An optical method recently proposed for non-invasive
in vivo blood glucose concentration (BGL) measurement, named “Pulse Glucometry”, was combined and compared with four multivariate analyses for constructing calibration models: Principal Component Regression (PCR), Partial Least Squares Regression (PLS), Artificial Neural Network (ANN), Support Vector Machines Regression (SVMsR). A very fast spectrophotometer for “Pulse Glucometry” provides the total transmitted radiation spectrum (I
λ) and the cardiac-related pulsatile component (ΔI
λ) superimposed on I
λ in human fingertips over a wavelength range from 900 to 1700 nm with resolution of 8 nm in 100 Hz sampling. From a family of I
λs measured, which include information relating to blood constituent such as BGL values, differential optical densities (ΔOD
λs, where ΔOD
λ=Log (1+ΔI
λ/I
λ)) were obtained and normalized by the ΔOD
λ values at 1100 nm. Finally, the 2nd derivatives of the normalized ΔOD
λs (Δ
2OD
λs) along wavelengths were calculated as regressors. Subsequently, calibration models from paired data sets of regressors (the values of Δ
2OD
λs) and regressand (the corresponding known BGL values) were constructed with PCR, PLS, ANN and SVMsR. The results show that each calibration model provides a relatively good regression with a modified 5-fold cross validation for total 95 paired data, in which the BGLs ranged from 100.7-246.3 mg/dl. The results were evaluated by the Clarke error grid analysis and all data points obtained from all calibration models fell within the clinically acceptable regions (region A or B). Among them, ANN and SVMsR calibration provided the best plot distributions (in ANN; Region A: 77 plots (81.1%), B: 18 plots (18.9%). in SVMsR; Region A: 78 (82.1%), B: 17 (17.9%)). Total calculation time of SVMsR is about 100 times shorter than ANN. These results suggest that a calibration model using SVMsR is highly promising for “Pulse Glucometry.”
View full abstract