Diabetes Risk Prediction using XGBoost

A FLASK-based web application that predicts the likelihood of the user having diabetes or prediabetes based on the user's responses to a questionnaire. The app currently uses an XGBoost (Extreme Gradient Boosting) model trained on a Kaggle Dataset (https://www.kaggle.com/datasets/alexteboul/diabetes-health-indicators-dataset?select=diabetes_binary_health_indicators_BRFSS2015.csv) of 253,680 survey responses to the Center for Disease Control's(CDC) Behavioral Risk Factor Surveillance System (BRFSS2015) in 2015.

Refer to the Jupyter Notebook to see preprocessing, training, and MLFlow tracking:

In this repo at notebooks/diabetes_project.ipynb

Link: https://github.com/sagartv/diabetes_risk_predictor/blob/main/notebooks/diabetes_project.ipynb

The Dataset was balanced using SMOTEENN, following which an XGBoost Classifier was trained on it, yielding a validation accuracy of 94.3%.

BIG UPDATE:

Model tracking and evaluation using MLFlow:

Refer to The jupyter notebook in notebooks/diabetes_project.ipynb for the new MLFlow Tracking integration.

To access the MLFlow Runs, go to the notebooks folder and type the following in your terminal:

mlflow server --host 127.0.0.1 --port 8080

Then open http://127.0.0.1:8080/ in your browser to access the MLFlow UI and all the results of the training and evaluation of the classifier.

Feature Selection using KBest + Chi Square:

After evaluation through MLFlow, 5 Features removed from questionnaire and training data: Sex, Fruits, Veggies, AnyHealthcare, and CholCheck.

Docker Image on Hub

Docker Image Pushed to Hub: https://hub.docker.com/repository/docker/sagartv/diabetes_risk_predictor/

To get this latest image use: docker pull sagartv/diabetes_risk_predictor:0.0.5.RELEASE

To run, expand optional settings and provide a Port Number for your system, this will map to the image's port 3000.

App Live on Render!

Docker Image RELEASE 0.0.5 is now Deployed and Live on Render. Access at https://diabetes-risk-predictor.onrender.com/

TODO:

Explore CI/CD Pipelines