Outcome prediction model for severe diffuse brain injuries: Development and evaluation
Background: Intensive care resources for the management of severe diffuse brain injury patients (SDBI) are limited. Their optimal use is possible only if we can predict at admission which patients are unlikely to improve. Aims: To develop a simple and effective model to predict poor outcome in patients with SDBI in order to help guide initial therapy. Material and Methods: The prognostic factors and outcomes of 289 patients with severe diffuse brain injury (GCS 3-8) were analyzed retrospectively. The prognostic factors analyzed were age, mode of injury, GCS at admission, pupillary reaction, horizontal oculocephalic reflex, and CT scan findings. Outcome at 1 month was classified as unfavorable—death or persistent vegetative state, or favorable—improvement with or without some disability. A stepwise linear logistic regression analysis was used to identify the most important predictors of poor outcome. A prediction model (NIMHANS model-NM) was developed using these factors. NM and several currently available outcome prediction models were prospectively applied in a separate group of 26 patients with severe diffuse brain injury managed with a different protocol. Results: The most important predictors of poor outcome were found to be the horizontal oculocephalic reflex, motor score of GCS, and midline shift on CT scan. NM was found to be more sensitive (75%) and specific (67%) than most other models in predicting unfavorable outcome. NM had high false pessimistic results (33%). Conclusion: Prediction models cannot be used to guide initial therapy.
The management of severe diffuse brain injury (SDBI) patients demands the dedication of expensive and limited intensive care resources for considerable lengths of time. The optimal use of such resources is possible if we can predict at admission which patients are unlikely to improve. An ideal prediction model should be easy to use and should have a high sensitivity and specificity even when used on patients managed by different protocols and at different times and places. Common predictors of outcome that have been used both individually and in combination include age,,,,, Glasgow Coma Scale Score (GCS),,,, pupillary reactivity,,,, early hypoxia and hypotension,,, brain stem reflexes,,,, and CT findings.,, However, in spite of using various combinations of predictors no model has satisfied all the requirements of an ideal model. This retrospective analysis was conducted in order to evolve a model to predict the outcome in a patient with SDBI in the emergency room and test its efficacy by prospectively in a different set of patients.
This study is based on the retrospective analysis of 289 patients admitted to NIMHANS between 1993 and 1998 with a post-resuscitation GCS12 of 8 or less and with CT scan of head showing evidence of diffuse brain injury without any operable mass more than 1 cm in diameter. Patients with midline shift of more than 5 mm and a significant mass effect due to a hematoma or contusion, even if slightly less than 1 cm thick were excluded because they were considered operable lesions. However, patients with significant midline shift due to asymmetric hemispheric edema or with minor tissue tear hemorrhages were included. Patients with hypotension at admission, injury to other organs, and those who were brain-dead at admission were excluded. Patients who did not survive for 24 hours and those without records of outcome at 1 month were also excluded. The factors analyzed were age, mode of injury, pupillary reaction, horizontal oculocephalic reflex, and GCS score. The CT scans of the patients were reviewed and effacement of basal cisterns and ventricles, presence of midline shift based on the position of the third ventricle, subarachnoid hemorrhage, and tissue tear hemorrhages were analyzed. Ventricles were said to be effaced if the lateral ventricles were not visualized and the basal cisterns were said to be effaced if the ambient cistern was obliterated bilaterally. All patients were managed with anti-edema measures like mannitol and glycerol. Steroids were not used. They were mechanically ventilated in the presence of respiratory embarrassment or if CT scan showed evidence of severe diffuse edema with or without midline shift. Intracranial pressure (ICP) was not monitored and they were nursed with the head end elevated. The outcome at 1 month was assessed using the Glasgow outcome score (GOS);13 patients who improved with or without some disability were categorized as having a favorable outcome, whereas those who died or remained in a persistent vegetative state were said to have an unfavorable outcome. All these factors were then analyzed individually using the Chi-square test and the independent - samples T-test in order to determine if they were significantly different in patients with a favorable outcome as compared to patients with an unfavorable outcome. Then a step-wise linear logistic regression analysis was used to identify the most significant factors. An equation was developed based on the relative significance of these factors to predict unfavorable outcome.
The equation developed (NIMHANS model) was subsequently applied prospectively in similar group of 26 patients who were managed using the Cerebral Perfusion Pressure (CPP) management protocol14 at NIMHANS between January 1999 and February 2000. They were all sedated and mechanically ventilated to maintain a targeted PaCO2 of 30-35 mmHg and PaO2 of 100 mmHg. All patients underwent central venous pressure (CVP) monitoring and maintaining the CVP around 10-12 cm H2O ensured adequate volume status of the patient. Mean arterial blood pressure (MAP) was monitored in all patients using an intra-arterial transducer in the radial or dorsalis pedis artery. ICP was monitored using an intra-ventricular catheter connected to a closed CSF drainage system. The patients were nursed in the flat-supine position. The CPP was calculated using the formula: CPP = MAP - ICP. The ICU nurses did the calculation and recording of MAP, ICP and CPP at half-hourly intervals. Constant attempts were made to keep the CPP above 70 mmHg using a step-wise increase in the intensity of therapy after ensuring the adequacy of sedation and euvolemia. ICP was first decreased by CSF drainage from the ventricle followed by the administration of bolus doses of mannitol up to a maximum dose of 200 g /day and finally by raising MAP using an intravenous infusion of Dopamine (5-15 µgm/kg/min). The predicted outcome of these 26 patients, using the NIMHANS model, the Narayan Logistic model (NLM),9 the Choi's Logistic model (CLM),1 Klauber's Logistic model (KLM),8 Glasgow-Liege model (GLM),2 and the Choi Classification and Regression model (CRM)6 was compared to the actual outcome at 6 months. The sensitivity, specificity, predictive value of unfavorable outcome, predictive value of favorable outcome, percentage of false optimistic outcome and percentage of false pessimistic outcome among the different models were calculated and compared.
The influence of the clinical factors on the outcome is shown in [Table - 1]. The majority of the patients were young adult males-the mean age of patients who had an unfavorable outcome was 31±16 years while that of patients who had a favorable outcome was 26±10 years and this difference was significant (t-test, P = 0.01). Almost all the patients older than 45 years (91%) had an unfavorable outcome as compared to those younger except for those younger than 10 years. Most of the patients were involved in road accidents, with only 61 patients being injured otherwise. The nature of trauma was not a significant predictor of outcome. The GCS sum score as well as the individual scores were evaluated as outcome predictors. [Table - 1] shows that the GCS sum score, the motor score and the verbal response score were all highly significant predictors of poor outcome. Ninety-six per cent of the patients with absent pupillary light reflex were found to have a poor outcome. The horizontal oculocephalic reflex when absent was also found to be a significant predictor of poor outcome, with 98% having a poor outcome.
The CT scans of 271 of the 289 patients were available for review and the result of this analysis is shown in [Table - 2]. Ninety-four per cent of the patients in whom the ventricles were effaced had an unfavorable outcome. A greater degree of midline shift on CT scan was associated with unfavorable outcome. Effacement of the basal cisterns also signified a poor prognosis. Traumatic subarachnoid hemorrhage (SAH) was present in 116 patients and their outcome was slightly worse than the patients without SAH. Presence of tissue tear hemorrhages did not significantly influence outcome.
When all these factors were evaluated using step-wise logistic regression analysis [Table - 3], the most significant factors were found to be the horizontal oculocephalic reflex (OCR), motor score of the GCS (MGCS) and the presence of midline shift on CT scan (MS).
In order to develop an outcome scoring system each of these factors was coded as follows:
OCR - Oculocephalic reflex - Absent: 1, Present: 2
MGCS - Motor score of GCS: 1 to 5
MS - Midline shift score- Absent: 1, <5 mm: 2, > 5mm: 3
The following prediction score was then developed based on the relative importance of the predictors.
Prediction Score = (3 x OCR) + (0.5 x MGCS) - (MS) - 6.6
If the value of the prediction score is less than zero then the outcome is likely to be unfavorable.
[Table - 4] shows the sensitivity, specificity, predictive value of unfavorable outcome, predictive value of favorable outcome, percentage of false optimistic outcome and percentage of false pessimistic outcome among the different models that were evaluated. It is obvious from this table that Narayan's Logistic model is the most sensitive prediction model followed by the NIMHANS model, though the latter scores over the former in terms of specificity.
This study has highlighted the fact that both clinical and CT scan findings are important in the prediction of outcome following severe diffuse brain injury. The factors which were found to have a significant bearing on the outcome at 1 month included GCS, pupillary reaction, oculocephalic reflex, and CT scan findings of effacement of ventricles and basal cisterns, presence of SAH and midline shift. Three of these, namely, oculocephalic reflex, motor score of GCS and the extent of midline shift on CT scan were found to be the important predictors and could be combined to develop a simple outcome prediction model. This model was externally validated and compared with other models, and was found to be moderately efficient.
The outcome following SDBI varies widely; from those whose scans are essentially normal to those with severe diffuse cerebral edema with midline shift. Vegetative states have been reported to be more common among patients with SDBI. Such patients are more likely to utilize limited intensive care resources for long periods of time. Hence, it becomes imperative that neurosurgeons dealing with them are able to prognosticate reliably. The model in this study was developed using the analysis of retrospectively collected data. The exclusion of patients who did not have outcome data at 1 month post-injury and poor follow-up data in the records introduced a bias in the selected cases. This bias resulted in the selection of a larger numbers of patients with more severe injury that resulted in mortality or a vegetative state in the hospital, because their outcome was readily available from the records. On the other hand, patients with less severe injury that led to an early discharge from the hospital were excluded to a greater extent because of the lack of follow-up data. This skew in the patient selection was responsible to a large extent for the high level (72%) of unfavorable outcome in this set of patients. The prospectively studied group, on the other hand, had complete follow-up and their level of unfavorable outcome was only 30%. We can only speculate whether the model would have been more accurate had it been derived for a more balanced data group. Another factor that can influence the outcome is the variation in the initial treatment given to a patient in a poor neurological grade based on the clinical bias of the treating surgeon. We excluded patients who died within 24 hours of injury to avoid such a bias. In both the retrospective and the prospective groups, the treatment intensity was greater for those patients with greater midline shift, and effaced cisterns or ventricles or with a worse GCS score. Since these are the markers of an unfavorable outcome, it is unlikely that inadequate treatment would be responsible for the unfavorable outcome. The reason why most prognostic models are not popular is that while they are effective and accurate in the group of patients from whom they were developed, they usually fail when applied to patients managed differently or at different centers. This study has attempted to overcome this flaw by verifying the efficacy of the model prospectively in an unrelated group of patients managed using a different protocol of treatment compared to the group from which the model was developed. Ideally, the validation of a prediction model should involve a large number of patients from different places and times. In that respect, the present study has not been comprehensively validated.
Marshall et al developed a grading system based on CT scan alone which helped predict outcome. The model developed in this study gives an overall picture of the status of the patient by combining the most significant clinical and CT factors in a simple and easy-to-apply manner. In Marshall's grading, Diffuse Brain Injury types I to III had midline shift ranging from 0 to 5 mm. Only type IV injury had midline shift more than 5 mm. We found that among all the CT scan features which Marshall et al used to derive the classification, midline shift is the most important factor that influences the outcome. Therefore we have used 3 categories for the midline shift score to further enhance its utility and significance in the model. The importance of midline shift in predicting the outcome following SDBI was also reported by Levine et al. Fearnside et al found that midline shift was an important predictor of mortality along with intra-ventricular blood and cerebral edema. Wardslaw et al reported that while the presence of SAH and the “overall appearance of the scan” (severe focal or diffuse injury as opposed to normal/mild/moderate injury) were very useful in predicting outcome, midline shift on CT scan did not have much significance. In children, the value of a CT scan in predicting outcome is controversial. Prasad et al found the number of lesions on the first CT scan to be of predictive value, but Pillai et al reported that an early or a single CT is not of predictive value in children with SDBI.
Increasing age, especially beyond 40-55 years, has been included as a valuable factor in several prediction models.,,,,, However, in children, age was not found to be a major predictor of outcome., Although age was an important predictor of poor outcome beyond 45 years in our study group, it did not form part of the final model probably because children and adults were combined together. The GCS, pupillary reaction and/or oculocephalic reflex form a part of most prediction models in both adults and children.,,,,,,,, Mamelak, interestingly, unlike in the present report, found that the oculocephalic reflex had predictive value only at 24 hours and that 24-hour data resulted in more accurate predictions for poor outcome as compared to earlier predictions. The motor score of GCS (even more than the total score) has been found to be a very good prognosticator by some authors and this has been corroborated in the present study.,
In the validation part of the study, the NIMHANS model achieved a sensitivity of 75% and a specificity of 67%. While this degree of sensitivity and specificity is far from ideal it is much better than Choi's logistic model (CLM), Choi's regression model (CRM), the Glasgow-Liege model (GLM), and Klauber's logistic model (KLM). Both CLM and CRM are simple to apply because they use relatively few factors-age, motor score of GCS, pupillary reaction in both and the presence of intracranial lesion in the latter-combined together as a set of graphs and a flow chart respectively. The GLM only requires the Glasgow-Liege score and age, but it uses a set of complex calculations that act to its disadvantage. The KLM is very unwieldy with a large number of factors and calculations. These 4 models, despite having a high predictive value for unfavorable outcome, are inadequate for routine use because of a poor sensitivity. Narayan's logistic model (NLM),on the contrary, has an excellent sensitivity of 86% and its predictive value of unfavorable outcome is acceptable at 71%. However, the high rate of false pessimistic results of 42% and the requirement of complex calculations prevents it from being put to routine use at the bedside. The chief advantage of the NIMHANS model is good sensitivity combined with ease of calculation using only 3 factors. Like the NLM, this model cannot be used to make a decision regarding withdrawal of treatment because its rate of false pessimistic results is 33%. This study therefore is in agreement with Waxman et al that prediction models cannot be used to decide on the initial course of treatment for a particular patient. In the Indian context, a similar model, though using clinical variables alone, was developed in a pilot study by Mukherjee et al. They too concluded that “the calculated prediction should not cloud clinical judgment”. The chief utility of efficient models like NLM and the NIMHANS model would be in the rational utilization of limited resources and during counseling of the relatives of patients.
The absence of the horizontal oculocephalic reflex, a poor motor score of GCS and the presence of midline shift on CT scan are the most important factors predicting poor outcome in patients with severe diffuse brain injury. These factors when combined in the equation developed in this study can predict poor outcome with 75% sensitivity and can be applied at the bedside in the emergency room after initial resuscitation. Prediction models cannot be used to guide initial therapy because of high false pessimistic results.