Diabetes Health Analytics Dashboard

Demo

output.mp4

An end-to-end, production-ready Streamlit application for analyzing diabetes health indicators and predicting diagnosed_diabetes using a trained XGBoost pipeline. The app provides:

• A polished, multi-page dashboard (Overview, Insights, Prediction)
• Clear visual analytics answering key questions about prevalence and risk factors
• An interactive prediction form that returns class and probability, with tailored guidance

Features

• Overview
  • Dataset summary (rows, columns, diabetes prevalence, average age)
  • Schema, dtypes, and numerical statistics
  • Feature importance (from XGBoost pipeline) with top features visualization
• Insights
  • Prevalence of diabetes (bar and pie charts)
  • Diabetes vs gender and smoking status
  • BMI distribution with clinical thresholds and category risk view
  • Demographic patterns (ethnicity, education, income, employment)
  • Clinical markers (HbA1c, fasting glucose) by diabetes status
  • Correlation heatmap of key biomarkers
• Prediction
  • Loads saved pipeline models/xgb.pkl
  • Collects 25+ inputs (demographics, lifestyle, vitals, labs)
  • Predicts class and probability; shows a gauge and tailored advice
  • Friendly, supportive UX and success/education messages

Quick Start

1. Clone and prepare environment

• Create a virtual environment and install dependencies:

python -m venv venv
# Windows: venv\Scripts\activate
# macOS/Linux:
source venv/bin/activate

pip install -r requirements.txt

2. Add required files

• Place the trained pipeline at:

models/xgb.pkl

• Place the dataset CSV at:

data/diabetes_dataset.csv

3. Run

• Launch the app:

streamlit run app.py

• Use the sidebar to navigate between Overview, Insights, and Prediction pages.

Training and Saving the Model

• Train an XGBoost-based pipeline with preprocessing and SMOTE inside a cross-validated pipeline.
• Save it as follows:

import joblib
joblib.dump(xgb_pipeline, "models/xgb.pkl")

• Ensure your pipeline’s ColumnTransformer expects the exact feature names used in the Prediction page input (see 3_Prediction.py build_input function). Align names 1-to-1 with your training schema.

Input Schema (Prediction Page)

The prediction form collects these features (adjust according to your training schema):

• Demographics & lifestyle: age, gender, ethnicity, education_level, income_level, employment_status, smoking_status, alcohol_consumption_per_week, physical_activity_minutes_per_week, diet_score, sleep_hours_per_day, screen_time_hours_per_day
• Medical history: family_history_diabetes, hypertension_history, cardiovascular_history
• Anthropometrics & clinical: bmi, waist_to_hip_ratio, systolic_bp, diastolic_bp, heart_rate
• Lipids: cholesterol_total, hdl_cholesterol, ldl_cholesterol, triglycerides
• Glycemic markers: glucose_fasting, glucose_postprandial, insulin_level, hba1c

Important:

• Exclude leakage features if predicting diagnosed_diabetes (e.g., diabetes_stage, diabetes_risk_score), unless you designed a different target.

Encoding & Preprocessing Recommendations

• Scale numerical features with StandardScaler
• One-hot encode nominal categories (gender, ethnicity, employment_status)
• Use OrdinalEncoder for genuine ordinal categories (education_level, income_level, smoking_status) if trained that way, otherwise one-hot is safe
• Put all preprocessing inside the Pipeline to prevent leakage
• Apply SMOTE inside the pipeline, after preprocessing and within CV folds

Model Choices

From benchmarking:

• Top performers: LightGBM, XGBoost, RandomForest (Test F1 around 0.93 in strong setups)
• XGBoost chosen for deployment due to the balance of accuracy, stability, and feature importance interpretability
• Start hyperparameter tuning with:
  • n_estimators, max_depth, learning_rate, subsample, colsample_bytree, min_child_weight, reg_lambda, reg_alpha, tree_method='hist'

Notes on Data Leakage

• Remove target-like columns (diabetes_stage) and risk composites (diabetes_risk_score) when predicting diagnosed_diabetes
• Ensure SMOTE and all preprocessing steps are inside the CV pipeline
• Use StratifiedKFold with shuffling for robust estimates
• Validate with DummyClassifier to sanity-check performance lift

Requirements

See requirements.txt:

• streamlit, pandas, numpy, plotly, scikit-learn, imblearn, xgboost, joblib

Customization

• Adjust Overview to show more metadata or profiling
• Extend Insights with additional clinical charts (e.g., LDL/HDL ratios, BP categories)
• Add download buttons for filtered data or charts
• Integrate model explainability (SHAP) if desired
• Internationalize labels and messages for different locales

Disclaimers

• This tool is for educational and decision-support purposes; it does not replace professional medical diagnosis or treatment.
• Always consult healthcare providers for clinical decisions and personalized care plans.

Acknowledgments

• Built with Streamlit and Plotly for fast, interactive analytics
• XGBoost pipeline for robust, high-performing tabular classification
• Inspired by standard diabetes risk indicators and clinical thresholds