Smartwatch Gesture-Based Meal Reminders Improve Glycemic Control

Background

Late and missed meal boluses are a significant problem in diabetes management. The rate of late/missed meal boluses is high for adults and adolescents, and missed meal boluses correlate with higher HbA1c^1–3. 38% of people with T1D report that they forget to bolus for a meal once a week or more⁴. We hypothesize a smartwatch application capable of detecting eating behavior and alerting the user will improve measures of glycemia by reducing the rate of late/missed meal boluses.

Methods

This was a pilot, randomized, crossover, open-label study to test the efficacy of the Klue^™ Health application^* (Klue) at two centers. The study was approved by the Stanford and University of Virginia Human Subjects Research Compliance Office, registered on ClinicalTrials.gov (NCT03970889/NCT03809858) and conducted in compliance with the standard of Good Clinical Practice and Declaration of Helsinki.

Klue detects eating events and delivers reminders intended to reduce missed/late meal boluses. The algorithm uses proprietary artificial intelligence trained to detect eating behaviors based on sensor data provided by the watch. Klue’s accuracy was previously tested using video recordings, and by subjects responding to alerts when gestures were detected. With 130 subjects there were 76,261 consumption events, and 544,084 gestures with recorded responses including sips, bites, eating, drinking, brushing teeth, combing hair, putting on earrings, lipstick or mascara, coughing, blowing their nose, talking on the phone, etc. Meals were detected 96% of the time, and drinking 94% of the time, and eating could be distinguished from drinking 98% of the time. The false positive rate for meal events was 1–2 events per week. In this study Klue was specifically used to detect the motion of eating and not drinking, so beverages with carbohydrates could have been consumed and not detected as eating. If Klue detects three motions that indicate the wearer is eating, a message appears on the smartwatch that asks, “Eating?,” to which the user could respond, “Yes, bolused,” “Yes, will bolus,” “Not eating,” “Hypo or no carbs,” or “Dismiss.” If there is no response to the prompt, the message reappears every five minutes while eating is detected.

Individuals age 13 and older with T1D were recruited and enrolled if they used a continuous glucose monitor (CGM) and an insulin pump or pen with downloadable memory. Participants with ≥four missed/late boluses in the previous two weeks were randomized to either usual care or use of Klue for the first six weeks. A missed/late bolus was determined by a significant rise in CGM values without a preceding insulin bolus. During Klue use, participants were asked to wear the Apple Watch on their dominant hand while awake. After six weeks, the groups crossed over and were on the other treatment for six weeks. Insulin pumps, connected pens, and CGM devices were downloaded, and HbA1c was measured at enrollment and after each six-week treatment period.

A comparison of the change in HbA1c during the treatments was conducted and evaluated using a paired t-test to assess the impact of Klue on glycemia. CGM readings from the Klue and usual care portions of the study were analyzed per established consensus outcomes⁵. A mixed-effects analysis was performed to analyze CGM outcomes for both treatment groups while minimizing the impact of missing values. Participants were included in the analysis if they had over half of available CGM data. The study was not statistically powered.

The accuracy of Klue meal detection was evaluated: eating detections were confirmed if there was a mealtime bolus or rise in CGM values >2mg/dL/min over 20 minutes in the prandial window between 30 minutes before (premeal bolus) and two hours after (late meal boluses) eating detections. Klue detections within 45 minutes of each other were considered one eating event. Detections with <75% CGM data availability in the next two hours were not considered in the results. Glucose rise detections within 15 minutes were considered one event for the results. Additionally, the number of days where Klue was worn (10 out of 24 hours) and the difference in timing between Klue and glucose rate of change detections were analyzed.

Results

34 individuals participated in this 12-week at-home study between October 2018-October 2019 at Stanford or the University of Virginia. 18 were female, and the average age was 23±14 years. 28 (82%) identified as White, 2 (6%) Asian, 1 (3%) Hispanic, and 3 (9%) of mixed race. Average height and weight were 69.9±16.0 kg and 168.7±10.7 cm. The average HbA1c at baseline was 7.9±1.0%.

A subset of 30 participants was used to analyze the change in HbA1c. Four participants were excluded because they dropped out or had missing HbA1c measurements. During the Klue portion of this study, these 30 participants had an HbA1c decrease of 0.3±0.4%, however in usual care they experienced an increase in HbA1c of 0.1±0.3%. This difference in the change in HbA1c during the treatments was 0.4% (p<0.001).

26 participants had sufficient CGM data during Klue use and 21 during usual care. Time-in-range (TIR) was significantly higher during the Klue portion of the study versus usual care (57.6±12.2% v. 54.2±14.7%, p=0.03). While using Klue, there was 3.8% less time when participants’ CGM values were >180mg/dL (p=0.03) and a 3.1% less time when CGM values were >250mg/dL (p=0.03). The amount of time when CGM was >300mg/dL was not significantly different between the two treatments. Furthermore, mean CGM values were lower while the participants used Klue (174±22mg/dL v. 181±30mg/dL, p=0.02). There was no difference in the amount of hypoglycemia the participants experienced as well as the standard deviation or coefficient of variation of the CGM values.

Detection accuracy was based on 2,987 eating events from Klue. 72% of detections were either associated with a meal bolus and/or a 2mg/dL/min increase in glucose in the two hours following the detection. 32% of the meals were detected by both a bolus and a glucose rise; 20% were detected by a bolus, and 48% had only a glucose rise. The instances where there was a glucose rise with no associated bolus that were detected by Klue were likely missed boluses. When a Klue detection preceded a rise in glucose, the rise in glucose occurred 37.8±32.7 minutes later. During the study’s Klue portion participants had the application active for more than 10 hours during 49.4±24.9% of the days.

An analysis was conducted that divided the participants with adequate CGM data during usual care into quartiles based on the number of glucose rises without preceding boluses (indicating missed boluses). Although this subsample was too small to provide statistical significance, the results may indicate how Klue may impact different subpopulations (Figure 1). Those with the highest number of glucose increases without boluses (Q4) showed the greatest improvement in TIR (+6%) and the greatest increase in number of boluses per day (+1.4).

Figure 1:

(A) A comparison of time-in-range (70–180mg/dL) for the Klue and usual care arms of the study.

(B) Time-in-range by baseline quartile of the number of glucose rises without a bolus.

(C) Boluses per day by baseline quartile of the number of glucose rises without a bolus. Usual care (dark gray) and Klue (light gray).

31 participants completed a survey regarding the use of Klue and rated the application on a scale of 1–10. The average rating was 7.7 with 12 (39%) giving it a perfect rating. 8 (26%) rated the application ≤6 and reported accuracy concerns (19%), dislike wearing the Apple Watch on their dominant hand (16%), general technical issues (13%), desire for additional features (10%), and notifications not being useful bolus reminders (3%).

Conclusions

The use of bolus reminders generated by Klue on the Apple Watch resulted in a statistically significant reduction in HbA1c in adults and adolescents with T1D and a history of missed/late boluses. The majority of Klue detections appeared confirmed with either a delivered bolus by the user and/or an increase in glucose. The remaining Klue detections may be the result of eating foods with low-carbohydrate intake (as seen with 20% of the meal boluses not having a glucose increase of >2mg/dL/min), treatment of hypoglycemia, or system false positives. Alarm fatigue is a significant consideration in the design of diabetes technology considering the proliferation of new devices and applications⁶. Therefore, it is reassuring that only 1 participant reported the notifications as not useful.

Strengths of the study include real-world data from two centers and a focus on participants of various ages likely to benefit from this technology. There are several limitations, including a limited sample size, differences in CGM use and lack of a gold standard for true meal detection. Additionally, if the user ate with their non-dominant hand, the meal would not be detected. Because there was no reliable record of when the participants ate, we used a simple meal detection algorithm based on glucose rise as a surrogate. False negatives could be caused by low carbohydrate meals, or well-timed and estimated insulin boluses before meals. False positives could be the result of counter-regulatory response from anaerobic exercise or stress. Neither HbA1c nor TIR in this study fully reflects glycemic control during the six weeks use. Indeed, results may have been even more favorable were all CGM data available or therapy used for a full three months for HbA1c assessment.

The technology holds promise as a bolus reminder and for meal detection in a variety of situations. Future integration with automated insulin dosing systems may improve fully closed-loop insulin delivery by increasing controller aggressiveness upon meal detection by Klue and subsequent confirmation by a positive CGM rate of change.