Development of novel image-based computational and artificial intelligence (AI) technologies for real time monitoring and event detection during in vitro dissolution testing

Abstract

To ensure safe and effective medications, in recent years the ever evolving pharmaceutical industry started to transform and to adopt novel technologies for continuous manufacturing, real time monitoring and quality control. From a development perspective, in vitro dissolution testing remains a vital drug performance test used to predict in vivo outcome. Strategies such as in silico simulations and animal models have been researched and applied in multiple situations and discussed broadly in the literature and were found suitable for early stage drug development. However, despite such promising results, their applicability for continuous manufacturing and process monitoring remains unknown. On the other hand, advances of imaging technologies like development of micro lens systems and non invasive imaging methods allow the implementation of such systems in routine manufacturing. They have been used as monitoring tools to ensure robust processing and medicinal products of high quality. The studies in this thesis were focused on development and application of image based analysis techniques to be used with standard dissolution apparatuses. The primary focus was on a non invasive imaging method for shadowgraph and direct imaging in the flow-through apparatus and some preliminary studies conducted using a commercially available imaging system in an alternative set up (paddle apparatus). Two imaging approaches were used with the non invasive system – a) shadowgraph imaging to characterise particulate behaviour in the images, and b) analysis of a whole frame using AI classification methods to characterise the dissolution process. Phenomena such as drug precipitation and agglomeration were observed and used for further analysis. A weakly acidic (pKa = 4.5) drug, ibuprofen, was used as a model drug to perform dissolution testing by variation of the sample mass and crystal habit of the drug powder. In the last part of this project, two brands of commercially available ibuprofen containing suspension formulations were used for dissolution in the flow through apparatus. Imaging and image analysis were used to capture differences in the dissolution rate, particle motion and suspension behaviour during testing. Imaging systems’ ability to detect deviations in the dissolution rate caused by changes in the environment such as addition of surfactant, change in the flow rate/paddle rotation speed, occurrence of drug precipitation, and the effect of excipients in the formulation, as well as changes related to the drug crystal habit and sample mass used for dissolution were tested. The work of D’Arcy and Persoons[1], [2] on image analysis based¬ on in house written code for shadowgraph imaging was expanded to investigate particle behaviour of sieved and un sieved particles with needle and plate like shapes and their agglomeration tendency. Agglomerates could have a negative impact on product stability, resulting in inaccurate dosing of formulations, e.g., oral, or parenteral suspensions. Such unfortunate events could require a recall of the drug product from the market and a withdrawal of the registration approval. As a proof of concept and by using the outputs of the image analysis, an automated agglomeration identification method was created. Such method showed potential to be further developed and utilized as an in line quality control tool as it was able to accurately capture agglomerates formed during dissolution testing. With the emergence of artificial intelligence (AI) smart monitoring of production lines could become a standard procedure. The United States Food and Drug Administration (FDA) recognizes the potential to apply AI in many areas of the pharmaceutical industry – from compound screening, through non clinical and clinical research, to post approval monitoring.[3] In this present work, an AI model for image analysis created and exploited for tablet surface defect detection was adapted for events detection in a dynamic environment, i.e., dissolution and precipitation processes. The method was successfully developed to provide quantitative and qualitative information that contributes to the general understanding of process and compound changes, which may affect the dissolution process. A recommended apparatus for dissolution testing of oral suspensions is the paddle apparatus[4], [5]. Although dissolution testing in the flow through apparatus showed a much slower dissolution rate compared to the recommended dissolution time in the paddle apparatus, the obtained results provided good discrimination between the two brands of suspensions. The image analysis of the dissolution tests identified differences in sample dispersal during testing which impacted on the dissolution profile. The analysis also showed notable differences in the suspensions’ dissolution behaviour, which could suggest different hydrophilicity or polarity of the excipients used in the formulation, or their impact on the balance between cohesion and adhesion between the formulation and the test vessel. Specific dissolution patterns were observed from each formulation. These product specifics could contribute towards the development of an AI tool able to distinguish between process and content changes, to assure batch to batch consistency in routine quality control, or even to participate in biowaiver decision making by providing high discrimination power and formulation related insights. All imaging and analysis approaches presented in this thesis provided additional insight into the dissolution process and drug behaviour, which could contribute to the general knowledge in the field. The imaging systems showed to be suitable to be used in combination with standard apparatuses in real time or for an offline image analysis. The shadowgraph imaging agglomerate identification method built during this project showed a potential to be further developed to recognize objects of interest. The AI based classification model was useful to provide qualitative information by using the pre trained settings. However, understanding of the AI process and opening the “black box” allowed further exploration of the system and the development of a quantitative classification method able to better describe the dissolution process. Thus, it could be suggested that there is unrealized AI potential remaining to be explored in the future.