence of treatment and heterogeneous pathology evolvement in TBI patients [4]. The fact that our best model has similar predictive power as the IMPACT data set independently is an encouraging result for a model that is trained on only 30 patients from a by nature heterogeneous group of TBI patients. A Random Forest classifier was chosen because of its ability to prevent overfitting and the model was trained using leave-one-patient-out, which also lowers overfitting. A limitation of Random Forest models for multifactorial data is that it excludes observations with missing data. For this reason, other possible relevant clinical parameters, which were not available in all patients, such as intracranial pressure, were not considered in modeling.

A limitation of our study is that primary outcome, the GOSE, does not take non-neurological cause of death (in case of polytrauma patients) into account and that patient survival depends on clinical decision-making. In addition, preferably, the poor outcome group would be defined as death to severe disability (GOSE 1–4) and good outcome as moderate disability to good recovery (GOSE 5–8). Due to a low number of patients with a GOSE of 5–8 and incomplete follow-up in 8 patients of the training set, we primarily choose to define poor outcome as GOSE 1–2. Model performances were indeed lower when predicting poor outcome defined as GOSE 1–4. A disadvantage of including the MAP as a parameter for prediction is that it could be artificially controlled because of treatment to optimize hemodynamics or cerebral perfusion pressure. Although MAP was included in the prediction model, its contribution was relatively low (not in the top 5 most relevant factors). As EEG features also depend on the MAP and CPP, this may explain its low additional value for outcome prediction. QEEG features were not corrected for the potential influences of decompressive craniectomy or epileptic activity. Although this might have increased the mean amplitude of the EEG, lower mean amplitude is a strong predictor for poor outcome.

Another limitation of the study is that the training and validation sets originate from different cohorts in time. However, we do not expect this to be a confounder, because there were no changes in treatment over this period and the ratio of poor outcome was relatively similar. The size of the training set and validation set are below the ideal size for a prognostic model, its validation and comparison to the IMPACT predictor alone. However, in this proof of concept study, we showed the potential of Random Forest models for this purpose.

Models should be further developed using larger datasets and use GOSE 1–4 and 5–8 to distinguish between poor and good neurologic outcome respectively. For future research, higher generalizability could be obtained by including more data from multiple time periods and medical centers [9] and by optimizing model parameters (i.e., amount of trees and nodes).

We showed that qEEG parameters provide useful information in prediction of outcome after moderate to severe TBI during the first days after trauma. Ultimately, using data from multiple modalities that monitor brain function might lead to better decision support tools in moderate to severe TBI patients at the ICU.

Multifactorial Random Forest models using qEEG features, clinical data, and radiological findings have potential to predict neurological outcome in patients with moderate to severe TBI.