PARAMETRIC AND NON-PARAMETRIC REGRESSION APPROACHES FOR NON-INVASIVE BLOOD GLUCOSE MONITORING

Abstract

Blood glucose monitoring systems (BGMSs) play a crucial role in health care applications. Invasive measurements are more accurate while non-invasive BGMS encourage self monitoring and reduce the cost of health care. Though multiple sensor data acquisition and suitable processing improve accuracy, self-monitoring becomes difficult in such non-invasive systems due to multiple signal acquisition. This paper investigates a non-invasive BGMS prototype that renders accurate measurements by statistically processing a single sensor data. The developed prototype is based on near-infrared (NIR) spectroscopy, which provides an electronic voltage that gets mapped to corresponding blood glucose level. This mapping is proposed using two different statistical regression approaches, parametric Bayesian Regression (BR) approach and the non-parametric Gaussian Process Regression (GPR) approach. Dataset is acquired from 33 subjects who visited Ramaiah Medical College Hospital, India. On each subject, voltage from the BGMS prototype and corresponding invasively obtained blood glucose level have been recorded. The BR and GPR approaches are trained with 75% of the data while the remaining 25% is used for testing. Test results show that BR approach renders root mean square error (RMSE) of 3.7[Formula: see text]mg/dL, while the mean absolute percentage error (MAPE) is around 2.5. The GPR with different radial basis function kernels revealed that a multiquadric kernel provides a lowest RMSE of 3.28[Formula: see text]mg/dL and lowest MAPE of 2.2, thus outperforming the parametric BR approach. Investigations also show that for a training data of less than 15 entries, BR renders better accuracy than the GPR approach.