Current disease risk prediction models, such as the Framingham risk score for coronary artery disease, rely on a limited set of demographic and clinical variables measured at a single time point, achieving a C statistic of 0.74. While this performance is good, there is significant room for improvement, particularly given the critical role of accurate risk prediction in precision medicine for guiding clinical decisions among multiple treatment options.
Disease progression typically follows temporal patterns that may be better captured through longitudinal analysis. This biological characteristic, combined with recent advances in machine learning, presents an opportunity to enhance prediction accuracy. Our research questions focus on determining the optimal historical time frame of health data and identifying which combinations of health status parameters contribute most significantly to disease risk prediction.
The objective is to develop and validate a machine learning algorithm that predicts future disease risk by analyzing temporal health data integrated with multi-omics information. This approach could potentially reduce the need for invasive diagnostic procedures like cardiac catheterization by accurately identifying low-risk patients through non-invasive methods.
Recent developments in deep learning, particularly transformer models, demonstrate extraordinary performance in sequence prediction tasks, making them well-suited for analyzing temporal health records. Furthermore, modern neural network architectures have proven capable of integrating diverse data types, such as text and images, suggesting their potential for combining clinical records with multi-omics data. This technical foundation provides a scientifically rational approach to developing a more comprehensive and accurate disease risk prediction system.