Multi-Ion-Sensing Emulator and Multivariate Calibration Optimization by Machine Learning Models

One paramount challenge in multi-ion-sensing arises from ion interference that degrades the accuracy of sensor calibration. Machine learning models are here proposed to optimize such multivariate calibration. However, the acquisition of big experimental data is time and resource consuming in practice, necessitating new paradigms and efficient models for these data-limited frameworks. Therefore, a novel approach is presented in this work, where a multi-ion-sensing emulator is designed to explain the response of an ion-sensing array in a mixed-ion environment. A case study is performed emulating the concurrent monitoring of sodium, potassium, lithium, and lead ions, in a medium representative of sweat samples. These analytes are relevant examples of sweat ion-sensing applications for physiology, therapeutic drug monitoring, and heavy metal contamination. It is demonstrated that calibration datasets output by the emulator explain accurately the experimental response of polymeric solid-contact ion-selective electrodes, where root-mean-squared error of 1.37, 1.44, 1.78,
$2\,mV$
are obtained, respectively, for
Na
⁺
,
K
⁺
,
Li
⁺
,
Pb
²⁺
sensor calibration in artificial sweat. Besides, synthetic datasets of custom size are generated to train, validate, and evaluate different types of multivariate regressors. A Multi-Output Support Vector Regressor (M-SVR) is proposed as a compact, accurate, robust, and efficient multivariate calibration model. It features 13.22% normalized root mean squares, and 20.29% mean root squares improvement compared to a simple linear regression model. It is an unbiased estimator for medium to large datasets, and its average generalization error is of 3.22%. Besides, M-SVR models have a lower computational complexity than single-output SVR or neural network models, making them a suitable solution for memory and energy-constrained edge devices used for continuous and real-time multi-ion monitoring.