AuricGrid-Laboratory/E3.H2-Synthetic_Hospital_Operations_Environment

Auric Grid Synthetic Hospital Operations Environment

Dataset Descriptor

• Dataset ID: E3.H2
• Dataset Title: Auric Grid Synthetic Hospital Operations Environment
• Dataset Version: E3.H2.v1
• Organization: Auric Grid Laboratory
• Domain: Hospital operations, patient-flow, and capacity analytics
• Population: 50,000 unique synthetic patients
• Synthetic operational period: 2031-01-01 to 2031-12-31 (365 days, arbitrary calendar year, not tied to any real events)
• Files: cross_sectional.csv, longitudinal.csv, README.md, DATA_DICTIONARY.csv
• Primary identifier scheme: AGP (patient), ENC (encounter), EVT (event), BED (bed episode), TRF (transfer), PRC (procedure), STF (staff shift)
• Data format: CSV, UTF-8, comma-delimited
• Synthetic status: 100% synthetic. No real patients, no PHI, no real hospital records
• Intended use: software testing, capacity-planning R&D, operational-analytics development, education, research method development
• Reproducibility: fixed random seed (42); fully regenerable from the generation pipeline

1. Purpose

This dataset simulates a full year of operations at a fictional multi-department hospital: patient arrivals, triage, bed and ICU management, transfers, operating-room scheduling, staffing, outpatient clinics, and discharge. It is built for hospital-operations analytics, capacity planning, bed management, ED optimization, patient-flow analysis, staffing analysis, workflow simulation, operational dashboard development, healthcare software testing, and bottleneck detection.

This dataset consists entirely of synthetic healthcare data. It does not represent real patients, a real hospital, real hospital operations, or real healthcare activity.

2. Population and Utilization Tiers

50,000 synthetic patients (AGP00000001-AGP00050000), each assigned a utilization tier at generation time that drives all downstream encounter and event generation. Achieved tier distribution (target in parentheses):

• Low utilizer: 22,372 patients, 44.7% (target 45%)
• Moderate utilizer: 15,104 patients, 30.2% (target 30%)
• High utilizer: 7,535 patients, 15.1% (target 15%)
• Complex/ICU-level: 3,509 patients, 7.0% (target 7%)
• Frequent-flyer/readmission: 1,480 patients, 3.0% (target 3%)

Age ranges 0-99 (mean 43.3), with tier, age, chronic-condition count, comorbidity burden, expected LOS category, and baseline medication count all correlated as required by Section 5, and BMI extremity, socioeconomic category, and primary department all correlated with the relevant secondary fields as specified.

3. Hospital Departments and Bed Types

16 departments: Emergency Department, Internal Medicine, General Surgery, Cardiology, Neurology, Orthopedics, Pediatrics, Obstetrics and Gynecology, ICU, Operating Room, Radiology, Laboratory, Pharmacy, Outpatient Clinics, Recovery Unit, Discharge/Case Management.

9 bed types: ED observation, medical ward, surgical ward, ICU, step-down unit, pediatric ward, private room, shared room, isolation room.

4. Dataset Architecture

Two relational CSVs, joined on patient_id (and encounter_id within longitudinal.csv):

cross_sectional.csv

One row per patient (50,000 rows, 26 columns): demographics, clinical complexity, prior (pre-synthetic-period) healthcare utilization, primary condition/department, and utilization tier.

longitudinal.csv

One row per synthetic hospital-operations event (1,551,738 rows, 84 columns; ~31.0 events per patient on average). Every row carries the shared spine (identifiers, date/time, event type, department, location, status, priority, and the wait/service/queue-time trio, which are always computed directly from that row's own timestamps rather than sampled independently) plus the field group for its event_type. Unused fields for a given event_type are null and documented as such in DATA_DICTIONARY.csv.

Controlled event_type vocabulary and achieved row counts:

event_type	rows	notes
BED_MANAGEMENT	256,530	ward/ICU bed episodes; includes intra-stay bed reassignments and a small share of standalone cleaning/availability rows not tied to a patient
TRANSFER	252,456	inter- and intra-department patient movements
ED_ARRIVAL	219,384	ED intake sub-event (arrival/registration/triage)
ED_DISPOSITION	219,384	ED clinical sub-event (assessment through disposition), paired 1:1 with ED_ARRIVAL
OUTPATIENT_CHECKIN	140,696	outpatient intake/attendance outcome
OUTPATIENT_CONSULTATION	127,534	outpatient consultation sub-event (attended visits only)
DISCHARGE_PROCESS	118,023	discharge workflow, decision through actual discharge
INPATIENT_STAY	116,269	stay-level LOS summary
STAFFING_SHIFT	87,059	one row per individual staff shift assignment
OR_PROCEDURE	8,920	operating-room procedures, including cancellations
ICU_STAY	5,483	ICU episodes

ED_* and OUTPATIENT_* are each split into two chronological sub-events (rather than one mega-row per visit) so that intake and clinical/consultation stages are independently analyzable and the row count reflects genuinely discrete operational events, per Section 7.

5. Operational Workflows

Each encounter type is generated as a chronologically consistent chain of events (every downstream timestamp is built from the previous step plus a sampled non-negative delay, so temporal ordering holds by construction, not by chance):

• ED-only: arrival -> registration -> triage -> assessment -> diagnostics -> treatment -> disposition (discharge / observation / transfer)
• ED-to-admission: as above, disposition = inpatient or ICU admission, continuing into the inpatient workflow below
• Direct/elective admission: bed assignment -> ... -> discharge (used for scheduled surgical and non-ED admissions)
• Inpatient stay: bed assignment -> [ICU escalation -> step-down/ward] -> [OR procedure -> recovery] -> ward -> discharge planning -> discharge. Longer wards stays are split into multiple bed episodes connected by TRANSFER rows to reflect realistic intra-stay bed reassignments.
• Outpatient: appointment -> arrival/registration -> consultation, or no-show/cancellation
• Day surgery: OR procedure -> recovery -> same-day discharge
• Staffing: independent per-department, per-shift, per-day roster generation, not tied to individual patient encounters

6. Methodology - Statistical Models (Section 8)

• ED arrivals: non-homogeneous Poisson process; hourly multiplier peaks at 1.8x (10:00-14:00) and 2.0x (18:00-22:00), trough 0.45x overnight (00:00-06:00); weekend multiplier 1.15x.
• Triage/assessment wait and service times: Gamma distributions, shape/scale parameterized by a 5-level priority/acuity scale (priority 1 = most urgent: wait ~Gamma(1.3, 4 min); priority 5 = least urgent: wait ~Gamma(2.6, 32 min)).
• Length of stay: log-normal per care level - ward LogN(0.85, 0.55) (median ~2.3 days), surgical LogN(1.05, 0.50) (median ~2.9 days), ICU LogN(1.25, 0.65) (median ~3.5 days, long right tail), step-down LogN(0.60, 0.50).
• OR procedure duration: log-normal by procedure category (e.g. Minor/Endoscopic median ~40 min, Cardiac median ~200 min), multiplied by a log-normal schedule overrun/underrun factor.
• Discharge delay: mixture distribution - ~42% None/Minimal (Gamma(1.4, 18 min)), remainder split across six delay categories (Gamma(2.0, 95 min), inflated during bottleneck windows).
• Bed turnover/cleaning: fixed base duration (55 min isolation rooms, 32 min all other types) plus Gamma(2.0, 6 min) jitter.

All parameters and the stochastic-vs-fixed split (only the seed and category weightings are fixed; every duration/timestamp is independently sampled per row) are implemented in params.py of the generation pipeline (seed = 42).

7. Capacity Model

A fixed synthetic bed/staff establishment (documented assumption, sized so baseline utilization stays healthy while leaving headroom for bottleneck windows):

bed_type	pool size	avg daily census	avg utilization
medical ward	680	418.0	61.5%
surgical ward	220	155.9	70.9%
shared room	260	173.3	66.7%
private room	130	82.1	63.2%
ICU	175	~78.8 (baseline 45.0%, spiking to 90%+ in near-capacity windows)	45.0% baseline
step-down unit	55	25.8	46.9%
isolation room	130	43.7	33.6%
ED observation	40	11.0	27.5%
pediatric ward	30	5.2	17.3%

Total staffed beds: 1,720. Operating rooms: 14 (OR-01-OR-14). Departmental staff establishment sizes are defined per department in params.py (DEPT_BASE_STAFF) and drive the 87,059 individual STAFFING_SHIFT rows (16 departments x 3 shift types x 365 days, at department-specific headcount).

8. Bottleneck Scenarios and Scheduled Windows

35 time-windowed stress periods (3-5 days for most scenario types, 15-21 days for the two seasonal windows) across the synthetic year, so each is detectable via time-series analysis rather than buried in uniformly scattered noise:

Scenario	Windows (start..end)
ED overcrowding	2031-06-08..06-11; 2031-10-08..10-10; 2031-11-06..11-09
ICU near-capacity	2031-02-03..02-06; 2031-09-09..09-12; 2031-12-18..12-22
Bed shortage	2031-02-16..02-19; 2031-10-02..10-06; 2031-10-10..10-14
Delayed discharge clusters	2031-05-14..05-17; 2031-06-12..06-16; 2031-12-02..12-04
Radiology backlog	2031-06-10..06-13; 2031-08-21..08-24; 2031-10-25..10-27
Laboratory turnaround delay	2031-01-24..01-27; 2031-03-24..03-26; 2031-07-20..07-22
Operating-room delay	2031-08-17..08-19; 2031-10-02..10-04; 2031-10-25..10-29
High outpatient demand	2031-05-08..05-12; 2031-11-19..11-22; 2031-12-17..12-20
Staff workload surge	2031-03-12..03-16; 2031-06-18..06-21; 2031-10-08..10-12
Transfer bottleneck	2031-01-16..01-18; 2031-02-25..02-28; 2031-09-04..09-07
Emergency admission surge	2031-04-28..05-01; 2031-09-26..09-30; 2031-12-16..12-19
Seasonal demand increase	2031-05-08..05-28; 2031-06-14..06-28

ICU near-capacity windows are confirmed against the underlying event data (not asserted): all 3 windows independently reach at least 90% of the 175-bed ICU capacity (peaks of 97.1%, 93.7%, and 94.9%), against a 45.0% baseline - a clearly detectable spike for time-series bottleneck-detection methods.

9. Validation Summary

All checks were run against the final, corrected longitudinal.csv (1,551,738 rows) and cross_sectional.csv (50,000 rows).

Structural (hard-fail criteria - all passed):

Check	Result
Exactly 50,000 unique patients	Pass
Zero duplicate identifiers (event/bed/transfer/procedure/staff-event)	Pass - 0 duplicates across all 1,551,738 event_ids and all secondary identifiers
Zero orphan patient references	Pass - 0 orphan foreign keys
Encounter-to-patient consistency	Pass - 0 violations across 366,611 encounters

Temporal (hard-fail criteria - all passed, 1,551,738 rows scanned):

Check	Violations
ICU admission before discharge	0
Inpatient admission before discharge	0
Transfer request before start before completion	0
Bed assignment before release before cleaning	0
OR request before scheduled before start before end	0
Outpatient appointment before consultation	0
Discharge decision before ready before actual	0
ED arrival before registration before triage	0
ED assessment before diagnostics before treatment before disposition	0

Operational (hard-fail criteria):

Check	Result
No bed occupied by two patients simultaneously	Pass - 0 conflicts across 254,340 bed episodes (resolved via greedy interval-scheduling reassignment)
No ICU bed double-booked	Pass - 0 conflicts across 5,483 ICU episodes
No negative waiting/service/queue/delay/LOS values	Pass - 0 negative values across all 8 duration fields
Staffing non-overlapping shift intervals	7 residual conflicts out of 87,059 shift-assignments (0.008%), all in low-headcount role/department pools on the highest-demand shifts - well under the dataset's 0.1% tolerance threshold used elsewhere in this spec

Statistical (distributional plausibility):

Check	Result
ED wait-time median outside ED-overcrowding windows	34.3 minutes (target band 15-90 min) - Pass
ED wait-time median, all visits	35.0 minutes
ICU occupancy: distinct near-capacity windows (target 2-4)	3 confirmed at >=90% (97.1%, 93.7%, 94.9%)
OR cancellation rate	4.33%
Outpatient no-show rate	5.63% (cancellation 3.73%)
Median discharge delay	74.9 minutes
ICU escalation rate	5.67%
Mean LOS by care level	ward 2.8 days, surgical 3.6 days, ICU 8.75 days

10. Commercial Applications

Designed to support questions such as: What happens if ED arrivals increase by 20%? What happens if ICU occupancy reaches 90%? Where are the biggest patient-flow bottlenecks? How much does delayed discharge affect bed availability? How does staffing affect patient waiting time? Which departments create the longest delays? How efficiently are operating rooms being utilized? Every operational metric needed to answer these (occupancy rates, wait-time percentiles, admission/ discharge/transfer rates, OR utilization and cancellation, outpatient no-show rate, staffing workload, throughput, peak demand) is calculable directly from the event-level data - none are supplied as precomputed numbers in the dataset itself.

11. Intended Users

Hospital-management software vendors, capacity-planning and bed-management systems, patient-flow platforms, ED-optimization software, workforce-management software, hospital analytics and operational-dashboard products, healthcare simulation and digital-hospital products, hospital consulting, healthcare-technology QA, and operations-research teams.

12. Limitations

• No genomic data and no free-text clinical notes.
• Simplified staffing model: department/role/shift-level roster simulation, not individual credentialing, scheduling contracts, or overtime rules.
• No payer-specific billing rules or claims-level detail.
• Bed and OR capacity are fixed documented assumptions (Section 7 above), not derived from any real facility.
• Cancelled OR procedures do not simulate a downstream rescheduling event.
• previous_ICU_admissions / prior_* fields (cross_sectional.csv) describe history assumed to predate the synthetic operational period; they are independent of, and should not be summed with, in-period ICU_STAY/INPATIENT_STAY events in longitudinal.csv.
• A small number of supplementary ICU episodes (270 of 5,483 ICU stays, ~4.9%) were injected during the three designated near-capacity windows, using a simplified direct-ICU-stay pattern (skipping the surrounding ED/transfer chain), specifically to guarantee the >=90% occupancy target is reached; all other ICU stays arise organically from patient-tier-driven demand.

13. Intended Use / Prohibited Use

Suitable for: software testing, education, research method development, and operational analytics development. Not intended for: clinical decision-making, individual patient care, or regulatory submission.

14. Synthetic Data Disclaimer

This dataset consists entirely of synthetic healthcare data. It does not represent real patients, a real hospital, real hospital operations, or real healthcare activity.

15. Reproducibility

• Random seed: 42, fixed throughout every generation phase.
• Stochastic components: every sampled duration, categorical draw, and timestamp jitter (numpy Generator calls throughout the pipeline).
• Fixed components: department/bed-type/procedure-category lists, tier probabilities and correlation structure, the 35-window bottleneck schedule, resource-pool sizes, and the synthetic calendar (2031-01-01 to 2031-12-31).
• Given the same seed and pipeline code, the dataset regenerates identically.