GPU Accelerated Estimation of a Shared Random Effect Joint Model for Dynamic Prediction.

Area of Science:

• Biostatistics
• Computational Biology
• Epidemiology

Background:

• Predicting clinical event risk using longitudinal data is crucial in cohort studies.
• Dynamic changes in risk factors and their associations over time complicate accurate prediction.
• Existing joint models for longitudinal and survival data face computational challenges, especially with large datasets and nonlinear trajectories.

Purpose of the Study:

• To develop a computationally efficient method for joint modeling of longitudinal and survival data.
• To improve the accuracy of risk prediction by incorporating nonlinearities in longitudinal predictors.
• To leverage Graphics Processing Unit (GPU) programming for faster model estimation.

Main Methods:

• A novel two-stage estimation procedure was developed.
• Graphics Processing Unit (GPU) programming, implemented via PyTorch, was used to accelerate computations.
• The proposed method was evaluated through numerical studies on large datasets.

Main Results:

• The proposed algorithm and software significantly reduced estimation time for joint models.
• Accounting for nonlinearity in longitudinal predictor trajectories enhanced prediction accuracy compared to models ignoring nonlinearity.
• The computational speedup was particularly notable for large datasets.

Conclusions:

• The developed method offers a computationally efficient and accurate approach for joint modeling in longitudinal studies.
• GPU programming and a two-stage estimation procedure effectively address the computational burden of complex joint models.
• Accurate modeling of nonlinear predictor trajectories is essential for improving clinical event risk prediction.