From algorithms to action: improving patient care requires causality.

Area of Science:

• Oncology
• Biostatistics
• Clinical Decision Support

Background:

• Outcome prediction models are crucial in cancer research for treatment decisions.
• Current validation methods for these models often overlook causal inference, leading to potential harm.
• Existing guidelines do not adequately address the causal validation of prediction models.

Discussion:

• Published prediction models, despite validation accuracy, can be harmful if causal aspects of treatment decisions are ignored.
• The discrepancy arises from validating models on associations rather than causal effects.
• This highlights a critical gap in the current methodology for developing clinically applicable prediction models.

Key Insights:

• Prediction models must be validated considering the causal impact of treatment decisions.
• Standard validation metrics do not guarantee clinical utility or prevent harm.
• A causal framework is essential for developing reliable decision-support tools in oncology.

Outlook:

• Future research should focus on causal inference methods for prediction model development and validation.
• Revised guidelines are needed to incorporate causal validation standards.
• This will enhance the reliability and safety of prediction models in clinical practice.