Post-Landing Orthostatic Intolerance Dataset

This upload contains the full technical documentation and a free evaluation sample (32 of 10,000 virtual astronauts, same schema as the full dataset) — not the full dataset. The full 10,000-subject dataset (crew, observations at 2-second resolution, factual/counterfactual outcomes, ground-truth physiology, benchmark script) is sold separately: https://sentineldata.com.ua/dataset/post-landing-orthostatic-intolerance ($490). Sample contents FileContentcrew.parquet32 subjects × 7 columns (demographics, mission, countermeasure)observations.parquet5,312 rows × 8 columns (noisy monitor readings at 2-second resolution)outcomes.parquet64 rows × 9 columns (factual + counterfactual outcomes per subject)truth.parquet9,632 rows × 11 columns (ground-truth physiology, unobservable in real HUT)sample_passport_EN.pdfSample data passport with schema and composition detailsworking_with_parquet.txtInstructions for pandas / R arrow / DuckDBsample_manifest.jsonSHA-256 checksums per file A synthetic cohort of 10,000 virtual astronauts exposed to a 10-minute 80° head-up tilt test immediately after landing. Each subject is simulated with a mechanistic cardiovascular model (Ursino-Magosso variant) run from a mission-specific baseline: short-duration (5–16 days) and long-duration (129–190 days) spaceflight stints, four countermeasure arms, and continuous hemodynamic output at 1-second resolution. The dataset ships a factual/counterfactual outcome pair per subject, noisy monitor observations at 2-second resolution, and the full unobserved ground-truth physiological state — the only synthetic dataset that exposes all three layers simultaneously. 1. Overview Post-landing orthostatic intolerance (OI) is the most common cardiovascular adaptation seen after spaceflight: astronauts returning from orbit experience transient inability to maintain blood pressure upright, ranging from mild lightheadedness to frank presyncope. The tilt-table or stand test is the standard clinical diagnostic. This dataset simulates that diagnostic at scale — 10,000 individual HUT episodes with full mechanistic physiology, realistic monitor noise and instrument dropout, and paired treatment outcomes — to support model development and countermeasure evaluation without access to restricted astronaut medical records. 2. Cohort & Simulation Design Parameter Value Subjects10,000 virtual astronauts Mission duration strataShort (5–16 days) ≈ 76.9 %, Long (129–190 days) ≈ 23.1 % Test protocol80° head-up tilt, up to 600 seconds (10 min), terminated at presyncope Presyncope incidence~24 % short-flight; ~84 % long-flight; ~31 % overall Outcomes per subjectFactual (assigned countermeasure) + 1 counterfactual (no countermeasure) — total 20,000 outcome rows Seed20261027 — bit-identical reproducible 3. Countermeasure Arms Arm Subjects Notes none4,000Baseline; no intervention fluid_loading2,500Oral fluid + salt load pre-test; reduces plasma volume deficit compression_garment2,500Modelled descriptively (see Known Limitations) lbnp1,000Lower body negative pressure in-flight conditioning; modelled descriptively 4. Data Schema Table Rows × Cols Content crew.parquet / .csv10,000 × 7Subject demographics and mission profile (age, sex, fitness, mission_days, prior_flights, countermeasure) observations.parquet2,183,530 × 8Noisy monitor readings at 2-second intervals: beat-to-beat BP, HR (ECG), symptoms, finger-cuff dropout flag, test-terminated flag truth.parquet3,010,000 × 11Ground-truth physiological state at 1-second resolution: MAP, HR, SV, CO, TPR, plasma volume deficit, contractility index, second_of_presyncope. Never observed by any monitoring device — use only for model training or evaluation, not as an observed feature. outcomes.parquet / .csv20,000 × 9One factual + one counterfactual row per subject: presyncope (boolean), tolerance_time_s, recovery_days, countermeasure_used, scenario label population_timeseries_summary.parquet / .csv1,204 × 23Per-second population-level aggregates (mean ± SD of MAP, HR, SV, etc.) across all 10,000 subjects — for cohort-level trajectory visualisation countermeasure_comparison_summary.parquet / .csv12 × 8Presyncope rates, mean tolerance time and recovery by countermeasure arm and mission-duration stratum validation_checks.parquet / .csv21 × 721 automated validation checks, all PASS: cohort-level incidence vs. calibration targets, monotonicity of tolerance time, observation noise distribution, and more audit_constants.csv / .parquet20 × 14All 20 model constants with verification status, source citation, value and units literature_registry.csv11 × 11Full bibliography of 11 literature sources cited in the parameter audit benchmark_results.parquet10 × 6Per-fold results from the benchmark script (5-fold CV, tasks A and B) for reproducibility 5. Hemodynamic Model The simulation uses an adapted Ursino-Magosso lumped-parameter cardiovascular model with baroreflex feedback. Subject-specific baselines (resting MAP, HR, plasma volume deficit, contractility) are drawn from literature-calibrated distributions and adjusted per mission duration and fitness score. During the HUT tilt, gravitational blood pooling is applied as a time-varying preload reduction; the baroreflex responds with sympathetic heart rate elevation and vasoconstriction. Presyncope is triggered when MAP falls below a threshold (55 mmHg) at any simulated second, consistent with the clinical definition used in ESA/NASA tilt protocols. 6. Observation Layer & Noise Model Monitor readings are generated from the ground-truth physiology with added instrument noise: beat_to_beat_bp — MAP + Gaussian noise (SD calibrated to finger-cuff Portapres variability); intermittent dropout segments modelled from observed Portapres failure rates. hr_ecg — HRtrue + smaller Gaussian noise (ECG-derived HR is more precise than photoplethysmographic); isolated spikes from motion artefact. finger_cuff_dropout — boolean; set during modelled signal loss windows (≈8% of all 2-second intervals in the full dataset). symptoms_reported — boolean, set probabilistically near presyncope; models patient self-report with false-positive and false-negative rates. test_terminated_early — boolean; set if presyncope occurred before the 600-second mark. 7. Validation & Calibration 21 automated validation checks, all PASS. Key calibration targets verified against literature: Target Literature Achieved Presyncope incidence, short-flight25.0%24.1% Presyncope incidence, long-flight83.3%84.5% 20 model constants audited: 4 VERIFIED (exact literature value), 3 ESTIMATE_SOURCED (literature anchor + modelled shape), 13 DERIVED (computed from VERIFIED constants). Full audit in audit_constants.csv with source citation per row. 8. Benchmark / Predictive Signal The dataset ships a logistic-regression benchmark (benchmark_script.py) with results in benchmark_results.parquet. Two tasks: Task A — during-test detection (AUC = 1.000, accuracy = 99.9%): Predicts presyncope from beat_to_beat_bp readings taken during the test. This score is a construction artifact, not a meaningful benchmark — presyncope is defined by MAP crossing a threshold, and beat_to_beat_bp is a direct noisy observation of MAP. The model recovers the label with near-zero error by construction. Reported honestly rather than suppressed. Task B — pre-test crew screening (AUC = 0.779, accuracy = 80.9%): Predicts presyncope using only pre-test crew covariates (age, sex, mission_days, prior_flights, fitness, countermeasure) — before any tilt-test reading exists. This is the commercially relevant benchmark: the flight-surgeon triage question. Real-world reference: Hart et al. (2021) reported >95% accuracy classifying orthostatic intolerance subtypes in real patients using their own feature set — not a like-for-like comparison but confirms this class of signal is physiologically predictable. 9. Literature Anchors 11 literature sources; key anchors: Buckey et al. (2001) — short-flight OI incidence (PMID 11719623) Platts et al. (2009) — long-flight OI incidence (PMID 21617145) Heldt et al. (2002) — cardiovascular simulation framework (CVSim, PhysioNet) Ursino (1998) — baroreflex model parameter base (PMID 4062772) NASA NTRS JSC-CN-22023 — tilt-test operational protocol Hart et al. (2021) — real-world OI classification benchmark (DOI 10.1007/s11517-021-02314-0) Full registry in literature_registry.csv. 10. Known Limitations Compression garment and LBNP countermeasures are modelled descriptively. No quantitative hemodynamic effect is applied for these two arms — their presyncope rates reflect the same underlying physiology as the no-countermeasure arm. The arms exist in the schema for structural completeness; do not use them to evaluate countermeasure efficacy. No cerebral autoregulation. The model has no intracranial pressure or cerebral blood flow component. Presyncope is triggered by MAP threshold alone, not by cerebral perfusion failure. Fitness is descriptive. The fitness variable modulates baseline hemodynamics via a simple linear scaling; no validated exercise physiology model is embedded. No CVSim cross-validation. The Heldt CVSim (PhysioNet) is cited as a structural reference; no numerical cross-check against that implementation was performed. task_terminated_early flag has ~9% overall rate. Because presyncope almost always occurs within the first 2 seconds of the tilt in affected subjects (second_of_presyncope ≈ 0, verified), the test-duration distribution is bimodal rather than continuous. Any feature that encodes test duration is a near-exact proxy for the label and was excluded from the benchmark (see benchmark_script.py comments). 11. Format & Delivery All tables available in Parquet (Snappy) and CSV (except truth.parquet which is Parquet-only due to size) Parquet files use pyarrow/pandas-compatible schema; no external dependencies beyond standard scientific stack Delivery: 7z archive (DATASET.7z) — all 10 tables plus data passport, sample passport, validation report, benchmark script and results, manifest_checksums.csv Generator script (orthostatic_kernel.py, sha256 b12645cc8f27012610ddcc21e188e57166e2e079f05e72bef82893da89750257) included for full reproducibility 12. Free Sample A free sample archive (SAMPLE.zip, 32 subjects) is available for download. Contents: File Rows × Cols Content crew.parquet32 × 7Subject demographics (same schema as full dataset) observations.parquet5,312 × 8Noisy monitor readings for the 32-subject sample outcomes.parquet64 × 9Factual + counterfactual outcomes (2 rows per subject) truth.parquet9,632 × 11Ground-truth physiology for the 32-subject sample sample_passport_EN.pdf—Sample data passport working_with_parquet.txt—Opening instructions for pandas / R arrow / DuckDB sample_manifest.json—SHA-256 per file, environment, kernel hash 13. Licensing The full dataset is sold under a commercial research license. The free sample is released under CC BY-NC 4.0 (non-commercial use with attribution). Fully synthetic data — not clinical evidence, not for clinical decisions. 14. Citation Code Sentinel Data (2026). Post-Landing Orthostatic Intolerance Dataset. Available at sentineldata.com.ua. [DOI to be assigned.]