A pH-Responsive colourimetric sensor array based on machine learning for real-time monitoring of beef freshness

A novel strategy based on a pH-responded colourimetric sensor array (CSA) was developed by combining it with machine learning models to effectively monitor the freshness of beef in real time through the detection of total volatile basic nitrogen (TVB-N) contents. A total of 168 colour features were calculated and effective colour features sensitive to the changes in TVB-N contents were extracted by sequential forward selection (SFS), random forest (RF), and principal component analysis (PCA). Besides, the linear regression model of partial least squares regression (PLSR) as well as nonlinear regression models of random forest regression (RFR) and support vector machine regression (SVR) were established to predict TVB-N values of beef at 28 degrees C based on full and effective colour features. The RF-SVR model had the best performance with a corresponding determination coefficient of prediction (Rp2) of 0.9596, a root-mean-square error of prediction (RMSEP) of 1.89 mg/100 g, and a relative prediction deviation (RPD) of 4.98. Moreover, the CSA based on the RF-SVR model was applied to the quanti-tative analysis of beef freshness at 4 degrees C storage, validated by using a standard method for the detection of TVB-N contents in beef samples. The results illustrated that the CSA as an objective and nondestructive tool can monitor TVB-N contents for the evaluation of beef freshness with the help of machine learning.

Automated summary

A novel strategy combining a pH-responded colourimetric sensor array (CSA) with machine learning models was developed to effectively monitor beef freshness in real time by detecting total volatile basic nitrogen (TVB-N) contents. Sequential forward selection (SFS), random forest (RF), and principal component analysis (PCA) were used to extract effective colour features sensitive to changes in TVB-N contents. Linear regression (PLSR), as well as nonlinear regression models (RFR and SVR), were established to predict TVB-N values based on full and effective colour features. The RF-SVR model showed the best performance, with a determination coefficient of prediction (Rp2) of 0.9596, a root-mean-square error of prediction (RMSEP) of 1.89 mg/100 g, and a relative prediction deviation (RPD) of 4.98. The CSA based on the RF-SVR model was applied to quantify beef freshness at 4 degrees with validation using a standard method. This study demonstrated that the CSA, combined with machine learning, can objectively and nondestructively monitor TVB-N contents for evaluating beef freshness.