Hand tremor-based hypoglycemia detection and prediction in adolescents with type 1 diabetes

Diabetes is a chronic disease where blood glucose (BG) concentrations are consistently high. A dangerous diabetes-associated condition is hypoglycemia, where BG level drops below the normal threshold. Hypoglycemia is often accommodated by tremors, sweating, fatigue, anxiety, lightheadedness, disorientation, irritability, and tachycardia. Very few studies are focused on detecting enhanced tremor as a peripheral physiological response to declining BG concentrations. This study undertakes a machine-learning approach to predict hypoglycemia using hand tremors. Tremors were detected and characterized by frequency and amplitude in non-hypoglycemic and hypoglycemic conditions. Accelerometers and continuous glucose monitoring devices recorded the tremor and BG datasets. A home study of 32 T1D adolescents in a free-living condition was conducted. Simultaneously recorded accelerometer and continuous glucose monitor (CGM) data of 194.6 h were collected from 15 participants. These data were utilized for training and testing the predictive model. Four lengths of the sliding window (1, 2, 5, and 10 s) and four machine learning algorithms (decision tree, support vector machine, k-nearest neighbor, and bagged trees ensemble classifier) were applied to classify tremor as non-hypoglycemic or hypoglycemic. The greatest accuracy (86.65%) was achieved by the ensemble classifier (bagged trees based on the subspace k-NN) for 1 s window length. Multiple prediction windows are merged to generate aggregate sequential predictions. Prediction accuracy of 86.05% was achieved for a 15 s batch (3 s window length). This study demonstrates the feasibility of detecting and predicting the onset of hypoglycemia based on hand tremor data collected by wrist-worn accelerometer sensors.