Pediatric Pneumonia Detection using Genetic Algorithm Feature Selection

A multimodal machine learning pipeline for detecting pneumonia and identifying its cause (Bacterial vs Viral) from chest X-rays combined with synthetic clinical data. A Genetic Algorithm (GA) is used for intelligent feature selection across the fused feature space.

---

Project Overview

This project demonstrates:

• Multimodal fusion: CNN image features (MobileNetV2) + clinical tabular features
• Genetic Algorithm for automated feature selection
• Binary classification: Normal vs Pneumonia
• Multi-class classification: Normal vs Bacterial Pneumonia vs Viral Pneumonia

Dataset

Chest X-Ray Images (Pneumonia) by Paul Mooney on Kaggle.

Note: Clinical features (temperature, WBC count, SpO2, etc.) are synthetically generated and correlated with the image labels to simulate a real multimodal dataset. They do not come from real patients.

---

Project Structure

Mini_Proj/
│
├── data/                        # Created after running download_data.py
│   └── chest_xray/
│       ├── train/
│       │   ├── NORMAL/
│       │   └── PNEUMONIA/
│       ├── test/
│       ├── val/
│       └── clinical_data.csv    # Auto-generated synthetic clinical data
│
├── outputs/                     # Created after running train_evaluate.py
│   ├── cm_pneumonia.png
│   └── cm_cause.png
│
├── data_loader.py               # Loads images + clinical data, preprocesses
├── download_data.py             # Downloads Kaggle dataset + generates clinical data
├── feature_extractor.py         # MobileNetV2 CNN feature extraction + fusion
├── genetic_algorithm.py         # GA-based feature selection
├── train_evaluate.py            # Main training & evaluation pipeline
├── requirements.txt
├── README.md
└── .gitignore

---

Setup & Installation

1. Clone the Repository

git clone https://github.com/<your-username>/pediatric-pneumonia-detection.git
cd pediatric-pneumonia-detection

2. Create a Virtual Environment (Recommended)

python -m venv venv
source venv/bin/activate        # On Windows: venv\Scripts\activate

3. Install Dependencies

pip install -r requirements.txt

4. Configure Kaggle API

1. Go to kaggle.com → Account → Create API Token
2. Place the downloaded kaggle.json in ~/.kaggle/ (Linux/macOS) or C:\Users\<user>\.kaggle\ (Windows)
3. Set permissions: chmod 600 ~/.kaggle/kaggle.json

---

Running the Project

Step 1 — Download Data

python download_data.py

This downloads the Kaggle chest X-ray dataset and generates clinical_data.csv.

Step 2 — Run the Full Pipeline

python train_evaluate.py

For a quick test with a small sample:

# In train_evaluate.py, change:
loader.load_data(max_samples=200)   # Use a small number
# GA settings:
GeneticAlgorithmFeatureSelection(population_size=10, generations=5)

For a full run, set max_samples=None and increase GA population/generations.

---

Pipeline Architecture

Chest X-Ray Images
        │
        ▼
  MobileNetV2 (pretrained, ImageNet)
  GlobalAveragePooling2D
        │
        ▼                    Synthetic Clinical Features
  CNN Features (1280-d)  +   (temperature, WBC, SpO2, ...)
        │                              │
        └──────────── Fusion ──────────┘
                          │
                          ▼
              Genetic Algorithm Feature Selection
                          │
                          ▼
              RandomForest Classifier
               /                    \
    Binary Classification       Multi-class Classification
    (Normal vs Pneumonia)    (Normal vs Bacteria vs Virus)

---

Results

Outputs are saved to the outputs/ directory:

• cm_pneumonia.png — Confusion matrix for binary classification
• cm_cause.png — Confusion matrix for cause classification

Results vary depending on max_samples and GA parameters. Use max_samples=None for best accuracy.

---

Key Design Choices

Component	Choice	Reason
CNN Backbone	MobileNetV2	Lightweight, pretrained on ImageNet
Image Size	128×128	Balance between speed and detail
Feature Selection	Genetic Algorithm	Handles mixed (image+clinical) feature spaces
Classifier	Random Forest	Robust, interpretable, handles high dimensions
Clinical Data	Synthetic	Demonstrates multimodal pipeline without real EHR data

---

Limitations

• Clinical features are synthetic and not clinically validated
• Small sample sizes significantly affect accuracy
• The GA is computationally expensive; increase generations for better results
• Unknown pneumonia cause samples (those without bacteria/virus in filename) are excluded from multi-class training

---

Requirements

• Python 3.8–3.10
• TensorFlow < 2.16
• See requirements.txt for full list

---

License

This project is for educational purposes. The chest X-ray dataset is subject to Kaggle's terms of use.