🩺 Explainable AI for Diabetes Risk Stratification

Multi-class diabetes prediction framework combining supervised learning (Logistic Regression, XGBoost, MLP, LSTM), unsupervised clustering (K-Means), and Explainable AI (SHAP).

📊 Key Features

• 3-class risk stratification: No Diabetes, Prediabetes, Diabetes
• 4 ML models compared: Logistic Regression, XGBoost, MLP, LSTM
• K-Means patient profiling: Identified high-risk cluster (Cluster 3)
• SHAP explainability: GenHlth, HighBP, Age, BMI top predictors
• Interactive UI: Real-time risk prediction with ipywidgets

🎯 Results

Model	F1-Score	AUC-ROC
Logistic Regression	0.47	0.82
XGBoost	0.47	0.82
MLP	0.46	0.82
LSTM	0.47	0.82

🚀 Quick Start

1. Clone repository

git clone https://github.com/BhanuTejaMalineni/Diabetes-Risk-Stratification-ML.git
cd Diabetes-Risk-Stratification-ML

2. Install dependencies

pip install -r requirements.txt

Required packages:

• pandas, numpy, scikit-learn
• xgboost, tensorflow
• shap (for explainability)
• matplotlib, seaborn
• jupyter, ipywidgets

3. Download dataset

Get Diabetes Health Indicators (BRFSS 2015) from Kaggle
→ Save CSV to ./data/ folder

4. Run notebooks

jupyter lab
# Open notebooks in Code Snippet/ folder

📁 Project Structure

├── Code Snippet/         # Jupyter notebooks (analysis & models)
├── Machine_Learning_Project.pdf  # Full research paper
├── requirements.txt      # Python dependencies
├── .gitignore
└── README.md

🔬 Methodology

Unsupervised Learning

• K-Means Clustering (K=4) to identify patient profiles
• High-risk Cluster 3: High BP, BMI>40, poor GenHlth, Age>60

Supervised Learning

• Binary Classification: Diabetes vs. No Diabetes
• Multi-class: No Diabetes (0), Prediabetes (1), Diabetes (2)
• Models: Logistic Regression, XGBoost, MLP, LSTM

Explainability

• SHAP values for feature importance
• Top predictors: GenHlth, HighBP, Age, BMI

📈 Key Findings

• Logistic Regression achieved best F1-Score (0.47) with balanced precision/recall
• All models reached ~0.82 AUC-ROC, indicating strong discrimination
• Unsupervised clustering successfully identified clinically relevant high-risk profiles
• SHAP analysis confirms predictions align with established medical risk factors

📄 Full Report

Detailed methodology, results, and analysis: Machine_Learning_Project.pdf

🛠️ Tech Stack

📚 Dataset

BRFSS 2015 - Behavioral Risk Factor Surveillance System (CDC)

• 253,680 survey responses
• 21 health-related features
• Class imbalance: 84% No Diabetes, 16% Prediabetes/Diabetes

🎓 Academic Context

MSc Software Engineering Project
University of Europe for Applied Sciences
Potsdam, Germany

📧 Contact

Bhanu Teja Malineni
LinkedIn | GitHub

---

⭐ If you find this project helpful, please star the repository!