Classification versus association models: should the same methods apply?

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Biomarker Research

Background:

• Association studies identify biomarker-disease links for etiological insights using measures like odds ratios.
• Classification studies assess biomarker utility for individual patient decisions, using metrics such as sensitivity and specificity.
• Current classification model development often uses association model criteria, which are not always optimal for clinical applications.

Purpose of the Study:

• To differentiate the fundamental objectives and methodologies of association versus classification models.
• To highlight the limitations of using association model criteria for classification model development.
• To propose an optimized approach for developing biomarker classification models.

Main Methods:

• Conceptual analysis of association and classification model objectives and measurements.
• Review of common statistical metrics used in both types of studies (e.g., odds ratio, sensitivity, PPV).
• Proposal of a novel method for classification model development focusing on specific receiver operating characteristic (ROC) curve regions.

Main Results:

• Association and classification models have distinct goals, measurements, and clinical relevance.
• Good association does not guarantee good classification performance.
• Existing methods for classification model development may not be clinically optimal.
• Focusing on application-specific ROC curve regions can optimize classification models.

Conclusions:

• Biomarker association and classification models are fundamentally different and serve distinct purposes.
• Classification models should be developed with clinical application specificity in mind.
• Optimizing classification models using relevant receiver operating characteristic (ROC) curve regions enhances their utility in clinical decision-making.