Transcript Risk scoring - Cardiff PICU

Risk scoring
Allan Wardhaugh
Why bother?
• Comparison of performance between units
• Used in RCT to adjust for case-mix
Standardised Mortality Ratio
• Measured mortality
• Predicted mortality – risk adjustment tool
• SMR = Measured/ Predicted
–  SMR > 1
– SMR < 1
performing poorly
performing well
Making a risk adjustment tool
Regression statistics
• Target variable – ‘dependent variable’
• Predictors – ‘independent variables’
• Regression statistics use association between
•
variables to predict one (DV) from another (IV).
Simplest form y = b0 + b1(x)
where y = predicted value, b0= regression constant, b1= regression
coefficient
• Multiple regression
y = b0+b1(x1)+b2(x2)+…bn(xn)
Regression statistics - logistic
• Linear multiple regression
– DV and IV quantitative
• For non-quantitative DV (e.g. dead/alive), logistic
regression is used
– Relationship with IV may be non-linear
• For each IV, odds are calculated for likelihood of having
•
DV
Odds very assymetrical
– very small number (0 – 1) if event unlikely
– very large if event likely (>1 - ∞)
• Rectified by using natural log off odds – called logit –
makes it a linear function
Regression statistics - logit
• Logit = b0+b1(x1)+b2(x2)+…bn(xn)
• Probability = odds/(1 + odds)
• Logit = ln odds
p
=
logit
e
/(1
+
logit
e
)
PRISM
• Pediatric Risk of Mortality Score
• 14 physiological variables
– Worst measurement in first 24 hours
– Now on PRISM III – relies on scores in first 12 or 24
hours
• Probability of PICU death
= eR/1 + eR
Where R = 0.207  PRISM – 0.005  age(mo) – 0.433  operative status – 4.782
PRISM – example
60 month old non-surgical patient
PRISM Score
Mortality Risk (%)
3
1.6
6
2.7
9
5
12
8.9
15
15.3
18
25.2
21
38.6
24
46.1
27
68.5
30
80.2
PRISM - disadvantages
• Data collection cumbersome (14 variables
over a 24 hour period)
• May diagnose death rather than predict it
(40% deaths occur in first 24 hours)
• Score may not allow comparison between
units – patients poorly managed in first 24
hours will develop high PRISM score, so
disease severity will appear to be greater
PRISM III
PIM – Paediatric Index of Mortality
– initial cohorts
• 678 consecutive admissions PICU
RCHM 1988
• 814 consecutive admissions RCHM
1990
• 1412 consecutive admissions 1994–5
RCHM
PIM – identifying variables
• Data collected for admission (for
most) and first 24 hours
• 34 Physiological Stability Index
measurements
• MAP, PIP, PEEP, and others
• Worst value in first 24 hours for all
PIM – derivation of model
• All PRISM data collected plus additional
information
• Univariate analysis carried out on all
factors to test for association with
mortality (Chi squared dichotomous variables, Copas
p by x plots continuous variables)
• Factors not associated (p>0.1) excluded
from further analysis
• Logistic regression analysis used to derive
preliminary model.
PIM – testing the model
• Learning and Test cohorts
– 1994 – 96
5695 patients in 8 PICUs (Australia, Birmingham)
• Enough patients in each unit to include 20 deaths.
• Learning sample data analysed to calculate regression
•
•
•
coefficients
Model then tested on test sample, and examined for
goodness of fit.
Regression coefficients re-estimated using all 8 units for
final model.
Risk of death assigned to 5 groups - <1%, 1–4%, 5–
14%, 15–29% and 30%
PIM - results
PIM – final equations
• elogit/(1+elogit)
• Logit = (2.357.pupils)
+(1.826.specified diagnosis)
+(–1.552.elective admission)
+(1.342.mechanical ventilation)
+(0.021.(SBP–120))
+(0.071.Baseex)
+(0.415.(100.FiO2/PaO2))
–4.873
UHW PICU PIM
1.2
Relative risk
1
0.8
0.6
0.4
0.2
0
<1%
1-4%
5-14%
15-29%
Probability of Death (by percentage risk)
>30%
PIM and PRISM compared
• Variables used by PIM
that are not used by
PRISM are
– presence of a specified
diagnosis
– use of mechanical
ventilation
– plasma base excess
• Variables used by PRISM
that are not used by PIM
– diastolic blood pressure,
heart rate
– respiratory rate, pCO2
– the Glasgow Coma Score
(three separate variables)
– prothrombin time, serum
bilirubin, serum potassium,
serum calcium, blood
glucose and plasma
bicarbonate
PRISM vs PIM
• PRISM predicted 66% more deaths in this
•
•
•
•
•
sample
Score altered by treatment in the first 24 hours
May diagnose rather than predict death
PRISM III data requires 96 measured variables
License required
Note that neither are adequate fro individual
case prediction – apply to populations only
PIM - recalibration
• PICU outcomes change with time
• Referral patterns change with time
• Attitudes to withdrawing and limiting care
may change with time
PIM 2
• 14 PICUs
– 8 Australia
– 4 UK
– 2 NZ
• 20 787 patients 1997-1998
• Units randomly assigned to be learning
sample or testing sample for new model
PIM 2
• PIM applied to new population (all units)
– Observed to expected deaths
• Poorly performing variables altered to make
•
•
prediction better
Re-tested by forward and backward logistic
regression to produce new model
New model applied to learning sample –
coeffciients adjusted and applied to testing
sample
Calibration findings
• Specific diagnosis
– Resp illness O:E 160:212
– Non-cardiac post-op O:E 48:82
• 293 coded diagnostic categories examined
– In-hospital cardiac arrest associated with increaed risk of death
– Asthma. Bronchiolitis, croup, obstructive sleep apnoea, DKA
associated with reduced risk
• New ‘high risk’ and ‘low risk’ categories introduced
• Post – op subdivided into with or without CBP.
• IQ <35 omitted (difficult to code reliably)
SMR
• Australia and New Zealand SMR 0.84
(0.76–0.92)
• UK 0.89 (0.77–1.00).
New coefficients
United Kingdom
Paediatric Intensive Care Outcome
Study
UK PICOS (phase I)
PIM mortality ratio (observed/expected unit deaths) by
unit. Generated using UK PICOS recalibration
2
1
.5
200
400
600
800
1000
Number of admissions
Your unit
Other units
Control limits
Mortality ratio calculated using the UK PICOS calibration of PIM in the UK.
Upper and lower control limits represent a 99.9% confidence interval around a mortality ratio of 1 based on the
UK PICOS overall mortality of 6.2%..
Phase I outcome
• PRISM III 24 hour score re-calibrated for
UK
• Performance of PIM-2 and PRISM III very
similar
• PIM – 2 recommended as model of choice
as data easier to collect
• DoH/ WAG funded
• Run from Universities of Sheffield, Leicester and
•
Leeds
First annual report March 2003 – February 2004
PELOD
• Death is relatively infrequent outcome
(6%) in PICU
– Sample sizes needed for trials need to be
large to detect different outcomes
• MODS more prevalent (11 – 27%)
– Correlates well with risk of death
– Good proxy outcome measure for risk of
death
PELOD
• Prospective study – 7 PICUS France,
Canada, Switzerland
• 18months 11998 – 2000
• 1806 patients (<18yrs)
probability of death=1/(1+exp [7·64–0·30PELOD score])
League tables
• Governments like them
• Journalists like them
• Local politicians like them
• Patients groups like them
Do any of the above understand them?
• 9 NICUs over 6 years
• Crude and risk adjusted (CRIB score)
mortality
• Hospitals ranked in league tables each
year according to W score
– W= 100  (observed - expected deaths)/No of
admissions.
– Mortality lower than expected if W < 0
Results
Conclusions
• Hospitals varied annually in their league position
• Confidence intervals for W scores overlapped for
•
all hospital every year except year 3
‘Overall, hospital 1 did perform significantly
better than expected but it is debatable whether
this makes it a model hospital since its
performance was inconsistent’.
Summary
• PIM/ PIM 2 data easy to collect
• Useful in comparing unit performance
• Interpret with care if number of deaths low
•
•
•
•
(especially <20).
Not for use as an individual prediction test
Important to complete as accurately as possible
PICANET randomly check to ensure data quality
League tables are unreliable