Traumatic brain injury (TBI) is a global
health problem with an approximate incidence of 0.2– 0.5% each year 1.  It is a leading cause of mortality, morbidity, and
socioeconomic losses in India. Approximately 1.6 million individuals sustain
TBI and seek hospital care annually in India. 2

Clinicians decide therapeutic interventions based on
their assessment of prognosis of TBI patients. Many doctors believe that an
accurate assessment of prognosis is important when they made decisions about
the use of specific methods of treatment which may be hyperventilation,
barbiturates, or hyperosmolar therapy.4 At the same time  the assessment help in deciding whether or
not to withdraw treatment. It plays as important role in counselling the patients
and relatives. 3

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!

order now

Commonly used clinical predictors of outcome of TBI both
individually or in combination are age, Glasgow coma scale score, pupillary
reactivity, extra cranial injury ,hypoxia and hypotension, brain stem re?exes .The investigational  predictors are based on CT ?ndings, like midline Shift, petechial
haemorrhages and obliteration of Third Ventricle or Basal Cisterns               4


There are many studies predicting the outcome
in severe head injuries taking into

consideration the both clinical and
investigational parameters. However there are not many

studies comparing the predictability of
clinical and investigational parameters based models.


 Clinical parameters are equally
important, particularly at the site of accident or at the place of

disaster, where investigation facilities 
are not available. We have contemplated a study to

evaluate the clinical and imaging parameters by developing logistic
regression models and a

assessing and comparing the models.



Subjects and Methods: After Institutional ethical and research committee
clearance 100 patient who had head injury at our emergency were enrolled in to
our study and followed up to 14 days.


Sample size: As suggested
by Miller and Kunce (1973) subject to predictor ratio is 10 to 1 the sample
size 100 were selected taking into consideration the clinical variable were 10


 The  clinical variables  and  CT
scan findings were recorded and transformed in to binary  data a following manner : age <40 as '1' ? 40 as '0' , female as '0'  and male as '1', systolic blood pressure (SBP) >100 mmHg as ‘1’ ? 100 as ‘0’ and diastolic 
blood pressure (DBP) >60’1′ ? 60as ‘0’ in mm Hg ,Pulse Rate (Beats/Minute) <100 as '1'and ? 100 as'0'  ,Respiratory Rate/Minute <20 as '1' and ? 20 as '0',GCS ?8  as'0' > 8 as ‘1’,Pupillary reaction to light  present as ‘1’ absent as’0′,Anisocoria
yes  as ‘1’ no as ‘0’,Extra Cranial
Injury  yes as ‘1’   no
as’0′ and CT findings , Midline Shift <5mm as '1' >5mm as ‘0’ Petechial
haemorrhages   yes  as ‘1’ 
no as ‘0’  Obliteration of Third Ventricle or Basal
Cisterns   yes as ‘1’  no as’0′, The outcome of head injury was
measured by Glasgow Outcome scale on 14th day as follows 1)
Discharges home without  neurological
Sequelae  2) Discharged home with  neurological sequelae  favorable 
as ‘1’ 3) Severe disability 4) Vegetative state 5)
Death as unfavorable as ,0, 



Initially all the variables were
analyzed and ranking was done by using the Tanagra datamining software 6
using Fisher filtering with P value of < 0.05.Table 2 The clinical variables selected were age, GCS, Pupillary reaction to light, reparatory rate, anisocoria and SBP, in CT findings petechial hemorrhages, obliteration of third ventricle or basal cisterns and midline shift depending upon 'p'value of <0.05. The clinical six variables age, GCS, pupillary reaction to light, respiratory rate, anisocoria  , SBP and the Imaging variables midline shift , petechial haemorrhages, obliteration of third ventricle or basal cisterns were taken and two models namely clinical and imaging were developed using binary logistic regression analysis. The two models were assessed by sensitivity, specificity and accuracy. Hosmer -Lemeshow test is used for calibration of models. (fig 1 1nd 2) and discrimination by AUC (area under the curve) of ROC (receiver operating characteristic curve) curve .9Fig3) Table3. The  ability  to  predict  correctly  is  one  of  the  most  important  criteria  to  evaluate  classifies in supervised  learning.  The preferred indicator is the error rate (1 ? accuracy rate).  It states the probability of misclassification of a classifier. The first estimator, the simplest, is the resubstituting error  rate. It calculates  the  percentage of misclassified  on the training  set  that  which  were  used  to  learn  the  classifier.  The resubstitution error rate is highly optimistic.  It underestimates the true error rate because the same datasets were used for training and testing the classifier. This  drawback  can be alleviated by splitting  the  data  into  2  parts:  the  first  called  training  (or  learning)  set  (e.g.  2 / 3) is used to create the classifier; the second, called the test set (e.g. 1 / 3), is used to estimate the error rate. It is unbiased. we have divided the datasets , into 60 % to train and 40% to test to find out the true error rate . (Table  5) when  we  deal  with  a  small sample  size,  dedicating  a  part  of  the  dataset  to  the  test  phase  penalizes  the  learning  phase,  and moreover the error estimation is unreliable because the test sample size is also small.   Thus,  in  the  small  sample  context,  it  is  preferable  to  implement  the resampling  approaches for error rate estimation. Like the cross validation (CV), bootstrap (boot). All of them are based on the repeated train test process, but in different configurations. The aim is to evaluate the error rate of the classifier created on the whole sample.  Along with the resubstitution error rate, the data was divided into train and test and the error rate calculated on the test datasets, cross validation (CV) and bootstrap (boot) were used in the two models. Table4 Finally the AUCs of the two model were compared by using Z test Table 4. (Ref NO 1) Table 5. 7 Results: Mean age the patients was 32.2(Standard Deviation=12.62) and ranging from 12 to 70 years. Male (88%) population had head injury more than female patients (12%). The CT findings and the diagnosis is shown in the table 1. Estimated Logistic Regression clinical Model for Neuro Outcome 1 = -2.73 + (0.028 x Age) – (0.99 x Anisocoria) + (4.47 x GCS) +  (0.50 x Pupil Reaction) + (0.31 x RR) + (0.94 x SBP) Estimated Logistic Regression Imaging   Model for Neuro Outcome 1= 3.21 + (0.63 x CT. Medline shift) – (3.30 x CT Petechial_Hemo) – (2.61x Obliteration of 3rd Ventricles or basal cisterns) The sensitivity, specificity, accuracy and AUC of ROC of the clinical model are  98%,76% ,94% and 0.94 respectively.  The sensitivity, specificity, accuracy and AUC of ROC of the imaging model are 85%, 100%, 99% and 0.91 respectively.  The sensitivity, accuracy and AUC of ROC are higher in clinical model whereas the specificity is more with imaging model. (Table 3.Fig1, 2, 3). The Hosmer –Lemeshow test was not significant in both the models, confirming good calibration in both the models. Thus both models have good calibration and discrimination.    The supervised learning of the both models were measured by error rate otherwise called internal validation. The Resubstitution error rate, and the test sample error are (TRUE) were similar in both the models. The error rates with resampling methods like cross validation and bootstrapping in both the model found have minimal difference. (Table)   When compared the AUCs of both the models by Z test, there is no statistical difference found (p>0.05) (Table).



Traumatic brain injuries (TBI) are a
real social problem because of industrialization and motorization 1. Two
million individuals each year sustain traumatic brain injury in the United
States, resulting in 56,000 deaths 2.

It is becoming a major cause of
death and disability. Establishing a reliable prognosis after injury is
difficult. On the other hand, clinicians treating patients often make
therapeutic decisions based on their assessment of prognosis. Many prognostic
models have been reported but none are widely used.7

7. Perel P,
Edwards P, Wentz R, Roberts I. Systematic review of prognostic models in
traumatic brain injury. BMC Med Inform Decis Mak 2006; 6:38.


We have developed two
prognostic models one is dependent on only clinical variable and other investigational
variables of CT imaging for predicting two clinically relevant outcomes in
patients with traumatic brain injury.. The models have excellent discrimination
and good fit with both internal validation.

The clinical model
variables were similar to crash study, however in our study the respiratory
rate, anisocoria, systolic blood pressure found to have significant effect on
TBI outcome, which were not there in many of the outcome studies of TBI. The
systolic blood pressure and respiratory rate were very good predictors in many
studies in trauma patients.8,9,10  The
extra-cranial injury which is one of the clinical factor in crash study, but in
our study it was not significantly effecting the  outcome (p value =0.944).

The clinical Model
sensitivity, accuracy. AUC of ROIC curve in fact are better comparted to

Imaging model. The error
rates by resubstitution, test set, cross validation and bootstrap
methods are almost similar .However the specificity is better with imaging
model. When compared with Z test there is no significant difference between the
two models. Thus the clinical model gives an equally good estimation of
prognosis particularly at site of injury or any natural disaster.


Limitations: Our
study is not a multicentre study. The external validation was not performed in
a different location or a centre.



I'm Krystal!

Would you like to get a custom essay? How about receiving a customized one?

Check it out