Skip to content

OUTPUT_REFERENCE.md

Latest commit

 

History

History
301 lines (237 loc) · 27.6 KB

OUTPUT_REFERENCE.md

File metadata and controls

301 lines (237 loc) · 27.6 KB
Raw

Output Reference

A complete map of every artifact under output/final/ and output/intermediate/, with the cohort filter that produced it and the module that generates it.

For the definitions of each cohort and stratification flag, see README.md §8.

Top-level layout

output/
├── final/
│   ├── overall/                Critical-illness cohort         (icu_enc OR death_enc)
│   ├── overall_ward/           Ward cohort, only if --ward     (ward_enc)
│   ├── strata/                 Stratified critical-illness subsets
│   │   ├── icu/                                                (icu_enc)
│   │   ├── advanced_resp/      Sub-strata land here as *_icu.csv / *_no_icu.csv
│   │   ├── nippv_hfnc/         Sub-strata land here as *_icu.csv / *_no_icu.csv
│   │   ├── vaso/               Sub-strata land here as *_icu.csv / *_no_icu.csv / *_ed_icu.csv / *_ed_ward.csv
│   │   ├── no_imv/             Sub-strata land here as *_icu.csv / *_no_icu.csv
│   │   └── deaths/                                             (death_enc)
│   │
│   │   Sub-strata are stored as filename suffixes inside the parent stratum
│   │   directory rather than as nested subdirectories. The flag column
│   │   columns map: advanced_resp/icu→high_support_icu_enc, advanced_resp/no_icu→
│   │   high_support_no_icu_enc, nippv_hfnc/icu→nippv_hfnc_icu_enc,
│   │   nippv_hfnc/no_icu→nippv_hfnc_no_icu_enc, vaso/icu→vaso_icu_enc,
│   │   vaso/no_icu→vaso_no_icu_enc, vaso/ed_icu→vaso_ed_icu_enc,
│   │   vaso/ed_ward→vaso_ed_ward_enc, no_imv/icu→no_imv_icu_enc,
│   │   no_imv/no_icu→no_imv_no_icu_enc. See modules/strata.py:19-23.
│   ├── validation/             Data quality assessment (cohort-agnostic — runs on the raw CLIF tables)
│   ├── stats/                  Per-stratum collection-coverage tables
│   └── meta/                   Execution reports, workflow logs, run metadata
│       └── configs/            Snapshot of the config files used for the run
└── intermediate/               Debug logs + scratch (critical-illness only)
    ├── vent_hours_debug.log
    └── imv_episodes.csv                           Per-episode IMV timestamps (repeated-measures)

Each cohort directory under overall/, overall_ward/, and strata/<name>/ shares the same internal layout. The next section explains it once. Sub-strata (advanced_resp/icu, vaso/no_icu, etc.) do not have their own directories — their artifacts are written into the parent stratum's tableone/, figures/, ecdf/, and bins/ subdirectories with _icu / _no_icu / _ed_icu / _ed_ward filename suffixes (see modules/strata.py:19-23, output_paths.py:parse_stratum).

Per-cohort artifacts

The same six subdirectories appear under every cohort/stratum directory. The "Cohort filter" column says which encounter blocks the data is restricted to before the artifact is computed.

Subdirectory Contents Cohort filter Generator
tableone/ Demographics, mortality, comorbidities, ventilator settings, medications, code status, SOFA, CCI, STROBE counts, upset data — see the per-file table below Encounter blocks where the parent flag = 1 modules/tableone/generator.py
figures/ CONSORT, sankey, venn, upset, area curves, ventilator-table PNG/HTML/PDF — see the per-file table below Same as tableone/ modules/tableone/generator.py (HTML medication area curves come from modules/medications/visualizer.py)
ecdf/{labs,vitals,respiratory_support}/*.parquet Empirical CDF — distinct (value, probability) pairs Parent cohort/stratum AND value timestamp ∈ [in_dttm, out_dttm] where the time window depends on the stratum: ICU stay windows for overall/, icu, deaths; first vasopressor → discharge for vaso family; first qualifying device → discharge for advanced_resp and nippv_hfnc families. See STRATUM_WINDOW_TYPE in generator.py. modules/ecdf/generator.py:compute_ecdf_compact
bins/{labs,vitals,respiratory_support}/*.parquet Quantile bins with auto-extreme splitting (first/last bin split into 5 sub-bins when configured) Same as ecdf/ modules/ecdf/generator.py via modules/ecdf/utils.create_all_bins
summary_stats/*.{csv,json} Per-category mean/median/IQR for labs / vitals / meds / patient_assessments / CRRT Critical-illness cohort (only generated under overall/) modules/mcide/collector.py
mcide/*.csv Minimum CDE value counts for categorical columns across CLIF tables Critical-illness cohort (only generated under overall/) modules/mcide/collector.py:155

Note: mcide/ and summary_stats/ are produced once for the critical-illness cohort and live under overall/ only. They are not regenerated per stratum.

What "cohort filter" means in plain English

  • overall/ → adults whose encounter block touched an ICU or discharged as expired/hospice or received advanced respiratory support or received vasoactive medications (excluding procedural/L&D-only encounters). ECDF window: ICU stay.
  • overall_ward/ → adults whose encounter block touched a ward at any point
  • strata/icu/ → encounters in the critical-illness cohort whose icu_enc == 1. ECDF window: ICU stay.
  • strata/advanced_resp/ → encounters in the critical-illness cohort that ever received imv / nippv / cpap (unconditionally) or high flow nc with lpm_set >= 30. ECDF window: first qualifying device recorded_dttm → discharge_dttm.
  • strata/advanced_resp/<files>_icu.* → above AND icu_enc == 1. ECDF window: same as advanced_resp/.
  • strata/advanced_resp/<files>_no_icu.* → above AND icu_enc == 0 (includes deaths and ward survivors with support). ECDF window: same as advanced_resp/.
  • strata/nippv_hfnc/ → encounters in the critical-illness cohort that ever received nippv (BiPAP) or high flow nc with lpm_set >= 30. ECDF window: first qualifying device recorded_dttm → discharge_dttm.
  • strata/nippv_hfnc/<files>_icu.* → above AND icu_enc == 1. ECDF window: same as nippv_hfnc/.
  • strata/nippv_hfnc/<files>_no_icu.* → above AND icu_enc == 0 (includes deaths and ward survivors with support). ECDF window: same as nippv_hfnc/.
  • strata/vaso/ → encounters in the critical-illness cohort that ever received norepinephrine, epinephrine, phenylephrine, vasopressin, dopamine, or angiotensin. ECDF window: first vasopressor admin_dttm → discharge_dttm.
  • strata/vaso/<files>_icu.* → above AND icu_enc == 1. ECDF window: same as vaso/.
  • strata/vaso/<files>_no_icu.* → above AND icu_enc == 0 (includes deaths and ward survivors on vasopressors). ECDF window: same as vaso/.
  • strata/vaso/<files>_ed_icu.* → encounters where the first vasopressor was administered in the ED and any subsequent ADT location includes ICU. ECDF window: same as vaso/.
  • strata/vaso/<files>_ed_ward.* → encounters where the first vasopressor was administered in the ED and any subsequent ADT location includes ward but not ICU. ECDF window: same as vaso/.
  • strata/no_imv/ → encounters in the critical-illness cohort that never received invasive mechanical ventilation (on_vent == 0). ECDF window: ICU stay.
  • strata/no_imv/<files>_icu.* → above AND icu_enc == 1. ECDF window: same as no_imv/.
  • strata/no_imv/<files>_no_icu.* → above AND icu_enc == 0 (includes deaths and ward survivors without IMV). ECDF window: same as no_imv/.
  • strata/deaths/ → encounters in the critical-illness cohort in ED/ward with death_enc == 1 (discharge_category in ('expired', 'hospice')). ECDF window: ICU stay.

The flag definitions live in modules/strata.py:24-41, the inclusion code is at modules/tableone/generator.py:1561 (ward) and generator.py:1837-1841 (critical-illness), and the stratum-to-window mapping is at modules/ecdf/generator.py:STRATUM_WINDOW_TYPE.

tableone/ per-file detail

Generated by modules/tableone/generator.py. The same set of files is written under each cohort directory (overall/tableone/, overall_ward/tableone/, strata/<name>/tableone/); some files are skipped under overall_ward/ and noted below.

File Contents
table_one_overall.csv Main demographic + clinical Table One for the parent cohort
table_one_by_year.csv Same, stratified by admission year
mortality_rates.csv In-hospital and discharge mortality counts/rates
strobe_counts.csv STROBE enrollment flow counts
upset_data.csv Cohort subset membership (ICU, advanced resp, NIPPV/HFNC, vaso, death) for the upset plot
comorbidities_per_1000_hospitalizations.csv Charlson/Elixhauser rates normalized per 1000 stays
comorbidities_per_1000_hospitalizations_summary.csv Summary stats for the above
code_status_counts_by_encounter_type.csv Code status value counts by encounter type
code_status_percentages_by_encounter_type.csv Same as %s
code_status_missingness_summary.csv Missing-data summary for code status
code_status_combined_summary.csv Combined code status summary
demographic_crosstab_race_ethnicity_sex.csv 3-way demographic crosstab
sofa_mortality_summary.csv SOFA scores by mortality outcome (skipped under overall_ward/)
hospice_trends_summary.csv Hospice admission trends over time
cci_hospice_mortality_comprehensive_summary.csv CCI + hospice + mortality
cci_mortality_hospice_trends_by_year_category_plotdata.csv Plot-ready CCI/hospice/mortality data by year
ventilator_settings_by_device_mode.csv Ventilator parameters by device + mode (skipped under overall_ward/)
ventilator_settings_counts_by_device_mode.csv Observation counts by device + mode
ventilator_settings_total_observations.csv Total ventilator observations
tidal_volume_volume_control_modes.csv Tidal volume stats (volume-control modes only)
tidal_volume_volume_control_modes_mean_sd.csv Same binned by hour with mean/SD
pressure_control_pressure_control_mode.csv Pressure control stats (pressure-control modes)
pressure_control_pressure_control_mode_mean_sd.csv Same binned by hour
mode_proportions_first_24h.csv Ventilator mode breakdown in first 24 h of mechanical ventilation
medications_hourly_data.csv Paralytic / sedative / vasoactive doses by hour-since-ICU-admit
medications_summary_stats.csv Mean/median/IQR for the above
pf_sf_summary_24h.csv Per-encounter PF/SF ratios in the first 24 h of respiratory failure onset. Only under strata/advanced_resp/ and strata/no_imv/ (with _icu/_no_icu suffixed variants).
pf_sf_aggregate_stats.csv Aggregate PF/SF statistics (n, mean, sd, median, Q25, Q75) segmented by onset device. Same strata as above.
adt_dwell_summary.csv Per location_category: total dwell hours/days, median stay hours with Q1/Q3, distinct encounters, number of stays. Overall cohort only.
adt_event_capture.csv Per (table, location_category): event counts, % of events, events per location-hour. Tables: labs, vitals, respiratory_support. Overall cohort only.

Additional rows in strata Table One CSVs:

Certain strata table ones (table_one_<stratum>_by_year.csv) include extra rows that do not appear in the overall Table One:

Row(s) Appears in strata Description
Resp. device onset, n (%) / Pre-device LOS (days) / Post-device LOS (days) advanced_resp, nippv_hfnc, no_imv (+ /icu, /no_icu splits) Time from admission to first respiratory device onset, and from onset to discharge. Only encounters with a detected onset are counted.
28-day VFD (IMV encounters), n (%) / VFD, median [Q1, Q3] Any stratum with IMV encounters 28-day ventilator-free days. VFD = 0 for death within 28 days (uses death_dttm or discharge_dttm when discharge_category is expired/hospice) or if still on IMV at day 28. Intermediate free days between reintubation episodes do not count.
Time to extubation (hrs), median [Q1, Q3] Any stratum with IMV encounters Hours from detected intubation to detected extubation for the first episode only, restricted to encounters with extubation_status == 'extubated'. Pre-admission IMV encounters are excluded (true intubation time unknown). Detection uses the two-lookback / two-lookforward pattern on device_category (clifpy #124). modules/tableone/extubation_calculator.py.
Extubation outcome, n (%) Any stratum with IMV encounters Distribution of extubation_status across IMV encounters: extubated, discharged_on_imv (left hospital still on IMV), death_on_imv (died while on IMV), unknown, failed_attempt (near-zero intub→extub gap).
Pre-admit IMV (excluded from time-to-extubation), n (%) Any stratum with IMV encounters (when non-zero) Encounters whose first respiratory_support row was already IMV — no detectable intubation transition. Reported as a separate count; excluded from the time-to-extubation median.
Intubated ≤24hr of admission, n (%) Any stratum with IMV encounters Encounters where the detected intubation event falls within 24 hours of admission_dttm.
Reintubation (≥2 IMV episodes), n (%) / Time to reintubation (hrs), median [Q1, Q3] / Extubation failure ≤48hr, n (% of extubated) Any stratum with IMV encounters Encounters with more than one detected IMV episode; hours between first-extubation-end and second-intubation-start; and the standard clinical threshold for extubation failure (time_to_reintubation_hours ≤ 48). Failed-attempt episodes (<5 min) are excluded from the episode count to avoid inflating reintubation incidence.
Norepinephrine equivalent (NEE), n (%) / Peak NEE / Median NEE vaso, vaso/icu, vaso/no_icu, vaso/ed_icu, vaso/ed_ward Vasopressor intensity per encounter in norepinephrine-equivalent mcg/kg/min. Peak = maximum concurrent intensity; Median = typical intensity. Weights: norepinephrine 1.0, epinephrine 1.0, phenylephrine 0.1, dopamine 0.01, vasopressin 2.5, angiotensin 10.0. Concurrent doses aligned by rounding to nearest hour.
Time to ICU after first pressor (hours) / Time to Ward after first pressor (hours) vaso/ed_icu, vaso/ed_ward (respectively) Hours from first vasopressor admin_dttm (in ED) to the first post-pressor ICU or ward in_dttm. Reported as median [Q1, Q3].

figures/ per-file detail

Generated by modules/tableone/generator.py (with HTML medication curves from modules/medications/visualizer.py). The same files appear under each cohort directory; ventilator/SOFA/medication-from-ICU plots are skipped under overall_ward/.

File Contents
consort_flow_diagram.png CONSORT enrollment flow chart
cohort_intersect_upset_plot.png UpSet plot of cohort overlaps (ICU, resp, NIPPV/HFNC, vaso, death)
venn_all_4_groups.png 4-way Venn of cohort intersections
code_status_stacked_bar_with_missingness_excl_missing_cat.png Code status stacked bar
comorbidities_per_1000_barplot.png Charlson/Elixhauser bar chart
sankey_matplotlib_icu.png Sankey: ICU → outcomes
sankey_matplotlib_high_o2_support.png Sankey: advanced respiratory support → outcomes
sankey_matplotlib_high_o2_proc_other.png Sankey variant for procedural-only encounters
sankey_matplotlib_vaso_support.png Sankey: vasopressor → outcomes
sankey_matplotlib_vaso_proc_other.png Sankey variant for procedural-only encounters
sankey_matplotlib_others.png Sankey: non-critical-illness encounters
sofa_mortality_histogram.png SOFA score distribution by mortality
tidal_volume_volume_control_modes.png Tidal volume trends (volume control)
tidal_volume_volume_control_modes_mean_sd.png Same with mean/SD overlay
pressure_control_pressure_control_mode.png Pressure control trends
pressure_control_pressure_control_mode_mean_sd.png Same with mean/SD overlay
mode_proportions_first_24h_vertical.png Ventilator mode proportions (vertical bar) for first 24 h
hospice_mortality_combined_trends.png Hospice + mortality trends over time
cci_mortality_hospice_comprehensive.png CCI + hospice + mortality 3-way analysis
paralytic_area_curve_7d.html Interactive 7-day area curve: paralytic dose by hour
paralytic_median_dose_by_hour.html Median paralytic dose by hour
sedative_area_curve_7d.html Interactive 7-day area curve: sedative dose by hour
sedative_median_dose_by_hour.html Median sedative dose by hour
vasoactive_area_curve_7d.html Interactive 7-day area curve: vasoactive dose by hour
vasoactive_median_dose_by_hour.html Median vasoactive dose by hour
ventilator_settings_table.png / .pdf Ventilator settings summary table rendered as image + PDF
pf_sf_comparison_overall_icu_noicu.png Box plot comparing PF/SF distributions across Overall/ICU/No-ICU splits. Only under strata/advanced_resp/figures/ and strata/no_imv/figures/.
km_time_to_extubation.png Two-panel Kaplan-Meier curve for time to extubation: overall + stratified by ICU vs no-ICU. Event = extubation; pre-admit IMV encounters excluded; censoring = death or discharge; right-censored at 28 days. Skipped in ward mode. Companion CSV: output/final/overall/ventilated_aggregates/km_time_to_extubation.csv. Generator: modules/tableone/extubation_plots.py.
min_pf_sf_per_day_post_intubation.png Two-panel daily minimum oxygenation plot post-intubation: P/F (PaO₂/FiO₂) and S/F (SpO₂/FiO₂), median + shaded IQR band, 28-day horizon. Reuses _calculate_concurrent_pf_ratios / _calculate_concurrent_sf_ratios from modules/sofa/calculator.py with a 4-hour FiO₂ lookback. Companion CSV: output/final/overall/ventilated_aggregates/min_pf_sf_per_day_post_intubation.csv. Skipped in ward mode. Generator: modules/tableone/extubation_plots.py.
adt_dwell_hours_by_location.png Bar chart of cumulative dwell hours per ADT location_category. Overall cohort only.
adt_los_distribution_by_location.png Pre-aggregated box plot of per-stay durations by location. Overall cohort only.
adt_event_capture_pct.png Grouped bar of % events (labs/vitals/respiratory_support) per location. Overall cohort only.
adt_events_per_location_hour.png 3-panel instrumentation density (events/hour) per location per table. Overall cohort only.

ecdf/ and bins/ parquet schemas

ecdf/{labs,vitals,respiratory_support}/<name>.parquet

Column Type Meaning
value float Distinct numeric value
probability float Cumulative P(X ≤ value)

Naming:

  • Labs: {lab_category}_{unit_safe}.parquet — e.g. albumin_g_dL.parquet
  • Vitals: {vital_category}.parquet — e.g. heart_rate.parquet
  • Respiratory support: {column_name}.parquet — e.g. fio2_set.parquet

bins/{labs,vitals,respiratory_support}/<name>.parquet

Quantile bins computed from the same filtered values. Schema matches the bin format produced by modules/ecdf/utils.create_all_bins — bin id, label, lower/upper edge, count, percentage.

summary_stats/ per-file detail

Generated by modules/mcide/collector.py. Each metric is written as both CSV and JSON. Lives under overall/summary_stats/ only (not regenerated per stratum).

File Contents
labs_summary_by_category.csv/json Mean/median/std/q25/q75/min/max per (lab_category, unit)
vitals_summary_by_category_and_name.csv/json Same per (vital_category, vital_name)
medication_admin_continuous_dose_by_category_and_unit.csv/json Continuous-medication dose stats per (med_category, dose_unit)
medication_admin_intermittent_dose_by_category_and_unit.csv/json Intermittent-medication dose stats per (med_category, dose_unit)
patient_assessments_summary_by_category.csv/json Numeric-assessment stats per assessment_category
crrt_blood_flow_rate_overall.csv/json + _by_mode.csv/json CRRT blood flow rate, overall and by crrt_mode_category
crrt_dialysate_flow_rate_overall.csv/json + _by_mode.csv/json CRRT dialysate flow rate, overall and by mode
crrt_ultrafiltration_out_overall.csv/json + _by_mode.csv/json CRRT ultrafiltration output, overall and by mode

mcide/ per-file detail

Generated by modules/mcide/collector.py:155. Each file is {table}_{columns}_mcide.csv and contains the value count of every distinct combination of the named columns. Lives under overall/mcide/ only.

Tables represented (one or more files each, depending on which categorical columns are summarized):

adt, clif_crrt_therapy, clif_microbiology_culture, code_status, hospitalization, labs, medication_admin_continuous, medication_admin_intermittent, patient, patient_assessments, position, respiratory_support, vitals.

validation/

Cohort-agnostic — runs on the raw CLIF tables, not the cohort-filtered views.

Path Contents Generator
validation/json_reports/<table>_dqa.json + supporting CSVs Per-table DQA results from clifpy (conformance, completeness, plausibility) clifpy invoked from run_analysis.py
validation/consolidated/consolidated_validation.csv One-row-per-table master status grid modules/reports/combined_report_generator.py
validation/consolidated/<table>_summary_summary.json Per-table summary stats modules/reports/combined_report_generator.py
validation/feedback/<table>_validation_response.json User-classified errors (Accepted / Rejected / Pending) saved from the web app server/routes/feedback_routes.py
validation/monthly_trends/*.csv Monthly admission / data-volume trends run_analysis.py
validation/pdf_reports/<table>_validation_report.pdf Per-table validation PDF modules/reports/
validation/pdf_reports/combined_validation_report.pdf All-tables-in-one PDF modules/reports/combined_report_generator.py

stats/

File Contents Generator
collection_statistics.csv Per (data_type, category, reference_unit): total stays, total observations, total distinct values, mean ICU LOS, whole-stay mean/median/IQR, first 24h / 48h / 72h count distributions. The cohort is the critical-illness cohort. modules/ecdf/statistics.py:159
collection_statistics_<stratum>.csv Same, restricted to each stratum (icu, advanced_resp, advanced_resp_icu, advanced_resp_no_icu, nippv_hfnc, nippv_hfnc_icu, nippv_hfnc_no_icu, vaso, vaso_icu, vaso_no_icu, vaso_ed_icu, vaso_ed_ward, no_imv, no_imv_icu, no_imv_no_icu, deaths) modules/ecdf/statistics.py

The 24h/48h/72h numbers here are coverage stats on observation counts, not the ECDF input filter. The ECDF/bins parquets cover the full time window for that stratum: ICU stay windows for overall/, icu, deaths; event-onset → discharge for vaso, advanced_resp, and nippv_hfnc families.

meta/

File Contents Generator
tableone_execution_report.txt Per-step memory checkpoints, timing, status modules/tableone/runner.py
tableone_ward_execution_report.txt Same, ward-mode run (only present when --ward was used) modules/tableone/runner.py
ecdf_execution_report.txt ECDF generation summary + structure modules/ecdf/runner.py
unit_mismatches.log Data rows whose lab_category or reference_unit is not accepted by the CLIF labs schema (clifpy/schemas/labs_schema.yaml) — site-side data-quality issues modules/ecdf/generator.py
ecdf_coverage_gaps.log Schema-valid categories that were skipped by the ECDF pipeline because lab_vital_config.yaml/outlier_config.yaml lacked a bin or outlier entry, plus any runtime errors — repo-side coverage gaps modules/ecdf/generator.py
lab_category_units.csv Every (lab_category, reference_unit) pair in the labs data with a row count, classified against the CLIF schema (schema_status = ok / unit_mismatch / not_in_spec, plus canonical_unit). Single source of truth for the data's lab vocabulary; drives the two logs above. modules/ecdf/generator.py:write_lab_category_units_csv
file_metadata.json Snapshot of config.json + tables_path for the run run_project.py
workflow_logs/workflow_execution_<timestamp>.log, workflow_execution_latest.log Full pipeline stdout/stderr per run run_project.py

meta/configs/

File Contents
config.json Snapshot of the config/config.json used for the run (site name, tables_path, file_type, timezone)
outlier_config.yaml Snapshot of the lab/vital outlier bounds applied during ECDF
lab_vital_config.yaml Snapshot of the lab/vital bin definitions used for the bins parquets

Nested under meta/ so that all run-provenance artifacts (configs, execution reports, logs, file metadata) live in one place. A downstream consumer (or a future you) can tell exactly which configuration produced a given output/final/ directory.

overall/ventilated_aggregates/ — cross-site shareable KM + daily P/F|S/F aggregates

Overall critical-illness cohort only (skipped in ward mode). Companion PNGs under overall/figures/ are site-specific renderings; the CSVs here contain the exact numeric values needed to re-render or to pool multiple sites into a single overlay without sharing patient-level data.

Path Contents Generator
overall/ventilated_aggregates/km_time_to_extubation.csv One row per (stratum, timepoint) from the Kaplan-Meier fit. Columns: stratum ∈ {overall, icu, no_icu}, timeline_days, survival_prob = S(t), ci_lower, ci_upper (95% log-log Greenwood band), at_risk = n_i (encounters still on IMV just before t_i), observed_events = d_i (extubations at t_i). KM is computed as S(t_i) = S(t_{i-1}) × (1 − d_i / n_i). Pre-admit IMV encounters excluded. Right-censored at 28 days. No PHI — concatenating this CSV across sites yields an overlayed or pooled cross-site KM. modules/tableone/extubation_plots.py:plot_km_time_to_extubation
overall/ventilated_aggregates/min_pf_sf_per_day_post_intubation.csv Long-format aggregate, one row per (ratio_type, day). Columns: ratio_type ∈ {PF, SF}, day ∈ [0, 27], median, q1, q3, n_encounters. For each IMV encounter, the daily minimum P/F and S/F ratio is computed (4-hour FiO₂ lookback on concurrent PaO₂ / SpO₂); then across encounters the per-day median and IQR are reported. Pre-admit IMV encounters excluded. modules/tableone/extubation_plots.py:plot_min_pf_sf_per_day_post_intubation

output/intermediate/ — debug logs + scratch

Sits outside output/final/ because the contents are debug traces / per-run scratch. All files here are critical-illness cohort only (skipped in ward mode).

Path Contents Generator
intermediate/vent_hours_debug.log Per-encounter waterfall sample, duration-calc sample, and IMV-hours distribution for the vent_duration_hours computation. Single-run debug trace; overwritten each run. modules/tableone/generator.py (~line 2801)
intermediate/imv_episodes.csv One row per detected IMV episode per encounter: encounter_block, episode_n, intubation_start_dttm, extubation_end_dttm, episode_duration_hours, is_synthetic. Synthetic rows (is_synthetic == True) mark pre-admission IMV where the first observation was used as an anchor. Use for repeated-measures analyses (one patient contributes one row per IMV episode). modules/tableone/extubation_calculator.py