Global and Episode-Specific Prediction of Recurrent Events Using Longitudinal Health Informatics Data.

Area of Science:

• Health Informatics
• Biostatistics
• Machine Learning in Healthcare

Background:

• Accurate prediction of recurrent clinical events is vital for managing chronic diseases like cancer and cardiovascular disease.
• Longitudinal health informatics databases are increasingly used for risk prediction models of repeated clinical events.

Purpose of the Study:

• To introduce a novel nonparametric framework for predicting recurrent events using survival tree ensembles.
• To address complexities in tree-based recurrent event prediction, including informative censoring and inter-event correlations.
• To offer a promising alternative to traditional Cox-type models by avoiding strong assumptions on event history.

Main Methods:

• Developed a nonparametric framework utilizing survival tree ensembles.
• Incorporated two predictive modeling strategies: episode-specific and global models.
• Addressed induced informative censoring and inter-event correlations using inverse probability of censoring weighting and modified resampling procedures.

Main Results:

• The novel framework demonstrated superior performance in predicting recurrent hospitalizations for breast cancer patients using SEER-Medicare data.
• Global models, which borrow information across events, significantly enhanced prediction accuracy for later hospitalizations.
• The proposed models avoided strong assumptions, offering a flexible alternative to existing methods.

Conclusions:

• The survival tree ensemble framework provides an effective nonparametric approach for recurrent event prediction.
• Borrowing information across events via global models is a key strategy for improving prediction accuracy in later events.
• This framework offers a valuable tool for enhancing the management of chronic conditions through improved risk prediction.