Noninvasive Blood Glucose Prediction from Photoplethysmogram Using Relevance Vector Machine

Diabetes mellitus is one of the predominant threats., hence monitoring the glucose levels is crucial. Clinically., it is measured using invasive blood tests. However., using photoplethysmogram (PPG) signals it is possible to find the blood glucose values., non-invasively. In this paper., PPG signals are used to predict BGL using relevance vector machine (RVM)., a supervised learning algorithm. Additionally., random forest (RF) is used to predict BGL for comparison. The BGLs obtained are validated using Cohen kappa statistics. RVM performs better with the kappa score of 0.955 with a computation time of 3.720 seconds; whereas., the kappa score of RF is 0.759 with computation time 1.94 minutes.