--- license: other viewer: false tags: - synthetic-data - healthcare - emr - survival-analysis language: - en task_categories: - text-generation - feature-extraction - summarization - tabular-to-text - table-to-text - text-retrieval --- # Clinically-Informed Synthetic Patient Populations ### (October 2025 Edition — 100 | 1 000 | 10 000 Patients) --- ## Overview These datasets represent **fully synthetic, privacy-free populations** that emulate the structure and statistical behavior of modern hospital data as of **October 2025**. Each population includes four inter-linked tables describing demographics, admissions, diagnoses, and laboratory measurements. **No real patient information** is used or referenced anywhere — every record is computer-generated, de-identified by design, and derived entirely from algorithmic logic. The populations reproduce the diversity of a contemporary healthcare system (≈ 30 – 95 years old in 2025) with realistic comorbidity networks, laboratory correlations, and healthcare-use patterns that collectively behave like authentic EMRs. --- ## Comparison: 2015 EMRBots vs. October 2025 Clinically-Informed Synthetic EMR | Feature | **EMRBots (2015)** | **Clinically-Informed Synthetic Population (October 2025)** | |:--|:--|:--| | **Data Source** | Randomized distributions seeded manually | Statistically generated using clinical logic and physiologic dependencies | | **Conditions** | ICD-coded, random frequency | 15 chronic conditions with realistic co-occurrence networks and age trends | | **Admissions** | Random encounters | Admissions scale with severity index and comorbidity load | | **Laboratory Data** | Independent random values | 30 + labs with coupled physiologic equations (Hct≈3×Hgb, AG = Na−(Cl+CO₂), TP = Alb+Glob) | | **Clinical Correlations** | None | Condition-specific lab shifts (↑ Glucose → Diabetes; ↑ BUN/Cr → CKD) | | **Severity Measure** | None | Transparent “Severity Index” = count of chronic conditions | | **Comorbidity Links** | Independent diseases | Realistic clusters – metabolic, cardiac, pulmonary, neuro-aging, psychosomatic | | **Validation** | Visual inspection only | Automated analytics, co-occurrence maps, and physiologic consistency tests | | **Clinical Plausibility** | Moderate (for demonstration) | High – patterns mirror true hospital data while remaining 100 % synthetic | | **Privacy Risk** | None (toy data) | None – no real patients, no protected health information | | **Intended Use** | Conceptual teaching | ML training, RWE validation, EHR pipeline testing, education, benchmarking | --- ### Summary The **October 2025 datasets** represent the **next generation of synthetic EMR design** — clinically coherent, statistically stable, and physiologically accurate. They bridge the gap between early random EMR simulators and research-grade synthetic populations that preserve medical realism without privacy concerns. --- ## Table Structure | Table | Description | |:--|:--| | **PatientCorePopulatedTable.txt** | One record per synthetic patient, including demographics, behavior, and a severity index (number of chronic conditions). | | **AdmissionsCorePopulatedTable.txt** | One record per simulated hospital stay with realistic start/end dates linked to each patient. | | **AdmissionsDiagnosesCorePopulatedTable.txt** | Chronic-condition descriptors attached to each admission (15 canonical diseases). | | **LabsCorePopulatedTable.txt** | Time-stamped laboratory values across 30 + analytes, each obeying physiologic constraints and disease-specific drifts. | --- ## Why the October 2025 Synthetic EMR Data Are High Quality These datasets are **entirely artificial yet clinically logical**. Their quality derives from coordinated realism across **demographics**, **disease architecture**, **laboratory dependencies**, and **severity-linked behavior**. ### Demographic and Behavioral Realism - **Age:** 30 – 95 yrs (peak ≈ 63 yrs), matching chronic-disease demographics. - **Gender:** Balanced male/female ratio. - **Socioeconomics:** Tri-modal poverty distribution mirroring urban–suburban–rural mix. - **Language / Race:** Predominantly English and Spanish speakers with diverse representation. - **Smoking:** Higher prevalence in lower-income groups – reflecting known public-health gradients. These parameters yield **statistically credible yet privacy-safe** population heterogeneity. ### Comorbidity Architecture and Severity Representation Fifteen chronic conditions form correlated clusters, producing realistic co-occurrence matrices: | Cluster | Typical Relationships | Observable Effects | |:--|:--|:--| | **Metabolic** | Diabetes ↔ CKD ↔ Hyperlipidemia ↔ IHD | Elevated Glucose and BUN/Cr pairs | | **Cardiac** | CHF ↔ Atrial Fibrillation ↔ IHD | More admissions + lower Albumin | | **Pulmonary** | Asthma ↔ COPD (overlap ≈ 12 %) | Mild neutrophilia and increased WBC | | **Neuro-Aging** | Alzheimer’s Dementia ↔ Age > 75 | Higher severity and longer LOS | | **Psychosomatic** | Depression ↔ Arthritis | Slightly elevated health-care use | The **Severity Index** (number of chronic conditions per patient) provides a simple, interpretable illness-burden metric. - Admissions rise monotonically with severity (r ≈ 0.9). - Lab abnormalities intensify with severity – ↑ Cr/BUN, ↓ Albumin and Hemoglobin – reflecting progressive multisystem stress. ### Physiologic Coherence in Laboratories Laboratory values are not independent; they respect medical equations and covariances: **Hematology** - Hct ≈ 3× Hgb (r ≈ 0.95) - Derived indices (MCV, MCH, MCHC, RDW) remain balanced. - CKD and Cancer patients show mild anemia and higher RDW. **Electrolytes & Acid–Base** - AG ≈ Na – (Cl + CO₂) ± 0.6 mmol/L. - Potassium independent within 3 – 6 mmol/L range. **Renal Panel** - BUN ↔ Creatinine (r ≈ 0.8). - CKD cases show BUN ≥ 20 mg/dL and Cr ≥ 1.1 mg/dL. **Proteins & Liver Function** - Total Protein = Albumin + Globulin (identity holds). - Albumin declines with cardiac disease severity. - Calcium tracks Albumin (Δ ≈ 0.1 mg/dL per 1 g/dL Alb change). **Inflammatory Markers** - WBC and Neutrophils rise in COPD / Asthma. - Absolute counts sum to total WBC exactly. **Urinalysis** - SG 1.014 – 1.028; pH 4.5 – 7.5. - CKD patients slightly higher Urine WBC. Across tens of thousands of records, < 0.5 % fall outside plausible ranges. ### Statistical Integrity Across Tables - **1 : 1 linkage** between PatientID and AdmissionID. - **Temporal logic:** no overlapping admissions per patient. - **Lab frequency:** 1 – 2 measurements per hospital day. - **Scalability:** distributions and correlations hold from 100 → 10 000 patients. ### Empirical Validation Visual and statistical checks confirm realism: - Age and severity histograms match population expectations. - Admissions vs Severity scatter shows monotonic trend. - Condition prevalence ≈ Hypertension > Hyperlipidemia > Diabetes > Arthritis > Depression. - Co-occurrence heatmaps show logical metabolic and cardiac clusters. --- ## Dataset Scales | Population | Patients | Admissions (≈) | Diagnoses (≈) | Labs (≈) | Avg # Conditions / Patient | Avg Admissions / Patient | |:--|:--:|:--:|:--:|:--:|:--:|:--:| | **Small** | 100 | ≈ 400 | ≈ 800 | ≈ 60 000 | 3.1 | 4.0 | | **Medium** | 1 000 | ≈ 4 000 | ≈ 8 000 | ≈ 600 000 | 3.2 | 4.0 | | **Large** | 10 000 | ≈ 40 000 | ≈ 80 000 | ≈ 6 000 000 | 3.2 | 4.0 | --- ## Analytics Gallery Example visuals in `analytics/` folder: - `age_hist.png` – Age distribution - `severity_hist.png` – Chronic condition counts - `condition_prevalence_bar.png` – Top 15 conditions - `condition_cooccurrence_heatmap.png` – Comorbidity matrix --- ## Applications - ML training and feature testing on synthetic EHR data - Educational modules for data science and epidemiology - Pipeline and dashboard validation for multi-table databases - Research on correlation preservation and synthetic data utility --- **Disclaimer:** All records are artificial. No real patient data were used, referenced, or replicated in any form. The datasets are generated for public research and educational purposes only. ---