BGEM PPG Explorer

All trials on one page · denoised & normalized PPG

Loading...
User Name Date & Time Actual (mg/dL) Ring Mid (mg/dL) Our Predicted (mg/dL) Ring Range Our Range Ring In Range Ours In Range Ring Error Our Error Winner
Per-user averages One row per user. Ranges shown as average lower–upper (mg/dL). Actual uses LAB → CGM → GLUCO priority. Within ±10% counts samples where our predicted point is within 10% of actual. Within ±10 mg/dL counts samples where |predicted − actual| ≤ 10.
User Name Samples Within ±10% Within ±10 mg/dL Avg Actual (mg/dL) Avg Our Predicted (mg/dL) Avg Our Range (mg/dL) Avg Ring Range (mg/dL)
Model components vs actual glucose Each column is a feature used by the model. Sort any column to see how values rise or fall with actual sugar. Headers stay visible while scrolling.
Entry-wise train / test split Model is trained on ~80% of labeled recordings. For users with multiple entries, some go to train and some to test (same user can appear in both sets).

Training set samples

User Actual (mg/dL) Our Predicted Our Error Ring Error Winner

Test set samples

User Actual (mg/dL) Our Predicted Our Error Ring Error Winner
User Date & Time Actual (mg/dL) Ring Mid Ridge Pred XGBoost Pred Ring Err Ridge Err XGB Err Closest
Old range vs confidence-based adaptive range Old = symmetric band from median LOO absolute error. Adaptive = per-sample confidence interval at 80%: bootstrap ensemble gives a conservative prediction; range = P100 ± |P100 − P80|. Both methods are clipped to ±10…±20 mg/dL of the predicted value. Existing tabs still use the old range.
User Date & Time Actual Predicted Old Range Adaptive Range Old Covers Adaptive Covers Old Width Adaptive Width Narrower