Liver Disease Prediction

This project focuses on predicting liver disease using the Indian Liver Patient Dataset. It follows a full machine learning workflow—from exploratory data analysis to model deployment using a FastAPI backend with an interactive HTML frontend.

Problem Statement

Early detection of liver disease is crucial for timely medical intervention. This project aims to build a robust binary classification model to predict whether a patient is likely to have a liver disease based on medical attributes.

---

Project Structure

Liver Disease Prediction/
│
├── app/
│   ├── static/                    # Static assets (if any)
│   ├── templates/                # HTML templates for the web interface
│   │   ├── form.html             # User input form
│   │   ├── result.html           # Prediction result page
│   │   └── metrics.html          # Model metrics (accuracy, confusion matrix, etc.)
│   ├── main.py                   # FastAPI backend logic
│   ├── Ada_Model.pkl             # Final trained model (AdaBoost)
│   ├── X_test.pkl                # X_test used for evaluation
│   └── y_test.pkl                # y_test used for evaluation
│
├── images/                       # Screenshots or plots (optional)
│
├── notebooks/                    # Jupyter Notebooks for experimentation
│   ├── ExplorationAndBaselineModel.ipynb
│   ├── DataPrepration.ipynb
│   └── FinalCode.ipynb
│
├── indian_liver_patient.csv      # Original dataset from Kaggle
├── cleaned_indian_liver_patient.csv  # Cleaned version after preprocessing
└── README.md

---

---

Workflow Summary

1. Exploratory Data Analysis & Baseline Modeling

Notebook: notebooks/ExplorationAndBaselineModel.ipynb

• Dataset: Indian Liver Patient Dataset (ILPD)
• Target Variable: Dataset (1: Liver disease, 2: No disease)
• Key tasks:
  • Null value analysis
  • Correlation heatmaps
  • Feature distributions by gender and class
  • Baseline models: DecisionTreeClassfier
• Insight: Initial model performance was limited due to noisy and mislabeled and Unbala data.

2. Data Cleaning and Integrity Check

Notebook: notebooks/DataPreparation.ipynb

• Detected and removed mislabeled rows through EDA and statistical inspection.
• Feature engineering steps:
  • Converted Dataset from {1, 2} to {1, 0}
  • One-hot encoding for Gender
• Exported cleaned dataset: cleaned_indian_liver_patient.csv

3. Final Modeling and Evaluation

Notebook: notebooks/FinalCode.ipynb

• Applied SMOTEN to address class imbalance in categorical features.
• Trained multiple classifiers; selected AdaBoostClassifier based on evaluation metrics.
• Evaluation included:
  • classification report
  • Confusion Matrix

---

Model Interpretability (LIME)

• Used LIME (Local Interpretable Model-agnostic Explanations) to explain individual predictions.
• Integrated into main.py to generate local explanation plots for user inputs.
• LIME visualizations are rendered dynamically inside result.html:
  • Users can view which features influenced the prediction and by how much.
  • Enhances transparency and trust in model predictions.

---

Deployment (FastAPI App)

• Directory: app/
• Frontend: HTML templates:
  • form.html: User input form
  • result.html: Displays prediction + LIME explanation
  • metrics.html: Shows model performance metrics
• Backend: FastAPI
  • Loads serialized AdaBoost model (Ada_Model.pkl)
  • Accepts user input and returns prediction with probability and explanation
  • Serves evaluation metrics on demand

Run the app locally

cd app
uvicorn main:app --reload

---

Workflow Summary

1. Exploratory Data Analysis & Baseline Modeling

Notebook: notebooks/ExplorationAndBaselineModel.ipynb

• Dataset: Indian Liver Patient Dataset (ILPD)
• Target Variable: Dataset (1: Liver disease, 2: No disease)
• Key tasks:
  • Null value analysis
  • Correlation heatmaps
  • Feature distributions by gender and class
  • Baseline models: Logistic Regression, Random Forest, SVM
• Insight: Initial model performance was limited due to noisy and mislabeled data.

2. Data Cleaning and Integrity Check

Notebook: notebooks/DataPreparation.ipynb

• Detected and removed mislabeled rows through EDA and statistical inspection.
• Feature engineering steps:
  • Converted Dataset from {1, 2} to {1, 0}
  • One-hot encoding for Gender
  • Removed outliers for numerical stability
• Exported cleaned dataset: cleaned_indian_liver_patient.csv

3. Final Modeling and Evaluation

Notebook: notebooks/FinalCode.ipynb

• Applied SMOTEN to address class imbalance in categorical features.
• Trained multiple classifiers; selected AdaBoostClassifier based on evaluation metrics.

---

Model Interpretability (LIME)

• Used LIME (Local Interpretable Model-agnostic Explanations) to explain individual predictions.
• Integrated into main.py to generate local explanation plots for user inputs.
• LIME visualizations are rendered dynamically inside result.html:
  • Users can view which features influenced the prediction and by how much.
  • Enhances transparency and trust in model predictions.

---

Deployment (FastAPI App)

• Directory: app/
• Frontend: HTML templates:
  • form.html: User input form
  • result.html: Displays prediction + LIME explanation
  • metrics.html: Shows model performance metrics
• Backend: FastAPI
  • Loads serialized AdaBoost model (Ada_Model.pkl)
  • Accepts user input and returns prediction with probability and explanation
  • Serves evaluation metrics on demand

Run the app locally

cd app
uvicorn main:app --reload

Visit: http://127.0.0.1:8000

Dependencies

pandas==2.2.2 numpy==1.26.4 scikit-learn==1.5.1 imbalanced-learn==0.12.3 matplotlib==3.9.2 fastapi==0.115.1 uvicorn[standard] joblib==1.4.2 lime==0.2.0.1

pip install -r requirements.txt

Model Insights

---

AdaBoost handled tabular and mildly imbalanced data effectively.

SMOTEN helped equalize class representation during training.

LIME added interpretability to individual predictions.

Key Learnings

Label quality is as important as model choice—data integrity matters.

Class imbalance distorts metrics; SMOTE-like techniques are essential.

Explaining predictions (via LIME) is crucial in sensitive domains like healthcare.

Deployment requires more than a good model: usability and interpretability are key.

---

Screenshots

1. Input Form

2. Model Metrics

3. Prediction Result