Cardiovascular disease (CVD), including coronary heart disease (CHD) and stroke, kills about 610,000 people per year in the United States and this represents about 25% of deaths. People are not born with the disease, but due to inherited traits such as sex and race, and due to environmental factors, some people in the population develop CVD as they age. Each person has some risk of developing the disease.
In 1957, the Framingham Heart Study was initiated to study a population’s incidence of CVD and to attempt to correlate it to ‘risk factors.’ Risk factors may be categorical, such as sex, or quantifiable by measurement, such as systolic blood pressure, cholesterol levels and age. The idea was to explain what causes CVD and formulate a public health recommendation to reduce the risk for developing CVD.
The correlation takes on the form of a CVD Risk model or formula, where a patient’s risk factor data are collected and then input into the formula to calculate the Risk for developing CVD in a 10 year timeframe. Each risk factor input can be varied to show the patient how recommended changes, such as taking hypertension medication to lower the blood pressure, or lowering cholesterol levels by lifestyle changes, can reduce the risk of developing CVD. Presumably, the risk factors that have the most influence would be good candidates to change. This same principle can be applied to risk factor variation or uncertainty. For example, a person’s systolic blood pressure will vary during the day or due to the ‘white coat effect’ that occurs when you see a doctor in an office and your blood pressure goes up. Every risk factor to be input to the formula can have a variation and this propagates to uncertainty in the calculated CVD Risk.
Over time, numerous CVD Risk models have been derived. There is a common core of risk factors that most models use, such as age, sex and blood pressure, but each model may have distinct risk factors such as ethnicity, high-sensitivity C-reactive protein (hsCRP), or body-mass-index (BMI). Current work, based on a review of 35 different models, requires consideration of 41 different input risk factors with the smallest model having 5 input risk factors and the largest having 17 input risk factors.
There are also different approaches to the math-functions that are used to relate the risk factors to the CVD Risk. Some studies report multiple versions such as four versions of the ‘Pooled-Cohort Model’, specific to sex (male and female) and race (white and black). For a given person, there may be a choice among the 35 models and the main question is which formula should be used to obtain the best assessment of CVD Risk. Often the model accuracy is documented but this requires a population of patients with known CVD outcomes to compare to the formula’s prediction. While one formula version may be superior for a population, this does not mean it is the ‘best’ CVD Risk formula for an individual patient.
The CVD Risk project was initiated to provide a novel assessment of model superiority by using the calculated uncertainty of CVD Risk. The uncertainty is calculated for an individual rather than a population so the model assessment is individualized. Also, the CVD outcome does not have to be known for the individual to determine accuracy because the uncertainty communicates the quality of the information. The duals arithmetic provides uncertainty components such that dominant inputs can be judged and this would indicate where patient data has to be improved. This also allows the novel terms of each model to be judged for its value. One would expect that including more specific patient data would support a more sophisticated calibration such that the formula ‘fits the data,’ better. However, there is no guarantee that a model with more risk factors will benefit the patient. Needing data for more risk factors can have large-scale healthcare implications. For example, adding hsCRP as a risk factor has to be demonstrated as a benefit because it may be an extra test to order for a patient and add to healthcare costs. With the duals arithmetic, the benefit would not be judged by the change of the CVD Risk number caused by adding a risk factor to the model. Instead the added risk factor should lower the uncertainty or increase the certainty of the calculated CVD Risk.