Blood Sugar Tracker

Full-Stack Clinical Architecture

The application follows a traditional MVC architecture using Flask as the web framework, SQLAlchemy as the ORM, and a pre-trained scikit-learn ElasticNet model served via joblib.

[Patient / Staff / Admin Browser]
   │
   ▼
[Flask Routes (app.py)] ── CSRF Protected WTForms
   │
   ├── [SQLAlchemy ORM] ── SQLite / PostgreSQL DB
   │       └── User, Patient, DailyLog models
   │
   └── [ML Prediction Pipeline]
           ├── ElasticNet model (.pkl)
           ├── Feature Scaler (.pkl)
           └── Feature Names (JSON)

• ML Pipeline: Patient vitals (weight, carbs, protein, fat, activity calories, heart rate, time-of-day) are scaled and passed to a trained ElasticNet regression model to predict blood glucose in mg/dL.
• Auth Layer: Flask-Login handles session management with role separation (admin vs. staff). All form submissions are CSRF protected via Flask-WTF.

Technical Details & Implementation

• ElasticNet Model Training: Trained on clinical patient vitals dataset using scikit-learn's ElasticNet regressor with cross-validated hyperparameter tuning (L1/L2 regularization ratio). Serialised with joblib for production serving.
• Daily Log System: Staff log patient vitals once per day (enforced by unique DB constraint on user_id + log_date). Net calories are computed server-side from macronutrient inputs.
• Alert System: Patients with predicted glucose readings above clinical thresholds are automatically flagged with `alert_flag=True` and highlighted in the dashboard.
• Report Export: PDF reports are generated per patient with their prediction history and health trends. CSV exports are available for bulk data analysis.

Project Outcomes & Results

• ML Accuracy: ElasticNet model achieved strong predictive performance on held-out validation data with low RMSE on blood glucose prediction.
• Clinical Utility: The alert system automatically identifies high-risk patients, enabling faster clinical intervention.
• Full Production Stack: Role-based auth, CSRF security, database migrations, PDF export, and a clean dashboard UI — all functional in a single deployable Python app.

Challenges & Key Decisions

• Challenge: Ensuring the ML model returns clinically meaningful predictions from sparse or incomplete patient data inputs.
• Solution: Applied feature scaling with a pre-fitted StandardScaler and used ElasticNet's regularization to prevent overfitting on the limited clinical dataset. Added server-side validation to reject incomplete vitals submissions before predictions are made.

Contribution Details

• 100% Solo Development: Designed and built the entire system — ML model training pipeline, Flask backend, Jinja2 frontend templates, SQLAlchemy database schema, Flask-Migrate migrations, role-based authentication, and report generation.