Fun project analyzing the recent CGMacros Dataset. Contained here is:

• exploratory analysis/visualization of CGM data for participants (including markers for meal times, postprandial spikes, healthy/unhealthy ranges)
• simple -> complex modeling attempts to use given markers to predict severity of postprandial spike, including:
  • naive linear regression
  • xgboost regression (+ classification on discretized set)
  • MLP (trained on just meal macros)
  • MLP (trained on meal macros + biomarkers)
  • hybrid LSTM+MLP (trained on past 75 min of CGM data + meal macros + biomarkers)
• current best result: MAE < 29 mg/dL, comparable to existing CGMs' margin

🍩