Background: Hypoglycaemia is a major barrier to achieving target glycaemia for people with diabetes. Microtremor involves involuntary resting rhythmic movements of an extremity, which our preliminary data suggests increases with hypoglycaemia.
Aims: To determine relationships between upper limb microtremor and nocturnal hypoglycaemia in adults with type 1 diabetes, and explore feasibility of a non-invasive device incorporating an algorithm utilising microtremor to detect nocturnal hypoglycaemia.
Methods: Participants wore the investigational device and continuous glucose monitoring (CGM) concurrently (masked) for 14 nights, calibrated via capillary blood glucose. The device recorded accelerometer and gyroscope measurements from wrist sensors, and accelerometer measurements from a finger sensor. Device data were amplified and transmitted to an artificial neural network (ANN) which determined hypoglycaemia likelihood. Nights were ‘assessable’ if the device was worn and CGM functioned. Data from a randomly-selected 75% of nights were used to ‘train’ the ANN (i.e. teach ANN to recognise hypoglycaemia from microtremor); remaining ‘novel’ nights were used to evaluate ANN performance. For the final 30 participants, the algorithm was exposed to device data with and without training.
Results: Eighty adults participated. Training phase analysis indicated microtremor amplitude at 19Hz and 5Hz frequencies was fundamental to hypoglycaemia recognition. For all assessable nights (74% of nights), hypoglycaemia detection sensitivity was 39% and 69% at CGM nadirs ≤4.0mmol/L and ≤2.2mmol/L, respectively; specificity 90%. Sensitivity during novel nights was 21% and 41% at nadirs ≤4.0mmol/L and ≤2.2mmol/L, respectively; specificity 78%. Sensitivity for novel nights at nadir ≤2.2mmol/L was 41% for non-novel and 25% for novel participants.
Conclusions: Changes in microtremor amplitude related to nocturnal hypoglycaemia severity. Substantial inter-individual microtremor variation suggests ANN training is important for hypoglycaemia recognition. Device performance is not currently adequate for utility as a stand-alone hypoglycaemia detection device; refinement is underway. This technology may provide an additional signal for nocturnal hypoglycaemia detection.