Diabetes Classification Using Gaussian Naive Bayes

1. Project Overview

This project focuses on predicting diabetes status based on a dataset containing blood pressure and glucose levels of 995 individuals. We utilized descriptive statistics, implemented statistical methods for data preprocessing, and developed a Gaussian Naive Bayes classifier from scratch. The model was compared with the results in the original dataset and validated using the Gaussian Naive Bayes classifier from standard Python packages.

2. Methods

2.1 Data Description

• Dataset: Obtained from Kaggle, comprising blood pressure, glucose levels, and diabetes status of 995 individuals.

• Features:

  • Glucose: Quantitative variable representing glucose levels.

  

  • Blood Pressure: Quantitative variable.

  

• Target Variable:

  • Diabetes: Binary indicating diabetic or non-diabetic status.

2.2 Data Preprocessing

• Converted "diabetes" column to Boolean format.
• Addressed outliers using the Interquartile Range (IQR) method.

• Standardized data for uniform scale.

2.3 Data Splitting

• Divided the dataset into 80% training and 20% testing data.

2.4 Descriptive Statistics

• Computed mean, standard deviation, and variance for "blood pressure" and "glucose."
• Standardized data for zero mean and unit variance.

2.5 Testing Normality

• Conducted Shapiro-Wilk test for normality on blood pressure and glucose features.

2.6 Conditional Plots

• Utilized conditional plots to visualize glucose and blood pressure distributions for diabetic and non-diabetic individuals.

2.7 Naive Bayes Model

• Developed a custom class called NB to implement the Gaussian Naive Bayes model.
• Trained the model using prior probabilities and estimated parameters for feature distributions.
• Predicted diabetes status for the test set.
• Compared results with the Gaussian Naive Bayes classifier from Python packages.

3. Results and Discussions

• Identified multimodal distribution for glucose and bimodal distribution for blood pressure.
• Tested non-normality of features using the Shapiro-Wilk test.
• Achieved 94.87% accuracy with the custom Gaussian Naive Bayes classifier.
• Verified identical results with the Gaussian Naive Bayes classifier from standard Python packages.

4. Conclusion

• Successfully classified individuals as diabetic or non-diabetic using blood pressure and glucose measurements.
• Acknowledged multimodal and bimodal distributions for glucose and blood pressure, respectively.
• Gaussian Naive Bayes model achieved a high accuracy of 94.87%.
• Noted the assumption of independence in the Gaussian Naive Bayes classifier and potential influences between glucose and blood pressure.