Transcript BUPA Presentation Template

Bupa Health Dialog
NHS Wales
Predictive Model Review
24th February 2010
Agenda
• Model Development Methodology
• Model Performance
• Potential Impact
• Discussion/Next Steps
Bupa Health Dialog
2
Development Methodology
Anonymised input data sets
(IP, OP, GP Practice)
50% Random ‘Development’
Sample
TEST
25% Random ‘Validation’
Sample
25% Random ‘Test’ Sample
Bupa Health Dialog
3
•
Modelling population of 298,077 patients
•
51 GP practices
•
Split sample methodology
•
Logistic regression to model unique
relationship between independent variables
in two years of patient history and
dependent variable (outcome) in third year
•
Inclusion of lag period
Operational Approach
Identify diagnoses, procedures, drugs
IP, OP, and GP
data for prior 24 months
Predict any emergency
admission
next 12 months
Dependent
Variable
Independent
Variables
Historical
Year 1
Bupa Health Dialog
Historical Lag period
3 month
Year 2
4
Year Following
Prediction
Prediction
Modelling Steps
•
Quality Check (QC) the IP, OP and GP files
•
Identify patients to include in the model building – ‘membership’
•
Identify patients with emergency admission outcome – dependent variable
•
Extract variables for inclusion in the model from IP, OP and GP files
–
~1300 variables created
•
Randomly partition the data into Development (50%), Validation (25%) and Test (25%)
samples
•
Data mining to identify significant variables for entry into the model
•
Logistic regression
•
Variable stabilisation and optimisation of model
•
Reporting of model performance on Test sample
–
Data presented in the comparative slides following are for the PRISM model vs. the Combined
Model run on the same Welsh patient test data set over the same time period to show
benchmarked model performance
Bupa Health Dialog
5
‘Member’ Selection
Creating the Model Development Population
Initial cleaning round
registration date is prior to date of birth
deceased date is prior to registration date
removed date is prior to or equal to registration date
missing pseudonymised NHS number
duplicated records
Continuous Membership
1stMay2004
History
30thApril2006
Lag
1stAug2006
Outcome
Patients were only included if they had continuous membership in any of the 51 practices
8 day gap was allowed for practice-to-practice transfer
Final membership total 298,077 from starting population of 534,955 (includes deceased)
Bupa Health Dialog
6
31stJul2007
Defining the Dependent Variable
An admission method with a HES value of (“21”,
“22”,”23”,”24”,”25”,”27”,”28” or “29”) and a patient
classification of ordinary admission (“1”)
Admission Method codes 25, 27 and 29 are
specific to Wales (i.e., not used in England)
and contribute to the higher overall
emergency admission rates than we are used
to seeing in England.
Code
Description
21
A & E or dental casualty department of the health care provider
22
GP: after a request for immediate admission has been made direct to a hospital provider (i.e. not
through a Bed Bureau) by a General Practitioner or deputy
23
Bed bureau
24
Consultant clinic of this or another health care provider
25
Domiciliary visit by Consultant
27
Via NHS Direct Services
28
Other means, including admitted from the A & E department of another provider where they had
not been admitted
29
’29’ is an internally derived code meaning Emergency Transfer from another hospital.
Admission Method “81” and Intended Management of “8”
Bupa Health Dialog
7
Variables in the Model - Demographic
Selected from ~1300 variables tested
Included Codes
Variable
Category
Baseline
Variable
Name
Description
Intercept
Demographic DEM_Age Age
Beta Coefficient
-3.1725
Bupa Health Dialog
Look
back
Read 2 ICD-10 Other
Clinical Rationale
N/A
-0.0428 Capped Current
at 100 Age
yrs
Demographic DEM_Age Age squared term 0.0005
_squ
Demographic DEM_Gen Gender ‘1’=male
der
‘0’=female
Range
Capped Current
at 100 Age
yrs
0.0906
N/A
8

Age is an important
indicator of clinical risk

Age is an important
indicator of clinical risk

Gender is an important
indicator of clinical risk
Variables in the Model - Diagnoses
Selected from ~1300 variables tested
Variable
Category
Diagnoses
Diagnoses
Variable Name
Description
Beta Coefficient
Included Codes
Range Look back Read 2 ICD-10 Variable
Category
GP_Neurosis_di Neurotic,
0.1797
sord
personality and
other
nonpsychotic
disorders
24
months

GP_Poisoning_d Poisoning
isord
24
months

Bupa Health Dialog
0.7806
9
Variable Name
Persons with mild to
moderate depression and
anxiety disorders are
significantly higher users
of health services, both for
specific mental health
issues and for issues
related to physical health
Intentional self poisoning
is associated with mental
health disorders, and thus
with service usage.
Unintentional poisoning is
associated with poor
socio-economic
circumstance, in turn a
predictor of overall clinical
risk, as well as of poorer
access to routine and
preventative services and
hence a greater
emergency services
utilization
Variables in the Model - Diagnoses
Selected from ~1300 variables tested
Included Codes
Variable
Category
Variable
Name
Description
Beta Coefficient
Range
Look
back
Read 2 ICD-10 Variable
Category
Diagnoses
GP_GI_Diso GI disorders
rder
0.1465
24

months
Diagnoses
GP_Sprain_ Sprains and strains 0.2590
disord
of joints and
adjacent muscles
24
months 
Diagnoses
GP_Mental_ Mental and
disord
behavioural
disorders
24

months
Bupa Health Dialog
0.2280
10
Variable Name
Evidence from our own
RCTs show significant
(impactable) increase in
overall services usage
from GORD and IBS.
Persons with IBD may
also face acute
exacerbations
Sprains and strains may
lead to mobility disorders
or problems in self care;
particularly among older
people
Persons with mild to
moderate depression
and anxiety disorders are
significantly higher users
of health services, both
for specific mental health
issues and for issues
related to physical health
Variables in the Model - Drug
Selected from ~1300 variables tested
Variable
Category
Variable Name
Description
Beta Coefficient
Range
Included Codes
Read 2 ICD-10 Variable
Category
24

months
Variable Name
Drug
GP_Cephalospo Cephalosporins & 0.1428
rins_rx_sqr
Cephamycins
Drug
GP_Corticostero Corticosteroid
id_rx
Clinical Use
0.2132
24

months
Drug
GP_Diuretics_rx Loop Diuretics
0.2085
24

months
CHF, CKD, Hypertension,
diabetogenic
Drug
GP_Macrolides Macrolides
_rx_sqr
0.0933
24

months
Chest infection, penicillin
allergy
Drug
GP_Analgesics_ Narcotic
rx
Analgesics
0.1696
24

months
End of life, lower back pain,
arthropathies, chronic
pain… all strong predictors
of service usage
Bupa Health Dialog
Square
root
Look
back
Square
root
11
Association with chest and
urinary infections which
can be associated with
debility or ltcs in older
people
Associated with asthma,
COPD, Inflammatory bowel
diseases etc. diabetogenic
Variables in the Model - Drug
Selected from ~1300 variables tested
Variable
Category
Variable
Name
Description
Beta Coefficient
Range
Look
back
Included Codes
Read 2 ICD-10 Variable
Categor
y
Variable Name
Drug
GP_Antid Other Antidepressant
epressan Drugs
t_rx
0.1828
24
months

Drug
GP_Penci Penicillinase Res
llin_rx
Penicillins
0.0561
24
months

Drug
GP_Sulp Sulphonamides &
honamid Trimethoprim
es_rx
GP_Ulcer Ulcer-Healing Drugs
_rx
0.1460
24
months

UTI,HIV, Crohn’s, UC
0.1049
24
months

GP_Vita
minB_rx
0.2976
24mon
ths

GORD, PUD. Significance
independent of GI disease,
possibly due to coding
issues
Alcoholism, Pernicious
Anaemia, malabsorption
Drug
Drug
Vitamin B Group
Bupa Health Dialog
12
Persons with mild to
moderate depression and
anxiety disorders are
significantly higher users
of health services, both for
specific mental health
issues and for issues
related to physical health.
Also used for chronic pain
and
neuropathies/neuralgias
Immunosupression,
nosocomial infection,
chest infection
Variables in the Model - Prescribing
Selected from ~1300 variables tested
Variable
Category
Variable
Name
Description
Beta Coefficient
Prescribing
GP_Polyp
harm
Polypharmacy
0.1518
Prescribing
GP_Polyp
harm_sq
u
Polypharmacy squared
term
-0.0062
Bupa Health Dialog
Range
Squared
13
Look
back
Included Codes
Read 2 ICD-10 Variable
Category
Variable Name
12
months

Associated with increased
adverse reactions and falls
and fractures
12
months

Associated with increased
adverse reactions and falls
and fractures
Variables in the Model – Chronic Condition
Selected from ~1300 variables tested
Variable
Category
Description
Chronic
Condition
GP_Num
_cc
Total number of
chronic conditions
0.1235
All
availabl
e GP
Data

Multiple comorbidities
greatly increase
unscheduled admission risk
Chronic
Condition
GP_Chf_
id
CHF (LTC)
0.1695

Heart failure is a strong
predictor of clinical risk
Chronic
Condition
GP_Copd
_id
COPD (LTC)
0.1053

COPD is a strong predictor
of clinical risk
Chronic
Condition
GP_Epile
psy_id
Epilepsy (LTC)
0.3869
All
availabl
e GP
Data
All
availabl
e GP
Data
All
availabl
e GP
Data

Epilepsy is a strong
predictor of clinical risk
Bupa Health Dialog
Beta Coefficient
Range
14
Look
back
Included Codes
Read 2 ICD-10 Variable
Category
Variable
Name
Variable Name
Variables in the Model – Clinical findings
Selected from ~1300 variables tested
Variable
Category
Description
Clinical
findings
GP_Current_
smoker
Patient has stated that
they are a current smoker
0.3043
All
available
GP Data
Clinical
findings
IP_CerebralP
_dx
0.6711
24
months

Clinical
findings
IP_Circulator
y_dx
0.2959
24
months

Clinical
findings
IP_Digestive
_dx
Inpatient admission with
diagnosis Cerebral Palsy
and other paralytic
syndromes
Inpatient admission with
diagnosis Symptoms and
signs involving the
circulatory and respiratory
systems
Inpatient admission with
diagnosis Symptoms and
signs involving the
digestive system and
abdomen
0.2978
24
months

Clinical
findings
IP_Urine_dx
Inpatient admission with
diagnosis Abnormal
findings on examination of
urine, without diagnosis
1.1327
24
months

Persons with CP have
complex needs and may have
challenges with regard to self
care
This includes acute MI,
pneumonia, acute severe
asthma, acute heart failure
and many high care need
conditions
Evidence from our own RCTs
show significant (impactable)
increase in overall services
usage from GORD and IBS.
IBD and PUD patients are
also high users of
unscheduled care
UTI, haematuria
Clinical
findings
IP_Alcohol_d
x
Inpatient admission with
Alcohol related diagnosis
0.8691
24
months

Alcohol related head injury
Bupa Health Dialog
Beta Coefficient
Range Look back
Included Codes
Read 2
ICD-10 Variable
Category
Variable Name
15

Variable Name
Smoking is an important risk
predictor
Variables in the Model - Inpatient
Selected from ~1300 variables tested
Variable
Name
Description
Inpatient
Utilisation
IP_Emer
g_admit
12
Emergency
admissions in last 12
months of history
period
0.4447
12
month
s

Previous emergency
admission is a strong
predictor of future
emergency admission
Inpatient
Utilisation
IP_Non
_emerg
_admit
Non-emergency
admission
0.0989
12
month
s

Elective admissions are
also a predictor of
emergency admissions
Inpatient
Utilisation
IP_Dayn
ight
Inpatient day & night
cases
0.7238
12
month
s

Ambulatory admissions
are also predictive of
emergency admissions
Inpatient
Utilisation
IP_Emer
g_admit
12
Emergency
admissions in last 12
months of history
period
0.4447
12
month
s

Previous emergency
admission is a strong
predictor of future
emergency admission
Bupa Health Dialog
Beta Coefficient
Range
16
Look
back
Included Codes
Read 2 ICD-10 Variable
Category
Variable
Category
Variable Name
Variables in the Model – OP and Deprivation
Selected from ~1300 variables tested
Variable
Category
Description
Outpatient
Utilisation
OP_Referr
al_Emerge
ncy
OP visit following an
emergency
admission
0.1959
24
months

Follow-up activity is an
additional indicator of
emergency admission risk
Outpatient
Utilisation
OP_Referr
al_GP
OP visit with referral
from a GP
0.1042
24
months

Outpatient
Utilisation
OP_Appoin OP visit with
tment
outcome 'Another
appointment given'
0.1765
24
months

Persons with illnesses
which require referral are
at higher risk of urgent
care need
Follow-up activity is an
additional indicator of
emergency admission risk
Deprivation
Index
DEM_Depr Deprivation
ivation
0.0055
Curren
t Dep
Index

Bupa Health Dialog
Beta Coefficient
Range
Included Codes
Look back Read 2
ICD-10 Variable
Category
Variable
Name
17
Variable Name
There is strong evidence
for an association
between deprivation,
health need and urgent
service usage
Agenda
• Model Development Methodology
• Model Performance
• Potential Impact
• Discussion/Next Steps
Bupa Health Dialog
18
Model Performance
Comparison with the Combined Model (England) on Test sample
The PRISM model consistently outperforms the Combined Model (CM) when looking at specific
cutpoints by ‘numbers’ of patients
CM
Top Cut Point
100
200
300
400
500
750
1000
2500
5000
PPV
Sensitivity N
73.00
1.46
61.50
2.46
57.33
3.44
51.75
4.14
49.20
4.92
43.73
6.56
40.00
8.00
31.16
15.58
25.56
25.57
PPV
Sensitivity N
60.00
1.20
55.00
2.20
50.33
3.02
46.50
3.72
45.80
4.58
40.27
6.04
37.10
7.42
29.44
14.72
23.88
23.88
PPV Top Cuts
73
123
171
207
245
328
400
778
1278
80
70
60
PPV (%)
PRISM
Top Cut Point
100
200
300
400
500
750
1000
2500
5000
50
CM
30
20
60
111
151
187
229
302
371
737
1195
10
0
Identified Patients out of 74,114
Test Sample N = 74,114
Bupa Health Dialog
PRISM
40
19
Model Performance
Comparison with Combined Model on Test sample
Gains vs Combined Model
Change in number of identified
patients with admission(gains)
90
80
70
60
50
40
30
20
10
0
100
200
300
400
500
750
1000
Identified Patients out of 74,114
Test Sample N = 74,114
Bupa Health Dialog
20
2500
5000
PRISM Model vs. Combined Model
0.5% segment – 370 patients
The PRISM model identifies a slightly younger Very High Risk population than the CM; LTC prevalences
are generally higher except hypertension
Age - Very High Risk Segment
(0.5% of Population)
Chronic Conditions - Very High Risk Segment
(0.5% of Population)
60%
50%
40%
Wales
30%
CM
20%
10%
Patients (% in Segment)
Patients (% in Segment)
60%
50%
40%
Wales
30%
CM
20%
10%
0%
Age 0 - 14
Age 15 - 44
Age 45 - 64
Age 65 - 75
Age 75+
0%
Asthma
Age Group
Test Sample N = 74,114
Bupa Health Dialog
21
COPD
Depression
Diabetes
Hypertension
Cancer
CAD
CHF
Agenda
• Model Development Methodology
• Model Performance
• Potential Impact
• Discussion/Next Steps
Bupa Health Dialog
22
Using segmentation and evidence-based clinical quality
indicators to target impactable patients
Risk Segment
Number of Patients
• Total = 4,230
Very High
470
• LTC = 2,582
(Top 0.5%)
High
(1%-5%)
Moderate
(6%-20%)
Low
(Bottom 80%)
4,230
• CHF =
439
• Gap =
290
14,100
75,273
High Risk CHF Beta Blocker
Gap ‘Campaign’
Bupa Health Dialog
23
Opportunity to Impact Admissions
A segmentation approach, using multiple commissioning and intervention
strategies aligned to risk, can significantly impact emergency admissions
Bupa Health Dialog
© 2008 Health Dialog UK Ltd – Commercial in Confidence
24
Agenda
• Model Development Methodology
• Model Performance
• Potential Impact
• Discussion/Next Steps
Bupa Health Dialog
25