Method—An Investigation Into Post-Hoc Analysis Methods for Electrochemical Biosensor Data

Abstract

Recently, researchers are exploring machine learning (ML) algorithms as post-hoc analysis tools to improve performances of electrochemical biosensors. Reported results are promising, yet comprehensive study on optimal methods for model development is still lacking. For improved efficiency, accuracy, and robustness, it is essential to optimise the relationships between feature extraction techniques and choice of training algorithms. Herein, this paper presents a comparative study between different feature extractions methods, namely principal component analysis (PCA), linear discriminative analysis (LDA), fast Fourier transform (FFT) ,and discrete wavelet transform (DWT), to compress and extract significant components from differential pulse voltammetry (DPV) and electrochemical impedance spectroscopy (EIS) datasets. Support vector regression (SVR) and multilayer perceptron (MLP) models were developed, and their performances were compared with conventional post-analysis methods. The best performing combination for DPV dataset was MLP with DWT, achieving an R2 of 0.995, and for EIS dataset was MLP with PCA, achieving an R2 of 0.960, on test set, respectively. The developed models had achieved an average of 0.61% improvement for real sample recovery tests. The presented approaches demonstrated the capabilities of optimised ML models to automate post hoc analysis for more robust outcomes, while eliminating tedium of post-analysis for end users.

Strategies to Enrich Electrochemical Sensing Data with Analytical Relevance for Machine Learning Applications: A Focused Review

“…To fully leverage Strategy II, it is critical to choose the best combination of the feature-extracting methods and training algorithms for specific data types. D. K. X. Teo et al demonstrated that predicting acetaminophen levels from DPV curves and carcinoembryonic antigen levels from EIS spectra varied significantly based on the feature extraction methods (principal component analysis, linear discriminative analysis, fast Fourier transfer, discrete wavelet transform) and training algorithms (linear regression, support vector regression, multilayer perceptron) used [ 23 ]. Based on performance metrics (coefficient of determination, root mean square error, mean absolute error, average relative error, F1 score), the authors found that multilayer perceptron with discrete wavelet transform was best for the DPV dataset and multilayer perceptron with principal component analysis was best for the EIS dataset.…”