ER Waiting Time Prediction

This project simulates and predicts the expected waiting time for patients in an Emergency Room (ER) using machine learning and large language models (LLMs). It integrates synthetic data generation, regression modeling, and explainability tools such as SHAP.

Overview

Hospitals often face challenges in managing patient flow in ERs. Accurately predicting wait times can improve patient satisfaction and resource allocation. This project demonstrates how to:

• Generate a synthetic ER dataset
• Model ER wait times using Random Forest and XGBoost
• Incorporate patient symptom descriptions using sentence embeddings from a pre-trained BERT model
• Visualize feature importance and interpret model decisions using SHAP

Project Structure

ER Waiting Time Prediction ├── er_waiting_time.py # Complete codebase ├── synthetic_er_wait_data.csv # Generated synthetic data └── README.md # Project documentation

Features

• Synthetic Data Generator Randomly simulates patient arrivals, symptom types, staff availability, and triage levels.

• Baseline Machine Learning Models Uses Random Forest and XGBoost for regression modeling of wait time.

• Text Feature Engineering Leverages BERT sentence embeddings to convert symptoms into numeric features.

• Explainability with SHAP
  Understands model predictions at both global and individual levels.

Model Evaluation

Model	MAE (min)	RMSE (min)	R² Score
Random Forest	~7.1	~9.2	~0.85
XGBoost (Base)	~6.7	~8.5	~0.88
XGBoost + LLM	~5.9	~7.4	~0.91

Visualizations

• Feature Importance
• Actual vs Predicted Wait Times
• SHAP Summary and Waterfall Plots

These plots help stakeholders understand which features (e.g., queue length, triage level, symptom) are influencing wait time predictions the most.

Technologies Used

• Python
• Pandas, NumPy, Scikit-learn
• XGBoost
• Sentence Transformers (BERT embeddings)
• SHAP
• Matplotlib

Use Cases

• Hospital wait time dashboard integration
• Resource planning for emergency departments
• Scenario testing under different ER conditions

Future Enhancements

• Deploy the model using Flask or FastAPI
• Build a front-end dashboard using Streamlit
• Use real-world datasets for improved generalizability
• Integrate alert systems for critical triage levels