Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches

Fig. 1. Graphical abstract of study. This study has four objectives: (1) Engineer features from non-invasive systems using a combined data-driven and domain-driven feature engineering approach. (2) Develop personalized glucose excursions definitions. (3) Classify glucose excursions using engineered features. (4) Build predictive models of glucose using both a population approach with leave-one-person-out cross validation and a personalized approach. Sensor placement for the study (Empatica E4 on the wrist and a Dexcom G6 continuous glucose monitor on the abdomen) is also shown.

Fig. 2. Case studies of CGM data over 24 h for 3 participants highlighting eating habits and subsequent glucose excursions. Green bands indicate a meal or snack, pink bands indicate consumption of a sugary beverage or soda, and orange bands indicate a sugary snack or dessert. Shown in red is the traditional hyperglycemic range (>180 mg/dL) and shown in yellow is the traditional hypoglycemic range (<70 mg/dL). Note that the participants do not exceed the traditional definitions of glucose excursions. A system that would alert prediabetes patients based on traditional definitions of glucose excursions would be inadequate in these cases. The personalized classifications we propose in this paper would better inform prediabetes patients so they can begin to self-manage their diet, exercise, and stress levels based upon this information. As shown, the personalized classes PersHigh (shown in plum) and PersLow (shown in teal) would provide more personalized information about fluctuating glucose. For example, sugary drinks and snacks result in higher glucose fluctuations. By reducing the amount and frequency of sugary drinks and snacks, participants would be less likely to experience a glucose…

Fig. 3. Personalized glucose excursions. Glucose excursions are classified on a personalized, rolling basis examining the previous 24 h of each participant’s historical data. a Boxplots of PersNorm, PersHigh, and PersLow for each participant in the dataset. b Histogram of all distributions over all participants. c PersNorm distribution. d PersHigh distribution. e PerLow distribution. Bin width for each histogram shown is 1 mg/dL.

Fig. 4. Confusion matrix: multiclass model. Confusion Matrix for multiclass decision tree model validated with a 70/30 train/test split. In this model, the class-accuracy for PersHigh glucose is 82.6%, the class accuracy for PersNorm is 81.3%, and the class accuracy of PersLow is 82.1%.

Fig. 5. Importance of 69 domain-driven and data-driven features in glucose predictions. Importance was determined from a random forest feature selection model using impurity-based features. The outer circle shows relative percent of importance for the categories food, circadian rhythm, activity, stress, gender, clinical metrics, personalization, electrodermal activity, heart rate, and skin temperature. The inner circle shows relative importance by source of features, including food logs, the wrist-worn wearable, user-defined features, and features defined for/by the model.