Using Scottish Administrative Data to Predict Mortality after Coronary Artery Bypass Graft Surgery

1
Using Scottish Administrative Data to Predict
Mortality after Coronary Artery Bypass Graft
Surgery
John Quinn, David N Clark, Adam Redpath Heart
Disease Stroke Programme, ISD
Scotland Saturday 19th September 2009 Exploiting
Existing Data for Health Research St Andrews
University
2
CABG Mortality Modelling

• ISD CHD Stroke Programme
• Coronary Artery Bypass Graft (CABG)
• Motivation
• Scottish Medical Record Linkage System
• Dataset Design
• Comparison with Published English Data (Aylin)
• Optimal Multivariate Model
• Summary Further Work

3
ISD CHD Stroke Programme Aims Objectives

• The Programme aims to provide a "patient-centred"
  data and information service to support the drive
  for improvements in care and services for
  patients with coronary heart disease or stroke.
• To lead the work with NHS QIS and national
  advisory groups in developing information
  systems to monitor and improve quality and
  performance of clinical services.
• To support the implementation of the new
  e-cardiology and e-stroke record.
• To provide a comprehensive, accessible and
  definitive information service to the SG, NHS
  HBs, MCNs and clinicians and the public.
• To develop research capability using routine
  data.
• Diversify into new areas- e.g. person level
  primary care data and community services for
  stroke.

4
CABG

5
Background/Motivation
6
Background/Motivation (contd)

• Research published by Aylin et al (BMJ 2007)
  based on English administrative data
• Concluded hospital episode statistics provide
  similar predictive discrimination to detailed
  clinical datasets
• Can Scottish data produce similar robust models?
• Can these be used to adjust outcome measures for
  case-mix?

7
Scottish Linked Acute Dataset All available
records back to 1981

• Inpatients and day cases
• Discharges from non-obstetric specialties
• Cancer registrations
• GRO death records
• Database content
• Patient identifiable
• Demographic socioeconomic
• Episode management
• Clinical
• Clinical support information
• Diagnoses (ICD-9 / ICD-10) 6 diagnostic
  positions
• Operations (OPCS-3 / OPCS-4) - 2/4 pair codes

Hospital Discharges SMR1 Psychiatric Inpatients
SMR4 Scottish Cancer Register SMR6 GRO Death
Records
8
Analysis Dataset Design

• 10,171 CABG main operation admissions in Scottish
  NHS hospitals (2003-2007)
• Excludes admissions with valve operations, PCI
  rescues, and other complicating procedures.
• Demographics Age, Sex, Area of Residence,
  Deprivation, Urban-Rural Classification
• Clinical/Episode Management Admission Type, Day
  of Admission, Revision CABG, Number of Arteries
  Replaced

9
Analysis Dataset Design (contd)

• Prior Admission (within past 90 days, including
  index) for AMI
• Prior Admissions (within past year) for IHD
• Prior Admissions (within past 5 years)
  Emergencies, Total Bed Days, Charlson
  Comorbidities (specific diseases index score )
• Prior Operations (since 1989) CABG, Heart
  Operations
• Outcome Death on discharge and/or within 30 days
  of admission

10
Comparison of Models
Logistic Regression was implemented to model
deaths occurring within 30 days or on discharge
from hospital.
Aylin Study
ISD Study

• 5 year 2003 2007
• 10,171 CABGs
• 208 (2.05) Deaths

• 8 year 1996/97 2003/04
• 152,523 CABGs
• 3,247 (2.13) Deaths

• Simple Model (Year, Age, Sex)
• Intermediate Model (Admission type, Deprivation
  category, Charlson comorbidity score, Previous
  emergency admission)
• Complex Model (Revision Procedures, Previous IHD
  Admission, AMI Admissions in the past 5 years,
  Previous Heart Op and Number of arteries
  replaced)

11
Comparison of Results

• The results of the study were very similar to
  those produced by Aylin. The models were compared
  against one another using the receiver operating
  characteristic (ROC) curve scores (c statistic).

Aylin Study
ISD Study

• Simple Model
• c statistic 0.67
• Hosmer-Lemeshow p0.75

• Simple Model
• c statistic 0.69
• Hosmer-Lemeshow p0.55

• Intermediate Model
• c statistic 0.72
• Hosmer-Lemeshow p0.01

• Intermediate Model
• c statistic 0.75
• Hosmer-Lemeshow p0.63

• Complex Model
• c statistic 0.77
• Hosmer-Lemeshow p0.13

• Complex Model
• c statistic 0.77
• Hosmer-Lemeshow p0.60

UK national adult cardiac surgical database c
statistic 0.78
12
Intermediate model Odds Ratios
13
Intermediate model Odds Ratios contd
14
Complex model Odds Ratios
15
Significant Predictors of Mortality
Multivariate Model
A full forward stepwise model was used to find
the significant variables.

• Charlson Co-Morbidity Index
• Admission Type
• Age
• Revision Procedures

• AMI Admissions in the past 5 years
• Not having a IHD admission within the last year

16
Stepwise optimal model Odds Ratios
17
Stepwise optimal model Odds Ratios contd
18
Optimal Model Multivariate Stepwise Model
Stepwise Model c statistic 0.76 Hosmer-Lemeshow
9.899 (0.272)
0.758
19
Conclusion Future work
The similarities between the two studies suggest
we are able to use the Scottish linked hospital
admissions and deaths data as an effective tool
to adjust hospital outcomes of CABG mortality for
case-mix. Further cross-validation required.
Write up a paper on the methodology and
results. Produce regular case-mix adjusted CABG
mortality by hospital.