Optimized Quantification of Multiple Drug Concentrations by WeightedMSE With Machine Learning on Electrochemical Sensor

Quantification of multiple drugs is of great importance and urgently needed in therapeutic drug monitoring (TDM) and personalized therapy. Especially, based on cyclic voltammograms (CVs) obtained by electrochemical sensors, the use of artificial neural networks (ANNs) has been widely attempted in the accurate quantification of drug concentrations, enabling the development of point-of-care and potentially system-level wearable devices. However, most of the work only considers the accuracy of how the predicted value is close to the actual value, which does not consider whether the predicted drug concentration is underestimated. In practical drug quantification, potential toxicity due to overexposure with underestimated quantification can lead to endangering the patient's body. Therefore, avoiding underestimating the concentration of drugs based on existing quantification models is required and necessary to optimize the conventional loss function at the output stage of ANN. In this letter, a novel loss function based on mean squared error (MSE), WeightedMSE, is proposed for avoiding underestimated quantification. It can be changed flexibly by adjusting parameters in order to adapt the acceptable overestimation range corresponding to the different types of drugs. A simulated dataset and a real dataset of etoposide and methotrexate are used as drug models, demonstrating that the proposed method can avoid underestimation in predicted values by over 98% in quantifying the concentration of multiple drugs and showing significant effectiveness for the development of point-of-care and wearable monitoring systems.