Cure modeling in real-time prediction: How much does it help?

Area of Science:

• Clinical Trials
• Biostatistics
• Survival Analysis

Background:

• Real-time prediction in time-to-event clinical trials utilizes various parametric and nonparametric modeling approaches.
• Parametric cure-mixture modeling, proposed by Chen (2016), addresses scenarios with a potentially cured patient fraction.

Purpose of the Study:

• To apply a Weibull cure-mixture model for predictions in the RTOG 0129 head-and-neck cancer trial.
• To compare predictions from a Weibull cure-mixture model against a standard Weibull model and a nonparametric Bayesian bootstrap model.

Main Methods:

• Application of a Weibull cure-mixture model to RTOG 0129 data for interim predictions.
• Comparison of realized data with predictions from Weibull cure-mixture, standard Weibull, and Bayesian bootstrap nonparametric models.

Main Results:

• The standard Weibull model predicted earlier events than the cure-mixture model, with divergence increasing as evidence for a cure emerged.
• Nonparametric predictions frequently resulted in undefined values or infinite intervals, especially early in the trial.
• Simulations indicated cure modeling provides better-calibrated prediction intervals for cured components but at the expense of interval width.

Conclusions:

• Weibull cure-mixture modeling is effective for time-to-event predictions in clinical trials with a cured fraction.
• Cure modeling offers improved prediction interval calibration when a cured component is present or suspected.
• While beneficial for accuracy, cure modeling may lead to wider prediction intervals compared to simpler models.