Near infrared spectroscopy as a tool to control the results of chemical analyses performed as single determinations

In laboratories, chemical analyses are generally performed in duplicate in order to confirm the results. When the difference between 2 duplicates is greater than a certain step, the analysis has to be repeated one more time to check the value. This procedure requires a huge number of individual measurements for large analytical series. In some laboratories, analyses are performed in single with use of a control sample for each series. However this procedure does not allow to control all individual results and therefore increases the risk of error. It was decided to investigate the potential of NIRS for the control of analyses performed in single in the laboratory. The procedure was to compare the analytical result to the prediction provided by an equation of known SEP (standard error of prediction). The gain in precision provided by this strategy is difficult to estimate directly since the distribution of the corrected values do not follow a normal (Gauss) law. A numeric simulation performed on 10 000 independent values gave the following relationship: scorr=s-0,12s3/SEP2 where s is the repeatability of analysis, scorr is the repeatability of the result after elimination of outliers detected by NIRS and SEP is the precision of NIRS calibration. This relationship concerns the detection of outliers of the normal error distribution (at p=5% level), but NIRS prediction can also help to track laboratory mistakes and sampling problems. An application of this strategy was performed on a real dataset. Fifty poultry feed samples were analysed in single for fibre content (WICW, water-insoluble cell walls), and NIR spectra were collected on a FOSS Nirsystem 6500 spectrometer. Laboratory results were checked against a prediction equation (PLS regression; R2=0.975; SECV=1.43%) and re-run if they differed by more than 1.15% (level calculated as 2.83 times the repeatability of reference analysis). This analysis strategy (SI) was compared to the "normal" duplicate analysis (S2). The agreement between the results obtained by the 2 strategies was R2=0.996 (intercept=0.22, slope=1.01). All outlier values were detected by NIRS. The number of samples to be re-analysed was 17 (10 in duplicate and 7 in triplicate) for SI and 12 for S2, leading to a total number of assays of 74 for SI and 112 for S2. The conclusion of this study is that the validation of single measurements by an existing NIRS equation for large series of samples is a powerful tool to reduce the number of measurements in the laboratories, while maintaining a control of results. (Résumé d'auteur)