Forecasting with Sparse but Informative Variables: A Case Study in Predicting Blood Glucose

In time-series forecasting, future target values may be affected by both intrinsic and extrinsic effects. When forecasting blood glucose, for example, effects can inferred from the history of signal alone (i.e. glucose), but accurately modeling impact requires auxiliary signals, like amount carbohydrates ingested. Standard techniques often assume that vary at similar rates. However, when signals are generated a much lower frequency than variable (e.g., glucose measurements made every 5 minutes, while meals occur once few hours), even well-known increase glucose) prove difficult to learn. To better utilize these sparse informative variables (SIVs), we introduce novel encoder/decoder approach learns per-timepoint effect SIV, (i) isolating it (ii) restricting its learned based on domain knowledge. On simulated dataset pertaining task SIV is recorded our outperforms baseline approaches in terms rMSE (13.07 [95% CI: 11.77,14.16] vs. 14.14 [12.69,15.27]). presence corrupted proposed still result error compared advantage reduced as noise increases. By their incorporating knowledge, makes possible SIVs forecasting.