Towards the advanced predictive modelling in epidemiology

Abstract Data-driven prediction systems used in epidemiological studies are still unsatisfactory from a practical point of view. Different pitfalls should be considered while transferring technologies from research to practice. The proposed k-Nearest Neighbors approach is designed to make disease-related predictions in a more holistic manner: we detect cases of novelty among unobserved subjects to identify situations when model predictions are not reasonably valid. Moreover, it copes with overlapping classes, finds new examples which cannot be labelled with the high confidence and reveals healthy subjects in the training data who might be at risk. Additionally, variable selection is built-in to select relevant predictors. The approach was applied to predict cardiovascular diseases based on the data collected within an ongoing follow-up study undertaken in Eastern Finland. According to the experimental results, our proposal allows increasing the accuracy of predictions made.