Logiciel Vibra4Fake 2020

Vibrational spectroscopy (NIR and Raman) is a family of powerful analytical chemistry techniques to determine the chemical composition of samples. It is now well-established that it provides precise information about the overall quality (composition, status, etc.) of chemical samples.

For the purpose of detecting potentially abnormal or suspicious quality of chemical samples with vibrational spectroscopy techniques, a new statistical strategy for outlier detection (belonging to the family of the so-called one-class classification models) has been developed. This strategy is two-fold. First and most important, it involves building a statistical tolerance band that is calibrated to contain with a high probability samples of the chemical having the required quality, while rejecting as much as possible sample of suspicious quality. Second, the strategy involves computing univariate tolerance intervals for some univariate statistical metrics or indexes (these are mapping of each spectrum or curve to a one-dimensional or univariate space). This second aspect is provided to cross-check or validate the decision suggested by the first method, the band. Several such univariate metrics or indexes are proposed.

The present program includes algorithms to implement the above described outlier detection strategy. The core code is written with the statistical software R and associated packages, and for now includes several scattered codes, each performing different tasks, that have been tested on given datasets. A further step involves combining all those in a coherent and intelligent function. The core model is still under development and will probably undergo refinements and adjusments with contributions from other partners of the consortium of the V4F project (Pharmalex).

Despite the algorithm has been developed and tested mainly on NIR and Raman technology, it is quite versatile and can be applied to robust outlier detection with other spectroscopic techniques