The POpulation HEalth Model (POHEM) is a health microsimulation model that was developed at Statistics Canada in the early 1990s. POHEM draws together rich multivariate data from a wide range of sources to simulate the lifecycle of the Canadian population, specifically focusing on aspects of health. The model dynamically simulates individuals’ disease states, risk factors, and health determinants, in order to describe and project health outcomes, including disease incidence, prevalence, life expectancy, health-adjusted life expectancy, quality of life, and healthcare costs. Additionally, POHEM was conceptualized and built with the ability to assess the impact of policy and program interventions, not limited to those taking place in the healthcare system, on the health status of Canadians. Internationally, POHEM and other microsimulation models have been used to inform clinical guidelines and health policies in relation to complex health and health system problems. This paper provides a high-level overview of the rationale, methodology, and applications of POHEM. Applications of POHEM to cardiovascular disease, physical activity, cancer, osteoarthritis, and neurological diseases are highlighted.

Background: The purpose of this study was to assess whether or not the change in coding classification had an impact on diagnosis and comorbidity coding in hospital discharge data across Canadian provinces. Methods: This study examined eight years (fiscal years 1998 to 2005) of hospital records from the Hospital Person-Oriented Information database (HPOI) derived from the Canadian national Discharge Abstract Database. The average number of coded diagnoses per hospital visit was examined from 1998 to 2005 for provinces that switched from International Classifications of Disease 9^th version (ICD-9-CM) to ICD-10-CA during this period. The average numbers of type 2 and 3 diagnoses were also described. The prevalence of the Charlson comorbidities and distribution of the Charlson score one year before and one year after ICD-10 implementation for each of the 9 provinces was examined. The prevalence of at least one of the seventeen Charlson comorbidities one year before and one year after ICD-10 implementation were described by hospital characteristics (teaching/non-teaching, urban/rural, volume of patients). Results: Nine Canadian provinces switched from ICD-9-CM to ICD-I0-CA over a 6 year period starting in 2001. The average number of diagnoses coded per hospital visit for all code types over the study period was 2.58. After implementation of ICD-10-CA a decrease in the number of diagnoses coded was found in four provinces whereas the number of diagnoses coded in the other five provinces remained similar. The prevalence of at least one of the seventeen Charlson conditions remained relatively stable after ICD-10 was implemented, as did the distribution of the Charlson score. When stratified by hospital characteristics, the prevalence of at least one Charlson condition decreased after ICD-10-CA implementation, particularly for low volume hospitals. Conclusion: In conclusion, implementation of ICD-10-CA in Canadian provinces did not substantially change coding practices, but there was some coding variation in the average number of diagnoses per hospital visit across provinces.

Prevention efforts are informed by the numbers of deaths or cases of disease caused by specific risk factors, but these are challenging to estimate in a population. Fortunately, an increasing number of jurisdictions have increasingly rich individual-level, population-based data linking exposures and outcomes. These linkages enable multivariable approaches to risk assessment. We demonstrate how this approach can estimate the population burden of risk factors and illustrate its advantages over often-used population-attributable fraction methods. We obtained risk factor information for 78,597 individuals from a series of population-based health surveys. Each respondent was linked to death registry (568,997 person-years of follow-up, 6,399 deaths).Two methods were used to obtain population-attributable fractions. First, the mortality rate difference between the entire population and the population of non-smokers was divided by the total mortality rate. Second, often-used attributable fraction formulas were used to combine summary measures of smoking prevalence with relative risks of death for select diseases. The respective fractions were then multiplied to summary measures of mortality to obtain smoking-attributable mortality. Alternatively, for our multivariable approach, we created algorithms for risk of death, predicted by health behaviors and various covariates (age, sex, socioeconomic position, etc.). The burden of smoking was determined by comparing the predicted mortality of the current population with that of a counterfactual population where smoking is eliminated. Our multivariable algorithms accurately predicted an individual’s risk of death based on their health behaviors and other variables in the models. These algorithms estimated that 23.7% of all deaths can be attributed to smoking in Ontario. This is higher than the 20.0% estimated using population-attributable risk methods that considered only select diseases and lower than the 35.4% estimated from population-attributable risk methods that examine the excess burden of all deaths due to smoking. The multivariable algorithms presented have several advantages, including: controlling for confounders, accounting for complexities in the relationship between multiple exposures and covariates, using consistent definitions of exposure, and using specific measures of risk derived internally from the study population. We propose the wider use of multivariable risk assessment approach as an alternative to population-attributable fraction methods.

Background: New guidelines for cardiovascular disease risk assessment and statin eligibility have recently been published in the United States by the American College of Cardiology and the American Heart Association ( ACC ‐ AHA ). It is unknown how these guidelines compare with the Canadian Cardiovascular Society ( CCS ) recommendations. Methods and Results: Using data from the Canadian Health Measures Survey 2007–2011, we estimated the cardiovascular disease risk and proportion of the Canadian population, aged 40 to 75 years without cardiovascular disease, who would theoretically be eligible for statin treatment under both the CCS and ACC ‐ AHA guidelines. The survey sample used (n=1975) represented 13.1 million community dwelling Canadians between the ages of 40 and 75 years. In comparing the CVD risk assessment methods, we found that calculated CVD risk was higher based on the CCS guidelines compared with the ACC‐AHA guidelines. Despite this, a similar proportion and number of Canadians would be eligible for statin treatment under the 2 sets of recommendations. Some discordance in recommendations was found within subgroups of the population, with the CCS guidelines recommending more treatment for individuals who are younger, with a family history of CVD , or with chronic kidney disease. The ACC ‐ AHA recommend more treatment for people who are older (age 60+ years). These results likely overestimate the treatment rate under both guidelines because, in primary prevention, a clinician–patient discussion must occur before treatment and determines uptake. Conclusions: Implementing the ACC ‐ AHA lipid treatment guidelines in Canada would not result in an increase in individuals eligible for statin treatment. In fact, the proportion of the population recommended for statin treatment would decrease slightly and be targeted at different subgroups of the population.

Artificial intelligence (AI) is being broadly integrated into processes to manage and analyze large amounts of data accurately and efficiently. In this work, we explored how AI methods, in particular machine learning (ML), are being implemented in human biomonitoring using a mixed methodology approach that consisted of: (1) a scoping literature review and (2) an international scan of biomonitoring programs to contextualize current practices and perceptions from researchers in the field. We synthesized findings from the review according to the year of publication, biomonitoring study and location, study outcomes, and the most frequent ML methods. We additionally categorized all published ML methods from the review according to three dimensions (paradigm, type of task, and model structure), mapped studies to biomonitoring themes and other applications, and provided details for the more commonly applied ML methods. The international scan was administered through a 30-question online survey and gathered information on current uses, perspectives, and barriers related to AI. Scoping review: We found 286 studies that applied a ML method to human biomonitoring data. Eighty-two ML methods were identified, with the most common being supervised approaches. ML was predominantly applied to predict health-related outcomes based on chemical exposure. International scan: The survey yielded 30 responses from programs across 15 countries. Approximately 27% of respondents reported implementing AI-related methods in the collection and analysis of biomonitoring data, and the primary barrier to adopting these methods was a lack of technical expertise (80%). This exploratory work provides an integrated understanding of ML applications in the human biomonitoring field. ML clearly holds promise for furthering our understanding of chemical exposure in people and will likely undergo continued growth in applications.