PredemicAI — Predicting the Epidemic Before it Begins

AI-powered epidemiological forecasting, hotspot detection, and outbreak risk mapping using COVID-19 global data.

---

Overview

PredemicAI is a machine learning pipeline that predicts the spread of infectious diseases using historical outbreak data, demographic factors, and government policy indicators. Built for the Epidemic Spread Prediction hackathon track.

What It Does

Component	Description
Case Forecasting	Random Forest + XGBoost models predict confirmed case counts 7 days ahead
Risk Classification	4-level outbreak risk system (Low / Medium / High / Critical) per country
Hotspot Detection	K-Means clustering groups 201 countries by epidemic trajectory and burden
SIR Transmission Model	Wave-fitted compartmental model estimates R₀ per epidemic wave
30-Day Forecast	Iterative future projection with 80% Monte Carlo confidence bands
Global Risk Map	Interactive Plotly choropleth maps of outbreak risk worldwide

---

Results

Metric	Value
Countries tracked	201
Date range	Feb 2020 → Mar 2023
Total records	225,321
Features engineered	21 epidemiological features
Random Forest R²	0.8921
Random Forest MAPE	~15%
Risk classes	Low (111) / Medium (54) / High (22) / Critical (14)
Wave-1 R₀ — India	~2.8
Wave-1 R₀ — US	~2.5
Wave-1 R₀ — Brazil	~2.6

---

Datasets

Primary — Johns Hopkins COVID-19 Time Series

• URL: https://github.com/CSSEGISandData/COVID-19
• Daily confirmed cases, deaths across 201 countries
• Used for: case forecasting, SIR modeling, risk classification

Secondary — Our World in Data COVID-19

• URL: https://github.com/owid/covid-19-data
• Vaccination rates, stringency index, population density, median age, hospital beds
• Used for: demographic vulnerability analysis, enriched feature set

---

Repository Structure

epidemic-ai/
├── README.md                        # This file
├── requirements.txt                 # Python dependencies
├── notebooks/
│   └── epidemic_prediction.ipynb   # Full analysis notebook (18 sections)
├── src/
│   ├── data_pipeline.py            # Data loading & preprocessing
│   ├── feature_engineering.py      # Epidemiological feature creation
│   ├── models.py                   # RF, XGBoost, SIR model classes
│   ├── forecasting.py              # 30-day iterative forecast
│   ├── clustering.py               # Hotspot detection
│   └── visualization.py            # All plot functions
├── dashboard/
│   └── index.html                  # Interactive epidemic dashboard
├── docs/
│   └── methodology.md              # Detailed methodology write-up
└── assets/
    └── screenshots/                # Dashboard and chart screenshots

---

Installation & Setup

# Clone the repository
git clone https://github.com/piyush4267/PredemicAI.git 
cd EpidemicAI

# Install dependencies
pip install -r requirements.txt

# Run the full notebook
jupyter notebook notebooks/PredemicAI.ipynb

# Or open the dashboard directly
open dashboard/index.html

---

Methodology

1. Data Pipeline

• JHU wide-format → long format via melt()
• Country-level aggregation (handles sub-national provinces)
• Left-join with OWID for demographic + policy features
• Time-ordered forward-fill for dynamic variables

2. Feature Engineering (21 features)

• Lag features: cases 1, 3, 7, 14 days ago
• Rolling statistics: 7-day average new cases
• Epidemiological metrics: Rₜ proxy, growth rate, doubling time, CFR
• Policy features: stringency index, vaccination coverage
• Demographics: population density, median age, hospital beds per 1000

3. Models

• Random Forest (200 trees, max_depth=15) — primary forecaster
• XGBoost (300 trees, lr=0.05) — comparison model
• Time-based split: trained on pre-2022 data, tested on 2022–2023
• SIR compartmental model: wave-by-wave fitting using scipy.optimize.curve_fit

4. Risk Classification

Multi-factor scoring across Rₜ proxy, 7-day growth rate, and cases per million population. Thresholds calibrated to WHO outbreak definitions.

5. Hotspot Clustering

K-Means (k=4) on per-capita burden features — separates Very High / High / Moderate / Low burden country groups with meaningful epidemiological interpretation.

---

Key Insights

1. Vaccination effect: High vaccination coverage (>60%) correlates with –40% reduction in cases per million in subsequent waves
2. Age vulnerability: Countries with median age >40 show 2.3× higher CFR
3. Policy lag: Stringency index increases precede Rₜ reduction by ~14 days
4. Wave patterns: All 4 tracked countries show 3–5 distinct waves, each with declining peak severity post-vaccination
5. R₀ range: Wave-1 estimates of 2.5–3.0 consistent with published COVID-19 literature

---

Limitations

• JHU dataset ends March 2023 — retrospective validation only
• SIR model assumes homogeneous mixing (no age/spatial structure)
• Mobility data not included (would improve hotspot detection)
• Clustering depends on data quality — countries with poor surveillance are underrepresented

---

Future Work

• SEIR model with vaccination compartment
• Integration of live data feeds (WHO, ECDC)
• Mobility data from Google/Apple COVID reports
• Sub-national (state/district level) predictions for India
• LSTM/Transformer-based temporal model for comparison

---

Meet the Team

---

Hackathon Submission — Codecure | SPIRIT '26 | IIT (BHU) Varanasi

---

• Piyush Anand
• Anshpreet Singh
• Gurkirat Singh
• Pratyush Sehgal

---

License & Acknowledgments

• License: Distributed under the MIT License
• Data Sources: Johns Hopkins University, Our World in Data.
• Special Thanks: The SPIRIT '26 Organizing Committee, IIT (BHU), and the Codecure Mentors for fostering innovation in health-tech.

---

Built with ❤️ for Codecure @ SPIRIT '26 | IIT (BHU) Varanasi.

---

Python