Daffaadityp/jkn-what-if-engine-synthetic

JKN What-If Engine: Synthetic Healthcare Causal & Policy Evaluation Dataset

Healthkathon 2026 BPJS Kesehatan - Subkategori #15: Tindakan Pengobatan Tidak Sesuai Indikasi Medis

Dataset Overview

This repository contains the synthetic relational benchmark dataset and ground-truth latent counterfactual outcomes developed for the JKN What-If Engine. The system evaluates medical treatment appropriateness, heterogeneous treatment effects (HTE), and counterfactual policy optimization within the Indonesian national health insurance (BPJS Kesehatan) scheme.

Relational Schema (data/synthetic/)

The observable dataset comprises 111,073 total rows across 8 relational tables (2025-01-01 to 2026-06-30):

• patients.csv (20,000 rows): Demographic profiles, BPJS membership segments, baseline risk.
• facilities.csv (60 rows): Healthcare facilities (FKTP & FKRTL), accreditation classes (A/B/C/D), capacity.
• diagnoses.csv (8 rows): Primary diagnoses across 4 major disease clusters (Kardiovaskular, Muskuloskeletal, Respiratori, Gastrointestinal) with locked severity levels (Ringan, Sedang, Berat).
• encounters.csv (30,000 rows): Clinical visits, admission/discharge timestamps, primary diagnosis, comorbidity count.
• treatments.csv (30,000 rows): Medical intervention codes (T01: Konservatif Standar, T02: Medis Intensif, T03: Prosedur/Operatif, T04: Alternatif/Eksperimental) and decision dates.
• claims.csv (30,000 rows): INA-CBG claim tariffs, submitted amounts, approved amounts, payment status.
• simulations.csv (5 rows): Multi-scenario simulation configurations for counterfactual policy testing.
• simulation_scenarios.csv (1,000 rows): Scenario evaluation outcomes and simulated health economics metrics.

Latent Ground Truth (data/latent/ground_truth_audit.parquet)

For benchmarking causal inference algorithms without unobserved confounding, complete latent ground-truth parameters are provided (30,000 encounters):

• Potential outcomes under all 4 treatment arms: $Y_i(T01)$, $Y_i(T02)$, $Y_i(T03)$, $Y_i(T04)$ for 30-day readmission risk, length of stay (LOS), and claim cost (IDR).
• True individual treatment effects (ITE) for all pairs.
• True propensity scores $P(T_k \mid X)$ derived from the Data Generating Process (DGP).

Important Methodological & Clinical Disclaimer

DISCLAIMER: This dataset is completely synthetic, generated via a transparent, auditable Data Generating Process (DGP) for algorithmic benchmarking and prototyping during Healthkathon 2026. It does NOT contain real personal health data (PHD) of BPJS Kesehatan participants. Findings, parameters, and counterfactual valuations reflect the synthetic DGP and must NOT be used for unrestricted real-world clinical claims or policy decisions without validation on real-world clinical databases (e.g., MIMIC-IV, MarketScan, or official BPJS Kesehatan sample data).