The table below shows the benchmarking results for blood glucose forecasting methods across various datasets and configurations. Scroll down to view plots comparing the methods.

Benchmark Results

Download CSV Download XLSX

Legend

• Algorithm: the algorithm used to forecast blood glucose levels
• Dataset: name of the dataset used for the evaluation (and training, if required)
• Horizon: the prediction horizon (observation length in number of samples + forecasted samples)
• Aug.: the data augmentation strategy used, if any

• Time features: list of time features used

  • %w

    Day of the week as a decimal number, where 0=Sunday and 6=Saturday

  • %d

    Day of the month as a decimal number

  • %m

    Month of the year as a decimal number

  • %Y

    Year with century as a decimal number

  • %H

    Hour of the day (24-hour clock) as a decimal number

  • %M

    Minute of the hour as a decimal number

  • %MoD

    Minute of day as a decimal number, range [0, 1440]

  • %S

    Second of the hour as a decimal number

  • %SoD

    Second of day as a decimal number

  • %j

    Day of the year as a decimal number

  • %U

    Week number of the year (Sunday as the first day of the week) as a decimal number. All days in a new year preceding the first Sunday are considered to be in week 0

• Norm.: data normalization strategy used
• Loss: loss function used during training
• RMSE: Root Mean Squared Error
• MAE: Mean Average Error
• MARD: Mean Absolute Relative Difference
• TG: Time Gain
• Sen Hypo: Sensitivity to Hypoglycemic events
• Sen Hyper: Sensitivity to Hyperglycemic events
• Spec Hypo: Specificity to Hypoglycemic events
• Spec Hyper: Specificity to Hyperglycemic events
• PCC: Pearson Correlation Coefficient
• PDE: Probability of predictions to end in the D or E zones of the Clarke’s Error Grid
• τ-lag: Tau lag

---

Plots

---

Error Grids and Confusion Matrix Comparison

Select two rows in the table to view and compare the Confusion Matrix, Clarke, and Parkes Error Grids.