New Isotonic Calibration Fixes Deep Cox Survival Models
A post-hoc fix that slashes miscalibration without hurting predictive accuracy.
Deep Cox models are widely adopted for time-to-event analysis in medicine and engineering because they handle censored data and can ingest unstructured inputs like clinical notes or genomic sequences. Despite their flexibility, these models often produce poorly calibrated survival probabilities — meaning the predicted risk of an event at a given time does not match the actual observed risk. This miscalibration limits their reliability in high-stakes decisions such as treatment planning or patient prognosis.
To address this, Jain, Zhang, and Bates introduce a novel post-hoc calibration technique that applies isotonic regression to the model's output survival probabilities. The method preserves the model's discriminative power (its ability to rank patients by risk) while aligning predicted probabilities with empirical frequencies. The authors prove strong theoretical properties, including double-robustness — calibration holds even if the base model is misspecified — and asymptotic consistency. Experiments on synthetic data and real clinical datasets confirm that the approach significantly reduces calibration error without degradation in concordance indices.
- Isotonic regression refines survival probabilities without hurting discriminative ability (c-index preserved).
- Double-robustness guarantee: calibration holds even under model misspecification.
- Validated on synthetic data and real clinical datasets, showing reduced calibration error.
Why It Matters
Better-calibrated survival models enable safer, more reliable clinical decisions and risk assessments from Deep Cox models.