Discrimination performance in illness-death models with interval-censored disease data

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Survival Analysis

Background:

• Illness-death models are standard for disease progression analysis.
• Clinical data often features interval-censored disease onset due to scheduled visits.
• Ignoring interval-censoring can bias model performance evaluation.

Purpose of the Study:

• To assess the impact of ignoring interval-censored disease data on illness-death model discrimination.
• To compare different estimation methods for interval-censored illness-death data.
• To evaluate discrimination using time-specific area under the ROC curve.

Main Methods:

• Simulation study using Weibull transition hazards with interval-censored data.
• Comparison of Cox model (ignoring interval-censoring) vs. interval-censored illness-death models (msm, SmoothHazard packages).
• Application to a high-grade soft tissue sarcoma patient dataset (n=2232).

Main Results:

• Ignoring interval-censoring significantly affects discrimination performance metrics.
• Interval-censored models (piecewise-constant, Weibull, M-spline) provide more accurate estimates.
• The choice of method impacts the evaluation of model discrimination.

Conclusions:

• Accounting for interval-censoring is essential for accurate illness-death model parameter estimation.
• Proper handling of interval-censored data is critical for reliable discrimination performance assessment.
• This study highlights the importance of robust statistical methods in clinical research involving time-to-event data.