🫁 LRI Risk Prediction in Chadian Children Under Five

Lower Respiratory Infection (LRI) is one of the leading causes of death in children under five in Chad.
This project builds a machine learning pipeline to predict LRI risk at the individual child level — enabling
targeted health interventions where resources are most scarce.

---

🌍 Why This Matters

Chad has one of the highest child mortality rates in the world. LRI alone accounts for a significant share of under-five deaths, yet healthcare infrastructure remains severely limited outside N'Djamena.

The core question this project answers:
"Given a child's demographic profile, household conditions, and nutritional status — how likely are they
to develop a Lower Respiratory Infection?"

Answering this with data can help NGOs, ministries of health, and field workers prioritize interventions before a child gets critically ill.

---

📊 Dataset

• Source: DHS Program — Chad Standard DHS 2014
• File: Children's Recode (KR) — household survey of mothers and children under 5
• Size: National representative sample across all regions of Chad
• Key Variables: Child age/sex, nutritional indicators (WHZ, BMI), mother's education, household wealth, access to healthcare, vaccination status, water/sanitation conditions

---

🔬 Methodology

The pipeline is designed to be leak-free and reproducible:

Raw DHS Data
    ↓
Target Construction  (H31, H31B, H31C columns → binary LRI label)
    ↓
Train/Test Split  (Stratified 70/30 — split BEFORE any feature engineering)
    ↓
Feature Screening  (Drop >60% missing | T-test + Chi-Square selection)
    ↓
Preprocessing  (Median/mode imputation | dummy encoding | "Don't know" handling)
    ↓
SMOTE Balancing  (Synthetic oversampling on training set ONLY)
    ↓
Model Benchmarking  (8+ algorithms evaluated)
    ↓
Evaluation  (Accuracy, Recall, Precision — with focus on Recall for health context)

---

🤖 Models Benchmarked

Model	Notes
Logistic Regression	Baseline
Decision Tree	Interpretability
Random Forest	Ensemble baseline
XGBoost	Best overall performance
LightGBM	Fast, competitive accuracy
CatBoost	Strong on categorical features
SVM	Tested with scaled features
KNN	Distance-based baseline

---

📌 Key Findings

• Nutritional status (Weight-for-Height Z-score) is the strongest individual predictor of LRI risk
• Mother's BMI shows significant correlation — maternal health directly impacts child vulnerability
• Household wealth index and access to clean water are among the top socioeconomic risk factors
• Ensemble tree models (XGBoost, LightGBM) consistently outperform linear models on this high-dimensional survey data
• SMOTE improved recall for the minority (positive LRI) class significantly without data leakage

---

🚀 How to Run

# 1. Clone the repo
git clone https://github.com/Derio001/lri-prediction-chad.git
cd lri-prediction-chad

# 2. Install dependencies
pip install pandas numpy matplotlib seaborn scikit-learn xgboost lightgbm catboost imbalanced-learn

# 3. Add the DHS dataset (requires free DHS program registration)
# Place chad_dhs_kr.csv in the project root

# 4. Run the full pipeline
python Mahamat_LRI_CHAD.py

# Or explore interactively
jupyter notebook Mahamat_LRI_Chad.ipynb

Note on Data Access: The DHS dataset requires a free registration at dhsprogram.com. Chad 2014 KR recode is publicly available upon request.

---

📁 Project Structure

lri-prediction-chad/
│
├── Mahamat_LRI_Chad.ipynb      # Full analysis notebook (EDA → modeling → evaluation)
├── Mahamat_LRI_CHAD.py         # Clean script version of the pipeline
├── verify_data.py              # Diagnostic: check target distribution
├── extract.py                  # Utility: extract code cells from notebooks
└── README.md

---

🔭 Future Work

• Incorporate 2023/2024 DHS data when available for Chad
• Add regional/geographic disaggregation (Sahel vs. southern regions)
• Build a lightweight Streamlit dashboard for field worker use
• Extend to other child health outcomes (malnutrition, malaria, diarrheal disease)
• Explore transfer learning from similar DHS datasets (Niger, Mali, Sudan)

---

👤 Author

Mahamat Hanga Derio
M.Tech Data Science — Christ University, Bangalore
Chadian national | Focused on data-driven solutions for Sub-Saharan African health & development

📬 Open to collaboration with NGOs, research institutions, and public health organizations
🔗 GitHub Profile

---

📄 Citation

If you use this work, please cite:

Mahamat Hanga Derio (2024). LRI Risk Prediction in Chadian Children Under Five.
GitHub. https://github.com/Derio001/lri-prediction-chad

---

This project is part of a broader effort to apply data science for public health impact in Chad and the Lake Chad Basin region.