A novel probabilistic spectral matching strategy based on wavelets regression and Bayesian modeling for product identification and quality verification

Near infrared (NIR) and Raman spectroscopy are powerful analytical tools approved by the EU and US pharmacopeias. They can provide an accurate description of the chemical composition of samples, and hence can be used to fingerprint a drug product. With the fast-development and miniaturization of handheld spectrophotometers, these vibrational spectroscopic techniques are more and more used in a large range of research and industrial applications including real-time identification, reaction monitoring, quality control and quality assurance.

While these applications are extremely appealing and promising, they have raised novel statistical analysis challenges: the needs for accurate, robust, risk-oriented, computationally efficient statistical decision-making tools to compare high dimensional spectra in order to identify or authenticate the quality of pharmaceutical products. Unfortunately, the current chemometric methods used to match spectral data do not meet most of these criteria. The majority of these methods (e.g. Hit Quality Indexes, multivariate equivalence tests, PCA-based distances) are either not enough sensitive to meaningful singularities in the spectra or are not robust to measurement errors.

We propose a novel and probabilistic (predictive) strategy based on spectroscopic fingerprints and newly emerging chemometric techniques such as the (Bayesian) functional data analysis [1,2], for the identification and verification of the quality of drugs. A representative set of spectra of a product is sampled from each batch at release using appropriate Raman or NIR techniques.

Based on this set and using Bayesian wavelet regression, a statistical tolerance region (band) is constructed so that it contains a high proportion, say at least 90% or 95% of future spectra of the product. The upper and lower limits of the tolerance band or region are used as thresholds or reference spectra that would enable testing the conformity of any unit from the product batch based on its spectrum, while controlling the risks of errors. Compared with existing spectral matching models, the proposed approach is fully risk-oriented. It enables the detection of very meaningful singularities or perturbations of the spectrum caused, for example by drug falsification, but also degradation, batch inversion and many other product manufacturing issues.

The computed tolerance regions or reference spectra for a product may be stored in a cloud server and can be accessed throughout a supply chain to check the conformity of the product quality. Hence, the proposed risk-oriented approach can potentially complement the so-called serialization of pharmaceutical batches launched by the European Union Falsified Medicines Directive to enhance the security of the pharmaceutical supply chain.