Techniques to improve precision of neural network-based biosensors under effect of noise

Abstract

Biosensors are devices for the detection and analysis of chemical compounds based on biochemical processes [1]. One of the most widespread biosensor types are enzyme-based amperometric biosensors [2], where the sample components are enzymatically converted into products, which produce an electric current that is measured. The biosensor is then calibrated using known samples to establish a dependence between component concentrations and electric current. Unknown samples can then be analyzed by measuring their current signal and, by using the inverse of the dependence, determining their composition i.e., solving an inverse biosensor problem. For a single substrate, this can be achieved very easily, but for multiple substrates its solution is complicated by the ill-posedness of the problem [3], especially when the biosensor response is under the effect of noise. This can negatively impact the precision of the device [4]. One promising method to alleviate the ill-posedness of such problems is to use neural networks, which have successfully been used to solve inverse problems in heat conduction, such as determining the initial condition of the heat equation from the temperature profile [5]. Although neural networks have been applied to solve the inverse biosensor problem [6], we were unable to find any reports dealing with biosensor precision under the effects of noise in this case. Since biosensors are used in various performance-critical applications, such as environmental monitoring and protection, food safety, medicine and so on [2], an investigation of effects of noise is necessary, in order to find ways to mitigate deterioration of biosensor performance. In this talk, we present our findings on noise effects and techniques currently in development to improve biosensor precision, such as neural network training set extension.