Transcript Slide 1

West TA, Rivara FP, Cummings P, Jurkovich GJ, Maier RV.
J Trauma 2000;49:530-541.

To develop a scoring system to better
estimate probability of mortality on the basis
of information that is readily available from
the hospital discharge sheet and does not
rely on physiologic data

There have been several attempts to develop a
scoring system that can accurately reflect the severity
of a trauma patient’s injuries, particularly with
respect to the effect of injury on survival

Current methodologies require unreliable physiologic
data for the assignment of a survival probability and
fail to account for the potential synergism of different
injury combinations

Four problems with current models:
1. Information often lost – physiologic/anatomic data combined into
intermediate scores, then combined to achieve final probability
of survival score
2. Injuries modelled as if effects are independent – but some
combinations more lethal than models predict
3. None account for pre-existing disease – widely acknowledged
contributor to outcome
4. Pre-hospital/emergency department physiologic data often
missing – making probability calculation of survival impossible

This injury severity classification method
attempts to explicitly address the possibility that
certain injury combinations might contribute to
mortality beyond their independent effects

In addition, it takes comorbid disease into
account when predicting mortality

This method uses data that are readily available
for all patients without relying on missing or
inaccurate physiologic data

Records from the trauma registry from Harborview
Medical Center (an urban Level I trauma centre) were
analysed using logistic regression

Information obtained for all trauma admissions and
emergency room deaths between 1st July 1985 and 31st
December 1997

No treatment-related variables were included in the
analysis; only those variables determined upon or before
the individual’s arrival at the hospital

Resulting data split into two roughly equal groups:
 A ‘design set’ to determine the best prediction model
 A ‘validation set’ to test the accuracy of the model on an independent set
of data
 Statistical analysis performed using Stata (College Station, TX)

ICD-9 codes representing injuries (codes 800-959.9; n = 2,034)
reclassified into 109 anatomically-similar injury categories
 ICD-9 codes that corresponded to Abbreviated Injury Scale (AIS) severity
scores <3 (e.g. minor injuries) were excluded
 Burns & burn-related injuries also excluded
Methods (2)

Included in the regression were International Classification
of Diseases-9th Rev (ICD-9-CM) codes for anatomic injury,
mechanism, intent, and pre-existing medical conditions, as
well as age.

Two-way interaction terms for several combinations of
injuries were also included in the regression model.

The resulting Harborview Assessment for Risk of Mortality
(HARM) score takes the form of a probability between 0
and 1 of in hospital mortality.
Methods (3)

HARM model compared to ICISS (ICD-9-CM Injury
Severity Score) and TRISS (Trauma and Injury Severity
Score) to discriminate between survivors and
nonsurvivors from this dataset, using an ROC curve.
 Area under the curve (AUC) calculated for each model and
compared using the Hosmer-Lemeshow (HL) statistic.

33,990 admissions recorded in Harborview Medical Center Trauma Registry
between 1/7/85-31/12/97. Excluded readmissions for same injury final data:
32,207 admissions
▪ 16,185  ‘design set’
▪ 16,122  ‘validation set’

Study population predominantly young and male. Most injuries resulted from
blunt trauma. No significant differences between design and validation sets
with respect to age, gender, mortality, etc.

Final logistic regression model contained 80 variables.
 51 injury categories included
 Six comorbid conditions included - cirrhosis, IHD, hypertension, psychoses,
alcohol/drug dependence, and congenital coagulopathy
Results (1)

HARM model calculated probability of mortality in 16,097 of
the 16,122 admissions (99.9%)
 ICISS only managed 15,820 (98.1%)
 TRISS only 9,923 (61.4%) because of missing physiologic
data

HARM had a better fit to the validation data (HL statistics =
21.37; p = 0.0315) than ICISS (HL = 712.4; p = 0.0005) and TRISS
(HL = 59.54; p = <0.005).
 NB smaller HL = better fit to actual data.

Specificity of HARM was 83.4%
 ICISS = 78.2%, TRISS = 72.1%
Results (2)

TRISS was the “standard” for mortality prediction among
trauma patients for many years, but has limitations:
 Most importantly, its applicability to patients with missing physiologic
data.

HARM score has excellent power in discriminating between
survivors and nonsurvivors, with better calibration than
either TRISS or ICISS

Comorbidities found to be important include:
 Cirrhosis
 IHD
 Congenital coagulopathy

Ten most lethal injuries according to HARM model
Independent Variable
Adjusted
Odds Radio
Loss of consciousness >24hrs (irreversible)
95.2
Full-thickness cardiac laceration
67.2
Unspecified cardiac injury
32.0
Complete spinal cord injury C4 or above
30.9
Superior vena cava or innominate vein
28.4
Pulmonary laceration
27.3
Cardiac contusion
22.4
Traumatic amputation above the knee
21.4
Major laceration of liver
14.6
Thoracic aorta or great vessels
13.5

Injury severity scores based on ICD-9 codes predict mortality
with as much or more accuracy than those based on
Abbreviated Injury Scale (AIS) scores, with considerably less
effort and expense

Further, predictive power of HARM does not require the use
of physiologic data

HARM is an effective tool for predicting in-hospital mortality
for trauma patients, outperforming both TRISS and ICDISS
with respect to discrimination and calibration, using
information readily available from hospital discharge coding,
without requiring physiologic data

Does not use physiologic data
 Two patients with the same injuries, mechanism of injury,
comorbidities and age have same score regardless of vital
signs on admission

BUT….point of HARM is to predict survival on bases
of factors established at time of injury itself.
 Avoids inherent problems of using physiologic data:
▪ E.g. often time elapsed since injury to admission to hospital is
unknown

Applicability of findings to other centres?
 Harborview Medical Center patient population may
be homogeneous when compared to other hospital
populations
 Accuracy of ICD-9 coding?
▪ Data usually coded by non-clinicians
▪ ICD-10 thought of as more accurate, with a more
comprehensive list of possible diagnoses and diagnostic
codes.

ICISS and TRISS models applied to vastly different
databases to that of HARM
 Ideally, should have derived TRISS coefficients and ICISS
risk ratios from Harborview dataset and then compared all
three models using either Harborview or an independent
dataset
 Calibration comparisons between the three models
inappropriate when underlying population mortality rates
are different (whole of N. america for TRISS, N. Carolina
for ICISS, Seattle, WA for HARM)