The Risks of Risk Assessment: Causal Blind Spots When Using Prediction Models for Treatment Decisions.

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Health Informatics

Background:

• Clinicians increasingly use prediction models for treatment guidance.
• Many models use observational data, including patients already treated, complicating interpretation.
• Causal inference is crucial for accurate prediction model development and application.

Purpose of the Study:

• To identify and illustrate
• causal blind spots
• in common prediction model development approaches.
• To advocate for extended guidelines ensuring accurate interpretation and application of prediction models.

Main Methods:

• Analysis of common prediction model development strategies (inclusion, restriction, exclusion of treatment).
• Illustration using real-world examples to demonstrate misinterpretation risks.
• Discussion of causal issues: confounding, selection, mediation, and time-varying treatments.

Main Results:

• Three common approaches to handling treatment in prediction models were found to have "causal blind spots."
• Misinterpretation of model-derived risks can lead to suboptimal clinical decision-making.
• Existing methods may not adequately account for confounding, selection, mediation, and treatment changes.

Conclusions:

• Prediction models intended for treatment decisions require "prediction under interventions," necessitating causal reasoning.
• Clear communication of model applicability, target population, and treatment conditions is essential.
• Adopting causal inference techniques and updated guidelines can prevent misinformed decisions and patient harm.