Machine learning used to create a multidimensional calibration space for sensing and biosensing data

Calibration curves are essential constructs in analytical chemistry to determine parameters of sensing performance. In the classification of sensing data of complex samples without a clear dependence on a given analyte, however, establishing a calibration curve is not possible. In this paper we introduce the concept of a multidimensional calibration space, which could serve as reference to classify any unknown sample as in determining an analyte concentration from a calibration curve. This calibration space is defined from a set of rules generated using a machine learning method based on trees applied to the dataset. The number of attributes employed in the rules defines the dimension of the calibration space and is established to warrant full coverage of the dataset. We demonstrate the calibration space concept with impedance spectroscopy data from sensors, biosensors and an e-tongue, but the concept can be extended to any type of sensing data and classification task. Using the calibration space should allow for the correct classification of unknown samples, provided that the data used to generate rules via machine learning can cover the whole range of sensing measurements. Furthermore, an inspection in the rules can assist in the design of sensing systems for optimized performance.