privacy-preserving-scd-pain-diagnosis

KO | EN

gitlite — search

#open-source #css #gemini #cross-platform #golang #typescript #claude #rag #chatbot #ai-agent

privacy-preserving-scd-pain-diagnosis

(★ 9)

Privacy-preserving ML system identifying causes of chronic pain in sickle cell disease using homomorphic encryption, with federated learning and differential privacy extensions. Research prototype, synthetic data only.

파일 탐색기

ZIP 다운로드

- - synthetic_sickle_cell_data_CLEANED.csv
- - synthetic_sickle_cell_data.csv
  - synthetic_sickle_cell_data_MESSY.csv
- 01_data_cleaning.ipynb
- generate_synthetic_data.py
- make_data_messy.py
.gitignore
README.md
requirements.txt

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Identifying the likely cause of a chronic pain episode in sickle cell disease
(SCD) patients, including vaso-occlusive crisis, infection, dehydration,
avascular necrosis, and acute chest syndrome, without ever exposing raw
patient data during computation.

## Status: Early-stage research prototype

> **This is not a validated clinical tool.** The model is currently trained
> and tested only on synthetic data designed to mirror plausible clinical
> patterns. It has not been validated on real patient data and is not
> intended for diagnostic use. Real-data validation is planned in
> collaboration with a clinical partner.

## Why this project

Sickle cell chronic pain can stem from several overlapping causes, and
telling them apart at the point of care is difficult, especially in
resource-constrained clinical settings. Existing machine learning work in
this space largely focuses on predicting *whether* a crisis will occur, not
*what* is causing a current pain episode. Patient health data is also
sensitive, so any system touching it should be built so that raw data is
never exposed to the system performing the computation.

This project combines:
- **Cause classification**: a multiclass model distinguishing likely causes
  of a current pain episode
- **Homomorphic encryption (HE)**: patient feature data is encrypted before
  any inference is performed, so the computing system never sees plaintext
  patient data
- **Federated learning (FL)** *(planned)*: enabling multiple clinics to
  contribute to model training without sharing raw records
- **Differential privacy (DP)** *(planned)*: protecting against information
  leakage through shared model updates in the federated setting

As far as we've been able to determine, this specific combination has not
been applied to sickle cell disease before (see Related Work below).

## Repository structure

```
privacy-preserving-scd-pain-diagnosis/
├── README.md
├── requirements.txt
├── .gitignore
├── data/
│   ├── raw/              # data as generated, never edited in place
│   └── processed/        # cleaned output
├── notebooks/            # exploration and data cleaning work
├── src/                  # reusable scripts (data generation, model, encryption, app)
└── models/               # trained model artifacts
```

## Setup

Using conda:
```
conda create -p venv python=3.13
conda activate ./venv
pip install -r requirements.txt
```

## How to run (current steps)

1. Generate the synthetic dataset:
   ```
   python src/generate_synthetic_data.py
   ```
2. (Optional, for practice or testing) generate a version with realistic
   data quality issues injected:
   ```
   python src/make_data_messy.py
   ```
3. Clean the data (see `notebooks/` for the working cleaning notebook)

## Roadmap

- [x] Synthetic dataset generation (2,100 records, 6 cause classes)
- [x] Synthetic "messy data" generator for data-cleaning practice
- [x] Data cleaning pipeline
- [ ] Baseline plaintext classification model
- [ ] Encrypted inference layer (homomorphic encryption via TenSEAL/CKKS)
- [ ] Federated learning extension across multiple sites
- [ ] Differential privacy on shared model updates
- [ ] Streamlit demo interface
- [ ] Real clinical data validation (pending clinical partnership)

## Related work

This project builds conceptually on prior work applying partial homomorphic
encryption to healthcare data analysis. It is distinct from the closest
related work we're aware of, a 2026 study applying federated learning
(without an encryption or differential privacy layer) to a sickle cell
classification task. This project differs by (1) adding a privacy-preserving
encryption and DP layer on top of FL, and (2) focusing on cause
differentiation for chronic pain rather than disease presence or
classification.

## Collaborators

- Abdul Kareem Adamu ([Khemshield](https://abdulkareem.khemshield.com)), research design,
  system architecture, implementation

## License

TBD

privacy-preserving-scd-pain-diagnosis

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README