Methods
Strength
The prediction strength metric is calculated by applying a non-linear (spline) correlation of each biomarker against a gender and age normalized dependent variable for each disease. The result of the normalization is a dependent variable with the form of a “risk multiplier” averaging a value of 1.0 for every age and gender combination. This is achieved by generating a set of weights using a bivariate tensor product spline fitting the relationship between age and each disease outcome for each gender. For each biomarker, we take the absolute value of the correlation, then average across men and women to create a gender-neutral metric. Next the value is divided by the maximum strength across all biomarkers, and then multiplied by 10, resulting in the final value for the metric where the strength is normalized for each individual disease outcome. The final presentation of the prediction strength metric is the result of sorting the individual biomarker averaged values, in descending order, and removing biomarkers redundant in the analysis based on similarity to others (e.g., ApoA, weight, Glucose, Cholesterol).
The normalization method discussed above avoids deceptive correlations with biomarkers correlating to outcomes simply because they correlate to age. Many of the outcomes also correlate highly with age.
VMR Values
The Voloridge Modeled Range or VMR, is the range of values of a marker yielding the lowest overall 15-year association with the seven disease outcomes. This does not represent a risk category, nor should it be interpreted as a recommendation of any kind. It is simply representing associations in the UK Biobank database of 500,000 participants and between their biomarker values and their 15-year outcomes. To determine the VMR, we begin by fitting a univariate spline for each of the seven targets against a biomarker. The seven splines are averaged into one final curve with respect to the biomarker. We define the VMR as the biomarker values corresponding to a risk of no more than 20% above the minimum mean risk. Additional constraints ensure the optimal range does not raise the risk for any individual disease beyond the greater of 0.9 or 30% above each disease’s minimum risk. We further restrict the range by placing limits at the top and bottom 1 percent of the biomarker distribution.
Population
The values in this report are calculated using data from the UK Biobank. While useful for its large amounts of data, the UK Biobank’s generalizability to the greater population is limited due to selection biases. The data come from primarily white, male and female, and more socioeconomically advantaged populations. However, research suggests the data may be generalizable to the larger population with regards to disease.[1]
In addition to biases inherent in the dataset, a common filter is applied to all disease models. This common filter limits the model inputs to a population between the ages of 40 to 70 containing data for height and weight and is observed in the dataset for at least 12 years. Participants who died within those 12 years are included. Additional filters are added to each disease model to exclude people with a prior diagnosis of each specific disease. For example, the Cardiac (MACE) model excludes anyone with prior MACE. Because the age bias for dementia is so strong, there is an extra filter which excludes people who died at any point without getting dementia. Additional age filters are placed on each disease to ensure the age distribution is balanced between genders for each disease target.
Age filters are applied for each disease. The database used had a uniform distribution of individuals aged 40 to 70. For some disease outcomes, (e.g. liver disease and metabolic, type 2 diabetes), this entire age range was utilized. For the remainder of the disease outcomes, a more limited age range was utilized as risk multipliers are not as clinically relevant on age groups with lower incidence of the disease. In our analysis for Cognitive (dementia), only 60- to 70-year-olds were included as this subgroup contained 86 percent of all cases in the database. For mortality, we analyzed 55- to 70-year-olds, as this age range included 87 percent of deaths in the database. For Cardiac (MACE), we analyzed 55- to 70-year-olds with no prior MACE, which included 82 percent of MACE events. Similarly, for Lung (COPD), we analyzed 50- to 70-year-olds with no prior COPD, which included 92 percent of measured Lung events. Finally, for kidney disease we analyzed 55- to 70-year-olds with no prior acute kidney disease, which included 86 percent of instances of the disease. For Metabolic (Type 2 diabetes), an additional filter requiring Hemoglobin A1C to be less than 6.0 was applied. This is due to the measure serving as the primary marker for diagnosis of the disease. This was done to avoid the inclusion of people who already had type 2 diabetes but were not yet diagnosed.
[1] Fry, Anna, Thomas J. Littlejohns, Cathie Sudlow, Nicola Doherty, Ligia Adamska, Tim Sprosen, Rory Collins, and Naomi E. Allen. “Comparison of sociodemographic and health-related characteristics of UK Biobank participants with those of the general population.” American journal of epidemiology 186, no. 9 (2017): 1026-1034.