Calibration and Uncertainty in Neural Time-to-Event Modeling

Area of Science:

• Machine Learning
• Statistics
• Biostatistics

Background:

• Accurate prediction of event timing is vital for risk assessment in various fields.
• Traditional time-to-event (survival) models primarily focus on relative risk, not absolute event times.
• Existing models often lack sufficient calibration and uncertainty estimation for absolute predictions.

Purpose of the Study:

• To develop neural time-to-event models that predict accurate absolute event times.
• To enhance model calibration and uncertainty quantification in time-to-event predictions.
• To introduce novel methods for estimating and comparing conditional survival distributions.

Main Methods:

• Proposed adversarial nonparametric models for estimating time-to-event distributions.
• Introduced survival-function matching estimators to improve model calibration.
• Utilized distribution matching techniques for probabilistic predictions.

Main Results:

• The proposed neural time-to-event models demonstrate improved calibration and concentration of predicted event times.
• Distribution matching methods significantly outperform existing approaches in experiments.
• The survival-function matching estimator effectively accounts for model calibration and predictive uncertainty.

Conclusions:

• Neural time-to-event models offer a promising approach for accurate risk assessment.
• The developed methods provide better calibration and uncertainty estimation for time-to-event predictions.
• These advancements are crucial for reliable decision-making in applications relying on event time predictions.