Deep learning for the dynamic prediction of multivariate longitudinal and survival data

Area of Science:

• Biostatistics
• Machine Learning
• Health Informatics

Background:

• Traditional joint models for longitudinal and survival data face limitations due to parametric assumptions and computational challenges with multiple outcomes.
• Accurate time-to-event prediction is crucial, particularly when incorporating evolving longitudinal health data.

Purpose of the Study:

• To explore and develop advanced machine learning methods for improved time-to-event prediction using multivariate longitudinal data.
• To introduce and evaluate a novel transformer-based neural network architecture, TransformerJM, for joint longitudinal and survival data modeling.

Main Methods:

• Utilized functional data analysis and convolutional neural networks for modeling longitudinal data, assessing their scalability with multiple outcomes.
• Proposed and implemented TransformerJM, a novel architecture integrating longitudinal and time-to-event data.
• Compared model performance using simulations and real-world Alzheimer's disease datasets, focusing on dynamic prediction updates.

Main Results:

• TransformerJM demonstrated improved predictive performance compared to existing methods across various scenarios.
• The proposed methods effectively handle multiple longitudinal outcomes, overcoming limitations of traditional joint models.
• Models were evaluated on their ability to dynamically update predictions as new longitudinal data became available.

Conclusions:

• TransformerJM offers a powerful and flexible approach for joint modeling of longitudinal and time-to-event data.
• Machine learning methods, particularly TransformerJM, provide significant advancements in time-to-event prediction accuracy and scalability.