Machine learning for optical chemical multi-analyte imaging : Why we should dare and why it's not without risks

Abstract

Simultaneous sensing of metabolic analytes such as pH and O₂ is critical in complex and heterogeneous biological environments where analytes often are interrelated. However, measuring all target analytes at the same time and position is often challenging. A major challenge preventing further progress occurs when sensor signals cannot be directly correlated to analyte concentrations due to additional effects, overshadowing and complicating the actual correlations. In fields related to optical sensing, machine learning has already shown its potential to overcome these challenges by solving nested and multidimensional correlations. Hence, we want to apply machine learning models to fluorescence-based optical chemical sensors to facilitate simultaneous imaging of multiple analytes in 2D. We present a proof-of-concept approach for simultaneous imaging of pH and dissolved O₂ using an optical chemical sensor, a hyperspectral camera for image acquisition, and a multi-layered machine learning model based on a decision tree algorithm (XGBoost) for data analysis. Our model predicts dissolved O₂ and pH with a mean absolute error of < 4.50·10^-2 and < 1.96·10^-1, respectively, and a root mean square error of < 2.12·10^-1 and < 4.42·10^-1, respectively. Besides the model-building process, we discuss the potentials of machine learning for optical chemical sensing, especially regarding multi-analyte imaging, and highlight risks of bias that can arise in machine learning-based data analysis.

Keywords: Decision tree algorithm; Dissolved oxygen; Intensity-based sensing; Supervised pattern recognition; XGBoost; pH.

Fig. 1

Schematic representation of the measurement setup used for measuring and calibrating the optical chemical dual-sensor for pH and dissolved O₂ (A) and real image of the fluorescence of the dual analyte optode upon excitation with a high-power LED (B)

Fig. 2

Spectral characterization of the single optode layers recorded on the ClarioStar Plus plate reader under different pH and O₂ conditions. The excitation spectra of A lipophilic HPTS as a pH-sensitive dye and C Pt-TPTBP as an O₂-sensitive dye are shown in the left panels, while the emission spectra of B the pH indicator and D the O₂ indicator are shown in the right panel. Note that the O₂-sensitive sensor layer also contains macrolex fluorescent yellow as the reference dye. While in A, C, and D, the fluorescence intensity is displayed relative to the maximum intensity, in B, the intensity is displayed relative to the isosbestic point at 530 nm

Fig. 3

pH and O₂ calibration of the dual analyte optode under constant condition of the respective other analyte. A pH calibration between pH 4 and 11 under anoxic (0 hPa) and air-saturated (195 hPa) conditions. B O₂ calibration is displayed as ratiometric intensity relative to the reference indicator, macrolex fluorescence yellow, while the pH is kept constant at either pH 4 or pH 8. The dashed curves in both panels represent the hypothetical calibration curves of the analytes if the respective standard calibration functions for the individual analytes, i.e., Boltzmann fit for pH calibration and simplified Stern–Volmer fit for O₂ calibration, were applied to the calibration points

Scheme 1

Overview of the workflow conducted to build up the multi-layered machine learning model for simultaneous detection of pH and dissolved O₂

Fig. 4

Dataset adjustment of the unbalanced calibration dataset by reducing the number of data points used where data are prevailing. The amount of data points used is adjusted to the general median. The adjustment is performed separately for each analyte. For each panel, the original distribution of the dataset is shown in light color, while the more balanced dataset is shown in dark colors, i.e., (A) in orange for pH and (B) in gray for O₂

Fig. 5

Overall model performance for predicting the pH (A) and dissolved O₂ concentration (B), respectively, assessed against test data that the algorithm has never seen before. The main plot compares the predicted and the respective target values for the entire calibration range using the multi-layered ML model based on XGBoost. The insets of each panel display the dispersion around the target value as black dotted markers. The target value is indicated as an orange solid line. C and D display examples of optode images before and after data analysis. C shows the absolute fluorescence intensity of the dual analyte optode is visualized at 773 nm, whereas D shows the chemical images in which the absolute fluorescence intensity has been translated into the corresponding pH and O₂ concentration (in hPa) to represent them in each pixel