Test Process and Results of Potential Masking of Sterilizing-Grade Filters.

Pre-use/post-sterilization integrity testing (PUPSIT) has been a widely debated topic for the last several years. To a large extent, the debate is because of the fact that scientific data were not available to provide additional clarity that could inform appropriate risk-based judgements and commensurate actions. To gain clarity, the Parenteral Drug Association (PDA) and BioPhorum Organizations Group (BioPhorum) formed the Sterile Filtration Quality Risk Management (SFQRM) consortium in late 2017. The consortium goals have been to fill existing gaps in scientific data as adequately as possible with studies and industry guidance that would provide professionals and license holders with the ability to make informed decisions about appropriate risk-management strategies. This paper is one in a series of publications that are the result of the collaboration, and these should be considered together and viewed holistically in order to determine the best course of action with regard to PUPSIT. In total, the four papers cover the following areas: (1) data mining to determine the influence of fluid properties on integrity test values, (2) filter masking studies and results (this publication), (3) risk assessment and management from filter production to end use, and (4) points to consider in the best practice of the use of PUPSIT. In total, 25 manufacturers and filter suppliers have contributed to the work of the Consortium, deploying their filtration experts and pooling their collective knowledge and applied science experience to address these questions. This effort has also been supported by many independent experts currently available who have contributed to and driven the Filtration Interest Group in the PDA for many years. Both PDA and BioPhorum have prioritized this program and combined their approaches to deliver this comprehensive body of work. We hope that collectively the publications aid decision making and create greater certainty and confidence and above all alignment between suppliers, manufacturers, and regulators alike on these important questions.

Data Mining to Determine the Influence of Fluid Properties on the Integrity Test Values.

Eudralex volume 4, Annex 1, the European Union Good Manufacturing Practice for sterile products, requires that "The integrity of the sterilised filter should be verified before use" (1). Implicit in this requirement for a PUPSIT is the rationale that the sterilizing filter could sustain damage during sterilization or use (i.e., subsequent to any pre-use test conducted prior to sterilization), causing a defect which would not be detected by the post-use integrity ("masked" during filtration). To assess whether a filter defect could be masked by partial filter plugging, we evaluated the impact of the bacterial challenge test (BCT) on the bubble point (BP) of the test filters. The BP tests that are conducted before and after the BCT have been collected and compared for 2086 filters (1571 × test filters and 515 × control filters), representing 531 BCTs on 518 different pharmaceutical products, buffers, and in-process fluids. These tests comprise a cross section of fluids from multiple firms spanning the pharmaceutical and biotechnology industry. A posttest to pretest BP ratio was calculated for each filter and the distribution of these ratios examined to determine whether there were cases of elevation of the BP because of bacterial loading to the point where masking of a filter defect could occur; that is, if a defective filter could pass integrity testing due to apparent reduction in filter pore size because of the bacteria retained during the BCT. Ratios were averaged across all tests for the same test fluid. The mean average ratio was 1.00 ± 0.15, indicating that on the average, elevation of the BP does not occur. To assess the risk of masking a filter defect, observed BP ratios were compared to the ratio of the minimum BP specification of a 0.2 µm filter to that of a 0.45 µm filter of the same membrane type. The lowest such ratio for any membrane type was 1.33. A BP ratio equal to or higher than this ratio was considered a risk for masking, because a 0.45 µm filter could appear to meet the specifications of a 0.2 µm filter. Out of 518 average BP ratios, only eight fluids (1.5%) produced BP ratios meeting this criterion for a masking risk. Potential risk factors associated with these cases are discussed. We conclude that filtration processes producing BP changes sufficient to present a risk of masking defects are not common, and are detectable during the routine BCT. The BP ratios observed during routine BCT are one means to assess the potential of a given filtration process to mask defects and can be considered when determining whether a PUPSIT should be implemented.

Semantic Fisher Scores for Task Transfer: Using Objects to Classify Scenes.

The transfer of a neural network (CNN) trained to recognize objects to the task of scene classification is considered. A Bag-of-Semantics (BoS) representation is first induced, by feeding scene image patches to the object CNN, and representing the scene image by the ensuing bag of posterior class probability vectors (semantic posteriors). The encoding of the BoS with a Fisher vector (FV) is then studied. A link is established between the FV of any probabilistic model and the Q-function of the expectation-maximization (EM) algorithm used to estimate its parameters by maximum likelihood. This enables 1) immediate derivation of FVs for any model for which an EM algorithm exists, and 2) leveraging efficient implementations from the EM literature for the computation of FVs. It is then shown that standard FVs, such as those derived from Gaussian or even Dirichlet mixtures, are unsuccessful for the transfer of semantic posteriors, due to the highly non-linear nature of the probability simplex. The analysis of these FVs shows that significant benefits can ensue by 1) designing FVs in the natural parameter space of the multinomial distribution, and 2) adopting sophisticated probabilistic models of semantic feature covariance. The combination of these two insights leads to the encoding of the BoS in the natural parameter space of the multinomial, using a vector of Fisher scores derived from a mixture of factor analyzers (MFA). A network implementation of the MFA Fisher Score (MFA-FS), denoted as the MFAFSNet, is finally proposed to enable end-to-end training. Experiments with various object CNNs and datasets show that the approach has state-of-the-art transfer performance. Somewhat surprisingly, the scene classification results are superior to those of a CNN explicitly trained for scene classification, using a large scene dataset (Places). This suggests that holistic analysis is insufficient for scene classification. The modeling of local object semantics appears to be at least equally important. The two approaches are also shown to be strongly complementary, leading to very large scene classification gains when combined, and outperforming all previous scene classification approaches by a sizable margin.