M17 - Machine Learning Applied to Electrochemical Data Processing for Improved Analyte Quantification in Complex Saliva

Introduction: : Biofluids that can be noninvasively and frequently collected, e.g., saliva, have great promise for real-time analyte monitoring at the point-of-care to inform on patient health. However, analyte quantification in these fluids can be challenging due to their complex composition that can reduce signal-to-noise ratio. For electrochemical sensing in saliva, saliva complexity can result in signal interference through a high and variable background such that analyte quantification is challenging. Conventional analysis algorithms that focus on a single peak feature may work well for analyte quantification in well-defined buffer backgrounds but may not be ideal for quantification from complex biofluids. Machine learning algorithms that measure analyte concentration based on multiple voltammogram features have high potential for improving accuracy. However, there is a scarcity of large voltammogram datasets required for their training and testing. Thus, we created multiple large saliva-based voltammogram datasets to develop and assess the performance of different types of machine learning regression models: k-Nearest Neighbors (KNN), Random Forest (RF), Support Vector Machine (SVM), Gaussian Process (GP), and linear multivariate (Table 1). Further, we assessed the performance of a model developed using one saliva background, on a different saliva background (same donor collected on a different day). We report our results in the context of our team’s development of a field-use device for the electrochemical detection of the antiseizure drug carbamazepine from whole human saliva for personalized drug titration, however, our analytical approach could be applied to the quantification of other analytes in complex fluids from voltammogram data to improve accuracy.

Materials and

Methods: : Electrochemical datasets: We created hundreds of voltammograms spanning five analyte concentrations (0–16 μM) of the anti-seizure drug carbamazepine spiked into whole human saliva. Saliva was collected at Oregon State University (OSU) using a protocol approved by the OSU Institutional Review Board. Passive drool saliva samples were donated by healthy volunteers not taking carbamazepine. Saliva from an individual donor was aliquoted and stored at -20oC until use. Saliva solutions contained 0.1 M KCl (electrolyte) and 0.4 mM sodium dodecyl sulfate (signal enhancement), in addition to the desired final carbamazepine concentration. A 40 μL sample volume was added to a disposable polymeric-laminate electrochemical flow cell with stencil-printed and Nafion-coated (for antifouling) electrodes. Square-wave voltammetry was performed using a benchtop potentiostat (0.5–1.2 V, 0.004 V step size, 0.025 V amplitude, 15 Hz frequency, 4 min. incubation time, 5 seconds quiet time).

Model development: The dataset used for model development was randomly split into a training-set (60%) and a test-set (40%). For each regression model type, we performed feature selection from nine quantitative features extracted from each voltammogram's shape in the vicinity of the analyte peak coupled with hyperparameter tuning using five-fold cross validation, identifying the optimal feature set and hyperparameter values which maximized a combined performance metric of R^2 and average % error. For unbiased model assessment, we applied each model to the test data and compared performance using R^2 and average % error. Further, paired permutation testing was conducted to statistically compare the performance of the different models.

Results, Conclusions, and Discussions:: Results of our systematic analysis provided several important insights regarding how to achieve improved accuracy in analyte quantification from voltammograms. First, our analysis identified one critical voltammogram feature associated with the carbamazepine analyte peak, namely the minimum rate of change of normalized current with respect to potential difference. This was selected as the top-performing feature across all models over the common peak features of height and area. Second, our analysis demonstrated that each model’s performance was substantially improved by including additional voltammogram features, indicating that for analysis of saliva-based electrochemical data, conventional linear algorithms that only use one peak feature are not optimal. Third, our analysis indicated that voltage-based and background features improve model accuracy (Figure 1 left). Test performance results showed that all models produced high R^2 values (0.84 and 0.86), but for the metric of average % error, KNN performed best (19%), SVM and RF performed moderately well (22%), and GP and the linear multivariate models performed worst (23%-26%) (Figure 1 right). To evaluate whether the difference in performance between KNN and the other models is statistically significant, we compared the models’ average % error with a permutation test and found a statistically significant difference (with 𝑝 < 0.05) between best performing KNN and each of the two worst performing models of GP and linear. Finally, further KNN model assessment on saliva data from the same individual but collected on a different day, showed comparable performance without additional training (Figure 2). SVM and GP also showed comparable performance across the datasets. In contrast, both RF and the linear model showed substantially degraded performance on the independent dataset. Overall, our results indicate the potential of machine-learning multivariate models to improve accuracy for the prediction of drug levels in saliva. In addition to performing the best, KNN has the advantages of being simple to both implement and interpret. This is the first study we are aware of, that has demonstrated the applicability of machine learning models trained for the quantification of analyte levels in one background of saliva, on a different saliva background.

Acknowledgements and/or References (Optional)::