1. Population Health Model (POHEM)
For educational / demonstration purposes
Not for distribution or citation
Contact:
July 2007

2. What can Micro-Simulation Do ?
project basic counts and distributions
population
prevalence of risk factors
cases eligible for primary interventions
disease incidence and progression; e.g. first AMIs, readmissions, and deaths
simulate interventions and their potential impacts
all these by age, sex, calendar year, geography, … any other modeled variables
July 2007

3. POpulation HEalth Model (POHEM)
case-by-case, Monte Carlo microsimulation
directly encompasses competing risks and comorbidity
longitudinal risk factor and disease sub-modules
generates plausible health biographies for synthetic individuals from empirical observations
population attributable fractions estimated through risk-factor deletion (ie, relative risk set to 1)
projects population forward in continuous time
population initialized in 2001 from Canadian Community Health Survey cycle 1.1
subject to cohort-specific mortality hazards based on age, sex and year of birth
new births and new immigrants generated in future years based on Census projections
July 2007

4. Main State Variables and Dependencies
births / immigration / emigration – vital statistics, immigration records and demographic estimates by sex, province and year
educational attainment - baseline = F (age, cohort, sex, …)
mortality = F (age, cohort, sex, AMI status)
Cancer OA
July 2007

5. POHEM example
Starting Population: Canadian Community Health Survey 2001 (CCHS)
cross-sectional representation of the Canadian population aged 18+
2001
……..
VARIABLE
age
sex
province
health region
immigration status
education level
income quartile
body mass index
smoking status
diabetic status
total cholesterol*
HDL*
blood pressure*
HUI
*imputed from Canadian Heart Health Surveys ( )
VALUE 44
male
Ontario
York
non-immigrant
post-secondary
Q4 (richest)
32.2 kg/m2 (obese)
smoker
non-diabetic
high
low
0.96
Every year on birthday, evaluate the hazard of developing disease (AMI, diabetes, cancer, osteoarthritis,...)
 no disease events in 2001
>Data Source: CCHS cycle Master file
>Database records: Start N=131535; final N= after exclusions including those less than 18
* variables imputed from Canadian Heart Health Surveys ( )
Included for printing – variables in blue indicate they are updated (initialize) in year shown
OTHER VARIABLES NOT included yet
heart disease status
alcohol level
nutrition level
ethnicity
July 2007

6. POHEM example
Starting Population: Canadian Community Health Survey 2001 (CCHS)
cross-sectional representation of the Canadian population aged 18+
2001
……..
age
sex
province
health region
immigration status
education level
income quartile
body mass index
smoking status
diabetic status
total cholesterol
HDL
blood pressure
HUI
Every year on birthday, evaluate the hazard of developing disease (AMI, diabetes, cancer, osteoarthritis,...)
 AMI in 0.3 years
2002
……..
AMI
AMI at age 45.3
Now at risk of 2nd AMI, CHF, UA, ...
variables in blue indicate they are updated (initialize) in year shown
illustrates case-by-case way of proceeding
July 2007

7. Congestive Heart Failue (comorbid with Congestive Heart Failure)
POHEM example
Starting Population: Canadian Community Health Survey 2001 (CCHS)
cross-sectional representation of the Canadian population aged 18+
2001
……..
age
sex
province
health region
immigration status
education level
income quartile
body mass index
smoking status
diabetic status
total cholesterol
HDL
blood pressure
HUI
apply Δ BMI model (function of age, sex education, income, region and BMI in 2001)
 Remains Obese
2002
……..
2003
…….. …
……..
Death
at age 71.2
AMI
CHF OA
Congestive Heart Failue
at age 66.1 in year 2023 OA
at age 69.4 in year 2028
(comorbid with Congestive Heart Failure)
Form of current BMI model (from NPHS):
Δ BMI(2002,2004)= α + ß1 Δ BMI(2000,2002) + ß2 Δ BMI(1998,2000) +ß3 Δ BMI(1996,1998) + ß4BMI(1996) + ß5IncQuartile(1996) + ß6Educ_cat(1996) + ß7Region(1996)
for 28 Sex-Age-BMI_categories defined in NPHS 1996
July 2007

8. POHEM >100,000 records on CCHS representing ~24 million Canadians
Starting Population: Canadian Community Health Survey 2001 (CCHS)
cross-sectional representation of the Canadian population aged 18+
2001
……..
2002
……..
2003
…….. …
……..
death
>100,000 records on CCHS representing ~24 million Canadians
(4 hours on a PC)
July 2007

9. July 2007

10. Simple Models The Workbook Approach (selected risk factors)
Cancer model
Heart disease Model
Smoking
Smoking
Obesity
Obesity
Cancers
Heart
Disease
Diabetes
These two diagram represent the traditional approach, using workbooks. Each risk factors is considered independently, and there is a separate model for each disease. While easy to represent, this is a considerable simplification of the reality.
As each component is decomposed (say by age group, sex, risk category), the number of bins (cells) and pathways between bins explode.
Nutrition
Alcohol
Alcohol
July 2007

11. Complex Causal Web Diagram: The Microsimulation Approach
Nutrition
Smoking
Cancers
Alcohol
Obesity
Heart
Disease
Simulation models can more precisely represent the complex web of causality.
A proportion of the population has more than one risk factor. One can not calculate the total population attributable fraction without taking the co-existence of risk factors into account. i.e. we can’t had them up.
There are multiple levels of interaction between risk factors and diseases: Nutrition is a risk factor for obesity, which is a risk factor for diabetes and cancer. Diabetes is also a risk factors for heart disease. And nutrition is also a risk factor for some cancers. It is very difficult, even sometimes impossible, to accurately represent such a complex web in workbooks.
Disease are of a competing nature. A new treatment that would improve the survival of cancer might not have a large impact on the life expectancy of smokers, if some of them suffer from heart disease instead.
Diseases share common risk factors. For example, smokers are at an increased risk of cancer and heart disease. Therefore, a proportion of people diagnosed with cancer will already be living with heart disease, and not in full health. Their loss of quality of life with consequently be less than for a person in full health.
Microsimulation models can represent such complex pathways, given the appropriate epidemiological data is available. While they also are a simplification of the complex biological processes at play, they are closer to reality than simple one-dimensional representation. And are a more useful tool to inform evidence-based decision making.
However they are more challenging to build and require more expertise and investment of the users. We are building both tools: Excel workbooks to document data (eg PHI cancer workbooks), compare to the work done by the WHO and Australia and generate first estimates of health-adjusted life years lost (HALYs). We build the microsimulation for the reason just stated (POHEM). We can then compare the results from both.
Diabetes
July 2007

12. Coronary Heart Disease: Acute Myocardial Infarction (AMI)
Causal pathway
Upstream health determinants
Intermediate risk factors
Intermediate diseases
Sequalae
Death
age (time)
initial values
Obesity
Smoking
Nutrition
Physical
activity
Alcohol
Income
Education
Region
Sex
CCHS 2001
transition models
NPHS
AMI*
Health Person-Oriented Information (HPOI)
(HIRD)
incidence rates by province, age and sex
2nd AMI
Congestive Heart Failure
Unstable Angina
Death
Registered Persons database for Ontario (ICES)
(CCORT I)
survival data for each transition
initial values & transition models
Diabetes
Total
cholesterol
&
HDL
blood pressure
Cdn Heart Health Surveys
S(t)
Kaplan-Meier survival analysis on observed data; fitted with 2-piece Weibull distribution; Weibull parameters fed into POHEM with inverted Weibull equation to generate waiting times. Since there was no cause of death on Ontario RPDB, the survival data was built by 5 year age groups (and by sex).
AMI bubble: represents incidence based on Framingham risk function with smoking, cholesterol, blood pressure, diabetes, age and sex as prediction variables (Wilson et al 1998)
An index AMI event was defined as cases on the Health Person-Oriented Information (HPOI) with a most responsible diagnosis of acute myocardial infarction (AMI) (ICD-9 410), and was defined for year 2001/02, by province.
Exclusions: Cases were excluded if they
- had a previous AMI within the previous 5 years (referred to as 5-year wash-out),
were not admitted to an acute care hospital,
were not between 20 and 105 years old,
were transferred from another acute care hospital,
the AMI was an in-hospital complication, or
if they had a date of death before the date of admission or discharge
probability of AMI incidence
combines individuals cardiac risk factor profile and baseline index AMI rates
The risk function is calibrated to province specific incidence rates by age, sex (the alpha term of the equation).
competing risk of death from other causes
Vital statistics
(and other POHEM disease modules)
July 2007
*incidence-risk equation based on Framingham risk function (Wilson 1998) for “index” AMI events

13. How POHEM Generates an Incident Case of AMI
POHEM selects a record from CCHS in simulation year 2001: male, age 44, smoker, non-diabetic, high total cholesterol, low HDL, medium blood pressure...
Lookup baseline risk and risk factor coefficients from input parameter table.
Evaluate the probability (p) of AMI
using the Framingham risk function
(with rescaling):
p = F/(1+F) = 0.877
Convert to annualized hazard (h):
h = -ln(1-p) = 2.09
Generate a a random number (u) between 0 and 1
u = 0.025
Transform the hazard to a waiting time:
t = -ln(u) / h = 1.76 years
AMI does not occur at age 44 in year The risk of AMI will be re-evaluated at the next birthday.
POHEM ages the person forward to next birthday
updates the person’s risk factors profile
re-evaluates risk of AMI (steps 2-7).
repeated every year until AMI occurs or death
other events are evaluated
Steps 1 to 8 are repeated for every record on CCHS
Data Analysis and Input to POHEM:
incidence rates (I) estimated from administrative data by age group, sex and province
incidence-risk equation obtained from the literature:
Framingham risk function
α represents the baseline risk (by age, sex, province) after accounting for the other risk factors
 coefficients vary by category for cholesterol, high density lipids, diabetes and smoking obtained from the study (Wilson, 1998)
the baseline risk (α) is calibrated such that the incidence-risk equation implemented in POHEM (F) reproduces the observed incidence rates by age, sex and province
takes into account the distribution of the risk factors (by category) in the population
d) values and models of change in risk factors based on data and trends from national surveys
α =
βsmoking= 0.523
βdiab = 0
βchol = 0.657
βHDL = 0.497
βBP = 0.283
In 2002 at age 45, the risk profile for this simulated person would remain unchanged but the random number this time generates a waiting time less than 1 year.
t = -ln(u) / h = -ln(0.5333) / 2.09 = 0.3 years
AMI would occur at age 45.3 in year 2002
*the alpha term is estimated by age group
the target rates are shown here, but actually we calibrate with the risk and distribution of BMI in the population in 2001 to obtain baseline rates (b). The baseline rates (by age and sex) are assumed to remain constant over time; the individual’s age and BMI changes over time, which changes the risk as they are projected forward in the simulation
July 2007

14. Data Sources Canadian Community Health Survey (2001)
starting population for POHEM (initialize age, sex, geography, BMI, smoking, diabetes)
National Population Health Survey ( )
models of change in BMI and smoking
Canadian Heart Health Survey (1986 to 1992)
joint distribution of other cardiac risk factors cholesterol, diabetes, blood pressure
HDL imputed
Health Person-Oriented Information (1992/93 to 2001/02)
hospital separations by province
rate of index AMI (5-yr wash-out) by province
managed at STC
Registered Person database (1988/89 to 2001/02)
Ontario hospital separations linked to vital statistics
Survival time from AMI event to subsequent AMI event or death
managed at ICES
July 2007

15. Geography Geography is an explanatory variable in the BMI model
ATLANTIC, QUEBEC, ONTARIO, PRAIRIES, BC
Geography is a dimension of the incidence rates for index AMI (by sex, age group, province groups
Geography was not used in the smoking model, and was not used in the joint risk factor transition model
July 2007

16. Model Input: Smoking transitions
Never smokers
Successful quitter in1996
Quit in1998
Smokers in all 3 years 0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
non-smoker 94
smoker 94
non-smoker 96 and 98
non-smoker 96 and smoker 98
smoker 96 and non-smoker 98
smoker 96 and 98
Example:
Smoking transitions capture in a 4th order hierarchical Markov chain by age group and sex
based on NPHS from 1994 to 2002 (after which time sample size is too small)
July 2007
Source: NPHS

17. Modeling Risk Factor Transitions
Initialization 2001
CCHS 2001
age = 55
sex = male
income
education
region
BMI
diabetes hypertension (y/n)
smoker (y/n)
CHHS (86-92)
total cholesterol
HDL
Blood pressure
Year
t= t t+10
Smokingt
Smokt+2
Smokt+4
Smokt+6
Smokt+8
Smokt+10
BMIt+10
BMIt
BMIt+2
BMIt+4
BMIt+6
BMIt+8
Diabt+5
Cholt+5
Hyptt+5
Diabt+10
Cholt+10
Hyptt+10
Diabt
Cholt
Hyptt
The transitions are activated when the person reaches the next age group (5 year age groups).
The joint transition modeled BMI (4 categories), diabetes (2 categories), cholesterol (5 categories), and blood pressure (5 categories) from one age group to the next (5 year intervals), by sex.
HDL (5 categories) is imputed from CHHS data based on the current values of sex, age group, total cholesterol level, diabetic status (Y/N), and BMI category as the simulation progresses.
Initial values for cholesterol level and blood pressure status imputed from CHHS by sex, age group, diabetic status (Y/N), and BMI category.
Smoking (not shown here) was modeled from NPHS and was based on sex, age group, and past history of smoking (1st, 2nd, and 3rd order Markov chains). It is updated every two years.
Legend:
transitions derived from NPHS ( )
transitions derived from CHHS
July 2007

18. Example of HDL distribution by cholesterol for male aged 55-59, overweight and non-diabetic
HDL (5 categories) is imputed from CHHS data based on the current values of sex, age group, total cholesterol level, diabetic status (Y/N), and BMI category as the simulation progresses.
2x12x2x4x5=960 combinations x 5 possible HDL levels.
So a person falls into 1 of the 960 possible configurations based on their sex, age, bmi category, diabetic status and total cholesterol.
A random number is used to then choose their HDL level from the distribution.
July 2007

19. Preliminary Results
July 2007

20. Acute Myocardial Infarction in Canada: Projection of risk factors, incidence and progression from 2001 to 2021
July 2007

21. Objectives
project the prevalence of risk factors most commonly associated with acute myocardial infarction (AMI) between 2001 and 2021
project the number of resulting AMI events over that period
estimate the contribution of each risk factor to AMI outcomes in future years
July 2007

22. Model Projection: Prevalence of Smoking
Prop. of Pop
Illustrative - not for distribution
Proportion smokers declining for all age groups.
July 2007

23. Model Projection: Prevalence of Diabetes
Prop of Pop
Illustrative - not for distribution
Prevalence of diabetes increase within age group over time, except the age group.
July 2007

24. Projected rate of new AMI cases per 1000
Illustrative - not for distribution
July 2007

25. Projected number of new AMI cases by province
Illustrative - not for distribution
The proportional increases are similar for BC, Alberta, Ontario and Quebec and Newfoundland. These are higher than the Canada average.
Percent increase relative to 5 years earlier
2006 vs vs vs vs 2016
British Columbia 18% 12% 12% 13%
Alberta 24% 10% 17% 12%
Saskatchewan 2% 12% 4% 2%
Manitoba 12% 8% 8% 8%
Ontario 18% 12% 14% 11%
Quebec 18% 9% 10% 10%
Nova Scotia 13% 8% 2% 8%
New Brunswick 19% 7% 7% 10%
P.E.I. -7% 28% 6% 6%
Newfoundland 22% 6% 13% 9%
Canada 17% 10% 12% 10%
July 2007

26. Projected number of new AMI events (from new Index AMIs only)
Illustrative - not for distribution *
Note: this model did not include a starting prevalence in This means too many at risk of Index AMI (initially).
This also explains the year over year “growth” in the other outcomes, because they depend on incidence cases.
Next step to include prevalence in 2001 (and progression from those cases).
* includes death from non-IHD causes
July 2007

27. Total Cholesterol / HDL
Projected Number Eligible for Statin Use in Ontario in 2001 – CMAJ 2000 Guidelines (’000s)
Total Cholesterol / HDL
<4
4-5
5-6
6-7
>=7
Total
0-5%
4 148
1 887
969
373
288
7 666
5-10%
201
278
238
116
111
945
10-15% 54
102 88 51 56
351
15-20% 14 40 41 28 27
150
20-25% 2 11 19 16 63
25-30% 1 5 9 8 33
>=30% 13
Illustrative - not for distribution
10-year predicted risk of AMI
This slide highlights POHEMs flexibility to identify high risk individual’s and then to evaluate interventions on them as will be shown in next slide
Note: Available for Canada too from POHEM, just not output yet.
Note: The guidelines have been updated but we have not generated updated results. Doug’s paper commented on the new guidelines and we could perform similar analysis using POHEM. (not done)
n.b predicted risk is based on POHEM, not the same variables as in the original CMAJ ariticle; uses constant RFs
Medication if target not reach after 6 months of lifestyle changes, n = 399,000
Medication if target not reach after 3 months of lifestyle changes, n = 162,000
Medication and lifestyle change, n = 109,000
Based on the recommendations for the management and treatment of dyslipidemia (CMAJ 2000)
July 2007

28. Illustrative “What-if ?” Scenarios
Statins: given to people at high risk according to guidelines from working group on dyslipidemias; reduces their AMI risk by 31% (La Rosa, 99).
BMI: 10% reduction for everyone overweight or obese (BMI ≥ 25) at baseline in 2001
Smoking: 20% of smokers permanently quit smoking at baseline in 2001
Cholesterol: 5% reduction of total cholesterol value for everyone at baseline in 2001
Note: interactions in RF dynamics  change in one at baseline affects subsequent levels of others
Statin given to people according to guidelines on previous page from previous slide (except for Canada).
July 2007

29. Illustrative - not for distribution
Cumulative number of index AMIs avoided by calendar year, by “what-if” scenario, Canada
Limitations:
statin coverage at baseline not modeled so this graph overestimates benefit;
uncertainty of benefit of statins not captured and this modeling exercised assumed relatively large benefit which may also over-estimate benefit;
no side-effects of statins were modeled
Illustrative - not for distribution
~40,000 new AMI cases in 2001,
~500,000 prevalence AMI cases after by 2011
July 2007

30. Projected fraction of AMI cases attributable to risk factors
Illustrative - not for distribution
62%
58%
47%
37%
Base case = when baseline (age and sex) risk and all risk factors are applied.
Background = all risk factors deleted (ie when all relative risks set to 1), so that only the baseline risk associated with age and sex apply.
Smoking scenario = the difference between baseline and the smoking-deleted scenario (ie when smoking relative risk set to 1) etc.
Results:
The incidence of AMI attributed to smoking dropped from about 4% to less than 1% in females, and from about 10% to 2% in males.
For females, the second largest contributor to incidence of AMI was elevated total cholesterol levels (17%);
For males, the second largest contributor was elevated blood pressure (24%).
The amount of AMI attributed to the other risk factors remained relatively constant over time.
July 2007

31. Additivity(?) of risk factors
Illustrative - not for distribution
Number of index AMI cases
Same information as previously slide, shown as counts by year, with the Base case curve superimposed to illustrate confounding.
Base case = when baseline (age and sex) risk and all risk factors are applied.
Background = all risk factors deleted (ie when all relative risks set to 1), so that only the baseline risk associated with age and sex apply.
Smoking scenario = the difference between baseline and the smoking-deleted scenario (ie when smoking relative risk set to 1)
etc.
July 2007

32. Summary of preliminary results
Number of AMI cases projected to increase, principally due to aging of the population
smoking projected to decline, reduces the overall increase in AMI
proportion of persons with diabetes projected to rise
approximately 10% of new index AMI cases attributed to diabetes
in males, 24% of new AMI cases attributed to elevated blood pressure
in females, 17% of new AMI cases attributed to elevated total cholesterol
** note ** Jack Tu is noticing a decline in AMI incidence in the last few years in their admin data **
July 2007

33. Future Work Revise, finalize and publish current work
revise / explore intervention scenarios
validation – e.g. recreate 1994 – 2004 history of incidence and mortality
Improve POHEM’s data foundations
update index AMI rates with most recent data (from 2001 to 2004)
update to CCHS cycle 3.1 (or pooled) to initialize POHEM
update with measured risk factor prevalence from CHMS (when available)
update survival with cause-specific mortality data (HPOI linked to vital stats)
Expand cardio-vascular disease model
develop more robust model of diabetes (Rosella and Manual, ICES)
add procedures (CABG, PCI, catheterizations) as consequence of AMI
relate procedures to survival outcomes – to the extent there are data
add CHF and UA as index events (if appropriate)???
add models of stroke and peripheral vascular disease
other CVD
Health-related Quality of Life
estimate health-adjusted life expectancy
Burden of disease
Build POHEM towards a comprehensive tool covering multiple diseases, risk factors and functional health status and other sequalae
LIMITATIONS:
Only models AMI identified with hospital admissions – misses sudden death
Risk factor prevalence and trends based on household population
Only have all cause of death after AMI means cannot model impact of treatment
Models only index AMI and consequences.
Other updates:
update models of smoking and weight change as more recent data becomes available from new cycles of NPHS
prevalence estimates to warm-up the simulation
health care utilization - physician visits, days of hospital stays, direct costs
health care access - additional variables from CCHS
Note: diabetes risk algorithm and complication algorithm being build by by Laura Rosella and Doug Manual at ICES
July 2007

34. Health-Related Quality of Life – Beyond Life Expectancy (LE)
LE = area under survival curve
HALE = “weighted” area under survival curve
where “weights” are levels of individual health status, ranging between zero (dead) and one (fully healthy)

35. POHEM: Overall causal flow
Upstream health determinants
Intermediate risk factors
Intermediate diseases
Diseases
Treatment
age and sex
Death
Health-related Quality of life (HUI)
Coronary
Heart
Disease
Alcohol
CAPG, PCI, CATH, Drugs, lifestyle
Ethnicity
Stroke
Smoking
ABS
Region
Peripheral
Vascular
Disease
Cholesterol
Amputation
hypertensive
Diabetes
Diabetic
Retinopathy
Income
Nutrition
Cataract surgery...
blood pressure
Kidney
Disease
Education
Dialysis
Obesity
POHEM Burden of Disease: multiple risks – multiple diseases
This is a “causal” map showing major health determinants, risk factors and diseases for POHEM modeling.
The overlapping of risk factors and diseases (comorbidity) illustrated in previous slides are implicit in this diagram since the POHEM model operates at the level of the individual, and are initialized with a comprehensive risk profile based on their CCHS record, and develop one or more diseases as they age in the simulation.
The diagram is incomplete in many respects:
It does not include certain diseases that we may well want to tackle (Mental Health) and others (Accidents/Injuries);
It does not include certain risk factors (fertility status, weather, air pollution), although for cancers some may be added (or modified).
It does not show data inputs and outputs.
It does not represent the longitudinal dynamic of microsimulation POHEM. For instance, there are longitudinal models of smoking and bmi.
Various bubble described in more detail in subsequent slides.
LEGEND FOR COLOUR CODING (to explain what is currently modeled in POHEM):
Thin arrow lines represent “causal” flow: solid implies survival times
Yellow Arrow: age and sex play a role in everything
Yellow bubbles : risk factors (health determinants) for which we have a longitudinal model in POHEM
Blue bubbles : no explicit longitudinal model of these items (risks/ health determinants), but they are given initial values based on CCHS and CHHS
Orange bubbles: represents intermediate diseases that are major risks for subsequent disease; hypertensive is directly a result of blood pressure and is not specifically a model; diabetes is currently not a disease model as represented here, but is modeled jointly with cholesterol, blood pressure and obesity (BMI) as a first order Markov chain between age groups from cross-sectional data (CHHS)
Pink bubbles: disease processes that have been modeled in POHEM (may include sequalae – see subsequent slides for expanded versions)
Pink Rectangles: treatments modeled (see subsequent slides for expanded versions)
White bubbles: have not been modeled in POHEM
White Rectangles: treatments not yet modeled
Osteoarthritis
Surgery
other risk factors
Physical
activity
other diseases
25 Cancers
Surgery, Radio/Chemo/Hormonal therapy
Initial state assigned from CCHS (+CHHS)
competing risk of death from other causes
July 2007