Consensus Framework for Cardiovascular Risk Prediction in a Clinical Setting

点击翻译按钮获取中文摘要

Project Summary/Abstract Accurate assessment of cardiovascular risk in clinical settings is important for the appropriate management and counseling of millions of patients. Current US treatment guidelines focus on a single risk score. This approach has good overall performance but has limitations when applied to clinical setting. For example, important new markers or specific missing patterns cannot be accommodated with established, guideline- endorsed prediction scores. Scores also often over or under predict in new populations or subgroups. These gaps call for exploring approaches to risk prediction that go beyond a single model paradigm and beyond classical Cox proportional hazards and logistic regression methods. At the same time, recent publications question the utility of uninterpretable models in prognostic settings especially in high-stakes situations in healthcare. Therefore we need to strike a balance between the sophistication of a model and its interpretability in order to avoid potentially tragic consequences for patients in the near future.The goal of this project is to address these issues. We will evaluate the improvement in discrimination and calibration of existing consensus models such as the Super Learner and eXtreme Gradient Boosting in clinical settings and develop a novel method called the Consensus Framework. This novel method has the consensus property because it combines multiple published and validated risk models to ensure not only good overall performance but also good performance in important subgroups of patients. The Consensus Framework is adapted to clinical practice because it can handle limited information or additional risk factors. We will also assess specific properties of prognostic risk prediction and how they inform the selection of the most appropriate class of models. This project is relevant to public health because 1) the interpretability of the consensus models that we propose to use ensures transparency in the assessment of their quality and limitations, which is of paramount importance in high-stakes decision making in healthcare, 2) their flexibility to accommodate missing data or the availability of known risk factors will produce more personalized treatment decisions, 3) a better understanding of the unique properties of prognostic risk prediction will dictate a more informative choice of prognostic model. All these factors will lead to better informed treatment decisions for millions of patients.