Diabetes Risk Prediction using Machine Learning

Diabetes screening starts with numbers a basic health checkup already collects — glucose, blood pressure, BMI, age — yet those numbers are usually read one at a time against fixed thresholds. Machine learning can weigh them together: the Pima Indians Diabetes dataset (768 patient records, 8 clinical features) is the classic teaching ground for exactly this, small enough to train in seconds but rich enough to teach imputation, class imbalance, model comparison and recall-oriented evaluation. Students often jump straight to fitting models and skip the data-quality work — this dataset punishes that, because several columns contain medically impossible zeros that must be treated as missing. This project does the full pipeline honestly: cleaning, scaling, SMOTE balancing, stratified evaluation of four classifiers, and a web demo that turns the trained logistic-regression model into an interactive risk calculator with per-factor contribution bars.

1. The Pima Indians Diabetes dataset (768 records, 8 features + Outcome) is loaded; zeros in Glucose, BloodPressure, SkinThickness, Insulin and BMI are flagged as missing and median-imputed.
2. Features are standardized; the training fold is balanced with SMOTE because only ~35% of records are diabetic.
3. Logistic regression, random forest, XGBoost and SVM train on a stratified 80/20 split with 5-fold cross-validation for hyperparameter selection.
4. Models are compared on the held-out test set using accuracy, precision, recall, F1 and ROC-AUC; logistic regression is selected for the demo for its interpretability and strong recall.
5. The winning pipeline (imputer → scaler → classifier) is serialized for deployment.
6. In the web demo, entered measurements flow through the identical pipeline and the app renders the risk probability, risk band and per-factor contribution bars.

Tech stack:

• Python 3, scikit-learn
• XGBoost (comparison model)
• Pandas, NumPy (data handling)
• Matplotlib, Seaborn (EDA plots)
• imbalanced-learn SMOTE
• Jupyter Notebook (training)
• HTML/CSS/JavaScript (web demo)
• Pima Indians Diabetes dataset

Dataset & model details

• Dataset: Pima Indians Diabetes Database — collected by the National Institute of Diabetes and Digestive and Kidney Diseases, via UCI/Kaggle. 768 records (female patients ≥ 21 years, Pima Indian heritage); 8 features: Pregnancies, Glucose, BloodPressure, SkinThickness, Insulin, BMI, DiabetesPedigreeFunction, Age; binary Outcome (diabetic ≈ 34.9%).
• Task: Binary classification; input = 8 standardized clinical measurements, output = probability of diabetes.
• Model: Logistic regression (selected for the demo) after comparing random forest, XGBoost and SVM; pipeline = median imputation → StandardScaler → SMOTE (train fold) → classifier.
• Metrics: Test accuracy (design target ≈ 80%), precision/recall/F1, ROC-AUC (design target ≈ 0.86); recall on the diabetic class reported explicitly. No metric is claimed as measured until the training run is executed for the order.

• [Risk prediction web demo] Enter the 8 clinical measurements; the app returns a live diabetes-risk probability, a low/moderate/high band, and which factors drive the score.
• [Per-factor explanations] Each prediction breaks down feature contributions (coefficient × standardized value), so the demo teaches as well as predicts.
• [Full training notebook] Missing-value treatment, scaling, SMOTE, stratified train/test split, and model training in one reproducible Jupyter notebook.
• [Four-model comparison] Logistic regression, random forest, XGBoost and SVM evaluated on identical splits with accuracy, precision, recall, F1 and AUC.
• [Recall-oriented evaluation] The report explains why recall on the diabetic class matters more than raw accuracy for screening, and how the threshold choice reflects it.
• [Sample profiles] One-click high-risk and low-risk patient profiles for demoing the app and discussing feature effects in the viva.
• [Exported trained model] Serialized pipeline (imputer + scaler + classifier) so the demo scores with the real trained model.

• Complete training & evaluation Jupyter notebook (EDA, preprocessing, SMOTE, model comparison)
• Trained pipeline file (imputer + scaler + classifier)
• Interactive risk-prediction web demo with factor explanations
• EDA plots, ROC curves and per-model metric tables
• Project report PDF (background, medical context, methodology, results, ethics note)
• PPT presentation for final review
• Viva Q&A preparation document (logistic regression, SMOTE, recall vs precision, imputation)