py_ntcpx v1.1.0 — reproduction package (anonymised inputs, compiled results, and figure-generation code)

Reproduction package accompanying the article "py_ntcpx: an open-source, overfitting-aware framework for evaluating radiobiological and machine-learning NTCP models in small cohorts — proof-of-concept in parotid xerostomia." This archive is the citable snapshot of the data and figure layer of py_ntcpx v1.1.0. It contains: (i) 54 anonymised parotid dose–volume histogram (DVH) exports (PT001–PT054); (ii) all compiled per-patient results reported in the manuscript, including four-tier NTCP predictions (classical, MLE-refitted, multivariable logistic, and machine-learning tiers), uncertainty-aware NTCP (uNTCP) with 95% confidence intervals, Cohort Consistency Score (CCS), calibration metrics, and SHAP/LIME attributions; (iii) the staged pipeline outputs (code0–code7 and four-tier summary tables); and (iv) a single script that regenerates every main and supplementary figure (Figures 1–4 and Supplementary Figures S1–S9) from those results. The full analysis-pipeline source code is maintained at the GitHub repository (release tag py_ntcpx_v1.1.0). Reproduction instructions, a full file manifest with SHA-256 checksums, and a data dictionary are included in README.md and MANIFEST.md. Cohort: 54 head-and-neck cancer patients treated with concurrent chemoradiotherapy; endpoint Grade ≥2 xerostomia (31 events, 57.4%). gEUD computed with volume-effect exponent n = 0.45 (a = 2.2). Ethics/privacy: retrospective study; the Research Ethics Committee (GLA University, Mathura) confirmed no ethical approval was required. All direct identifiers were removed prior to deposition and replaced with study anonymisation codes. Licence: source code is released under the MIT Licence; anonymised data files are released under CC-BY-4.0.