SIDHA is an open reference data architecture for hidradenitis suppurativa (HS) research: it turns fragmented, episodic data — clinic records, public patient discussions, and (next) wearable-sensor streams — into a continuous, research-ready signal for early detection and flare-risk prediction.
HS patients routinely wait the better part of a decade for a correct diagnosis while rich early signals scatter across EHR codes, registries, and patient forums. SIDHA ingests those heterogeneous sources, maps everyday patient language to clinical taxonomies (ICD/SNOMED), normalizes everything to open health-data standards (OMOP CDM, FHIR R4), and scores early-detection risk with transparent, explainable models — all runnable on synthetic data with zero PHI.
git clone https://github.com/developer-rpai/sidha.git
cd sidha
pip install -e . # PyPI release planned for Phase 5Requires Python ≥ 3.10. The core is stdlib-only — no heavy dependencies.
End-to-end on synthetic data (deterministic, no PHI, no setup):
from sidha.datasets import generate_cohort
from sidha.ehr import build_timelines
from sidha.clinical import summarize_cohort
from sidha.diagnostics import EarlyDetectionScreener, misdiagnosis_report
# Synthetic cohort -> patient timelines -> cohort summary
events = generate_cohort(n_patients=200, seed=99)
timelines = list(build_timelines(events).values())
print(summarize_cohort(timelines))
# How long does HS take to be diagnosed, and what is it confused with?
print(misdiagnosis_report(timelines)["median_delay_days_overall"])
# Early-detection screening on one patient — score plus the reasons why
screener = EarlyDetectionScreener() # reference weights
print(f"risk score: {screener.score(timelines[0]):.3f}")
print(screener.feature_contributions(timelines[0])[:3])A fuller pipeline (FHIR/OMOP loaders, differential ranking, calibrating the screener on your own cohort) follows below — keep reading.
Hidradenitis suppurativa carries one of the longest diagnostic delays in dermatology — patients routinely wait the better part of a decade between first symptoms and a correct diagnosis, cycling through abscess, acne, and pilonidal-disease codes along the way. Meanwhile, rich signals about early symptoms, flare patterns, and treatment experience sit scattered across clinical records, registries, and public patient discussions, in formats no researcher can query together.
SIDHA is a data architecture for closing that gap: ingest heterogeneous HS data, map everyday patient language to clinical taxonomies, quantify diagnostic delay, score early-detection risk from EHR patterns, normalize everything to open health-data standards (OMOP CDM, FHIR R4), and evaluate analytic approaches reproducibly.
There is currently no actively maintained HS-specific open-source project. SIDHA aims to be that shared foundation.
- Load records from flat CSVs, FHIR R4 Bundles, or OMOP CDM extracts
(
sidha.ehr). - Phenotype cohorts with versioned, citable HS case definitions,
comorbidity profiles, Hurley staging helpers, and HiSQOL/HSSD instrument
scoring (
sidha.clinical). - Measure diagnostic delay from first suggestive code to confirmed
diagnosis, and generate cohort-level misdiagnosis-pattern reports
(
sidha.clinical,sidha.diagnostics). - Screen for pre-diagnosis risk with
EarlyDetectionScreener, a transparent logistic-style scorer over documented EHR features — with reference weights for exploration and afit()method to calibrate on your own labeled cohort (sidha.diagnostics). - Rank differential diagnoses with an evidence-accumulating, fully
explainable baseline (
sidha.diagnostics). - Work on synthetic data — a seeded synthetic HS cohort generator means
every tutorial, test, and demo runs with zero PHI (
sidha.datasets). - Normalize to OMOP CDM and FHIR R4 builders (
sidha.normalize).
pip install sidha # once published; for now: pip install -e .
python -c "import sidha; print(sidha.__version__)"End-to-end on synthetic data — no PHI, no setup:
from sidha.datasets import generate_cohort
from sidha.ehr import build_timelines
from sidha.clinical import summarize_cohort
from sidha.diagnostics import (
EarlyDetectionScreener, evaluate, misdiagnosis_report, rank_differentials,
)
# 1. Synthetic cohort (deterministic) -> patient timelines
events = generate_cohort(n_patients=200, seed=99)
timelines = list(build_timelines(events).values())
# 2. Cohort summary: case counts + mean diagnostic delay
print(summarize_cohort(timelines))
# 3. What gets misdiagnosed, and how long does it take?
print(misdiagnosis_report(timelines)["median_delay_days_overall"])
# 4. Early-detection screening on one patient
screener = EarlyDetectionScreener() # reference weights
print(f"risk score: {screener.score(timelines[0]):.3f}")
print(screener.feature_contributions(timelines[0])[:3]) # why
# 5. Differential ranking with evidence reasons
for cand in rank_differentials(timelines[0])[:3]:
print(f"{cand.name}: {cand.score:.1f} — {cand.supporting[0] if cand.supporting else 'no evidence'}")Bring your own records:
from sidha.ehr import load_csv, load_bundle, load_omop
events = load_csv("records.csv") # documented schema, see module docs
events, report = load_bundle("bundle.json") # FHIR R4
events, report = load_omop("omop_tables/") # OMOP CDM v5.x CSVsCalibrate the screener on your own labeled cohort:
screener = EarlyDetectionScreener().fit(timelines, labels) # 1 = HS case
print(evaluate(screener, timelines, labels))┌─────────────┐ ┌──────────────┐ ┌───────────────────┐
│ ingest │────▶│ clinical │────▶│ diagnostics │
│ public │ │ ICD-10/11, │ │ screener, │
│ social + │ │ SNOMED, │ │ differential │
│ episodic │ │ phenotypes, │ │ ranking, │
│ records │ │ instruments │ │ misdiagnosis │
└─────────────┘ └──────────────┘ └────────┬──────────┘
│
▼
┌─────────────┐ ┌──────────────┐ ┌───────────────────┐
│ eval │◀────│ judge │◀────│ normalize │
│ metrics + │ │ LLM-as-judge │ │ OMOP CDM + │
│ synthetic │ │ differential-│ │ FHIR R4 output │
│ golden sets │ │ dx harness │ │ │
└─────────────┘ └──────────────┘ └───────────────────┘
See docs/ARCHITECTURE.md for the full design and docs/BIG-PACKAGE-PLAN.md for the v1.0 build plan.
sidha/
├── sidha/ # Python package (stdlib-only core)
│ ├── ingest/ # public social-listening collectors + record loaders
│ ├── ehr/ # EHR loaders: flat CSV, FHIR R4 bundles, OMOP CDM
│ ├── clinical/ # ICD-10/ICD-11/SNOMED crosswalks, computable
│ │ # phenotypes, diagnostic-delay analytics, instruments
│ ├── diagnostics/ # early-detection screener, differential ranking,
│ │ # misdiagnosis reports
│ ├── datasets/ # dataset catalog + synthetic cohort generator
│ ├── taxonomy/ # patient-language → clinical-taxonomy mapper
│ ├── normalize/ # OMOP CDM + FHIR R4 normalization layer
│ ├── judge/ # LLM-as-judge differential-diagnosis eval harness
│ └── eval/ # metrics + synthetic golden test sets
├── notebooks/ # exploratory analyses
├── data/ # synthetic/sample data ONLY (never real data)
├── docs/ # architecture + roadmap + big-package plan
└── tests/ # 52 tests, all on synthetic data
- Phase 1 (done):
ehr+clinicalcore — timelines, loaders, crosswalks, versioned phenotypes, diagnostic-delay analytics, instruments, synthetic cohort generator. - Phase 2 (done):
diagnostics— early-detection screener, differential ranking, misdiagnosis reports. - Phase 3: public dataset loaders (GEO transcriptomics), dataset catalog expansion.
- Phase 4:
sensors+nlp+viz+privacy— full sensor-to-insight pipeline. - Phase 5: PyPI release, docs site, v1.0 tag, Zenodo DOI.
The detailed checklist lives in docs/ROADMAP.md.
good-first-issue labels mark bite-size entry points for outside contributors.
Contributions are welcome — code, documentation, dataset ideas, and evaluation design. See CONTRIBUTING.md. Ground rules: public/synthetic data only, no medical-advice framing, standards (OMOP/FHIR) first.
If you use SIDHA in research, please cite it (see CITATION.cff).
Apache-2.0 — see LICENSE.
SIDHA is a research reference implementation, not a medical device and not medical advice. The early-detection screener's reference weights are illustrative defaults, not fitted on real patient data — calibrate on your own labeled cohort before any research use. This repository contains no real patient data — only synthetic or sample data. Any clinical application requires appropriate validation, oversight, and data-use agreements.