Phantom glucose calibration models from simulated noninvasive human near-infrared spectra

The validity of published reports claiming to have successfully measured in vivo blood glucose from noninvasive near-infrared spectra collected in a time-dependent manner is challenged on the basis of results obtained from a phantom glucose spectral data set. An in vitro model is used to simulate noninvasive human near-IR spectra. The phantom glucose data set is created by purposely omitting glucose in these modeled samples. Glucose values are then assigned to successive phantom glucose spectra, and multivariate calibration models are generated for glucose based on partial-least squares regression. As expected, calibration models are incapable of predicting glucose values when the glucose assignments are made randomly. Apparently functional models are obtained, however, when glucose assignments are made in a nonrandom, time-dependent manner. Prediction errors from these nonrandom models are essentially identical to those published by others as evidence of successful noninvasive blood glucose measurements. Chance temporal correlations between assigned glucose concentrations and some uncontrolled experimental parameter are responsible for this apparent model functionality.