🩺 Diabetes Risk Prediction — ML Mini Project

An end-to-end machine learning pipeline for early diabetes screening using the CDC BRFSS 2015 Health Indicators dataset. Achieves 88% recall on diabetic cases with a clinically-calibrated ensemble model.

---

📌 Table of Contents

• Overview
• Dataset
• Project Structure
• Methodology
• Results
• Installation
• Usage
• Key Findings
• Technologies Used

---

Overview

This project builds a binary classification model to predict diabetes risk using survey-based health indicators. Given the real-world cost of missing a diabetic diagnosis (false negatives), the model is intentionally calibrated for high recall (sensitivity), making it suitable as a first-pass clinical screening tool.

The pipeline covers the full ML lifecycle: exploratory data analysis, feature engineering, class imbalance handling, model benchmarking, ensemble construction, and threshold tuning.

---

Dataset

Source: CDC Behavioral Risk Factor Surveillance System (BRFSS) 2015
File: diabetes_binary_health_indicators_BRFSS2015.csv

Property	Value
Original rows	~253,680
After deduplication	~229,474
Features	21 health indicators
Target	Diabetes_binary (0 = No Diabetes, 1 = Diabetes)
Class imbalance	~85% Non-Diabetic / ~15% Diabetic

Key features include: HighBP, HighChol, BMI, GenHlth, Age, PhysActivity, Smoker, Stroke, HeartDiseaseorAttack, Income, Education, and more.

Note: The dataset is not included in this repository. Download it from Kaggle and place the CSV in the project root directory.

---

Project Structure

├── MLminiProject.ipynb         # Main notebook (EDA + Modeling)
├── fig1_overview.png           # Class distribution & correlations
├── fig2_deepdive.png           # Feature deep-dives (BMI, Age, GenHlth, etc.)
├── fig3_combinations.png       # Risk factor combinations & engineered features
└── README.md

---

Methodology

1. Exploratory Data Analysis

• Visualized class imbalance, feature correlations, and diabetic rates across all categorical/ordinal variables.
• Identified GenHlth, HighBP, and BMI as the strongest predictors.
• Created 3 comprehensive dark-themed visualization panels covering 10+ charts.

2. Feature Engineering

Feature	Description
BMI_log	Log transform of BMI to reduce skew
BMI_cat	Ordinal BMI category (Underweight → Extreme Obese)
is_obese	Binary flag: BMI ≥ 30
physhlth_log / menthlth_log	Log transforms for health-day counts
BP_Chol	Interaction: High BP × High Cholesterol
risk_count	Sum of vascular risks (BP, Chol, Stroke, Heart Disease)
lifestyle_score	PhysActivity + Fruits + Veggies − HvyAlcoholConsump
age_high_risk	Flag for high-risk age groups (55–74)

• Dropped: AnyHealthcare, NoDocbcCost (near-zero signal)
• Selected: Top 22 features via SelectKBest with Mutual Information

3. Class Imbalance Handling

Used SMOTE (Synthetic Minority Oversampling Technique) with sampling_strategy=0.4 inside a pipeline to prevent data leakage during cross-validation.

4. Models Trained

Model	Description
Logistic Regression	Baseline linear model
AdaBoost	Adaptive boosting on shallow decision trees
XGBoost	Gradient boosted trees (primary model)
Weighted Ensemble	Soft-voting ensemble of AdaBoost + XGBoost

5. Threshold Tuning

Two thresholds were evaluated:

• Optimal F1 threshold (0.248): Balances precision and recall.
• Medical threshold (0.208): Maximizes recall ≥ 70% to minimize missed diagnoses.

---

Results

Model Comparison (5-Fold Stratified Cross-Validation)

Model	CV ROC-AUC	CV PR-AUC
Logistic Regression	0.8070	0.4180
AdaBoost	0.8032	0.4094
XGBoost	0.8120	0.4290
AB + XGB Ensemble	0.8156	0.4399

Final Ensemble (Medical Threshold = 0.208)

Metric	Value
ROC-AUC	0.8156
PR-AUC	0.4399
Recall (Diabetic)	88%

The PR-AUC of 0.44 represents a ~3× improvement over the naive baseline of ~0.15.

Top Feature Importances (XGBoost)

1. HighBP
2. GenHlth
3. BMI_cat
4. BP_Chol (engineered)
5. Age

---

Installation

# Clone the repository
git clone https://github.com/your-username/diabetes-risk-prediction.git
cd diabetes-risk-prediction

# Create a virtual environment (recommended)
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Requirements

pandas
numpy
matplotlib
seaborn
scipy
scikit-learn
imbalanced-learn
xgboost
jupyter

Or install directly:

pip install pandas numpy matplotlib seaborn scipy scikit-learn imbalanced-learn xgboost jupyter

---

Usage

1. Download the dataset and place diabetes_binary_health_indicators_BRFSS2015.csv in the project root.
2. Launch Jupyter:

   jupyter notebook MLminiProject.ipynb

3. Run all cells in order. The notebook is structured in sections:
   • Cell 1–4: Data loading and EDA setup
   • Cell 5–7: Visualizations (Figures 1–3)
   • Cell 8–9: Feature engineering and train/test split
   • Cell 10–11: XGBoost pipeline + threshold tuning
   • Cell 12: Feature importance analysis
   • Cell 13: Medical threshold calibration
   • Cell 14–16: Cross-validation and model comparison
   • Cell 17–18: Ensemble construction and final evaluation

---

Key Findings

• High Blood Pressure + High Cholesterol together raise diabetic rates dramatically — from ~6% (neither) to ~27% (both).
• BMI is the strongest continuous predictor; obese individuals (BMI ≥ 30) show nearly 3× the diabetic rate of normal-weight individuals.
• Age shows a clear monotonic relationship: diabetic rates rise from ~2% in the 18–24 group to ~24%+ in the 65–74 group.
• Income and Education are inversely correlated with diabetes — lower socioeconomic status is associated with significantly higher risk.
• General Health self-rating is a surprisingly strong predictor: respondents who rate health as "Poor" show ~40% diabetic rates vs. ~4% for "Excellent."

---

Technologies Used

---

License

This project is open-source and available under the MIT License.

---

Built as part of a Machine Learning mini-project exploring real-world health data and clinical model calibration.