Diabetes Risk Prediction (with Explainable AI)

A machine-learning project that predicts an individual's risk of diabetes from routine health and lifestyle indicators, then explains every prediction with SHAP. Built on the CDC's BRFSS survey of 253,680 US adults.

Why it matters: a screening model is only useful in healthcare if clinicians can trust and interpret it. This project goes beyond accuracy, it handles class imbalance honestly, checks probability calibration, and uses SHAP to show exactly which factors drive each person's risk.

---

📊 Results

Model	ROC-AUC	PR-AUC
Logistic Regression (baseline)	0.822	0.397
Random Forest (best)	0.827	0.421

At a recall-tuned decision threshold (0.61), the model flags ~62% of true diabetes cases, useful for a screening tool, where missing a case is costlier than a false alarm.

Diabetes prevalence in the data is ~14%, so the model is evaluated with PR-AUC (0.42 vs. a 0.14 random baseline), calibration, and a recall-focused threshold, not accuracy, which is misleading on imbalanced data.

What the model learned (SHAP)

The strongest drivers of predicted risk are general health, high blood pressure, BMI, age, and high cholesterol, consistent with the clinical literature, which is a good sign the model is learning real signal rather than noise.

---

🎯 What this project demonstrates

• End-to-end ML workflow, data → EDA → modelling → evaluation → explainability.
• Honest evaluation on imbalanced data, PR-AUC, calibration curve, and a deliberately chosen decision threshold (not the naive 0.5).
• Explainable AI (SHAP), global drivers and per-patient explanations.
• Responsible framing, clear statement of what the model can and can't be used for.
• Reproducibility, one script regenerates every figure and metric.

🗂️ Repository structure

diabetes-risk-prediction/
├── analysis.py                 # full pipeline: load -> EDA -> model -> SHAP
├── data/
│   └── diabetes_health_indicators.csv   # CDC BRFSS 2015 (253,680 rows, 21 features)
├── outputs/                    # all generated figures + metrics.json
└── README.md

📦 The data

CDC Diabetes Health Indicators (BRFSS 2015), 253,680 survey responses, 21 features (e.g. HighBP, BMI, GenHlth, Age, PhysActivity, Smoker), binary diabetes target. Source: UCI ML Repository #891.

🔁 Reproduce it

pip install pandas numpy scikit-learn matplotlib seaborn shap ucimlrepo
python analysis.py        # writes all figures + metrics.json to ./outputs

⚠️ Responsible-use note

This is a screening / educational model trained on self-reported survey data. It is not a diagnostic tool and must not be used for individual medical decisions. Predictions reflect statistical association, not causation, and survey data carries reporting and sampling biases.

🛠️ Tools

Python · pandas · scikit-learn · SHAP · matplotlib / seaborn

👤 Author

Kingsley Amegah, Health Data Scientist · GitHub: @Kingsley-amg