This project aims to develop, validate and interpret a predictive deep learning (DL) model for use in patient-centered CER, leveraging methods that specifically incorporate patient partner and clinical stakeholder perspectives throughout model development. The primary focus is developing on Human-Centered AI for Clinical Decision Support The project will develop early opioid risk prediction models for Opioid Use Disorder (OUD) and Overdose (OD), and generalize the framework for other diseases such as heart failure.
This project addresses the gap between ML model development and clinical adoption through:
- Transparent Models: Interpretable predictions with SHAP-based explanations
- Clinician-in-the-Loop Validation: Interactive dashboards for clinical review
- Scalable Pipeline: Data processing for large EHR datasets
- Multi-Model Support: LightGBM, Random Forest, LSTM, GRU
| Field | Value |
|---|---|
| Contract | PCORI 23C3 |
| PI | Fusheng Wang, PhD and Richard Rosenthal, MD |
| Institution | Stony Brook University |
| Start Date | February 1, 2025 |
| Reporting | Interim reports every 6 months |
| Milestone | Due Date | Status |
|---|---|---|
| Study Protocol Submission | 2025-03-01 | Complete |
| IRB Documentation | 2025-03-01 | Complete |
| Cohort Identification | 2025-05-15 | Complete |
| Initial Model Design | 2025-07-15 | Complete |
| Progress Report #1 | 2025-07-31 | Complete |
| Model Release 1 (GitHub) | 2025-08-31 | Complete |
| Interpretability Methods | 2025-10-31 | In Progress |
- AHDT Meetings: Bi-monthly All Hands Design Team meetings
- Quarterly Surveys: Stakeholder engagement surveys
- PCORI Oversight: Reports via PCORI Online portal
- Protocol: IRB2023-00456 (Stony Brook University IRB)
- Status: Not Human Research
- Data Source: Cerner Health Facts under executed DUA
- De-identification: HIPAA Safe Harbor compliant
- Access Control: Role-based, audit-logged
- PHI: No direct identifiers in any dataset
| Dataset | Records | Features | Purpose |
|---|---|---|---|
| Health Facts | ~70M encounters | 500+ | Training/validation |
| Synthetic | 10,000 patients | 50 | Development/testing |
| Model | AUROC | AUPRC | Accuracy |
|---|---|---|---|
| LightGBM | 0.82 | 0.61 | 0.78 |
| LSTM (T=10) | 0.79 | 0.58 | 0.75 |
| Logistic Reg. | 0.75 | 0.52 | 0.72 |
External validation planned for Q2 2026.
- Subgroup analysis by demographics (age, gender, race/ethnicity)
- Calibration curves across groups
- Disparate impact assessment planned
- Research and model development
- Educational purposes
- Stakeholder-in-the-loop validation prototyping
- Not for clinical decisions: Models not validated for direct patient care
- Population specificity: Trained on specific populations
- No FDA clearance: Requires separate regulatory review for deployment
┌──────────────────────────────────────────────────────────────┐
│ PCORI ML Infrastructure │
├──────────────────────────────────────────────────────────────┤
│ Raw EHR Data → Pipeline (ETL) → Feature Store (Parquet) │
│ ↓ │
│ Model Training: LightGBM | LSTM | GRU | Logistic Reg. │
│ ↓ │
│ Explainability: SHAP values, feature importance │
│ ↓ │
│ SITL Dashboard: Cohort Builder | Training UI | AI Chat │
└──────────────────────────────────────────────────────────────┘
| Component | Description | Docs |
|---|---|---|
| SITL Dashboard | Clinician validation interface | README |
| Pipeline | Training pipeline | README |
| Feature Selection | NTK, LightGBM, Elastic Net methods | Docs |
# Clone and setup
git clone https://github.com/StonyBrookDB/PCORI.git
cd PCORI
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
# Run dashboard
cd SITL_Dashboard_PCORI
pip install -r requirements.txt
python -m uvicorn backend.main:app --port 8503
# Train model
cd ../pipeline-pcori
python train.py --dataset ./data/synth --model lightgbm@software{pcori_sitl_2026,
title = {PCORI: Human-Centered AI for Clinical Decision Support},
author = {Wang, Fusheng and Liu, Yinan and Ding, Zihan},
year = {2026},
publisher = {Stony Brook University},
url = {https://github.com/StonyBrookDB/PCORI},
note = {Funded by PCORI Contract 23C3}
}Funding: Patient-Centered Outcomes Research Institute (PCORI) Contract 23C3
Research Team: Department of Biomedical Informatics, Stony Brook University
Data Partner: Cerner Corporation (Health Facts)
MIT License - see LICENSE
This software is for research purposes only. Clinical deployment requires regulatory approval.