BIFX546_diabetes-indicators

Diabetes Prediction Using Health Indicators

BIFX-546: Machine Learning for Bioinformatics — Spring 2026

Student:
Duong Nguyen (dtn2@hood.edu)

Instructor: Dr. Sarangan Ravichandran

🎯 Project Goal

The goal of this project is to develop interpretable machine learning models for predicting diabetes status using health indicator data from the CDC Behavioral Risk Factor Surveillance System (BRFSS). The project also aims to identify the strongest risk factors associated with diabetes and evaluate how different preprocessing and imbalance-handling techniques affect predictive performance.

📊 Dataset

Field	Details
Name	Diabetes Health Indicators Dataset
Source	Kaggle / CDC BRFSS 2015
URL	https://www.kaggle.com/datasets/alexteboul/diabetes-health-indicators-dataset
Size	253,680 observations with 21 feature variables
Target Variable	`Diabetes_binary` (0 = No Diabetes, 1 = Diabetes/Prediabetes)

The dataset contains demographic, lifestyle, and clinical health indicators related to diabetes risk.

🧠 Techniques Used

Phase	Technique
EDA	Summary statistics and class distribution analysis
EDA	Histograms, boxplots, countplots, correlation heatmap
Data Cleaning	Duplicate removal
Data Cleaning	BMI outlier removal using IQR
Modeling	Logistic Regression
Modeling	Random Forest
Imbalance Handling	Class weighting
Imbalance Handling	SMOTE oversampling
Imbalance Handling	Random undersampling
Feature Selection	Reduced-feature logistic regression
Evaluation	Accuracy, Precision, Recall, F1-score
Evaluation	ROC-AUC, PR-AUC, Confusion Matrix

📁 Repository Structure

BIFX546_diabetes-indicators/
├── data/
│   ├── diabetes_012_health_indicators_BRFSS2015.csv
│   ├── diabetes_binary_5050split_health_indicators_BRFSS2015.csv
│   └── diabetes_binary_health_indicators_BRFSS2015.csv
├── notebooks/
│   └── EDA_diabetes_indicators.ipynb
├── results/
│   ├── age_distribution_diabetes_status.png
│   ├── bmi_distribution.png
│   ├── bmi_vs_diabetes_boxplot.png
│   ├── correlation_heatmap.png
│   ├── correlation_with_diabetes_barplot.png
│   └── diabetes_status_distribution.png
├── src/
│   └── .gitkeep
├── LICENSE
├── Project Proposal.docx
└── README.md

⚙️ How to Run

Option 1 — Google Colab

Open Google Colab
Click File → Open notebook → GitHub
Paste the repository URL:

https://github.com/dtn2/BIFX546_diabetes-indicators

Open the notebook:

EDA_diabetes_indicators.ipynb

Run all notebook cells:

Runtime → Run all

Required Python packages are installed within the notebook as needed.

Option 2 — Local Jupyter Notebook

# Clone repository
git clone https://github.com/dtn2/BIFX546_diabetes-indicators.git

# Move into repository folder
cd BIFX546_diabetes-indicators

# Launch Jupyter Notebook
jupyter notebook

Open the notebook:

notebooks/EDA_diabetes_indicators.ipynb

Install any missing Python packages directly in the notebook if needed.

📈 Key Results

Model	Accuracy	Diabetes Recall	ROC-AUC
Logistic Regression	73%	0.78	0.827
Random Forest	86%	0.16	0.799
Logistic Regression + SMOTE	72%	0.74	0.802
Logistic Regression + Downsampling	72%	0.75	0.804
Reduced Feature Logistic Regression	72%	0.76	0.813

Main Findings

General health, BMI, high blood pressure, age, and mobility issues showed the strongest associations with diabetes
Logistic Regression outperformed Random Forest for detecting diabetes cases
Random Forest achieved higher accuracy but missed many diabetes cases due to class imbalance
Duplicate removal and BMI outlier removal had minimal impact on performance
SMOTE and undersampling produced only small performance changes
A reduced-feature model achieved similar performance using only 6 variables

📝 Conclusion

Logistic Regression provided the most effective and interpretable model for diabetes prediction in this imbalanced healthcare dataset. Although Random Forest achieved higher overall accuracy, Logistic Regression achieved substantially higher recall for diabetes cases, making it more clinically useful for screening purposes. The results also demonstrated that a simplified model using only a few key health indicators can still effectively predict diabetes risk.

📦 Dependencies

Core Python packages:

pandas
numpy
matplotlib
seaborn
scikit-learn
imbalanced-learn
jupyter

📜 References

CDC Behavioral Risk Factor Surveillance System (BRFSS)
https://www.cdc.gov/brfss/
Kaggle Diabetes Health Indicators Dataset
https://www.kaggle.com/datasets/alexteboul/diabetes-health-indicators-dataset

BIFX-546 · Hood College · Spring 2026

This site is open source. Improve this page.