AuricGrid-Laboratory/E3.H1-Synthetic_Hospital_Interoperability_Environment

Auric Grid Synthetic Hospital Interoperability Environment

Dataset ID: E3.H1 Prepared for: Auric Grid Laboratory Version: 1.0 Population: 50,000 unique synthetic patients Data generation window: primarily January 1, 2024 – June 30, 2026 (a small number of long inpatient stays and pending lab results that started near the end of the window extend into early July 2026; no generated timestamp falls after the current date)

---

1. Synthetic-Data Disclaimer

This dataset consists entirely of synthetic healthcare data. It does not represent real patients, a real hospital, real healthcare transactions, or real clinical events.

No real patients, real medical records, protected health information (PHI), real patient identifiers, or identifiable individuals were used, referenced, or represented in any part of this dataset. All identifiers, clinical values, timestamps, and system events were generated programmatically. This dataset must not be used for actual patient care or clinical decision-making.

---

2. Purpose

This dataset is a realistic, fully synthetic hospital interoperability environment that simulates how patient, clinical, diagnostic, medication, and administrative information moves through interconnected hospital information systems:

PATIENT → EHR → LABORATORY → RADIOLOGY → PHARMACY → CLINICAL SERVICES → BILLING

It is designed to function as a synthetic software-testing environment for a modern hospital, not merely as a synthetic patient database.

3. Commercial Use Cases

• EHR integration testing
• Hospital information system testing
• Healthcare API development and testing
• Interoperability / message-exchange testing
• Clinical data pipeline testing
• Laboratory, radiology, pharmacy, and billing integration testing
• Healthcare dashboard and analytics software development
• Data migration testing
• Automated QA and workflow-simulation testing
• Event-processing system testing
• Software demonstrations and sales engineering

4. Intended Users

Health-tech engineering and QA teams, interoperability/integration engineers, healthcare analytics and dashboard developers, EHR/HIS implementation teams, sales/solutions engineers running product demonstrations, and students or researchers learning healthcare data engineering.

---

5. Dataset Architecture

File	Grain	Description
cross_sectional.csv	One row per patient (50,000 rows)	Stable patient demographic, clinical-background, utilization-history, and hospital-context fields.
longitudinal.csv	One row per hospital event (680,515 rows)	The core interoperability event log — every registration, clinical, diagnostic, medication, transfer, procedure, billing, and system-to-system interoperability event.
DATA_DICTIONARY.csv	One row per variable (88 rows)	Full documentation of every column in both data files.
README.md	—	This document.

Identifier relationships

Identifier	Assigned by	Links
patient_id	Patient generation	cross_sectional.csv ↔ every row of longitudinal.csv for that patient
encounter_id	Encounter start	Groups all events within a single outpatient visit, ED visit, or inpatient stay (including an ED stay that converts to an inpatient admission, which retains one encounter_id across both phases)
order_id	Lab, imaging, or medication order	Links an order to its results / dispensing / administration / interoperability-message rows
result_id	Laboratory result	Uniquely identifies a laboratory result, referencing its order_id
procedure_id	Procedure order	Links a procedure_ordered row to its procedure_completed row
transfer_id	Patient transfer	Uniquely identifies a single transfer event
event_id	Every row	Primary key of longitudinal.csv

Every laboratory result references a valid laboratory order; every medication administration references a valid medication order; every imaging report references a valid imaging order; every procedure and every transfer belongs to a valid encounter; every longitudinal event belongs to a valid patient. There are no orphan records (see §9, Validation Results).

---

6. Cross-Sectional Data (cross_sectional.csv)

One row per synthetic patient (patient_id format AGP00000001–AGP00050000), covering:

• Patient information — age, sex, race/ethnicity, geographic region, socioeconomic category, insurance category, smoking status, BMI
• Clinical background — chronic condition count and seven condition flags (cardiovascular disease, hypertension, diabetes, chronic kidney disease, chronic liver disease, chronic respiratory disease, cancer history), allergy count, baseline medication count
• Healthcare utilization — prior hospitalization, ED visit, surgery, and outpatient visit counts; chronic-care status
• Hospital context — primary hospital (one of 12 synthetic Auric Grid network facilities across four US regions), hospital type/region, primary department, and an overall patient complexity category (Low / Moderate / High / Very High) that drives longitudinal utilization intensity

Complexity category is not decorative — it is the primary driver of how many encounters and events a patient generates downstream:

Complexity	Share of population	Mean longitudinal events/patient
Low	55.0%	9.1
Moderate	28.0%	14.1
High	13.0%	24.5
Very High	4.0%	36.7

7. Longitudinal Data (longitudinal.csv)

One row per synthetic hospital event, 680,515 rows total (mean 13.6 events/patient; median 8; 90th percentile 24; a small number of complex patients exceed 200 events). Distribution across domains:

event_category	Rows
registration	183,281
pharmacy	174,244
laboratory	141,270
billing	82,982
emergency	33,050
radiology	30,732
inpatient	16,900
interoperability	13,068
procedure	2,696
transfer	2,292

496 patients (1%) are registered but never present for care — an intentionally included edge case (an empty/near-empty patient record) that QA teams commonly need to test against.

Encounter volume

66,872 unique encounters (~1.34 per patient), including 6,610 ED visits, 4,225 inpatient admissions (203 of which are readmissions within 2–28 days of a complicated discharge), 1,075 ICU transfers, and 1,348 completed procedures.

---

8. Hospital Information Systems & Interoperability Workflows

Seven systems are represented, with realistic system-to-system message flow recorded in every row via source_system/destination_system:

EHR · Laboratory Information System (LIS) · Radiology Information System (RIS) · PACS · Pharmacy Information System · Hospital Information System · Billing System

Representative flows implemented: EHR→LIS/LIS→EHR (lab orders/results), EHR→RIS→PACS→EHR (imaging), EHR→Pharmacy/Pharmacy→EHR (medications), EHR→Billing/Billing→EHR (charges).

Clinical workflows

• Registration: patient_registered → encounter_created → (occasional demographic_updated, insurance_updated, encounter_updated)
• Emergency Department: arrival → triage (with acuity) → physician assessment → orders/investigations → treatment → disposition (discharge / inpatient admission / ICU admission / transfer)
• Inpatient: admission → ward assignment → clinical assessment → investigations → medication ordering/dispensing/administration → optional ICU transfer, procedure, or complication → discharge, with pending results occasionally finalizing just after discharge (a realistic and commonly tested scenario)
• Laboratory: order → specimen collection → processing → result → transmission to EHR, across a 14-test panel (CBC, hemoglobin, WBC, platelets, creatinine, eGFR, sodium, potassium, glucose, AST, ALT, bilirubin, CRP, troponin); ~22% of results are flagged abnormal
• Radiology: order → scheduling → acquisition → report, across X-ray/CT/MRI/ultrasound/echocardiography
• Pharmacy: order → (occasional modification) → verification → dispensing → administration (repeated across an inpatient stay) → discontinuation, including cancellations and missed-administration events
• Procedures: request/scheduling → completion, with a ~6% complication rate
• Transfers: department-to-department movement with reason and bed category

---

9. Data-Integrity / Interoperability Test Scenarios

An explicit message_transmission interoperability event is logged for ~13% of eligible order/result exchanges, drawn from a weighted mix of QA scenarios, plus several scenario types embedded directly in the clinical workflow. Final counts in this generation run:

Scenario	Count
Successful transmission	3,816
Delayed transmission / delayed result	1,596
Duplicate message	1,031
Corrected result	888
Failed transmission	992
Message retry	1,049
Missing acknowledgement	694
Incorrectly routed message	529
Successfully recovered transmission	554
Cancelled medication order	2,310 (see medication_order_cancelled in longitudinal.csv)
Order modification	1,766 (medication_order_modified)
Duplicate encounter event	835 (encounter_created_duplicate)
Missed medication administration	2,983
Readmission	203
Procedure complication	69

All scenarios are identifiable through event_type, event_status, acknowledgement_status, processing_status, and error_status — see DATA_DICTIONARY.csv for the full value sets.

---

10. Synthetic Data Generation & Validation Methodology

Generation approach

The dataset was generated with a custom, fully programmatic Python pipeline (NumPy 2.4 / pandas 3.0) built specifically for this specification — not a general-purpose synthesizer. No third-party synthetic-health-data package (e.g. Synthea) and no generative model (e.g. TimeGAN, CTGAN, SDV) was used or executed at any stage; none of those tools are claimed anywhere in this document. Generation proceeded in the following stages:

1. Data-model definition — entities, relational keys, and system/workflow reference tables were defined before any records were generated.
2. Patient generation (cross_sectional.csv) — vectorized NumPy sampling with conditional/correlated distributions rather than independent randomization per field: age uses a two-component (younger-adult / older-adult) mixture; chronic-condition flags are age-weighted Bernoulli draws with explicit cross-condition coupling (e.g. diabetes probability rises given hypertension); medication burden and utilization counts are Poisson draws scaled to condition burden and age; hospital assignment is region-consistent (88% within-region); a composite complexity score (condition count, prior utilization, age, small random term) is bucketed into the four complexity tiers.
3. Encounter/event generation (longitudinal.csv) — a sequential, per-patient stochastic simulation. Each patient's pathway (outpatient / ED / direct inpatient / readmission counts) is sampled from distributions whose parameters scale with that patient's complexity tier, directly producing the mean-events-by-complexity relationship in §6. Encounters are scheduled chronologically per patient with an explicit "busy until" cursor that prevents a patient from having two overlapping inpatient-type stays. Within each encounter, leaf-level workflow generators (lab, imaging, pharmacy, procedure, transfer, billing) construct each sub-event's timestamp as a positive offset from the previous timestamp in its own chain, so chronological ordering (order ≤ collection ≤ processing ≤ result; order ≤ scheduled ≤ performed ≤ report; admission ≤ transfer ≤ discharge) holds by construction rather than by post-hoc correction.
4. Interoperability-scenario injection — for a controlled ~13% of eligible order/result exchanges, one scenario is drawn from a fixed weighted distribution (§9) and rendered as an explicit message_transmission (or paired follow-up, for duplicates/corrections) row.
5. Missingness injection — applied last, as an independent randomized pass over specific fields only (never over identifiers or fields required for relational/temporal integrity): cross-sectional demographic/insurance/lifestyle fields (1.5–6.1%), plus longitudinal provider_role (2.5% of all rows), and, within their structurally-eligible row populations, result_value (3.0%), verification_time (4.9%), processing_time (2.1%), body_region (1.0%), and acknowledgement_status (2.0%). DATA_DICTIONARY.csv reports both the structural population each field applies to and the additional injected blank rate within that population.
6. Reproducibility — the pipeline is deterministic given its fixed random seed and code; re-running the generation scripts against the same seed reproduces this dataset exactly.

Validation performed

Before export, the full generated dataset was programmatically checked against the following rules, all of which passed with zero violations on the final 50,000-patient / 680,515-event dataset:

• Exactly 50,000 unique patient_id values, no duplicates
• No duplicate event_id values; no orphaned patient_id references
• Every encounter_id-bearing row maps to exactly one patient
• No orphan order_id references (every lab result / imaging report / medication dispensing / administration / discontinuation / cancellation traces to a placed order)
• No orphan procedure_id or interoperability order_id references
• Lab chronology: collection ≥ order, processing ≥ collection, result ≥ processing (0 violations across 70,635 lab results)
• Imaging chronology: scheduled ≥ order, performed ≥ scheduled, report ≥ performed (0 violations across 15,366 imaging studies)
• Admission ≤ discharge for every inpatient encounter; no encounter missing a discharge
• No overlapping inpatient-type stays for the same patient

This validation is statistical and relational, not a claim of clinical guideline conformance — see Limitations below.

---

11. Limitations

• This is a fully synthetic dataset. Clinical values, correlations, and workflows are statistically plausible approximations designed for software/interoperability testing, not outputs of a validated clinical simulation engine, and are not derived from or validated against any real-world clinical dataset or guideline.
• The interoperability layer models realistic HL7/FHIR-style workflow patterns and failure modes (delays, duplicates, retries, routing errors, missing acknowledgements) but does not implement literal HL7v2 or FHIR message payloads/transport — it is a denormalized, relational event-log representation intended for workflow, pipeline, and relational-integrity testing rather than wire-protocol conformance testing.
• The 12-hospital synthetic network, department list, medication/lab/procedure catalogs, and reference ranges are simplified, fixed reference sets, not a comprehensive real-world formulary or nomenclature (e.g. not mapped to LOINC/RxNorm/SNOMED codes).
• Correlations between demographics, chronic conditions, and utilization are intentionally simplified (a curated set of clinically directionally-correct relationships) rather than an exhaustive epidemiological model.
• Must not be used for actual patient care, clinical decision-making, or as a substitute for real-world evidence.

---

12. File Summary

File	Rows	Columns	Size
cross_sectional.csv	50,000	29	~8.4 MB
longitudinal.csv	680,515	59	~167 MB
DATA_DICTIONARY.csv	88	11	—