3. Can have several complex injuries
Often a relatively large number of
diagnoses/patient
A relatively large number of injured body areas
Can have high injury severity
High costs
An increased need for expensive resources
4. Classification and severity of injuries
◦ NISS- New Injury Severity Score
The NISS sums the severity scores for the
three most severe injuries, regardless of body
region
5. The Israel National Trauma Registry is maintained by
the Israel National Center for Trauma and Emergency
Medicine Research
It contains data on hospitalized patients at 10 trauma
centers in Israel- all 6 level 1 Trauma Centers in the
country and 4 regional trauma centers
◦ A level 1 Trauma Center provides total care for every aspect
of injury, and conducts research.
◦ A level 2 Trauma Center also provides comprehensive
care, but may not have all the specialties of a level 1
center, and is also not committed to conducting research.
6. Over 200 data fields are included in the
registry including demographic information
about the patients, details on the injury which
includes diagnoses (up to 20 per
patient), severity indicators, details on the
external causes of the
injury, treatment, length of hospital stay and
outcome.
The Registry, which has been maintained
since 1997, accumulates approximately
20,000 records per year.
7. Israeli hospitals are currently compensated
for trauma patients by some function of the
duration of hospital stay, according to the
average per diem rate, and not injury severity
Trauma patients incur much higher costs:
duration of hospital stay does not accurately
reflect these costs
Is preference given to patients whose
treatment will be less expensive?
9. A separate fairer classification system for
trauma patients to be used as a new
management tool
Setting up of “equal cost groups” based on
length of stay, but also taking other variables
into account e.g. those dealing with resource
diagnoses
Ability to forecast costs for trauma patients
10. Variables used to create the groups must be
based on data collected routinely by hospitals
The total number of groups should be
“reasonable” and together should include all the
patients
Each group should include patients with a similar
resource utilization
Each group has to be logical from a clinical point
of view
11. Prospective payment systems (PPS’s) for
hospital reimbursement have been
established in many countries. Integral to the
PPS is the use of pre-defined diagnoses and
diagnosis related groups (DRG’s)
But- most PPS’s based on DRG’s were not
created specifically for trauma cases
12. A system which classifies hospital cases into
1 of many groups (DRG’s) developed as part
of a PPS. Based on ICD
diagnoses, procedures, age, sex and the
presence of complications or comorbidities
Inappropriate for severe and complex trauma
cases?
13. Used Length of Hospital Stay as the measure
of resource utilization
Made definitions for multiple trauma patients
based on the primary and secondary
diagnoses
-or required two or more substantial injuries
in different body systems
-or used the full clinical profile of the patient
14. Logical choice- creation of homogenous groups
No problems with missing variables
Handles large data sets with ease
Easily interpretable=> convenient management
tool
Once trained, can be used for forecasting?
15. Usual measure in these models is length of stay (LOS)
in hospital
We constructed a new measure- “cost days” as a sum
of two components:
◦ (1) number of hospitalization days not spent in the ICU
◦ (2) 4.33 times the number of days in the ICU. The cost of a
day of hospitalization in the ICU is estimated as equal on
average to the cost of 4.33 days in other hospital units
In addition, we obtained 3 years of actual costs for
trauma patients in 2 hospitals
16. However, the distribution of “costdays” and
actual costs is highly skewed to the right-
there are many lower values and far fewer
higher values
In order to avoid creation of many sub-
groups with high values of “costdays” or
actual costs, a log transformation was used
17. Based on admittance data only
Divided into 6 groups;
◦ patient characteristics-demographics: age
◦ Characteristics of the trauma center: Level 1 or Level 2
◦ Characteristics of the injury:
The severity score- NISS
The body regions which were injured
Circumstances of injury (ecode)
◦ Diagnoses of hip fracture
◦ Diagnoses of burns
18. Vital status “at discharge”
Operations
Race
Number of ICD9 codes
19. Not readily available
For this project, real costs were evaluated for
each trauma patient in 2 hospitals (one for level 1
and one level 2) for a period of 3 years
Although costs were calculated for a variety of
hospital resources e.g. laboratory
costs, operative costs etc., we only included the
total actual cost incurred by a patient
Calculation of costs were subcontracted to an
external organization
20. Trees built for each hospital separately and
together
Training set built on a random two thirds of
the data set, and the learning set on the
remaining third
Trees built both for actual costs and costdays
Log transformation used for dependent
variables- more sensitive to lower numbers
31. Trees built on actual costs and costdays
For each terminal node, for both types of
tree, the deviance based on actual costs is
calculated;
◦ Σ(log(actual cost)-log(predicted cost))^2 and these
deviances are summed over all the terminal nodes for
each tree (sum of sum of squares or deviances)
This results in 2 measures- one from each type
of tree
If the ratio between these 2 measures is close to
1, we claim that these 2 trees are comparable
32. Each set of data was resampled 200 times, and
both types of tree built on new training sets
For each new pair of trees, comparability was
looked at by comparing the ratio of the sum of
deviances for each pair of trees tested on the
new learning set
The mean ± s.d of the ratios; both hospitals
combined; 1.016±0.023
◦ Level 1 Hospital; 1.025±0.028
◦ Level 2 Hospital; 1.009±0.031
33. T-tests for log means
2 sample t-tests were conducted between the
true means (from the training set) and the
predicted means (from the learning set) at
each terminal node. Since there are some
schools of thought that prefer to retain the
original log transformation used during
analysis, t-tests were carried out using the
log values of the costs. The results showed
some significant values.
35. The true and predicted total costs at each node were
calculated for the predicted trees
The predicted total “actual cost” or total “costdays”
were calculated by multiplying the mean number of
“actual costs” or “costdays” at each node in the
regression tree by the number of trauma patients in
the learning set who were classified at that node, and
then summing over all the terminal nodes
The ratio of true (T) to predicted (P) total costs (T/P)
was used to see how far the true costs deviated from
the predicted.
39. It is expected that deviant resource utilization in
each group resulting from untypical and extreme
cases, such as those with unusually short or long
hospital stays, costs, transfers, expensive
medical or surgical procedures, etc. will be
encountered
Excluding cases which are defined as outliers,
enhances the statistical homogeneity with the
group
Definition- Mean ± 3 x s.d.
44. An important aspect of this project is whether
the models can be used to predict future costs
To this end, 6 years of data were extracted from
the trauma registry for 8 hospitals- the first 3
years (1998-2000) were used for training the
model, and the following 3 years were used as
learning datasets
No actual cost data is available for this dataset-
costdays used as the dependent variable
49. Costdays is a good proxy for actual costs
The proposed models benefit from the
removal of outliers
The predictive ability of models for future
years must be interpreted carefully
The proposed models are a step towards
establishing a more equitable prospective
payment system