Popis: |
For decades, researchers have been putting effort in advancing the prediction of common diseases to improve the identification of at-risk groups for preventive interventions, support physicians in medical decision making, and inform individuals about their risk or progression of disease. Ultimately, this would lead to health gain for many individuals. Current prediction models for common diseases typically include clinical, demographical, environmental and lifestyle predictors, but due to the multifactorial etiology of these diseases and the discoveries of many common single nucleotide polymorphisms (SNP; a variation occurring at a single nucleotide of the genome) over the past decades, there has been a great interest in adding polygenic risk scores (PRSs) to the clinical models. PRSs quantify the combined contribution of multiple SNPs to the risk of common diseases. Since prediction models are developed with the aim of applying them in healthcare, and hence medical decisions are based on the risk estimates, adequate risk predictions are of great importance. Therefore, prediction studies are needed to evaluate the predictive performance of prediction models and provide the necessary evidence for claims about the clinical validity and utility. This thesis describes methodological studies on (genetic) risk prediction of common diseases and aims to improve understanding and use of traditional and newer metrics of model performance and to provide insight into key concepts and considerations in prediction research. We conclude that the intended use of prediction models has a pivotal role in the design and interpretation of prediction studies. As the predictive ability of prediction models varies between populations and settings, the prediction study should be conducted with the targeted healthcare setting in mind, and claims about the readiness of PRSs for implementation in clinical care should be supported with evidence of well calibrated models and improved discriminative ability of the model compared to currently used prediction models. For the assessment of the discriminative ability of prediction models we have shown that the area under the receiver operating characteristic curve (AUC) is the separation between the risk distributions of patients and nonpatients. For the evaluation of all metrics, including the AUC, net reclassification improvement and integrated discrimination improvement applies that the interpretation should not only rely on the statistical significance, but also on their values in context of the intended use. The field of prediction research could be improved by using the intended use as guidance and by explaining prediction metrics more intuitively so that more researchers could have a greater understanding of them. Whether it is time to consider the implementation of PRSs in health care does not depend solely on the predictive performance of prediction models, but proof of sufficient predictive performance is essential before executing further studies on the usability, usefulness, and meaningfulness of PRS in healthcare. |