Isotonic Regression Calibrates Deep Cox Survival Models
A novel post hoc calibration technique for Deep Cox models employs isotonic regression to enhance predicted survival probabilities while maintaining discriminative capability. Introduced in arXiv:2605.16571, this method tackles the frequent issue of inadequate calibration in deep survival models. It provides theoretical assurances such as double-robustness and asymptotic calibration. Empirical tests on both synthetic and actual clinical datasets showcase its effectiveness. This method is adept at managing censored time-to-event data within life sciences and engineering, making it suitable for unstructured data types like clinical narratives, genomic sequences, and pathology images.
Key facts
- Method uses isotonic regression for post hoc calibration
- Does not affect discriminative power
- Theoretical guarantees: double-robustness and asymptotic calibration
- Tested on synthetic and real-world clinical data
- Applicable to unstructured data (clinical text, genomics, pathology images)
- Addresses poor calibration in Deep Cox models
- Handles censored time-to-event data
- Published on arXiv with ID 2605.16571
Entities
—