A Product Partition Model With Regression on Covariates.

Area of Science:

• Biostatistics
• Machine Learning
• Computational Biology

Background:

• Accurate prediction of patient outcomes is crucial in clinical trials.
• Existing clustering methods may not fully leverage patient covariate data.
• Predicting time to progression in breast cancer requires sophisticated analytical tools.

Purpose of the Study:

• To develop a model-based clustering algorithm that incorporates patient covariates for improved prediction.
• To extend Product Partition Models (PPM) to include regression on covariates.
• To provide a flexible framework for analyzing diverse covariate types in clinical data.

Main Methods:

• Proposed a probability model for random partitions with covariate integration.
• Developed a weighted averaging approach based on covariate similarity for predictions.
• Extended PPM by modifying the cohesion function to favor clustering of similar covariates.

Main Results:

• The model a priori clusters patients with similar covariates.
• Posterior predictive inference provides a formal mechanism for prediction.
• The approach is adaptable to various covariate types (continuous, categorical, count, ordinal).

Conclusions:

• The developed model effectively utilizes covariates for patient clustering and outcome prediction.
• This covariate-inclusive approach offers enhanced predictive power for clinical trial data.
• An R-package implementation is available, facilitating practical application of the model.