Calibrated predictions for multivariate competing risks models.

Area of Science:

• Biostatistics
• Epidemiology
• Medical Informatics

Background:

• Time-to-event prediction models are crucial for individual disease risk assessment, particularly for diseases with a strong family history component.
• Family history reflects genetic susceptibility, shared environment, and behavioral patterns, influencing disease risk.
• Accurate models aid in identifying high-risk individuals, estimating disease burden, and guiding patient care.

Purpose of the Study:

• To develop and evaluate prediction models for time-to-event data that incorporate family history and account for competing risks.
• To address the limitations of current methods that may misestimate risk when competing events are present.

Main Methods:

• Utilized frailty models to accommodate family history as a risk factor.
• Introduced a novel approach to incorporate competing risks (e.g., other diseases, mortality) into prediction models.
• Conducted simulation studies to compare the proposed method with traditional approaches that ignore competing risks.

Main Results:

• Naively treating competing risks as independent censoring events leads to non-calibrated predictions, overestimating the expected number of events.
• Discrimination performance of prediction models is not significantly affected by ignoring competing risks.
• The proposed methodologies that correctly account for competing events demonstrate excellent calibration and are easy to implement.

Conclusions:

• Accurate disease risk prediction requires explicit modeling of competing risks, especially when family history is a factor.
• The proposed frailty models with competing risks provide well-calibrated and practical tools for risk assessment.
• Ignoring competing risks can lead to biased risk estimates, impacting clinical utility and population health assessments.