Enhancing the accuracy of blood-glucose tests by upgrading FTIR with multiple-reflections, quantum cascade laser, two-dimensional correlation spectroscopy and machine learning

[Display omitted] •Upgraded FTIR technology improves the accuracy of noninvasive blood-glucose test.•Multi-passes setup and QCL boost signal-to-noise ratio in the upgraded FTIR system.•2D correlation spectroscopy enhances the analysis of spectral features for glucose.•Machine learning enables blood-glucose screening reliability. The accuracy of screening diabetes from non-diabetes is drastically enhanced by strategically upgrading the bench-marking infrared spectroscopy technique for non-invasive tests of blood-glucose, both with state-of-the-art instrumentation-retrofits and with intelligent spectral-datamining tools. First, the signal-to-noise performance of FTIR in measuring the spectral features of a glucose solution containing bovine serum albumin is improved by 2–3 times with the common single-pass attenuated total-reflection setup replaced by a multi-passes-reflections setup. Second, replacing the ordinary infrared lamp with a quantum cascade laser further improves the signal-to-noise by 3 times. The performance of the upgraded spectrometer in screening hyperglycemia is gauged by the accuracy of such screens derived from 100 repetitive spectral-measurements per glucose concentration, for 24 glucose concentrations spanning the range of 70–300 mg/dL, a range which covers the blood-glucose contents of all non-diabetic and diabetic human-subjects. Third, intelligent datamining methods are exploited to implement decision trees for screening hyperglycemia. Their decisions are mapped into a confusion matrix and the matrix-elements are used to calculate the accuracy merits of each method. Evidently, the accuracy of the multi-passes-FTIR with the standard principal-components datamining method is 80 %. The adoptions of the quantum cascade laser and two-dimensional correlation spectroscopy datamining technique raises this to 96.3 %. Finally, a novel machine learning method, which comprises three different decision-tree tools to generate trial screening decisions and a “majority-voting” datamining tool to reach a final screening decision, yields the best accuracy of 98.8 % ever reported in the literature.