Novel Time-to-Event Machine Learning Approach for Integration of Longitudinal Data and Image Data

Two-sentence Summary of Public Health Significance: Current available risk prediction methods in medicine are limited in their ability to deal with high-dimensional, heterogeneous, and longitudinal data. Methods that not only capable of handling such complex data but also are more accurate, more interpretable, and robust are needed in the rising age of precision medicine.

Summary of Proposed Research: We will first validate current state-of-the-art machine learning and statistical methods for survival time prediction on the same datasets to justify for their clinical use. We will then propose a method for better survival time prediction and disease trajectory forecasting using a novel approach that incorporate the rich information from longitudinal data and image data, which current methods have limited capability to handle.

• Aim 1: To evaluate machine learning and deep learning survival methods against commonly used, statistical methods on the same datasets.
• Aim 2: To develop a method that could make use of longitudinal data in survival prediction.
• Aim 3: To develop a method that could combine variables automatically extracted from image-type data with the information from non-image data to enhance survival prediction.

The work we are proposing could result in much better, more suitable methods to analyze longitudinal, complex, and multi-modal data from cohort studies as well as electronic health records. The results could also be applied for survival analysis in many other domains outside of medicine.