Machine Learning Experiments with Noninvasive Sensors for Hypoglycemia Detection

Accurate hypoglycemia detection would enable people with type 1 diabetes (T1D) to treat this dangerous condition promptly, improving health and safety. This paper presents machine learning experiments that aim to improve hypoglycemia detection by leveraging data from noninvasive sensors found in fitness bands. A middle-aged subject with T1D provided blood glucose and fitness band data for two months. Sensor data included heart rate, galvanic skin response, and skin and air temperatures. Statistical tests identified features derived from this data that could differentiate hypoglycemic from non-hypoglycemic states. Support vector machines (SVM) were then trained, using only these features, to classify instances as hypoglycemic or non-hypoglycemic. An SVM with a linear kernel was able to outperform two simple baselines. Results show proof of concept; however, system performance was limited by the size and nature of the dataset. Results are being used in ongoing work to improve the performance of overall blood glucose prediction models that use blood glucose, insulin, and life-event data.