This project predicts Emergency Department (ED) outcomes using machine learning (ML) models on data from MIMIC-IV-ED.
We used demographic info, vital signs, and chief complaints to predict outcomes like hospital admission, discharge, or mortality. The models help with early triage decisions and resource planning.
Emergency Departments are resource-constrained and need fast decisions. Using ML models trained on structured and unstructured data, we aim to:
- Predict disposition (e.g., discharge, admit)
- Estimate mortality risk
- Improve triage support using early-available data
- Dataset: MIMIC-IV-ED v2.2 (PhysioNet)
- Tools: Python 3, Google Colab
- Libraries: Pandas, NumPy, Matplotlib, Plotly, Scikit-learn, XGBoost
- Demographics: age, gender, race
- Chief complaint (text, processed using NLP)
- Vital signs: temperature, heart rate, BP, SpO₂, respiratory rate
- Social factors: marital status, insurance, language
- ED details: arrival method, triage acuity, length of stay
-
Preprocessing
- Filter ICD-10 cases only
- Cleaned triage, vitals, and meds tables
- Categorical encoding + scaling
- NLP categorization of chief complaints (reduced to 19 groups)
-
Modeling
- Feature selection + PCA
- K-Means clustering (4 clusters)
- Classification models:
- Logistic Regression
- Decision Tree
- XGBoost (best performer)
- LLM (Me-LLaMA)
-
Performance
- Best AUC-ROC: 0.88
- Evaluated with accuracy, precision, recall, F1-score
- Chief Complaints + Vital Signs were the most predictive
- XGBoost outperformed all baseline models
- NLP helped convert free-text into structured inputs
📦 Predictive-Analysis ├── Clustering_and_Classification.ipynb # ML model and clustering ├── ED_exploratory_analysis.ipynb # EDA & preprocessing ├── Final Report.pdf # Detailed report └── README.md # Project summary
Gaurav More
📧 more.56@buckeyemail.osu.edu
🔗 LinkedIn