Compact Quantum Cascade Laser-Based Noninvasive Glucose Sensor Upgraded with Direct Comb Data-Mining.

Mid-infrared spectral analysis has long been recognized as the most accurate noninvasive blood glucose measurement method, yet no practical compact mid-infrared blood glucose sensor has ever passed the accuracy benchmark set by the USA Food and Drug Administration (FDA): to substitute for the finger-pricking glucometers in the market, a new sensor must first show that 95% of their glucose measurements have errors below 15% of these glucometers. Although recent innovative exploitations of the well-established Fourier-transform infrared (FTIR) spectroscopy have reached such FDA accuracy benchmarks, an FTIR spectrometer is too bulky. The advancements of quantum cascade lasers (QCLs) can lead to FTIR spectrometers of reduced size, but compact QCL-based noninvasive blood glucose sensors are not yet available. This work reports on two compact sensor system designs, both reaching the FDA accuracy benchmark. Each design commonly comprises a mid-infrared QCL for emission, a multiple attenuation total reflection prism (MATR) for data acquisition, and a computer-controlled infrared detector for data analysis. The first design translates the comb-like signals into conventional spectra, and then data-mines the resultant spectra to yield blood glucose concentrations. When a pressure actuator is employed to press the patient's hypothenar against the MATR, the sensor accuracy is considered to reach the FDA accuracy benchmark. The second design abandons the data processing step of translating combs-to-spectra and directly data-mines the "first-hand" comb signal. Beyond increasing the measurement accuracy to the FDA accuracy benchmark, even without a pressure actuator, direct comb data-mining upgrades the sensor system with speed and data integrity, which can impact the healthcare of diabetic patients. Specifically, the sensor performance is validated with 492 glucose absorption scans in the time domain, each with 20 million datapoints measured from four subjects with glucose concentrations of 3.9-7.9 mM. The sensor data-mines 164 sets of critical singularity strengths, each comprising 4 critical singularity strengths directly from the 9840 million raw signal datapoints, and the 656 critical singularity strengths are subjected to a machine-learning regression model analysis, which yields 164 glucose concentrations. These concentrations are correlated with those measured with a standard finger-pricking glucometer. An accuracy of 99.6% is confirmed from the 164 measurements with errors not more than 15% from the reference of the standard glucometer.