A Coverage Debt Ledger for CDSS Deployment Review: Auditing How Conformal Uncertainty Is Distributed Between Risk Strata in Two National Hospital Systems

Analysis code and derived outputs accompanying the manuscript "Coverage debt: an auditable ledger for conformal prediction in clinical decision support" (Enríquez López J, Chuga Martínez E, Arias T. SA). The study evaluates a mortality-prediction clinical decision support system trained on 3,258,421 harmonized hospital discharge records from Ecuador (INEC, 2024) and the United States (SPARCS New York, 2023). Pooled conformal calibration met its 90% nominal coverage target almost exactly (89.99% empirical) while distributing coverage inversely to mortality risk: patients aged 70 or older, 25.1% of the test population and the highest-risk stratum, carried 95.7% of the shortfall. This deposit contains the coverage debt ledger implementation, which computes the number of patients by which each subgroup's realised coverage falls short of nominal. The metric requires only subgroup size and realised coverage, so it runs from published subgroup tables without access to the model or to individual-level predictions. Also included are the harmonization script for the two discharge registries, a missingness audit of the Ecuadorian source across all 1,132,667 records, and the calibration decile data behind Figure 1. Neither source registry is redistributed. Both are publicly available: INEC Registro Estadístico de Camas y Egresos Hospitalarios 2024 (https://anda.inec.gob.ec/anda5/index.php/catalog/1151) and SPARCS Hospital Inpatient Discharges, De-Identified, 2023 (https://health.data.ny.gov/Health/Hospital-Inpatient-Discharges-SPARCS-De-Identified/46xm-urtu). The ICD-10-chapter to APR-MDC lookup table used in the original harmonization is not included; it is not recoverable from the archived outputs. The harmonization script documents the mapping it applies. See README.md for a file-by-file description.