Clinical prediction models using machine learning in oncology: challenges and recommendations.

Area of Science:

• Oncology
• Biostatistics
• Health Informatics

Background:

• Clinical prediction models are vital in oncology for personalized diagnosis and prognosis.
• Machine learning is increasingly used, but methodological flaws hinder implementation.

Purpose of the Study:

• To outline key considerations for developing robust and equitable cancer prediction models.
• To identify challenges in model development, evaluation, and implementation.

Main Methods:

• Review of critical steps: systematic reviews, protocol development, registration, end-user engagement, sample size, and data representativeness.
• Addressing technical challenges: missing data, fairness, complex data structures (censoring, competing risks, clustering).
• Comprehensive evaluation: statistical performance (discrimination, calibration) and clinical utility.

Main Results:

• Most cancer prediction models remain unimplemented due to methodological and translational challenges.
• Barriers include limited stakeholder engagement, insufficient clinical utility evidence, workflow integration issues, and lack of post-deployment monitoring.

Conclusions:

• Addressing development-to-practice gaps requires attention from study design to post-implementation monitoring.
• Developing trustworthy tools is essential for realizing the potential of personalized cancer care.