System Architecture
1Dept. of Mechanical Engineering, University of Maryland, College Park, MD 20742, USA 2Fujitsu Research of America, Santa Clara, CA 95054, USA 3Fujitsu Limited, Nakahara-ku, Kawasaki, Kanagawa 211-8588, JAPAN
We present an end-to-end pipeline that converts healthcare policy documents into executable, data-aware Business Process Model and Notation (BPMN) models using large language models (LLMs) for simulation-based policy evaluation. We address the main challenges of automated policy digitization with the following four contributions: data-grounded BPMN generation with syntax auto-correction, executable augmenta-tion, KPI instrumentation, and entropy-based uncertainty detection. We evaluate the pipeline on diabetic nephropathy (DN) prevention guidelines from three Japanese municipalities using three LLM backends, generating 100 candidate models per backend per municipality to characterize output variability. Our ex- perimental results show that well-structured policies yield deterministic models matching human baselines exactly, while complex policies with implicit temporal dependencies produce high-entropy distributions across all backends, confirming that the flagged ambiguity is an intrinsic document property rather than an LLM artifact.
Original diabetic nephropathy prevention guidelines used as input to the pipeline — one document per Japanese municipality
Stage 2 or 3 diabetic nephropathy patients
Meeting ① or ② and also meeting ③ or ④:
A. Diabetes
B. Renal function is impaired
Individuals with type 1 diabetes or cancer should be excluded from the program.
Meet any of the following criteria A ①–③ and any of B ①–④:
A. Diabetes
B. Impaired renal function
Individuals with type 1 diabetes or cancer should be excluded from the program.
① Eligible for urinary albumin testing (quantitative)
② Eligible for Kidney Screening Interview
Individuals with type 1 diabetes or cancer should be excluded from the program.
Guidance for Preventing Severe Disease (Kidney Screening Interviews)
End-to-end pipeline: from uploading a healthcare policy PDF to generating executable BPMN models and simulation KPIs
KPI distributions across cities and LLM variants — click any chart to enlarge
Agreement rate, F1, recall, balanced accuracy, and Cohen's κ across all backends and cities (1,000 patients each)
Clinical attributes used as input to simulate BPMN models (synthetic records, 1,000 patients)
| # | Column | Type | Description |
|---|---|---|---|
| 1 | ID | Integer | Patient identifier |
| 2 | Sex | Binary (0/1) | Patient sex (0 = female, 1 = male) |
| 3 | Age | Integer | Patient age in years |
| 4 | Health_Check | Binary (0/1) | Annual health checkup conducted this year |
| 5 | Fasting_Blood_Glucose | Numeric (mg/dL) | Fasting blood glucose level |
| 6 | HbA1c | Numeric (%) | Glycated haemoglobin level |
| 7 | Urinary_Protein | Ordinal (0–5) | Urinary protein level (0 = −, 1 = ±, 2 = 1+, 3 = 2+, 4 = 3+, 5 = 4+) |
| 8 | Urine_lbumin | Numeric (mg/g·Cr) | Urinary albumin-to-creatinine ratio |
| 9 | eGFR | Numeric (mL/min/1.73m²) | Estimated glomerular filtration rate |
| 10 | Diabetes | Binary (0/1) | Current diabetes diagnosis flag |
| 11 | Diabetes_History | Binary (0/1) | Prior-year diabetes diagnosis on record |
| 12 | Type_1_Diabetes | Binary (0/1) | Type 1 diabetes diagnosis (exclusion criterion) |
| 13 | Type_2_Diabetes_Prior_Year_Jan_to_Dec | Binary (0/1) | Type 2 diabetes claims in prior fiscal year |
| 14 | Diabetes_Under_Treatment | Binary (0/1) | Currently receiving oral or insulin treatment for diabetes |
| 15 | Cancer | Binary (0/1) | Cancer diagnosis flag (exclusion criterion) |
| 16 | Health_Guidance | Binary (0/1) | health guidance decision choice |
| 17 | Specific_Health_Guidance_Target | Binary (0/1) | Specific health guidance program target flag |
Select a city, LLM, and generated sample — then choose a patient to see their path highlighted in the process model
Artifacts and models from this work