Support vector machines as multivariate calibration model for prediction of blood glucose concentration using a new non-invasive optical method named Pulse Glucometry.

A novel optical non-invasive in vivo blood glucose concentration (BGL) measurement technique, named "Pulse Glucometry", was combined with a kernel method; support vector machines. The total transmitted radiation intensity (I lambda) and the cardiac-related pulsatile changes superimposed on I lambda in human adult fingertips were measured over the wavelength range from 900 to 1700 nm using a very fast spectrophotometer, obtaining a differential optical density (delta OD lambda) related to the blood component in the finger tissues. Subsequently, a calibration model using paired data of a family of delta OD lambda s and the corresponding known BGLs was constructed with support vector machines regression instead of using calibration by a conventional partial least squares regression (PLS). Our results show that the calibration model based on the support vector machines can provide a good regression for the 183 paired data, in which the BGLs ranged from 89.0-219 mg/dl (4.94-12.2 mmol/l). The resultant regression was evaluated by the Clarke error grid analysis and all data points fell within the clinically acceptable regions (region A: 93%, region B: 7%).