PRIM versus CART in subgroup discovery: when patience is harmful

Area of Science:

• Clinical informatics
• Machine learning in healthcare
• Data mining for medical research

Background:

• Subgroup discovery is crucial for identifying patient cohorts with unique characteristics in high-dimensional clinical data.
• Established algorithms like CART and PRIM are commonly used, but their comparative performance in real-world clinical settings is not fully understood.

Purpose of the Study:

• To systematically compare the performance of CART and PRIM for subgroup discovery on a large, high-dimensional clinical database.
• To evaluate the effectiveness of these algorithms in identifying clinically relevant patient subgroups, particularly when dealing with ordinal variables.

Main Methods:

• A systematic comparison of CART and PRIM algorithms was conducted.
• The algorithms were applied to a large, real-world, high-dimensional clinical database for subgroup discovery.
• Performance was evaluated based on the ability to identify significant patient subgroups.

Main Results:

• Contrary to prevailing assumptions, CART generally demonstrated superior performance compared to PRIM.
• PRIM's "peeling off" strategy often disregarded valuable data segments associated with ordinal variables, leading to the omission of important subgroups.
• CART proved more effective in navigating the complexities of the clinical dataset.

Conclusions:

• CART is a more effective algorithm than PRIM for subgroup discovery in high-dimensional clinical databases, especially those with ubiquitous ordinal variables.
• Improvements in PRIM's utility could be achieved by better leveraging global information of ordinal variables and incorporating mechanisms to track alternative solutions.