Our study aims to:
1. Develop and validate predictive models for cardiovascular disease (CVD) and associated mortality using three distinct data modalities:
a) Bio-signaling data, including photoplethysmography and accelerometer measurements
b) Imaging data, including cardiac MRI, and carotid sonography.
c) Genetic sequencing data, focusing on known CVD-related polymorphisms
2. Evaluate and compare the predictive performance and economic implications of each machine learning model.
3. Investigate the potential for integrating these diverse data types into a unified, multi-modal predictive framework to enhance overall CVD risk stratification accuracy.
We intend to use the comprehensive data resources of UK Biobank, including accelerometer derived predictors, self-reported physical activity from short-form international physical activity questionnaire (IPAQ), PPG waveform variables, cardiac MRI imaging, carotid sonography imaging, genetic sequencing data, and demographic factors.
We aim to develop machine learning-based survival analysis models for each data type to estimate the probability of CVD events and mortality over time. Model performance will be evaluated using time-dependent Area Under the Receiver Operating Characteristic curve (AUROC).
Additionally, we will conduct a comprehensive analysis of the economic implications of implementing each model in clinical practice. This analysis will consider factors such as data acquisition costs, computational requirements, and potential healthcare savings.
Our primary hypothesis posits that models utilizing bio-signaling data and imaging will demonstrate superior efficacy in predicting short-term CVD incidence, while models incorporating genomic data will exhibit enhanced performance in forecasting longer-term CVD incidence and mortality.
