+44 1625708989Application of the Cross-Validation Method to the Calibration Functions in the Validation Process of an Indirect Analysis Method: Determination of Organic Compounds in Water by UV-Visible Spectrophotometry
Author(s): Chaymae Jermouni, Mohamed Nohair, Sara Azmi, El Mati Khoumri, Abdelhak Elbrouzi and Mohssine ElmorrakchiIn the process of validation of analytical methods, the guides NF T90-210 or NF V03-110 [1,2] are most often used, they are respectively used in chemical analysis of water and the agro-food field. For indirect analysis methods, which is frequently the case and especially for spectroscopic analysis methods, validation takes place after verification of the performance criteria of an appropriate calibration model. The validation or verification of the calibration function in a measurement area depends on the guide used.
For example, an MAD (Maximum Accepted Deviation) can be proposed for the adjustment between the experimental values and those adjusted by the calibration function. One can also use a statistical test of suitability by a variance study, which compares the residual variance with the variance explained by the proposed model. Statistical indicators such as the coefficient of determination and the standard error (R2; s) of the model are then calculated.
However, these verification methods are insufficient because they do not provide an experimental protocol to test the performance of the proposed model in the prediction phase. It is easy to find a coefficient of determination R2>0.95 for a fit between the observed and calculated values, but it does not guarantee the quality of the results that will be obtained. It is easy to erroneously accept an analytical method as valid in one's field of study because the instrumental response function is not properly defined. As a result, the performance evaluation of the characteristics of the analytical method leads to biased results. In this work, we propose a simple and robust protocol for the verification of the performance of the instrumental response function (calibration model) of an indirect analytical method.
The approach consists in the use of the cross-validation method (80-20%), which is a method commonly used in statistical modeling in chemistry and biology (QSAR and QSPR) [3,4]. Statistical indicators recalculated using this method are more robust and reflect better the suitability of the calibration function and its performance. The analysis method used in this work consisted of measuring the quantity of organic matter using a UV-Visible spectrum between 200 and 300 nm. We used three calibration functions by comparing the statistical indicators, as the coefficient of determination and the residual error (R2, s), which are derived from the cross-validation method of these functions. The fit between the observed and predicted values with respect to an acceptability deviation of 20% was additionally verified.
The obtained results lead to the adoption of a three-wavelength multiple regression model as a fair and robust calibration model. Unfortunately, the limited data available did not allow us to test other more advanced functions as neural networks.