Diabetes Prediction Using Machine Learning 🩺📊

This project aims to predict the likelihood of an individual having diabetes based on various health indicators using multiple machine learning models: Random Forest, Logistic Regression, and Gradient Boosting. The dataset used comes from the CDC's 2015 Behavioral Risk Factor Surveillance System (BRFSS), with health-related features such as BMI, smoking status, and physical activity.

Table of Contents 📚

• Overview
• Dataset
• Models Used
• Results
• Feature Importance
• Visualizations
• How to Run the Code
• PowerPoint Presentation
• License

Overview

In this project, multiple machine learning models were developed to predict whether an individual is likely to have diabetes or prediabetes based on 21 features. Additionally, undersampling was applied to address class imbalance, and univariate logistic regression analysis was used to select the most important features.

Objective:

• Can we predict diabetes risk based on health-related survey data? 🧠
• Which factors are most predictive of diabetes?

Dataset 📂

The dataset used for this project is the Diabetes Health Indicators Dataset from Kaggle. It contains over 250,000 survey responses from the CDC’s 2015 BRFSS survey.

• Diabetes_012: Target variable with 3 classes:
  • 0: No diabetes or diabetes during pregnancy
  • 1: Prediabetes
  • 2: Diabetes

For this project, the target was binarized into two classes:

• 0: No diabetes
• 1: Prediabetes or diabetes

To handle the class imbalance (more people without diabetes), undersampling was applied to reduce the majority class, ensuring balanced training data.

Key Features:

• BMI (Body Mass Index)
• Age
• Income
• Smoking Status 🚬
• Physical Activity
• General Health
• High Blood Pressure, Cholesterol, etc.

Models Used

1. Random Forest Classifier 🌲:

   • An ensemble method that builds multiple decision trees to make predictions.
   • Provides feature importance to understand which factors have the greatest influence on diabetes risk.
   • Achieved an accuracy of 73.2%, precision of 71.3%, and ROC-AUC of 0.80.

2. Logistic Regression:

   • A simple and interpretable linear model for binary classification.
   • Achieved an accuracy of 74.2%, precision of 73.4%, and ROC-AUC of 0.81.

3. Gradient Boosting:

   • A powerful ensemble method that iteratively builds decision trees to improve performance.
   • Added after initial models to explore whether boosting would increase predictive performance.
   • Achieved the best accuracy of 74.7%, precision of 73.0%, and ROC-AUC of 0.82.

Univariate Logistic Regression 📊

Before applying the models, a univariate logistic regression analysis was performed to assess the statistical significance of each feature. Only features with a p-value < 0.05 were selected for model building to reduce noise and improve accuracy.

Selected Features:

• HighBP: Odds ratio 2.18, p-value < 0.001
• Age: Odds ratio 1.78, p-value < 0.001
• DiffWalk: Odds ratio 1.74, p-value < 0.001
• PhysHlth: Odds ratio 1.56, p-value < 0.001
• GenHlth: Odds ratio 2.49, p-value < 0.001

Results

• Random Forest:

  • Accuracy: 73.2%
  • Precision: 71.3%
  • ROC-AUC: 0.80

• Logistic Regression:

  • Accuracy: 74.2%
  • Precision: 73.4%
  • ROC-AUC: 0.81

• Gradient Boosting:

  • Accuracy: 74.7%
  • Precision: 73.0%
  • ROC-AUC: 0.82

Visualizations 📈

The following visualizations help interpret the models' performance:

• Confusion Matrix: Shows the true vs. predicted classifications for both models.
• ROC Curve: Evaluates the trade-off between true positive rate and false positive rate for both models.

Confusion Matrix and ROC Curve:

The confusion matrices and ROC curve for all models are shown below:

Random Forest

Logistic Regression

Gradient Boosting

ROC Curves

How to Run the Code

To run this project on your local machine:

1. Clone the repository:

   git clone https://github.com/swimmingbrain/diabetes-prediction.git

2. Install the required Python packages:

   pip install -r requirements.txt

3. Run the Python script:

   python project_files/decoding_diabetes_model.py

4. Check the output for model performance, visualizations, and feature importance.

PowerPoint Presentation 🎓

You can find a PowerPoint presentation and a PDF export for my university conference summarizing this project here.

License

This project is licensed under the MIT License - see the LICENSE file for details.