Interpretable machine learning for high-dimensional trajectories of aging health

Area of Science:

• Computational biology
• Gerontology
• Biomedical informatics

Background:

• Aging is a complex process influenced by numerous interconnected health factors.
• Existing models often lack the ability to capture the dynamic and high-dimensional nature of aging trajectories.
• Personalized prediction of health and survival outcomes remains a challenge.

Purpose of the Study:

• To develop a novel computational model for predicting individual aging trajectories of health and survival.
• To create an interpretable network that reveals interactions between health variables.
• To assess the model's performance against existing methods and explore its utility in various applications.

Main Methods:

• Development of a dynamic joint interpretable network (DJIN) model.
• Integration of modern machine learning techniques with interpretable interaction networks.
• Utilizing longitudinal data from the English Longitudinal Study of Aging (ELSA) for model training and validation.

Main Results:

• The DJIN model accurately predicts individual health trajectories and survival.
• The model infers an interpretable network of directed interactions between health variables, revealing physiological connections.
• DJIN demonstrated superior performance compared to dedicated linear models and comparable performance to latent-space models.

Conclusions:

• The DJIN model offers a scalable and accurate approach to modeling individual aging.
• It provides valuable insights into the complex interplay of health variables during aging.
• The model has potential applications in generating synthetic aging data, imputing missing health information, and simulating future health outcomes.