Development and validation of a model for prediction of mortality in

Report
Neophytos Stylianou






First thought that mortality must be related
to Burn Surface Area (BSA) 1860
Mortality Prediction Models for Burn Injury
exist since 1961 (BSA + Age)
1982 inhalation injury was incorporated in
model (ABSI model)
More than 40 models exist now
In the UK only 3 were developed
Only 3 models used nationwide data world
wide



Predict the probability of an outcome for a
condition given a specific amount of input
data
Various types of models eg. ANN, Logistic
regression
A model should be:
◦
◦
◦
◦
Based on objective criteria
Accurate and reliable
Easy to use
Should be dynamic







Burn lead to premature deaths
Reduction in mortality is good endpoint
Well defined
Good surveillance coverage
Easily measured
Change is easily detected
Other outcomes:
 LOS
 Functional status
 Quality of Life





Mortality in burn injuries has dropped
significantly in the last decades
They can aid in clinical decision making
Quality control/performance indicator
Burn management is one of the most
expensive conditions to treat
Resource allocation







BOBI model
Refined Ryan model: TBSA, Age, Inhalation
injury
Belgian nation wide data from 1999-2004,
6227 patients
1999-2003 data was used to derive the
model 5246 patients
Validation:2004 data of 981 patients
AUC:0.94 (CI: 0.90-0.97)
Calibration:0.452 not on the published model
Risk factors
Score
Age (years)
TBSA (%)
Inhalation injury
0
<50
<20
No
1
50-64
20-39
2
65-79
40-59
3
>80
60-79
4
Predicted
mortality
(%)
YES
>80
0
1
2
3
0.1
1.5
5
10
Total score
4
5
6
20
30
50
7
8
9
10
75
85
95
99






Wide categorisation of BSA
Does not compare with continuous variables
Derivation of scoring system
Arbitrary scale-up/down of predicted
probabilities
H-L test based on continuous model and not
categorised
No logistic regression formula published

Data from iBID 20032011
no of observed predicted
Score patient mortality mortality
s
(%)
(%)
52,942 0.00
0.01
0
6,865
0.01
0.015
1
3,173
0.03
0.05
2
2,315
0.10
0.1
3
400
0.33
0.2
4
211
0.43
0.3
5
165
0.56
0.5
6
92
0.88
0.75
7
34
1.00
0.85
8
7
1.00
0.95
9
1
1.00
0.99
10
Total
66,205
0.01
BOBI model

Since no logistic
equation the model
published it had to be
recreated
Recreated BOBI
Odds ratio
P>z
Odds Ratio
age<50
1.0
0
1.0
age 50-64
8.7
0
4.4
0
age 65-79
15.6
0
22.8
0
age>=80
66.4
0
85.7
0
BSA <20
1.0
0
1.0
BSA 20-39
4.1
0
28.1
0
BSA 40-59
11.6
0
94.4
0
BSA 60-79
27.0
0
257.2
0
BSA >=80
Inhaltion injury
abscent
135.2
0
875.3
0
1.0
0
1.0
inhalation present
6.8
0
7.6
0
0.0
0
_cons
P>z
1.00
0.75
0.50
0.25
0.00

AUC:0.96 (CI 0.95-0.96)
H-L(4) 4.88 P>x2 0.300
Sensitivity

0.00
0.25
Area under ROC curve = 0.9565
0.50
1 - Specificity
0.75
1.00




Our model: Age+Age2+TBSA(categorised in
10%)+inhalation injury + number of existing
disorders + type of burn injury
AUC 0.971 (CI 0.965-0.977)
HL(10) 7.02 P>x2 0.7235
Comparing the two gave a x2 (1) of 31.4 thus
the models are different
Any questions?

similar documents