dahutapea/doctortool-clinic-data-suite

DoctorTool clinic data suite

An end-to-end case study built around a real problem stated by DoctorTool (a Hacktiv8 hiring partner): "capturing large volumes of clinic data."

The answer comes in two connected parts:

1. Market intelligence - where is the uncaptured clinic market? Analytics on 40,653 SATUSEHAT-integrated health facilities (undocumented public JSON API behind the ministry's healthcare-list page), joined against Kemenkes facility counts. Adoption waves, geographic concentration, and whitespace analysis.
2. Record digitization - how do you capture a clinic once you win it? An OCR + NLP pipeline that turns (synthetic) paper clinic forms into structured, SATUSEHAT-ready records, with a measured accuracy story.

The two halves interlock: the synthetic form generator in part 2 samples the real district and facility-type distribution from part 1's dataset, and both halves end at the same destination - SATUSEHAT-shaped clinic data.

Status

• Data acquisition: 40,653 facilities + 301 verified RME vendors (2026-07-19 snapshot)
• Phase 1: market intelligence notebook + dashboard page
• Phase 2: digitization pipeline + demo page

Repository layout

app.py                     Streamlit entry point (home)
pages/
  1_market.py              dashboard: adoption, geography, whitespace
  2_digitize.py            demo: upload form -> structured record (phase 2)
market_intel/
  download_satusehat.py    snapshot downloader (rerunnable; VPN must be OFF)
  doctortool_notebook.ipynb  analysis notebook (h8_env kernel)
  data/                    CSVs + raw JSON pages (raw/ is gitignored)
digitize/
  generate_forms.py        synthetic clinic forms + ground-truth labels
  degrade.py               scan-noise augmentation
  ocr_extract.py           image -> raw text (PaddleOCR)
  parse_fields.py          raw text -> structured fields + ICD-10 mapping
  evaluate.py              field accuracy + CER/WER vs ground truth

Environments

• h8_env (conda): notebook + market dashboard (pandas, matplotlib, rapidfuzz, streamlit)
• ocr_env (conda): digitization pipeline (paddleocr, reportlab, jiwer, ...) created separately so the heavy OCR stack cannot disturb coursework installs

Data sources and etiquette

• SATUSEHAT platform facility list: public JSON endpoint discovered behind https://satusehat.kemkes.go.id/platform/healthcare-list (category=fasyankes, perPage=100). Rate-limited to ~1 request/second; raw responses are kept untouched under market_intel/data/raw/<date>/.
• Kemenkes sites sit behind Cloudflare and reject non-Indonesian IPs: turn the VPN OFF before running the downloader.
• Vendor pricing fields exist in the API but are zeroed at source for all 301 vendors (no vendor publishes pricing) - documented as a known limit.
• No real patient data anywhere in this project. All medical records in part 2 are synthetic (Faker id_ID); only market structure is real.

Run

conda activate h8_env
streamlit run app.py