Interstitial glucose prediction improvement through wearable-derived features

Abstract

In this work, we evaluated different sets of features derivedfrom wearable physiological signals to assess their feasibility in predictinginterstitial glucose (IG) values for a patient with Type 1 DiabetesMellitus (T1DM) over a 30-minute prediction horizon. These featureswere tested both independently and in combination with data from aContinuous Glucose Monitoring (CGM) sensor. In addition, we analyzedpredictions based solely on CGM data. For each prediction horizon, amachine learning regression model was trained using the correspondingset of characteristics with a 10-fold cross validation scheme. Model performancewas assessed using analytical metrics and the Clarke Error Gridwas used to evaluate clinical relevance. The best results were obtained forthe five-minute prediction horizon when combining wearable sensor featureswith CGM data, achieving a MAE of 5.18 mg/dL, a RMSE of 7.62mg/dL, and an R2 of 0.95, all predicted values falling within clinicallyacceptable zones. These findings highlight the potential of incorporatingwearable sensor data to improve glucose prediction accuracy.