Predicting health changes with a wearable tracker

  • Personal project in R, after consultation with a Harvard lab
  • Tested how well resting heart rate and heart-rate variability predict short-term changes in health
    • Conclusion: they predicted poorly
  • Extensive data preparation:
    • Imputation
    • Moving averages of baseline values
    • Percentage change from recent baseline
    • Difference from recent baseline
    • ROSE data-balancing
  • Assessed multiple models using cross-validation
    • Logistic regression (best)
    • SVM
    • Random forest
    • K nearest neighbors
    • XGBoost with Bayesian parameter estimation
    • K-means
    • Hierarchical clustering