Non-intrusive Blood Glucose Monitoring with Multi-task Deep Learning
Inferring abnormal glucose events such as hyperglycemia and hypoglycemia is crucial for the health of both diabetic patients and non diabetic people. However, continuous or regular blood glucose monitoring can be invasive and inconvenient in everyday life. We present SugarMate, a non-intrusive blood glucose detection system solely based on smartphones. In addition to the records of food, drug and insulin intake, it leverages smartphone sensors to measure physical activities and sleep quality automatically. Provided with the imbalanced, personalized, and often limited measurements, the most challenging part of SugarMate is the inference of blood glucose level at a fine-grained time resolution. To this end, we propose Md3RNN (multi-division deep dynamic recurrent neural network), a learning paradigm that is able to make full use of the available blood glucose information. Specifically, Md3RNN (1) captures complex, multi-scale blood glucose dynamics via deep neural networks, (2) extracts grouped feature representations with a multi-division learning structure, and (3) preserves user-specific characteristics using personalized output layers. Evaluations on 112 users show that Md3RNN yields an average accuracy of 82.14%, significantly outperforming previous learning methods that are either shallow, generically structured, or oblivious to grouped behaviors.
Speaker: Weixi (Gavin) Gu
Department of Civil Engineering & Department of Electrical Engineering and Computer Science, UC Berkeley
Weixi Gu is a Ph.D. candidate in the Tsinghua-Berkeley Shenzhen Institute (TBSI) of Tsinghua University (THU). His advisors are Prof. Lin Zhang, Prof. Khalid Mosalam and Prof. Costas Spanos. His principal research interests lie in the areas of designing theoretical machine learning and statistical methodologies and developing reproducible large-scale mobility solutions to tackle complex applications in the context of high-dimensional, multimodal, and dynamic data. His current research work involves, 1) foundations of statistical learning, including information theory and algorithm on the automatic temporal and spatial feature engineering; 2) deep learning frameworks on high-dimensional mobility data modelling, to uncover inherent dynamic associations among a set of random variables; 3) large-scale applications of statistical learning in health sensing, transportation, and human dynamics. Weixi Gu has received several awards including the Best Paper Runner-Up Award in Mobiquitous 2016, the Best Paper Award in Trustcom 2014, and the National Scholarship, China at 2014, 2016,2017.