Dynamic models to predict health outcomes: current status and methodological challenges.

Area of Science:

• Medical Informatics
• Biostatistics
• Health Services Research

Background:

• Clinical prediction models degrade over time due to evolving disease patterns and healthcare systems.
• Static models become outdated, necessitating adaptive solutions for sustained accuracy.
• Dynamic prediction models offer a potential solution by evolving with observed changes.

Purpose of the Study:

• To review the state-of-the-art in dynamic prediction modeling.
• To identify current methodologies and their applications.
• To highlight unresolved methodological challenges in the field.

Main Methods:

• Systematic literature search of MEDLINE, Embase, and Web of Science.
• Extraction of data on model updating techniques, update windows, decay factors, and validation.
• Identification of reported limitations and future research recommendations.

Main Results:

• Eleven papers identified seven distinct dynamic prediction modeling methods.
• Methods categorized into frequentist (discrete updates), Bayesian (continuous updates), and time-varying coefficient approaches.
• Limited empirical comparisons and applications across diverse healthcare problems were noted.

Conclusions:

• Dynamic models address calibration drift inherent in static models.
• Challenges remain in selecting decay factors and managing abrupt data shifts.
• Validation strategies for dynamic prediction models require further development and exploration.