Random survival forests with multivariate longitudinal endogenous covariates.

Area of Science:

• Biostatistics
• Personalized Medicine
• Machine Learning in Healthcare

Background:

• Predicting individual clinical event risk from complete patient history is crucial for personalized medicine.
• Existing analytical methods struggle with numerous time-dependent predictors that have irregular, error-prone measurements and are subject to informative truncation by the event.

Purpose of the Study:

• To extend competing-risk random survival forests to predict event probabilities using endogenous longitudinal predictors.
• To develop and evaluate a method, implemented in the R package DynForest, capable of handling complex patient history data for dynamic risk prediction.

Main Methods:

• Extended competing-risk random survival forests (DynForest) to incorporate time-dependent longitudinal predictors.
• Internally transformed time-dependent predictors into time-fixed features using mixed models at each tree node.
• Computed individual event probabilities by averaging leaf-specific Aalen-Johansen estimators across trees.

Main Results:

• DynForest demonstrated effective prediction of clinical events, outperforming a joint modeling approach with two predictors.
• The method proved superior to regression calibration when handling a large number of longitudinal predictors, accounting for informative truncation.
• An application in dementia research showcased DynForest's utility in developing dynamic prediction tools from multimodal markers.

Conclusions:

• DynForest provides a robust framework for dynamic risk prediction in personalized medicine, effectively managing complex longitudinal data.
• The method allows for the quantification of marker importance in prediction models, aiding clinical interpretation.
• DynForest offers a valuable tool for developing advanced clinical prediction models, particularly in areas like dementia research.