Patients with chronic disorders may experience challenges achieving effective medication dosing throughout the day. In the case of people with epilepsy, optimal dosing requires minimizing the adverse side effects of overprescription and the seizure risk of an under prescription. Currently, blood samples and lab-based testing are required of patients to measure drug levels, so measurements are made infrequently. Point-of-care electrochemical sensing from saliva provides a non-invasive alternative to laboratory blood testing that would support frequent and portable testing. However, the complexity of saliva and differences in protein and chemical composition of an individual’s saliva can be variable between persons and may interfere with effective electrochemical analyte quantification. Simple electrochemical analysis methods may be insufficient for analyte quantification in complex biofluids. To develop machine learning models for improved analyte quantification, electrochemical datasets were collected for hundreds of saliva samples containing different spiked-in concentrations of the drug carbamazepine. Three regression models were investigated: k-nearest Neighbors (KNN), Random Forest (RF), and linear multivariate. For each regression model type, we performed feature selection from nine voltammogram features coupled with hyperparameter tuning, using a performance metric that combined the coefficient of determination R2 and average % Error. For unbiased model assessment, we applied each model to test-set data, using metrics of R2 and % Error. Our analysis (i) identified one critical voltammogram feature associated with the analyte peak that was common across models, but that is not commonly used in voltammogram analysis; (ii) indicated that both voltage-based and background current features can improve model accuracy; and (iii) demonstrated non-parametric models returned improved model performance on independent voltammogram data over linear multivariate models. Overall, our results indicate the high impact potential of machine-learning models to substantially improve accuracy for the quantification of drug levels in saliva over conventional linear regression models.
