A 72-Hour Machine Learning Prognostic Model for ARDS Patients Requiring Invasive Mechanical Ventilation

Aligns the repository with the revised Supplementary Tables S8 and S14, and makes the three input workbooks self-explanatory by name.

Added

• python/resampling/optimism_all8_reselect.py --arm oob — the out-of-bag variant reported in Table S8, section D. The selectors are re-run inside every resample exactly as in the reselect arm, but each replicate is scored only on the patients left out of that resample, so n_test (94–118, median 105) is written to the per-replicate detail file.

Changed

• Both selectors now run with the seeds reported in Table S14, row 6: the resampling seed is 499 + b (bootstrap replicates 500–699) and Boruta and LASSO both use 500. The out-of-bag arm and the reselect arm therefore share one set of resamples.
• The input workbooks are named in plain language: training set.xlsx, internal validation set.xlsx, external validation set.xlsx. Only the development cohort is read by the selection, tuning, fitting and resampling scripts; the two validation workbooks are read when the frozen model is scored and need the predicted probability as column p.
• Every command-line example is quoted, because the file names now contain spaces.
• README.md and data/README.md describe the role of each workbook and the columns it needs.
• CITATION.cff and the citation block in README.md bumped to v1.0.4.

Reproducibility

Patient-level data are not distributed. The four committed analysis steps (Boruta + LASSO, grid search, final fitting, bootstrap optimism) run unchanged on the development cohort when data/training set.xlsx and data/candidate_predictors.txt are in place.