Poster W9 - Progress in the development of a point-of-care device for anti-seizure medication titration: electrochemical signal quantification in saliva using multivariate and machine learning methods

Introduction: The noninvasive monitoring of analytes in complex biofluids at the point of care holds great promise for improving healthcare for numerous applications including therapeutic drug monitoring. Our team is developing a field-use device for the electrochemical detection of the anti-seizure drug carbamazepine from whole human saliva. Challenges of saliva-based electrochemical monitoring include that saliva is a complex matrix with components that can lead to a high and variable background current such that robust quantification of the analyte signal can be difficult. Thus, simple electrochemical signal processing algorithms that work well for quantification in well-defined buffer backgrounds may not be ideal for analysis in complex biofluids. Further, most studies in complex fluids are based on data sets with only a handful of measurements. The focus of the current study is on developing robust signal analysis methods using a large voltammogram dataset, with over 250 samples, of carbamazepine at defined levels in whole human saliva. Although applied to our specific context of a point-of-care device for anti-seizure medication titration, we expect our result to have high relevance for electrochemical signal analysis more generally.

Materials and

Methods: Square wave voltammetry data was acquired using our system of polymeric laminate electrochemical cells containing carbon (working and counter) and Ag/AgCl (reference) stencil-printed electrodes. Key features of our electrochemical system for sensing in saliva included a non-fouling electrode coating of Nafion and carbamazepine signal enhancement using sodium dodecyl sulfate. Human saliva was collected using an IRB approved protocol and mock samples created by spiking know amounts of carbamazepine into the saliva. Carbamazepine concentrations of 0, 8, and 16 micromolar, in a range relevant to carbamazepine monitoring in saliva, were used. We then comprehensively investigated different peak features using univariate, multivariate, and the machine learning-based methods, Random Forest, k-Nearest Neighbors, and Gaussian Process. Performance metrics used were the coefficient of determination (R2) and average absolute percent difference between predicted and nominal concentrations. Five-fold cross validation, on 70% of the dataset, was performed to select the best peak features and (for the machine-learning algorithms) hyperparameter values. Final comparison of the methods was then based on performance in an unbiased validation using the previously reserved 30% of the dataset.

Results, Conclusions, and Discussions: Our comparison of different peak features using univariate linear analysis (Figure 1 shows representative results from one of the folds) indicated a substantial difference in performance (as assessed by coefficient of determination R2) across peak features with peak area and peak height among the best performing (two leftmost light gray bars in Figure 1). Derivative-based peak features performed moderately well and potential-based peak features performed the poorest (white bars in Figure 1). Interestingly, peak curvature, which has shown promise in other contexts, also performed relatively poorly (light gray bar third from the left in Figure 1), while less commonly used peak features such as the mean peak height and the standard deviation of the peak height also performed well. A key finding of this work is that multivariate linear analysis (dark bar on the right in Figure 1), using the five top performing peak features, produced a significant improvement over univariate linear analysis using any of the peak features as assessed by the five-fold cross validation and a permutation paired t-test of fold-level R2 performance. A second key finding of this work is that three machine learning methods, Gaussian Process (using the default kernel of radial basis function and a regularization strength of 1.5), Random Forest (using 100 trees), and k-Nearest Neighbors (using five nearest neighbors and Euclidian distance) showed improved performance compared to multivariate linear analysis (see Figure 2). Strategies in progress for further performance gains include whole voltammogram processing using Gaussian process and semi-supervised techniques. Our work has high relevance for improving the robustness of signal quantification of voltammograms in other complex sample matrices.

Acknowledgements (Optional):