MIMIC-RiskLab is a research-grade pipeline designed to process Electronic Health Records (EHR) to predict patient outcomes in Intensive Care Units.
Predicting mortality risk in the ICU is critical for resource allocation and clinical decision support. This project implements a deep learning approach to analyze patient vitals over time using the MIMIC-III dataset. 🚀 Launch Temporal Risk Dashboard
- Data Cleaning: Automated handling of irregular time-series data and missing clinical vitals.
- Feature Engineering: Calculation of clinical scores (SAPS II, SOFA) and rolling statistics.
- Deep Learning Architecture: Utilizes Long Short-Term Memory (LSTM) networks to capture temporal dependencies in patient health.
- Extraction: Querying PostgreSQL tables (Admissions, Patients, LabEvents).
- Preprocessing: Outlier detection and mean-imputation.
- Training: Model optimization using Adam optimizer and Cross-Entropy loss.
- Evaluation: Performance measured via AUC-ROC and PR-Curve.
Access to the MIMIC-III dataset is restricted. You must complete the CITI training and request access via PhysioNet. No PHI (Protected Health Information) is included in this repository.
| Model | AUC-ROC | Accuracy |
|---|---|---|
| Random Forest | 0.78 | 81% |
| LSTM (Current) | 0.88 | 89% |
- Install dependencies:
pip install -r requirements.txt - Run the main evaluation:
python src/main.py