Regarding “Detection of Hypoglycemia and Hyperglycemia Using Noninvasive Wearable Sensors: ECG and Accelerometry”

We read with great interest the recent article by Dave et al entitled “Detection of Hypoglycemia and Hyperglycemia Using Noninvasive Wearable Sensors: ECG and Accelerometry.” ¹ In this article, the authors discussed the potential for using machine learning models to analyze nine features of heart rate variability (HRV) measured by an electrocardiogram (ECG), and three domains of acceleration measured by an accelerometer. Based on the known physiologic responses to a low blood glucose level, this novel approach could be used to qualitatively and accurately detect the onset of clinically significant hypoglycemia. ² We commend Dave and colleagues for their work. However, we would like to point out two additional factors to consider in any discussion of technology for qualitative detection of hypoglycemia. These include the potential for (1) diabetes-related autonomic neuropathy to be a confounding factor to using ECG and accelerometer data, and (2) incorporation of multiple data streams to personalize qualitative hypoglycemia detection.

Cardiac autonomic neuropathy (CAN) is common in patients with diabetes although it is often asymptomatic. ³ Diabetes-related CAN is characterized by widespread neuronal degeneration of small nerve fibers of both sympathetic and parasympathetic tracts, which can impair heart rate responses to several physiological stimuli including hypoglycemia. A common effect is loss of HRV. ⁴ Therefore, individuals with diabetes-related CAN will not experience the same cardiac responses that would otherwise be reflected in changes to HRV metrics, despite the presence of hypoglycemia. This makes it difficult for the detection of hypoglycemia to be accurate based on ECG data alone. Furthermore, impairment of the autonomic nervous system is also associated with loss of other typical early warning symptoms of hypoglycemia, including adrenergic sweating, increasing the risk of more profound hypoglycemia.^5,6 As a corollary, the machine learning–based regression model characterized in Dave’s study already has the capability to recognize small changes in ECG metrics, and therefore the algorithm could potentially be adapted and trained to diagnose autonomic dysfunction. Given the impact of diabetes-associated CAN on the anticipated cardiovascular responses to a low blood glucose level, we would like to suggest that patients with autonomic dysfunction should be excluded from the training data set, and that the algorithm cannot be relied upon, in its current form, to make a diagnosis of hypoglycemia where autonomic dysfunction is present.

Finally, we believe that greater accuracy of hypoglycemia detection can be achieved through incorporating multiple data streams into an aggregate multiple sensor-data platform, ² rather than relying on information from a single sensor. Dave et al created their model by modifying ECG data with contextual information provided by an accelerometry sensor. ¹ However, incorporation of data from additional sensors for hypoglycemia detection, such as electroencephalogram (EEG) and volatile organic compounds (VOCs), could further improve detection and prediction accuracy. Therefore, adoption of a multi-data stream approach may be suitable for identifying hypoglycemia in patients with diabetes, with the caveat that the impact of preexisting autonomic dysfunction will need consideration.