ERLyAlert is a predictive analytics platform that forecasts emergency room (ER) visit surges using a hybrid machine learning model and real-time contextual data. Our goal is to help hospitals, emergency planners, and public health agencies proactively allocate staff and resources to prevent ER overload.

ERLyAlert provides:

• 7-Day Forecasts for ER admissions across selected counties in NC.
• Risk Classification using flu activity, local events, and weather.
• Factor Breakdown explaining what causes surge risks (e.g. events, flu, heat spikes).
• Interactive Streamlit Dashboard with region selector, filters, confidence intervals, and insights.
• Natural Language Forecast Assistant that answers user queries in English.
• Exportable Reports in CSV or JSON formats.

• FastAPI: for real-time API endpoints (/forecast)
• Hybrid Forecasting Model:
  • Prophet captures seasonality, trends, and holidays.
  • LSTM models recent residuals (non-linear error patterns) to increase accuracy.
  • Both models are trained per-county and combined for a hybrid prediction.
• Risk Scoring:
  • Calculates risk levels based on flu index, event count, and temperature anomaly.
  • Generates human-readable explanations and factor impact metadata.
• External Data Sources:
  • Flu: CDC FluView API
  • Events: Ticketmaster (via API)
  • Weather: OpenWeatherMap
• Libraries: pandas, scikit-learn, tensorflow/keras, joblib, Prophet

• Streamlit: Interactive Dashboard
• Plotly: For charting predicted visit volumes and confidence
• Interactive Visualization: Line and bar charts with dynamic filters

• .env - API keys and environment config
• trained_models/ - saved Prophet, LSTM, and scaler files for each county

• streamlit_app.py - the main dashboard application

• main.py - FastAPI app with /forecast route, returns hybrid forecast improved with contextual risk factors

• fetch_events.py - fetches NC event listings from Ticketmaster
• fetch_weather.py - pulls 5-day forecast data for a specified city
• fetch_flu.py - queries CDC FluView for weekly flu activity
• process_event_counts.py - converts raw event list to daily event count JSON
• process_weather_anomalies.py - calculates temperature anomalies vs monthly baselines
• external_data.py - merges flu, weather, and event data into a unified context dictionary

• train_hybrid_model.py - trains or loads per-county Prophet+LSTM models and returns 7-day forecasts
• predict_hybrid_models.py - runs inference using saved hybrid models
• retrain_all.py - retrains hybrid models for all counties

• risk_scoring.py - calculates risk scores from contextual data and generates explanations
• preprocessing.py - handles ER data cleaning

• events_daily.json, flu_NC.json, weather_daily_charlotte.json - contextual inputs used for risk scoring

We are using a hybrid time series model:

• Facebook Prophet is trained on historical ER visits for each county
• Forecast residuals (errors) are fed into an LSTM neural network
• The LSTM learns non-linear deviations that Prophet misses
• Final forecast = Prophet prediction + LSTM residual adjustment
• High-risk days are flagged when predictions exceed mean + 1 standard deviation

The contextual risk scoring further analyzes each day by matching the date with real-world external factors and produces:

• risk_level: Low, Medium, High
• explanation: why it’s risky (e.g., "flu and events")
• contributing_factors: flu index, event counts, temperature anomalies