Exercise Prediction System

Predicting daily exercise habits using machine learning

Project summary

This exploratory health analytics project predicts whether a person exercises daily using self-reported lifestyle and vitals. It served as an early sandbox for building interpretable classification models that could later inform wellbeing recommendation systems.

Approach

  • Cleaned and imputed a Kaggle medical lifestyle dataset, engineering features such as BMI buckets, stress level encodings, and physician visit frequency.
  • Trained logistic regression, gradient boosting, and random forest models with stratified cross-validation while tackling severe class imbalance via SMOTE.
  • Generated SHAP explanations and counterfactual recourses so that end users could understand how incremental habit changes affect the prediction.

Outcomes

  • Achieved 89% accuracy and a Matthews correlation coefficient of 0.62 on the held-out set while maintaining calibrated probabilities.
  • Produced clinician-facing dashboards that highlight top contributing features and provide “next best action” suggestions.
  • Shared the pipeline with the UCSD Health recommendation team to guide features feeding into our production experimentation framework.

Tools

  • Python, pandas, scikit-learn, imbalanced-learn, SHAP, Plotly Dash.
  • Deployed via Docker for reproducible experimentation and collaboration with public health teammates.

View on GitHub