Transcript Dr Foster High-Impact Users Methodology

Dr Foster High-Impact Users
Analysis
• October 05: Paper supplied by Mansfield &
Ashfield PCTs
• Note Dr Foster use “High Impact” as
synonomous with “Frequent Flyer” (FF)
• Definitions
– High impact patients: basically  3 emergency
admissions / year (or 3+ spells in 12 months)
– Very high impact patients:  9 emergency
admissions between Apr 1st 2001 and Mar 31st 2003
– ACS (Ambulatory Care Sensitive) high-impact
users: as per high impact but with a PDX belonging
to an ICD classification (don’t know if this is a Dr
Foster classification ??)
ACS classification
Appendix giving ICD10 codes for Ambulatory Care Sensitive conditions
ACS group name
ICD10 codes
Description
Influenza and pneumonia
J10
Influenza due to identified influenza virus
J11
Influenza, virus not identified
J13
Pneumonia due to Streptococcus pneumoniae
J14
Pneumonia due to Haemophilus influenzae
J15.3
Pneumonia due to streptococcus, group B
J15.4
Pneumonia due to other streptococci
J15.7
Pneumonia due to Mycoplasma pneumoniae
J15.9
Bacterial pneumonia, unspecified
J16.8
Pneumonia due to other specified infectious organisms
J18.1
Lobar pneumonia, unspecified
J18.8
Other pneumonia, organism unspecified
A35
Other tetanus
A36
Diphtheria
A37
Whooping cough
A80
Acute poliomyelitis
B05
Measles
B06
Rubella [German measles]
B16.1
Acute hep B with delta-agent (coinfectn) without hep coma
B16.9
Acute hep B without delta-agent and without hepat coma
B18.0
Chronic viral hepatitis B with delta-agent
B18.1
Chronic viral hepatitis B without delta-agent
B26
Mumps
G00.0
Haemophilus meningitis
Other vaccine preventable
Other group names include:
Asthma, Congestive Heart Failure, Diabetes Complications, COPD, Angina,
Iron Deficiency Anaemia, Hypertension, Nutritional Deficiencies, Dehydration
and Gastroenteritis, Pyelonephritis, Peforated/Bleeding Ulcer, Cellulitis, Pelvic
Inflammatory Disease, Ear Nose and Throat Infections, Dental Conditions,
Convulsions and Epilepsy, Gangrene
Other Aspects of Analysis
• 01/02 to 03/04 use HES; unique ID is HESID
• 04/05: use NWCS – unique id  dob,sex,pcode
• Use routinely available data off HES like
– Age, Sex, Source of Admssn, Ethnicity
• Append other info from external datasets
– Mosaic, Deprivation Quintile (based on IMD 04),
Charlson Index of Comorbidity etc, HRG plus v3.5
tariffs (for costing purposes)
• Analysis focuses on
– Patients, spells, superspells (join up spells when
there is a transfer to another provider), beddays, cost
based on HRG tariffs, breakdown by practice
Modelling Steps
• Built a logistic regression model for 1st
admissions in 03/04
• Outcome variable = FF patient (Yes/No;
1/0)
• Predictor Variables
• These factors came out as significant
predictors of FF patients
1.
2.
3.
4.
5.
6.
Age (five-year age band up to 90+)
Sex
Mosaic type (based on postcode)
Source of admission (various categories, such as from own home, from nursing/residential home, from another hospital etc)
Deprivation quintile (IMD2004, based on super output area)
Charlson index of comorbidity (for the patient’s first spell in 2003: this is a score from 0 to 6 reflecting the number and potential
seriousness of recorded secondary diagnoses: see Appendix)
7. Whether the patient had a spell in the previous year (not counting any spell included in the three spells in 12 months of FF patients)
8. ACS condition on admission
9. Ethnic group
10. Log of the age- and sex-standardised admission ratio for the statistical ward (this is to allow for differing admission thresholds between
hospitals)
• All these predictors significant
• Most significant: previous emergency admission (not
including any spell included in the 3 spells in 12 months of
FF patients)
• Next most: Charlson index of comorbidity
• Least important: source of admission, sex
Modelling Steps
• predictive model built and validated
• Using information on the predictor variables for
each patient the probability of becoming a FF
can be calculated
• Then applied to a dataset where FF status not
known
• Choose arbitrary thresholds
– eg if probability (FF)  0.3  class as FF
– eg if probability (FF) < 0.3  class as not a FF
Results for England
• sensitivity means: if you are a FF what is the
chance the model predicts you are a FF
• specificity means: if you are not a FF what is the
chance the model predicts you are not a FF
• positive predictive value means: if the model
predicts you are a FF; what is the chance you really
are a FF