Sensing using SERS-substrate and machine learning approaches

Background and Objectives: Accurate and duly determination of glucose levels is critical for the diagnosis and control of diabetes. Recently, optical methods for glucose determination have become the subject of increased interest due to their potential cost-effectiveness, portability, and low invasiveness. Raman spectroscopy coupled with surface-enhanced Raman scattering (SERS) substrates demonstrates outstanding sensitivity through signal amplification and high specificity due to the unique vibrational spectra of target molecules. However, direct detection of glucose using SERS is complicated by the weak adsorption of glucose on metal surfaces and its low scattering cross section. Materials and Methods: Glucose sensors were constructed on the basis of polycaprolactone(PCL) scaffolds which have been modified using vaterite microparticles or filter paper (FP), both of which were then decorated with silver nanoparticle aggregates. The surface of the created substrates was assessed using scanning electron microscopy (SEM) on a MIRA II (Tescan, Czech Republic). To make the sensors specific for glucose detection, they were coated with a layer of glucose oxidase (GOx). To analyze the SERS spectra obtained as a result of measurements of aqueous solutions of glucose with various concentrations on sensors, classification models developed using the ensemble method RandomForestClassifier were used. Confusion matrices were obtained to assess the ratio of truly classified spectra. Results: Carrying out three cycles of modifying the surface of PCL fibers with microparticles of calcium carbonate leads to uniform overgrowth of the entire treated area. Additional immobilization of glucose oxidase (GOx) onto the surface of a matrix of PCL fibers with grown vaterite particles and a reduced layer of silver aggregates has provided selectivity for glucose detection when examining samples using SERS spectroscopy. The highest sensitivity in determining low glucose concentrations (1 mM) has been obtained for substrates with three sequential modifications of PCL fibers with vaterite and the reduction of aggregates of Ag nanoparticles from 5 M solutions of silver nitrate and ammonia hydrate with overall accuracy of 92.2%. Filter paper was considered as an alternative to using PCL-based scaffold. The reduction of silver was carried out without vaterite particles growth by varying the concentration of the reagents used (AgNO3, NH3·H2O). Sensors based on filter paper after the reduction of silver on the surface from salt solutions with concentrations of 2 M have shown overall accuracy of 90.2% and the ratio of truly classified 1 mM glucose solution of 88%. Conclusion: Increasing the number of cycles of sequential modification of the polycaprolactone surface with vaterite microparticles makes it possible to obtain a more uniform overgrowth, which was observed in SEM images, and, as a consequence, greater ratio of truly classified spectra at lower glucose concentrations. The PCL-based sensor (PCL/(CaCO3)3/Ag (5 M)) have outperformed FP/Ag (2 M) both with overall accuracy of classification (92.2% versus 90.2%) and 100% of truly classified spectra of 1 mM glucose solution.

glucose sensor, SERS spectroscopy, silver coating, machine learning, random forest