Abstract :
[en] Since last decades, the world has known significant changes in the pharmaceutical products sale. The emergence of the internet trade is an important issue because it is easy to sell medicines without passing any control. Moreover, in low- and middle-income countries (LMIC), the number of local pharmacies has grown, increasing the risk to have substandard or falsified medicines. Indeed, according to the World Health Organization (WHO) it is difficult to ensure quality medicines due to areas conflict, corrupted governments and poor health system [1]. For that reason, several analytical techniques have been developed since last decade. One of the most interesting tool is the Minilab, developed by the Global Pharma Health Fund (GPHF), which is a mobile mini-laboratory for fast drug quality control. Moreover, Raman spectroscopy has gained a great interest because it can be used at any step of analytical chain or on the fieldwork with handheld devices. However, spectroscopic data implies development of chemometrics models to gather relevant information. Several unsupervised techniques have already been used to authenticate drug products [2-3]. Due to the intrinsic properties of classification methods, class modeling is more appropriated to this kind of analysis. Indeed, falsified medicines can be quite different from the calibration set, so that, it does not have sense to attribute a class meanwhile it is dissimilar to the calibration set. In this study, the performances of two class-modeling techniques will be evaluated on handheld Raman spectra, to separate falsified medicines from authentic drugs.
Three different generics of paracetamol, ibuprofen and artemether-lumefantrine with different dosage, dosage form and formulations were analyzed. Most of them were gathered in local Belgium pharmacies and other were gathered in Africa (artemether-lumefantrine formulations). Samples were analyzed with a handheld Raman spectrophotometer Pharma 21CFR part 11 qualified (Truscan RM, Thermo Scientific, USA) directly through the blister. In order to have a good representativeness of intra-batch variability, 10 tablets were analyzed per sample. The acquisition parameters were set to default. Two models were tested: one-class PLS (OC-PLS) [4] and the data driven-soft independent modeling class analogy (DD-SIMCA) [5]. All the computations were done in Matlab® (R2017b). The calibration and validation set was the same for each model and composed of 60 spectra for each, with different batch number.
Because of the nature of each algorithm, there is a significant difference in terms of separation. Looking at Figure 1, the separation of dosage form for ibuprofen is much different between the two models. For the DD-SIMCA, it is more difficult to separate the long acting release from the soft capsule/coated tablet compared to the OC-PLS model. For the other API, similar results are obtained for the dosage and for the brand. It seems that the OC-PLS is more sensitive to the small spectral variabilities.
In the case of artemether-lumefantrine formulations, the separation between samples is much more difficult. Indeed, the lumefantrine is a high Raman scatterer. This can explain that the signal of artemether and excipients is difficult to access. Elsewhere, in terms of development, the DD-SIMCA is much harder to optimize because there are more tunable parameters than for OC-PLS. Furthermore, both models are really influenced by spectral pretreatment. An optimization has to be done for each.
The authentication of pharmaceutical products by handheld Raman spectroscopy has been possible thanks to class modeling. Both tested algorithms shown interesting results regarding the separation of samples depending on their characteristics. The optimization of data processing and pre-processing is the key-step to improve as sensitivity as specificity of both class modeling methods.
