EVA Decompression Sickness Risk Dataset

FREE SAMPLE INCLUDED. This record includes a free 40-subject sample (SAMPLE.zip) together with the full dataset passport and validation report. The full 10,000-subject dataset (30,000 rows in observations and outcomes) is available for purchase at sentineldata.com.ua. Sample contents crew.parquet — 40 subjects × 10 columns (demographics, PFO status) observations.parquet — 120 rows × 7 columns (branch assignments, TR_peak, VGE grade) outcomes.parquet — 120 rows × 9 columns (DCS outcome, risk stage, flags) profile.parquet — 1,053 rows × 7 columns (decompression profile, full branch key) truth.parquet — 1,053 rows × 8 columns (audit constants per profile step) sample_passport_EN.pdf — schema, cohort design, model architecture, limitations sample_manifest.json — file hashes and metadata working_with_parquet.txt — quickstart guide About the Dataset The EVA Decompression Sickness Risk Dataset contains 10,000 synthetic astronaut crew members each evaluated under all three EVA pre-breathing protocol branches — a true within-subject counterfactual design generating 30,000 paired observations. All data are generated from a calibrated single-compartment Haldane nitrogen-washout physiological model with empirically grounded parameters and full audit trail. Physiological Model Tissue nitrogen saturation is computed via a perfusion-limited Haldane half-time model. The peak tissue ratio TR_peak drives a calibrated two-stage logistic: first predicting VGE (venous gas emboli) grade 0–4, then converting grade to binary DCS outcome. PFO (patent foramen ovale) presence multiplies DCS odds by RR=2.65 (Germonpré 2021), consistent with published clinical evidence. All 59 audit constants are recorded in the audit_constants table. Three Experimental Branches control — no nitrogen loading; DCS rate ≈ 17%, high-VGE rate ≈ 42.2% group2_pre_only — pre-EVA oxygen pre-breathing only; DCS rate ≈ 15%, high-VGE rate ≈ 38.3% group1_during_supersaturation — pre-breathing during the supersaturation peak; DCS rate ≈ 57%, high-VGE rate ≈ 19.5% Key Tables crew (10,000 × 10) — subject demographics and PFO status; PFO prevalence 24.4% observations (30,000 × 7) — one row per subject × branch; TR_peak, VGE grade, pre-EVA O₂ time outcomes (30,000 × 9) — DCS flag, risk stage, counterfactual linkage profile (1,053 × 7) — full decompression profile with ambient pressure schedule truth (1,053 × 8) — step-by-step nitrogen washout audit values audit_constants (59 × 14) — all model parameters with sources and verification status literature_registry (12 × 11) — calibration references with DOIs and effect sizes benchmark_results (11 × 7) — two-task ML benchmark with fold-level results Validation 9 automated checks — ALL_PASS = true. Calibration reproduction deviation < 0.0023. Task A benchmark (binary high_vge, logistic regression): accuracy 73.9%, AUC 0.752 — matches the Bayes ceiling to 4 decimal places (expected: these two covariates fully determine the generating probability). Task B (5-class vge_grade, gradient boosting): accuracy 41.8%, weighted kappa 0.344 — bounded by construction because fine grade within each binary side is an undisclosed random draw requiring Doppler bubble-detection signal to resolve. Permitted Use Non-commercial research, education, and algorithm development under CC BY-NC 4.0. Commercial use requires a separate license from CODE SENTINEL DATA.