Abstract
The two leading causes of mortality worldwide are cardiovascular disease (CVD) and cancer. The annual total cost of CVD and cancer is an estimated $844.4 billion in the US and is projected to double by 2030. Thus, there has been an increased shift to preventive medicine to improve health outcomes and development of risk scores, which allow early identification of individuals at risk to target personalised interventions and prevent disease. Our aim was to define a Risk Score R(x) which, given the baseline characteristics of a given individual, outputs the relative risk for composite CVD, cancer incidence and all-cause mortality.
A non-linear model was used to calculate risk scores based on the participants of the UK Biobank (= 502548). The model used parameters including patient characteristics (age, sex, ethnicity), baseline conditions, lifestyle factors of diet and physical activity, blood pressure, metabolic markers and advanced lipid variables, including ApoA and ApoB and lipoprotein(a), as input. The risk score was defined by normalising the risk function by a fixed value, the average risk of the training set. To fit the non-linear model >400,000 participants were used as training set and >45,000 participants were used as test set for validation. The exponent of risk function was represented as a multilayer neural network. This allowed capturing interdependent behaviour of covariates, training a single model for all outcomes, and preserving heterogeneity of the groups, which is in contrast to CoxPH models which are traditionally used in risk scores and require homogeneous groups. The model was trained over 60 epochs and predictive performance was determined by the C-index with standard errors and confidence intervals estimated with bootstrap sampling.
By inputing the variables described, one can obtain personalised hazard ratios for 3 major outcomes of CVD, cancer and all-cause mortality. Therefore, an individual with a risk Score of e.g. 1.5, at any time he/she has 50% more chances than average of experiencing the corresponding event. The proposed model showed the following discrimination, for risk of CVD (C-index = 0.8006), cancer incidence (C-index = 0.6907), and all-cause mortality (C-index = 0.7770) on the validation set. The CVD model is particularly strong (C-index >0.8) and is an improvement on a previous CVD risk prediction model also based on classical risk factors with total cholesterol and HDL-c on the UK Biobank data (C-index = 0.7444) published last year (Welsh et al. 2019). Unlike classically-used CoxPH models, our model considers correlation of variables as shown by the table of the values of correlation in Figure 1.
This is an accurate model that is based on the most comprehensive set of patient characteristics and biomarkers, allowing clinicians to identify multiple targets for improvement and practice active preventive cardiology in the era of precision medicine.
Figure 1. Correlation of variables in the R(x)
Funding Acknowledgement
Type of funding source: None
Machine Learning Prediction of Biomarkers from SNPs and of Disease Risk from Biomarkers in the UK Biobank