Exploring conceptual preprocessing for developing prognostic models: a case study in low back pain patients.

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Prognostic Modeling

Background:

• Developing accurate prognostic models is crucial for patient care.
• Large datasets with numerous potential prognostic factors present challenges.
• Conceptually oriented preprocessing aims to simplify models by grouping factors.

Purpose of the Study:

• To compare conceptually oriented preprocessing with statistical variable selection and expert preselection for prognostic models.
• To determine the optimal approach for developing prognostic models using extensive baseline data.
• To evaluate the prognostic capacity and model complexity of different preprocessing methods.

Main Methods:

• Utilized an existing dataset of low back pain patients with 112 baseline factors.
• Compared two forms of conceptually oriented subgrouping against expert preselection.
• Applied seven statistical methods, including Lasso-based variable selection, to predict six outcome variables.

Main Results:

• Conceptually oriented subgrouping showed limited prognostic capacity.
• Lasso-based variable selection on all factors or principal component scores yielded the best results.
• Expert preselection offered a favorable balance between model complexity and prognostic ability.

Conclusions:

• Purely statistical approaches, like Lasso, are often superior to conceptually oriented preprocessing for enhancing prognostic capacity.
• Expert preselection of established factors in a simple linear model is a viable strategy for large-scale prognostic rule development.