Applications of Fluorescence Spectroscopy and Machine Learning Methods for Monitoring of Elimination of Carbon Nanoagents from the Body

Abstract

The study considers the application of artificial neural networks to solve the inverse problem of fluorescence (FL) spectroscopy for monitoring the elimination of carbon nanocomplexes’ components from the body – the problem of determining the concentrations of carbon dots (CD) and the anticancer drug doxorubicin (Dox) excreted from the body with urine. The problem was solved in three ways using three sets of FL spectroscopy data: spectral data obtained by exciting FL of urine together with CD and Dox with radiation of 405 nm, 532 nm wavelength as well as spectra obtained by combining these data. Multilayer perceptrons (MLP) were applied to the obtained spectral data, which enabled the determination of the concentrations of CD and Dox in urine. To increase the accuracy of monitoring the excretion of CD and Dox with urine, principal component analysis and autoencoders were additionally used. The conducted studies showed that the best results of solving this problem are provided by the application of a MLP to spectral data compressed using an autoencoder. This approach allows us to determine the concentration of CD in urine with MAE of 39 ng/mL (3.3% of the upper limit of the concentration range) and the concentration of Dox with MAE of 27 ng/mL (2.8% of the upper limit of the concentration range). The proposed approach shows results comparable with analogues, however it lacks several significant drawbacks such as rigid fixation of the CD concentration in the suspension, and it can be used for simultaneous rapid monitoring of a number of substances.