Predictions are difficult, especially about the future. - -Yogi Berra
Along with diagnosis and treatment, prognosis is one of the fundamental duties of all clinicians. It is especially critical in traumatic brain injury (TBI) where uncertainty about the future compounds the suffering of patients and their families. Indeed, families have identified information about prognosis as one of their most important needs after (TBI). Yet studies have shown that this need is rarely met.2,3 One of the reasons we as clinicians have difficulty providing prognostic information is that our abilities to predict outcome are limited. Many of the traditionally used prognostic indicators (such as the initial Glasgow Coma Scale [GCS] score or the results of computed tomography [CT] imaging) are simply not powerful enough to allow us to make very specific prognoses. And more powerful predictors such as the duration of posttraumatic amnesia are often unavailable for weeks or months after the injury.4
Given the limitations of traditional prognostic variables, there is a need for newer measures that can provide more definitive information about outcome. In particular, there is hope that recent developments in technology will provide us with more powerful prognostic predictors. This issue is devoted to such developments and explores the promise and limitations of recent technological advances. In particular, it focuses on the implications of neural networks, magnetic resonance spectroscopy, serum biomarkers, newer forms of electrophysiological studies (such as event-related potentials), and genetic factors for prognosis.
The articles in this issue are meant as introductions to each of these modalities. Therefore, they all begin by reviewing basic information about the mechanisms and rationale of the technology being discussed. They then review the literature supporting the use of these modalities in TBI. In doing so, some of the authors present data from their own work (either in the form of an original study or by summarizing previous work). Finally, they all draw conclusions about the promise and limitations of these various technologies for prognostication after TBI.
Although we tend to think of technology as involving material objects such as machines or instruments, this does not always have to be the case. Advances in formal techniques such as new algorithms or forms of statistical analyses also count as technological achievements. An example of this is artificial neural networks (ANN). In this context, neural networks refer to a form of statistical analysis, playing the same role in the analysis of data that, for instance, multiple regression plays. Based both on theoretical reasons as well as empirical work, it is believed that neural networks, using the same predictor variables that are currently being used (such as the initial GCS score or CT imaging), might be more accurate than traditional statistical methods in making predictions about outcome. That is, neural networks might be able to extract more information from these variables, thereby rendering them more powerful as prognosticators. If true, it means that there may be more value to classical predictors than we thought and that reanalyzing their predictive capacity with neural networks might improve our accuracy in predicting outcome.
The results of studies on ANN in other fields is mixed. While some studies have found ANN to be more accurate than traditional statistical methodologies, most have not found ANN to be superior.5 In TBI, ANN has been studied only in the acute care context where it was found to be superior to multiple regression in predicting short-term outcome.6,7 In this issue, we publish what we believe is the first study of the use of ANN in the TBI rehabilitation setting, where it was used to predict long-term functional outcome. In their study, Segal et al analyze the same set of predictor variables with 3 different statistical methodologies: multiple regression (MR), ANN, and classification and regression trees (CART). They then compare the accuracy of the outcome predictions made by each method.
They found that ANN was no more accurate than MR in predicting long-term outcome after TBI. However, the ability to generalize from this finding is limited, as they point out, by the form of MR that they used. In particular, they used a sophisticated form of MR and data imputation that may not be readily available to all researchers. Therefore, the authors conclude that ANN may still be found to be more accurate than traditional MR in predicting long-term outcome after TBI. This is an area that will require further research, and it is hoped that some readers of this article will be inspired to pursue this.
Our second article reviews the promise of serum biomarkers in TBI prognostication. These biomarkers are substances found in the brain that are released into the bloodstream after injury. It is hypothesized that the presence and level of these substances would correlate with the degree of injury severity and, thus, with outcome. One of the advantages of these biomarkers is that they are easy to obtain (requiring just the drawing of blood) and are available within the first hour or two after injury. Thus, if they prove to be of prognostic value, they would provide a powerful tool to clinicians working in the acute care or rehabilitation settings. Berger is one of the leaders in investigating the role of these biomarkers in pediatric brain injury and, in her article, she provides a comprehensive review of the principles underlying the use of serum biomarkers as well as the literature supporting their role in severe, mild, and pediatric TBI.
Berger focuses on the most widely studied and promising of the biomarkers, S100B. Although the literature on the use of S100B after mild TBI is inconsistent, Berger finds that studies on severe TBI have been more uniform in their findings. In particular, the studies to date demonstrate that S100B is consistently correlated both with severity of injury and, more importantly, outcome. In addition, studies that compared S100B to other predictive variables (eg, GCS and CT) found that the serum biomarker provided an incremental improvement in predictive accuracy compared to these traditional measures. These are promising findings and, as Berger points out, have led, in certain settings, to the routine clinical use of S100B in Europe. She cautions, however, that more work needs to be done before these serum biomarkers can be recommended as standard practice.
The third article, by Shutter et al, focuses on the possible role of magnetic resonance spectroscopy (MRS) in TBI prognostication. While traditional magnetic resonance imaging [MRI] utilizes signals from the proton nuclei of water, MRS (performed on the same machines as conventional MRI) measures signals from other metabolites in the brain. Although not used to construct an anatomical image of the brain, the information obtained about the relative concentrations of these metabolites can shed light on the underlying pathophysiology of TBI and, it is hoped, outcome.
Shutter's group is one of the leading investigators on the use of MRS in predicting outcome after TBI and in their article, they summarize their work to date. In doing so, they present in publication form some data that had previously been available only in posters or abstracts. Their article also represents one of the first published reports of the use of multivoxel, 2-dimensional MRS imaging in predicting outcome after TBI.
Among other findings, Shutter's group found that the early cerebral concentrations of certain metabolites differed significantly between those individuals with a poor outcome versus those with a good outcome. More importantly, when her group used logistic regression to compare MRS measures to more traditional variables (such as age, initial GCS score, and somatosensory-evoked potentials), they found that the MRS variables were the most powerful, achieving an 89% predictive accuracy in forecasting 6- to 12-month Glasgow Outcome Scale outcomes. One of the virtues of the way in which Shutter's group presents their data is that they calculate the sensitivities and specificities of their variables in predicting outcome, thereby making their results much more useful to clinicians looking to assess the clinical implications of their findings.
In our fourth article, Lew et al revisit a modality that may seem familiar to many in the field: evoked potentials (EPs), which have been studied in the setting of TBI for more than 2 decades. EPs represent the passive electrophysiological response of the brain to sensory stimuli. EPs have different components, most often distinguished by their time of onset after the stimulus (latency). As most rehabilitationists are aware, the early (short latency) responses of EPs (such as somatosensory-evoked potentials) are most useful in predicting negative outcomes during the acute stage of brain injury, most notably the failure to emerge from coma.8 Lew et al suggest that other components of EPs (such as the middle latency responses) might provide us with additional information, allowing us to make more comprehensive predictions. In particular, they suggest that we may be able to go beyond simply anticipating negative outcomes and start predicting which individuals in the acute stage might have good outcomes.
They also describe the potential value of another, less familiar, electrophysiological measure: event-related potentials (ERPs). Unlike EPs, which usually represent a passive response to stimuli, ERPs generally reflect the electrophysiological state of the brain when actively performing a cognitive task. As a result, ERPs are often considered a potential measure of more complex cognitive processing and may improve our ability to understand and predict the return of higher order cognitive functions. As an example of the role of ERPs in assessing more elaborate mental processing, Lew et al cite their work noting reduced or delayed responses in the P300 (a component of ERPs) when people with TBI were asked to judge facial emotion. This example highlights the fact that the role of EPs and ERPs is not limited to prognostication, and Lew et al admirably address the other possible uses of these modalities in TBI rehabilitation.
In the concluding article, Diaz-Arrastia and his coauthor review the implications of recent advances in genetics for our field. They provide an accessible introduction to the role that genetic factors play in the recovery from TBI, underscoring the fact that even when 2 brain injuries seem identical in severity and location, the different genetic constitutions of the individuals could lead to very different outcomes. As an example, they review the recent work on the link between the apolipoprotein E4 allele (APOE-e4) and outcome after TBI. Although they point out that not all studies support an association between APOE-e4 status and outcome, the weight of the evidence does support the assertion that patients with an APOE-e4 allele were more likely to do worse than those without the allele (with an odds ratio approximately between 3 and 13).
However, although Diaz-Arrastia and Baxter acknowledge the potential relevance of genetic findings to prognostication, they believe that the most important impact of genetic advances lie elsewhere. Specifically, advances in genetics will teach us much more about the pathophysiology underlying the different stages of the brain's response to injury (from the acute neurochemical reaction to the more chronic response of regeneration and repair). And that information, in turn, could potentially lead to new therapies for treating TBI at these different stages. Thus, although there may come a time when clinicians will estimate prognosis using a patient's APOE-e4 status along with other information (such as length of coma or duration of posttraumatic amnesia), this should be considered more of an ancillary benefit of genetic advances, rather than its main reward.
In the field of TBI, our ability to prognosticate has been significantly constrained by the limited power of the information available to us, leading to frustration both for ourselves as well as for those we care for. The articles in this issue describe technological advances that have significant implications for improving our ability to prognosticate, raising the hope that the situation may soon change, as these technologies of prognostication bear fruit and begin to be adopted into general practice.
Sunil Kothari MD, Issue Editor
REFERENCES