In this study, we have attempted to propose a method to translate the head injury AIS codes into the Marshall CT Classification. This involves two steps; first to cross-tabulate various AIS codes with the Marshall Classes and secondly to select the single Marshall Class allocated to a case of TBI through an algorithm. In order to perform this transformation some assumptions had to be made.
Limitations/assumptions
Although both the Marshall Classification and the AIS dictionary group CT features according to their severity, one important difference between the two systems relates to their purposes. The main aim of the Marshall Classification is to identify those TBI patients who are at higher risk of deterioration or mortality, whereas the AIS scoring system is used to classify injuries based on their anatomy rather than physiological merits. These different approaches to CT classification mean that certain assumptions have to be made when trying to reconcile the two systems. Ideally the two systems would be completely interchangeable and no assumptions would be required. Since this is not the case an important question is whether or not mapping AIS codes onto the Marshall Classification is worthwhile. We believe adoption of the conceptual approach we have proposed allays some concerns in that, instead of strictly meeting the definition of each Marshall Class, the objective and rational surrounding that class are also employed to spot the appropriate AIS codes. A disadvantage of the Marshall Classification is that it is not a reliable classification to be used in the retrospective research settings in which the access to the real CT obtained during acute phase of therapy is often not possible in case the Marshall Class is not recorded in the existing dataset.
The Marshall Classification should be ideally performed by the expert who views the CT. However, Marshall Class II, unlike other classes, contains a broad range of heterogeneous injury types or severities. Considering the different objective of AIS dictionary which is to anatomically classify injury severities, mapping AIS codes with Marshall Classes is in fact alike ignoring many valuable individual pieces of information by pooling them into one class such as class II. This leads to them all being treated similarly in prognostic analysis despite potentially having varied individual prognostic merit on their own. This defect is substantial when class II contains, for instance, infarction beside laceration which are different in nature and perhaps in prognostic strength.
The method proposed in this study is based primarily on the assumption that the descriptions in the AIS dictionary can be substitutes for CT reports, but this is not always the case. As well as including CT reports, the sources of information to document injury descriptions also encompass MRI, surgery, x-ray, angiography, post-mortem examinations or clinical diagnosis. Nevertheless, it seems reasonable to assume majority, if not all the information for AIS coding, is obtained from CT scans since CT is the commonest modality for diagnosing structural brain damage for every patient suspected to have sustained severe trauma in the developed world and several other developing nations.
As AIS coding does not rely only on one CT report, the dynamic nature of brain injury as to progressing or regressing over time (which results in the evolution of CT findings) is reflected in AIS codes unlike the Marshal Classification. This is because the Marshal Classification is collected from CT images/reports at a certain point in time (oftentimes on admission) whilst AIS codes contain information after discharge or death. As such, using our algorithm to obtain the Marshal Class with AIS codes intermediation would inherit the information on dynamic nature or CT findings evolution as well. This poses a problem since evolution of structural brain damages per se is a prognostic factor indicating higher chance of unfavorable outcome [
11]. Servadei et. al. have shown that the worst CT classification has more prognostic value than less severe CT classification (s) [
11]. Consequently, if the Marshal Classification is obtained from AIS codes, there may be some overestimation of its negative prognostic role in TBI as compared to the Marshal Classification using CT images/reports. This may be particularly an issue with patients who sustain more severe brain injuries as they are more subject to various means of investigations such as MRI or operation.
As well as the above fundamental assumption regarding AIS descriptions as substitutes for CT reports, there are two other important assumptions related to the brain swelling and the mass lesion. Unfortunately neither the Marshall Classification nor the AIS dictionary describe precisely the severity of brain swelling. The degree of swelling in AIS dictionary is only determined by cistern/basal cisterns status whereas the degree of midline shift is also an important determining factor. Likewise, although midline shift or cisterns status is important in the Marshall Classification of brain swelling, other causes of midline shift, such as mass lesion, are disregarded. With respect to the size of mass lesions, future research is required to determine the precise size cut-offs for categorising such lesions, in spite of the already-known fact that larger lesions are associated with poorer outcome [
12]. Comparing the cut-offs, those for subdural and epidural haemorrhage in the AIS dictionary are larger than those in the Marshall Classification by 25 cc although this difference may be negligible for contusion and intracerebral haemorrhage which is only 5 cc. The evidence base for lesion size in both classifications appears to be limited, despite claims that the cut-offs are backed by substantial experience and are not merely arbitrary [
5,
13].
In Table
3, we assumed that codes indicating hypoxic or ischemic brain damage are related to normal CT scan. This may not always be the case as some patients may develop brain swelling as a secondary damage to hypoxia/hypotension. Whilst our assumption of normal CT for hypotension/hypoxia may not be acceptable in our cross-tabulation, we believe the algorithm would address this problem. For example, if a patient develops brain swelling following hypoxia, then two codes of hypoxia and brain swelling will be allocated. As the injuries are taken through the algorithm, the brain swelling Marshal Class of III or IV (depending on the severity) is allocated prior to the final step of the algorithm as Marshal Class I (normal CT).
The position of various steps of the algorithm is based on the assumption that the Marshal Classification is ordinal in severity. Despite being the case from Class I to IV, this is not true for class IV versus class V as patients in class IV demonstrate lower likelihood of favorable outcome or survival than those in class V [
5,
14]. Since the Marshall Classification is mutually exclusive, it is conspicuously necessary to prioritise which type of injuries is more relevant for allocation of the proper Marshal Class in case of multiple brain injuries. Whilst, according to the adverse outcome frequency, the brain swelling with compressed cisterns may have to be placed prior to mass lesion in the algorithm, we believe the current position of each step is more reflective of what occurs in real life of Marshal Classification through observing the actual CT. In fact, the current position of various steps of the algorithm are according to what has been suggested by Mass et. al. [
5]. In Mass's algorithm, the Marshal Classification was taken as ordinal but still class V represented lesser degrees of brain injury than class IV in their subsequent prognostic analysis.
Implication
The Marshall Classification has prognostic value to make predictions on the outcome of the TBI patient [
4‐
7]. AIS coding is also important from prognostic viewpoint [
15] but the severity scores (ranging from 3 to 6 in TBI) encompass a wide variety of different injuries that the relationship of the score and CT findings can not be easily made particularly in clinical settings where the AIS dictionary is not a familiar tool. Hence, it is important for trauma registries to ` avoid exclusive reliance on AIS coding for the sake of better communication with the clinical audience. As the Marshall Classification holds comparable prognostic value to age, GCS, pupillary reactivity, SAH etc. [
4,
7] and trauma registries commonly do not have record of this classification, the translation of AIS codes to the Marshall System opens up the possibility for multivariate prognostic analysis of large series of TBI subjects saved in trauma registries. In fact, the internationally known IMPACT prognostic models [
7] in TBI employ the Marshall Classification for outcome prediction and using our proposed translation not only permits running the IMPACT models in trauma registries, derivation of new prognostic models including the Marshall Classification becomes feasible. Furthermore, as other TBI series accrued in clinical studies (observational or clinical trials) often do not have AIS coding, our proposed translation facilitates mergence of datasets from trauma registries and clinical studies to conduct more powerful studies or performance of comparative analysis across datasets when data recording is not uniform.
Future direction
The design of the algorithm is such that at the end of the allocation, there must be no cases left with no Marshall Class assigned. We tested this in a dataset of 802 TBI cases from the Trauma Audit Research and Network (TARN) with positive results (unpublished data). However, we acknowledge that our proposed allocation still requires three possible forms of validation in the future. First, it is yet to be determined how accurate our method is when the Marshall Classification is performed using the AIS codes. In this manner, AIS codes are applied as substitutes for CT reports and in case all the assumptions are followed, 100% accuracy should be met. The second form of validation is when the allocation is performed with actual CT images at hand. In this manner, the allocations are compared across two groups. In one group, the classification is done through observing the CT and in the other group the Marshall Class is obtained following assignments of AIS coding and subsequently using our proposed cross-tabulation and algorithm. This form of validation is not expected to yield 100% accuracy and it examines how strong the assumptions are. The third form of validation is to compare the Marshall Classification at certain time point with that collected from AIS codes obtained from any available source including CT, MRI, operation notes etc. This form of validation would examine the influence of multiple sources of information or the temporal effect of events on the cross-tabulation and algorithm.