J-Score: A new joint parameter for PLSR model performance evaluation of spectroscopic data

Since its beginnings, many parameters have been proposed to evaluate the goodness of Partial Least Squares Regression (PLSR) models and thus help chemometricians to choose the most appropriate one. This article proposes a new performance evaluation parameter for regression models based on spectroscopic data, the JScore, which combines some of the most commonly used model evaluation parameters (Ratio of Performance to Deviation, Calibration and Validation Root Mean Square Errors and Regression Vector) into a single indicator. The J-Score can help non-experienced analysts select both the adequate number of Latent Variables (LVs) and the best preprocessing technique for their dataset in an automated way. The performance of the J-Score has been compared to other evaluation methods with different datasets, demonstrating that it can be used for different types of samples and spectroscopic data; that it is stable and objective, and offers an easy way to select the optimal number of LVs.

Automated summary

This article proposes a new performance evaluation parameter, JScore, which combines commonly used model evaluation parameters and can assist non-experienced analysts in selecting the appropriate number of LVs and preprocessing technique in an automated way.