Application of Artificial Neural Networks in the Process Analytical Technology of Pharmaceutical Manufacturing-a Review

Iterative Regression of Corrective Baselines (IRCB): A New Model for Quantitative Spectroscopy

In this work, a new model with broad utility for quantitative spectroscopy development is reported. A primary objective of this work is to create a novel modeling procedure that may allow for higher automation of the model development process. The fundamental concept is simple yet powerful even for complex spectra and is employed with no additional preprocessing. This approach is applicable for several types of spectroscopic data to develop regression models that have similar or greater quality than the current methods. The key modeling steps are a matrix transformation and subsequent feature selection process that are collectively referred to as iterative regression of corrective baselines (IRCB). The transformed matrix (Xtransform) is a linearized form of the original X data set. Features from Xtransform that are predictive of Y can be ranked and selected by ordinary least-squares regression. The best features (rows of Xtransform) are linear depictions of Y that can be utilized to develop regression models with several machine learning models. The IRCB workflow is first detailed by using a case study of Fourier transform infrared (FTIR) spectroscopy for prepared solutions of a three-component mixture. Next, IRCB is applied and compared to benchmark results for the 2006 "Chimiométrie" near-infrared spectroscopy (NIR) soil composition challenge and Raman measurements of a simulated nuclear waste slurry.

UV imaging for the rapid at-line content determination of different colourless APIs in their tablets with artificial neural networks

This paper presents a novel high-resolution and rapid (50 ms) UV imaging system, which was used for at-line, non-destructive API content determination of tablets. For the experiments, amlodipine and valsartan were selected as two colourless APIs with different UV induced fluorescent properties according to the measured solid fluorescent spectra. Images were captured with a LED-based UV illumination (385-395 nm) of tablets containing amlodipine or valsartan and common tableting excipients. Blue or green colour components from the RGB colour space were extracted from the images and used as an input dataset to execute API content prediction with artificial neural networks. The traditional destructive, solution-based transmission UV measurement was applied as reference method. After the optimization of the number of hidden layer neurons it was found that the relative error of the content prediction was 4.41 % and 3.98 % in the case of amlodipine and valsartan containing tablets respectively. The results open the possibility to use the proposed UV imaging-based system as a rapid, in-line tool for 100 % API content screening in order to greatly improve pharmaceutical quality control and process understanding.

Exploring the Potential of Artificial Intelligence as a Facilitating Tool for Formulation Development in Fluidized Bed Processor: a Comprehensive Review

This in-depth study looks into how artificial intelligence (AI) could be used to make formulation development easier in fluidized bed processes (FBP). FBP is complex and involves numerous variables, making optimization challenging. Various AI techniques have addressed this challenge, including machine learning, neural networks, genetic algorithms, and fuzzy logic. By integrating AI with experimental design, process modeling, and optimization strategies, intelligent systems for FBP can be developed. The advantages of AI in this context include improved process understanding, reduced time and cost, enhanced product quality, and robust formulation optimization. However, data availability, model interpretability, and regulatory compliance challenges must be addressed. Case studies demonstrate successful applications of AI in decision-making, process outcome prediction, and scale-up. AI can improve efficiency, quality, and cost-effectiveness in significant ways. Still, it is important to think carefully about data quality, how easy it is to understand, and how to follow the rules. Future research should focus on fully harnessing the potential of AI to advance formulation development in FBP.

Comparing the Performance of Raman and Near-Infrared Imaging in the Prediction of the In Vitro Dissolution Profile of Extended-Release Tablets Based on Artificial Neural Networks

In this work, the performance of two fast chemical imaging techniques, Raman and near-infrared (NIR) imaging is compared by utilizing these methods to predict the rate of drug release from sustained-release tablets. Sustained release is provided by adding hydroxypropyl methylcellulose (HPMC), as its concentration and particle size determine the dissolution rate of the drug. The chemical images were processed using classical least squares; afterwards, a convolutional neural network was applied to extract information regarding the particle size of HPMC. The chemical images were reduced to an average HPMC concentration and a predicted particle size value; these were used as inputs in an artificial neural network with a single hidden layer to predict the dissolution profile of the tablets. Both NIR and Raman imaging yielded accurate predictions. As the instrumentation of NIR imaging allows faster measurements than Raman imaging, this technique is a better candidate for implementing a real-time technique. The introduction of chemical imaging in the routine quality control of pharmaceutical products would profoundly change quality assurance in the pharmaceutical industry.

Application of an AI image analysis and classification approach to characterise dissolution and precipitation events in the flow through apparatus

Imaging and artificial intelligence (AI) approaches have been used with increasing frequency in pharmaceutical industry in recent years. Characterisation of processes such as drug dissolution and precipitation is vital in quality control testing and drug manufacture. To support existing techniques like in vitro dissolution testing, novel process analytical technologies (PATs) can give an insight into these processes. The aim of this study was to create and explore the potential of an automated image classification model based on image analysis to identify events (dissolution and precipitation) occurring in the flow-through apparatus (FTA) test cell, and the ability to characterise a dissolution process over time. Several precipitation conditions were tested in a USP 4 FTA test cell with images recorded during early (plume formation) and late (particulate re-formation) stages of precipitation. An available MATLAB code was used as a base to develop and validate an anomaly classification model able to detect different events occurring during the precipitation process in the dissolution cell. Two variants of the model were tested on images from a dissolution test in the FTA, with a view to application of the image analysis system to quantitative characterization of the dissolution process over time. It was found that the classification model is highly accurate (>90%) in detecting events occurring in the FTA test cell. The model showed potential to be used to characterise the stages of dissolution and precipitation processes, and as a proof of concept demonstrates potential for deep machine learning image analysis to be applied to kinetics of other pharmaceutical processes.

Real-time in-line prediction of drug loading and release rate in the coating process of diclofenac sodium spheres based on near infrared spectroscopy

The preparation of diclofenac sodium spheres by fluidized bed is a common production mode for the pharmaceutical preparations at present, but the critical material attributes in the production process is mostly analyzed off-line, which is time-consuming and laborious, and the analysis results lag behind. In this paper, the real-time in-line prediction of drug loading of diclofenac sodium and the release rate during the coating process was realized by using near infrared spectroscopy. For the best near infrared spectroscopy (NIRS) model of drug loading, R²_cv, R²_p, RMSECV, RMSEP were 0.9874, 0.9973, 0.002549 mg/g, 0.001515 mg/g respectively. For the best NIRS model of three release time points, the R²_cv, R²_p, RMSECV and RMSEP were 0.9755, 0.9823, 3.233%, 4.500%; 0.9358, 0.9965, 2.598%, 0.7939% and 0.9867, 0.9927, 0.4085%, 0.4726% respectively. And the analytical ability of these model was verified. The organic combination of these two parts of work constituted an important basis for ensuring the safety and effectiveness of diclofenac sodium spheres from the perspective of production process.

Predictive Models for Compound Binding to Androgen and Estrogen Receptors Based on Counter-Propagation Artificial Neural Networks

Endocrine-disrupting chemicals (EDCs) are exogenous substances that interfere with the normal function of the human endocrine system. These chemicals can affect specific nuclear receptors, such as androgen receptors (ARs) or estrogen receptors (ER) α and β, which play a crucial role in regulating complex physiological processes in humans. It is now more crucial than ever to identify EDCs and reduce exposure to them. For screening and prioritizing chemicals for further experimentation, the use of artificial neural networks (ANN), which allow the modeling of complicated, nonlinear relationships, is most appropriate. We developed six models that predict the binding of a compound to ARs, ERα, or ERβ as agonists or antagonists, using counter-propagation artificial neural networks (CPANN). Models were trained on a dataset of structurally diverse compounds, and activity data were obtained from the CompTox Chemicals Dashboard. Leave-one-out (LOO) tests were performed to validate the models. The results showed that the models had excellent performance with prediction accuracy ranging from 94% to 100%. Therefore, the models can predict the binding affinity of an unknown compound to the selected nuclear receptor based solely on its chemical structure. As such, they represent important alternatives for the safety prioritization of chemicals.

Interpretable artificial neural networks for retrospective QbD of pharmaceutical tablet manufacturing based on a pilot-scale developmental dataset

As the pharmaceutical industry increasingly adopts the Pharma 4.0. concept, there is a growing need to effectively predict the product quality based on manufacturing or in-process data. Although artificial neural networks (ANNs) have emerged as powerful tools in data-rich environments, their implementation in pharmaceutical manufacturing is hindered by their black-box nature. In this work, ANNs were developed and interpreted to demonstrate their applicability to increase process understanding by retrospective analysis of developmental or manufacturing data. The in vitro dissolution and hardness of extended-release, directly compressed tablets were predicted from manufacturing and spectroscopic data of pilot-scale development. The ANNs using material attributes and operational parameters provided better results than using NIR or Raman spectra as predictors. ANNs were interpreted by sensitivity analysis, helping to identify the root cause of the batch-to-batch variability, e.g., the variability in particle size, grade, or substitution of the hydroxypropyl methylcellulose excipient. An ANN-based control strategy was also successfully utilized to mitigate the batch-to-batch variability by flexibly operating the tableting process. The presented methodology can be adapted to arbitrary data-rich manufacturing steps from active substance synthesis to formulation to predict the quality from manufacturing or development data and gain process understanding and consistent product quality.

Application of Machine-Learning Algorithms for Better Understanding the Properties of Liquisolid Systems Prepared with Three Mesoporous Silica Based Carriers

The processing of liquisolid systems (LSS), which are considered a promising approach to improving the oral bioavailability of poorly soluble drugs, has proven challenging due to the relatively high amount of liquid phase incorporated within them. The objective of this study was to apply machine-learning tools to better understand the effects of formulation factors and/or tableting process parameters on the flowability and compaction properties of LSS with silica-based mesoporous excipients as carriers. In addition, the results of the flowability testing and dynamic compaction analysis of liquisolid admixtures were used to build data sets and develop predictive multivariate models. In the regression analysis, six different algorithms were used to model the relationship between tensile strength (TS), the target variable, and eight other input variables. The AdaBoost algorithm provided the best-fit model for predicting TS (coefficient of determination = 0.94), with ejection stress (ES), compaction pressure, and carrier type being the parameters that influenced its performance the most. The same algorithm was best for classification (precision = 0.90), depending on the type of carrier used, with detachment stress, ES, and TS as variables affecting the performance of the model. Furthermore, the formulations with Neusilin® US2 were able to maintain good flowability and satisfactory values of TS despite having a higher liquid load compared to the other two carriers.

Data-Driven Prediction of the Formation of Co-Amorphous Systems

Co-amorphous systems (COAMS) have raised increasing interest in the pharmaceutical industry, since they combine the increased solubility and/or faster dissolution of amorphous forms with the stability of crystalline forms. However, the choice of the co-former is critical for the formation of a COAMS. While some models exist to predict the potential formation of COAMS, they often focus on a limited group of compounds. Here, four classes of combinations of an active pharmaceutical ingredient (API) with (1) another API, (2) an amino acid, (3) an organic acid, or (4) another substance were considered. A model using gradient boosting methods was developed to predict the successful formation of COAMS for all four classes. The model was tested on data not seen during training and predicted 15 out of 19 examples correctly. In addition, the model was used to screen for new COAMS in binary systems of two APIs for inhalation therapy, as diseases such as tuberculosis, asthma, and COPD usually require complex multidrug-therapy. Three of these new API-API combinations were selected for experimental testing and co-processed via milling. The experiments confirmed the predictions of the model in all three cases. This data-driven model will facilitate and expedite the screening phase for new binary COAMS.