COBS — The Cost of Binarizing Survival

Code and data for "The Cost of Binarizing Survival Outcomes in Clinical Prognostic Modeling" (Yadav, Routman, Foong), Machine Learning for Healthcare (MLHC) 2026.

The paper introduces the Survival-Aware Bayesian Network (SA-BN), which replaces the binary scoring in Bayesian-network feature selection with the Cox partial log-likelihood and quantifies how much prognostic signal is lost when a survival endpoint is binarized.

Setup

pip install numpy pandas scikit-learn scikit-survival lifelines statsmodels pgmpy networkx

Reproduce

bash reproduce.sh

Results are written to survival/notebooks/; reference outputs (*.json, *.md) are included for comparison.

Layout

common.py, surv_common.py, surv_bn_aware.py   # SA-BN implementation
shah_replication.csv, hancock_rebuilt.csv     # head-and-neck cohorts (radiation, surgical)
tcga_extra/                                    # TCGA clinical files (breast, colorectal, kidney)
survival/                                      # analysis scripts + reference outputs

Data

• RADCURE (radiation): https://www.cancerimagingarchive.net/collection/radcure/
• HANCOCK (surgical): https://hancock.research.fau.eu/
• TCGA (BRCA / COAD / KIRC): public clinical data via cBioPortal (included in tcga_extra/).