X-CAL: Explicit Calibration for Survival Analysis.

Area of Science:

• Computational Biology and Bioinformatics
• Machine Learning in Healthcare
• Statistical Modeling

Background:

• Survival analysis models time-to-event data, crucial for predicting outcomes like hospital discharge or ICU admission.
• Model calibration, ensuring predicted event counts match observed ones, is vital for reliable survival predictions.
• Traditional calibration methods are applied post-training, limiting their integration into the modeling process.

Purpose of the Study:

• To develop a novel, integrated approach for optimizing survival model calibration during the training phase.
• To introduce explicit calibration (X-CAL) as a differentiable objective function for survival models.
• To evaluate the impact of X-CAL on model calibration and its trade-off with predictive performance.

Main Methods:

• Formulated distributional calibration (D-CALIBRATION) as a differentiable objective (X-CAL).
• Integrated X-CAL with maximum likelihood estimation for joint optimization of calibration and prediction.
• Applied X-CAL to various shallow and deep survival models using simulated, MNIST-based, MIMIC-III, and TCGA brain cancer datasets.

Main Results:

• Demonstrated that commonly used survival models can exhibit miscalibration across diverse datasets.
• Showcased X-CAL's effectiveness in significantly improving D-CALIBRATION across all tested models and datasets.
• Confirmed that X-CAL enhances calibration without substantial degradation of model concordance or likelihood.

Conclusions:

• Explicit calibration (X-CAL) provides a practical and effective method for directly optimizing survival model calibration.
• X-CAL enables a balanced approach, improving reliability without compromising predictive accuracy.
• This differentiable objective facilitates the development of more trustworthy and well-calibrated survival prediction models.