This document describes the external validation of the CRASH prediction model for mortality after traumatic brain injury (TBI) using data from the NIMHANS ICU in Bangalore, India. The CRASH model was validated on 150 TBI patients admitted to the ICU. Three variables from the CRASH model - pre-ICU Glasgow Coma Scale, obliteration of the third ventricle/basal cisterns on CT scan, and complications during ICU stay - were identified as significant predictors of mortality. The validated model demonstrated excellent discrimination and calibration for predicting mortality in this population. However, the study notes that prediction models may need to be customized for individual ICUs as patient characteristics and services change over time.
1. External Validation of a Prognostic
Model to Predict Mortality After
Traumatic Brain Injury
Dhaval Shukla*, Akhil Deepika*,
GS Umamaheshwar Rao#, DK Subbukrishna@
Departments of Neurosurgery*, Neuroanesthesiology#, and
Biostatistics@
NIMHANS, Bangalore
2. Prediction Models
• Statistical models that combine two or more
items of patient data to predict outcome
• Two requirements
– clinically valid
– methodologically valid
• More reliable than what doctors can foretell
• Influence patient management
3. Hierarchy of Prediction Models
• Univariate Analysis
• Multivariate Analysis
• Logistic Regression Analysis
• Discriminant Analysis
• Web Based Calculator
7. 102 Prediction Models of TBI
• Small Sample Size
• Logistic Regression
• 93% High Income Countries
• 11% External Validation
• 19% User-friendly
Perel, et al. BMC Medical Informatics and Decision Making 2006
10. External Validation
King: Tell me about my future
Soothsayer 1: Your all relatives will DIE in front
of your eyes
King punishes him.
Soothsayer 2: You will LIVE longest
King rewards him.
Its only human to cross check if
someone predicts bad about you
11. External Validation
• The performance of a model on a different
population
(‘generalizability’ or ‘transportability’)
• CRASH model not validated in middle/ low
income country
BMJ 2008;336:425
12. • Review of clinical and CT Scan
data of consecutively admitted
TBI patients in ICU over 6 months
DATA
COLLECTION
• Univariable Logistic Regression
Analyses (LRA)
• Multivariate LRA
MODEL
CONSTRUCTI ON
• Discrimination
• Calibration
MODEL
PERFORMANCE
• Bootstrap MethodVALIDATION
Indicates how closely predicted
outcomes match observed
outcomes
Resample from the sample
data at hand for approximating
sampling distribution of a
statistic and bias correction
Describes how well a model
distinguishes between those
who die from those who survive
0.90-1 is Excellent
13. Demographics
• Total no of patients: 150
• Male : Female :: 5.5 : 1
• Age range: 1 to 85
• Mean ICU stay: 8.3±7.2 days
• Mortality: 15.3%
• Time to death: 7.52±4.56 days
14. Variable CRASH
OR (95%CI)
NIMHANS
OR (95%CI)
p Value
Age 1.46 (1.39 to 1.54) 0.99 (0.95 to 1.04)
GCS 1.27 (1.24 to 1.31) 2.14 (1.27 to 3.60) 0.004
Pupil Reaction
Both
One
None
1
1.45 (1.14 to 1.86)
3.12 (2.46 to 3.97)
1
1.30 (0.3 to 43.7)
1.23 (0.4 to 32.66)
Extracranial Injury 1.08 (0.91 to 1.28)
CT Scan
Petechial Hemorrhages 1.26 (1.07 to 1.47) 0.81 (0.16 to 4.00)
Obliteration of 3rd
Ventricle/ Basal Cisterns
1.99 (1.69 to 2.35) 7.32 (1.27 to 42.14) 0.026
SAH 1.33 (1.14 to 1.55) 0.98 (0.23 to 4.17)
Midline Shift 1.78 (1.44 to 2.21) 0.42 (0.06 to 2.63)
Non Evacuated
Hematoma
1.48 (1.24 to 1.76) 0.70 (0.00 to 1.70)
Complications in ICU 0.04 (0.00 to 0.29) 0.001
15. Model Construction
• CRASH Variables
• 3 variables from univariate analysis
– (Pre ICU GCS, Intubation, Complication )
Result
• Pre ICU GCS (P_GCS)
• Obliteration of 3rd ventricle/ Basal Cistern (OB)
• Complication during stay (CD)
19. Internal Validation
Predicted group * OUTCOMECrosstabulation
13 3 16
72.2% 4.0% 17.2%
5 72 77
27.8% 96.0% 82.8%
18 75 93
100.0% 100.0% 100.0%
Count
% within OUTCOME
Count
% within OUTCOME
Count
% within OUTCOME
DEAD
SURVIVED
Predicted
group
Total
DEAD SURVIVED
OUTCOME
Total
Overall Accuracy 91.4%
20. Explanation
• CRASH model never validated for developing
countries
• Only ICU patients were sampled
• Many patients underwent surgery
• Pre ICU GCS
21. Conclusions
• Prediction models based on large population
studies may not be valid for a selected group
of patients
• Each intensive care should have their own
prediction models, which should be revised
when services improve
Editor's Notes
For a prognostic model to be clinically useful it
should fulfil two requirements: it must be clinically valid
and methodologically valid
Same thing happened to me.
When I ran variables of our patients in ICU in CRASH model, I go exceptionally high mortality, which did not seem true with our observation, hence I went to Dr. Subbukrishna to ask the question? Whether this model is wrong?
He said, “This model is correct, probably your patients are wrong”.
CRASH model gives double mortality for developing country
This prompted me to validate this model for our patients
IMPACT (international mission for prognosis and
clinical trial) dataset).
Resample from the sample data at hand for approximating the sampling distribution of a statistic and bias correction
Discrimination:
Calibration:
Bootstrap method : use the data of a sample study at hand as a “surrogate population”, for the purpose of