By using classification methods from statistical learning, we developed new models that can predict the risk of cardiovascular disease at least as well as the models used by the healthcare system today. We also uncovered several variables that may be important risk factors, but which are not included in today's prediction models - including exercise.

As a basis for the new models, we linked data from 50,000 healthy participants in the Nord-Trøndelag Healty Study with data on cardiovascular disease from various health registers. We found the 20 variables that had the greatest connection with the risk of suffering a heart attack or stroke during the next ten years. We developed three prediction models for each gender, and then evaluated how well these models can predict cardiovascular disease, compared to both the Norwegian NORRISK 2 model and the British Framingham model.

The models we developed for men performed significantly better than both NORRISK 2 and Framingham, even though two of our three models also included all seven variables that make up NORRISK 2. The models we came up with for women, on the other hand, did not perform much better than NORRISK 2. In contrast to the NORRISK 2 model, neither HDL cholesterol nor heart disease in near relatives were included as variables in any of our prediction models for women.

The inflammatory marker CRP and the enzyme ALP in the blood were among the less established risk factors that were included in some of our models for both sexes. Chronic illness or injury was also linked to increased risk in our machine learning analysis, but is not included in today's existing models for assessing cardiovascular risk.

Read the full Master's thesis:
Risk Prediction of Cardiovascular Disease with Statistical Learning Methods