Optimizing dynamic predictions from joint models using super learning.

Area of Science:

• Biostatistics
• Medical Informatics
• Computational Biology

Background:

• Joint models for longitudinal and time-to-event data are crucial for dynamic individualized predictions in precision medicine.
• Model accuracy depends on longitudinal trajectory shape and the link between outcome history and event hazard.
• Developing a single, accurate joint model for diverse subjects and follow-up times, especially with multiple outcomes, is challenging.

Purpose of the Study:

• To introduce a super learning approach for joint models to enhance predictive accuracy.
• To avoid selecting a single, potentially suboptimal, joint model specification.
• To provide a robust method for dynamic individualized predictions in precision medicine.

Main Methods:

• Utilized super learning by creating a weighted combination of dynamic predictions from a library of joint models.
• Optimized weights using V-fold cross-validation to maximize a chosen predictive accuracy metric.
• Employed expected quadratic prediction error and expected predictive cross-entropy as accuracy measures.

Main Results:

• The super learning approach demonstrated predictive performance comparable to the Oracle model (best-performing model on test data).
• This ensemble method effectively handles complexities in longitudinal data and event time associations.
• The proposed methodology is implemented in the R package JMbayes2.

Conclusions:

• Super learning offers a robust alternative to single model selection for joint models.
• This method improves the accuracy of dynamic individualized predictions, advancing precision medicine applications.
• The JMbayes2 R package provides accessible implementation of these advanced statistical techniques.