Robustness of a continuous direct compression line against disturbances in feeding

Rinnakkaistallenteet Terveystieteiden tiedekunta

2020

Robustness of a continuous direct

compression line against disturbances in feeding

Karttunen, A-P

Elsevier BV

Tieteelliset aikakauslehtiartikkelit

© Elsevier B.V.

CC BY-NC-ND https://creativecommons.org/licenses/by-nc-nd/4.0/

http://dx.doi.org/10.1016/j.ijpharm.2019.118882

https://erepo.uef.fi/handle/123456789/7958

Downloaded from University of Eastern Finland's eRepository

Robustness of a continuous direct compression line against disturbances in feeding

A.-P. Karttunen, J. Poms, S. Sacher, A. Sparén, C. Ruiz Samblás, M.

Fransson, L. Martin De Juan, J. Remmelgas, H. Wikström, W-K. Hsiao, S.

Folestad, O. Korhonen, S. Abrahmsén-Alami, P. Tajarobi

PII: S0378-5173(19)30927-5

DOI: https://doi.org/10.1016/j.ijpharm.2019.118882

Reference: IJP 118882

To appear in: International Journal of Pharmaceutics Received Date: 18 September 2019

Revised Date: 12 November 2019 Accepted Date: 13 November 2019

Please cite this article as: A.-P. Karttunen, J. Poms, S. Sacher, A. Sparén, C. Ruiz Samblás, M. Fransson, L.

Martin De Juan, J. Remmelgas, H. Wikström, W-K. Hsiao, S. Folestad, O. Korhonen, S. Abrahmsén-Alami, P.

Tajarobi, Robustness of a continuous direct compression line against disturbances in feeding, International Journal of Pharmaceutics (2019), doi: https://doi.org/10.1016/j.ijpharm.2019.118882

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

© 2019 Elsevier B.V. All rights reserved.

Robustness of a continuous direct compression line against disturbances in feeding

A-P. Karttunen^a*, J. Poms^b, S. Sacher^b, A. Sparén^c, C. Ruiz Samblás^c, M. Fransson^c, L. Martin De Juan^c, J. Remmelgas^c, H. Wikström^c, W-K. Hsiao^b, S. Folestad^c, O. Korhonen^a, S. Abrahmsén-Alami^c, P. Tajarobi^d

aSchool of Pharmacy, University of Eastern Finland, Kuopio, Finland

bResearch Center Pharmaceutical Engineering GmbH, Graz, Austria

cPharmaceutical Technology & Development, AstraZeneca, Gothenburg, Sweden

dPharmaceutical Sciences, AstraZeneca, Gothenburg, Sweden

*Corresponding author:

Anssi-Pekka Karttunen

Address: University of Eastern Finland, School of Pharmacy, P.O. Box 1627, FI-70211, Kuopio, Finland

E-mail address: anssi-pekka.karttunen@uef.fi

Abstract

The aim of the current study was to characterize the robustness of an integrated continuous direct compression (CDC) line against disturbances from feeding, i.e. impulses of API and short step disturbances. These disturbances mimicked typical variations that can be encountered during long- term manufacture. The study included a primary formulation, with API of standard particle size, which was manufactured at 5 and 10 kg/h production rates, and a modified formulation with API of large particle size which was manufactured at 5 kg/h production rate. Overall, the CDC line smoothened all the disturbances, fulfilling the USP uniformity of dosage units (UDU) limit for single tablets. However, runs with the modified formulation failed the pharmacopoeia UDU requirements for the entire run due to high variation between tablets. The primary formulation passed the requirements in all cases. The residence time distribution (RTD) results indicated that the primary formulation allowed better smoothening ability, and an increase in production rate led to poorer smoothening due to shorter RTD. The RTDs revealed that a substantial part of back-mixing took place after the blender. Thus, the tablet press has an important role in smoothening disturbances longer than the mean residence time of the blender, which was very short.

Keywords

Continuous Blending, Direct Compression, Feeding disturbances, Process Analytical Technology, Near-Infrared Spectroscopy, Content Uniformity

1. Introduction

Nearly all pharmaceuticals are currently produced using batch manufacture schemes. However, there is a strong investment ongoing in the pharmaceutical industry to adopt continuous manufacture (CM) to boost the efficiency and quality of manufacture, especially amongst tablet products (Almaya et al., 2017; Brennan, 2016; DiProspero, 2018; Lee et al., 2015; Nasr et al., 2017;

Schaber et al., 2011; Yin and Clayton, 2014). To list some of the benefits, CM offers shorter lead- time in manufacture, enhanced and better monitored quality, deeper process understanding, flexibility in manufacture and smaller equipment footprint. When designed and implemented properly, the continuous process is operated under state of control (Nasr et al., 2017). This means that process parameters and quality attributes stay within previously approved ranges. Deviations from these ranges can occur, but they must be detected and non-conforming intermediate and product must be rejected.

One of the most critical quality attributes for solid oral dosages, such as tablets, is content uniformity (CU). Active substance content in all final dosages must be within a narrow distribution around the label claim (European Pharmacopoeia, 2018a; U. S. Pharmacopoeia, 2018). An adequate uniformity of content is essential to guarantee safety and efficacy for every dose (Karande et al., 2010). One challenge that can have implications on the CU in CM is the variation in feeding of the raw materials. Some realistic scenarios are sticking and later dislodging of material from e.g. feeder outlet or from a hopper at blender inlet, or deviation in feed rate caused by refill (Engisch and Muzzio, 2016, 2015). To provide a robust process, it is beneficial if the CM line possesses back-mixing to reduce these variations. We have previously shown that large short-term variations in the feed stream only had minor effect on the content of the final dosages (Lakio et al., 2017).

The requirement for uniformity of dosage units (UDU) testing has traditionally relied on two-tier procedure of measuring either 10 or 30 tablets and has been harmonized over the three major compendia (USP, Ph.Eur and JP). However, this traditional method is no longer supported as batch release method by FDA (FDA, 2016). The reason is that the traditional method does not include a statistical sampling plan and should therefore not be expanded to larger populations. New recommendations have been published to tackle this issue, such as the proposal from ISPE working group (Garcia et al., 2014) and a further technical discussion paper (Bergum et al., 2014). In addition to providing a statistical plan to assess UDU, this novel approach supports the use of larger sample sizes from e.g. Process Analytical Technology (PAT) tools. An alternative approach to asses CU has been proposed using a parameter tolerance interval test where a typical test assures with e.g. 95 % confidence that 87.5 % of dosages are within 85-115 % label claim (Goodwin et al., 2018; Shen et al., 2014; Tsong and Shen, 2007). An approach to enable the use of PAT tools for UDU assessment has also been provided by Ph.Eur with the creation of chapter 2.9.47. (European Pharmacopoeia, 2018b).

The implementation of PAT tools is usually combined with the adoption of continuous processes.

PAT makes it possible to increase the amount of information gained on the process during development, provides product quality monitoring during manufacture and ultimately enables real- time release testing for the product (Allison et al., 2015; FDA, 2004; Lee et al., 2015; Pawar et al., 2016). Spectroscopic techniques, such as near-infrared (NIR) and Raman spectroscopy, are some of the usual PAT tools that can be implemented to continuous processes (Alam et al., 2017; De Beer et

al., 2008; De Leersnyder et al., 2018; Goodwin et al., 2018; Järvinen et al., 2013; Martínez et al., 2013; Nagy et al., 2017). Both techniques have also proven their benefits as off- and at-line measurements compared to wet chemistry analysis as the measurements can be performed without time-consuming sample preparation, allowing to analyze larger amounts of samples (Bonawi-tan and Williams, 2004; Dave et al., 2017; Paudel et al., 2015). NIR-chemical imaging (NIR- CI) is a further application of NIR that has been used to monitor final tablets (Khorasani et al., 2015;

Wahl et al., 2017). In both NIR and Raman spectroscopy, the basic principle relies on measuring the absorption bands in the spectrum, caused by molecular vibrations (Siesler, 2002). However, NIR is a single point spectroscopic technique that does usually not directly reflect distribution of physical or chemical properties, because a bulk NIR spectrum represents an average composition of the sampled area. Implementation of NIR-CI together with chemometric methods as process monitoring solution provides the information necessary to develop a fast and accurate approach for both qualitative and quantitative characterization of physical and chemical properties of measured samples. The pixel information of NIR-CI makes it possible to assess the distribution of the physical and chemical properties in a sample (Reich, 2005). On the other hand, Raman shows higher structural specificity to different materials since it measures the fundamental vibrations leading to sharp peaks in the measured spectrum. Applications of Raman spectroscopy in pharmaceutical analysis have surged in the past three decades as laser sampling and detector technology has improved, therefore it has become a practical analysis technique inside and outside the laboratory (Esmonde-White et al., 2017). Naturally, the best choice for each application should be evaluated and implemented for each individual case.

The aim of the current study was to characterize the robustness of a continuous direct compression (CDC) line against disturbances in feeding. The disturbances, i.e. pulses of additional API and steps in API feeding, were aimed to mimic typical variations that could be encountered during manufacture and their magnitude was chosen accordingly. The pulses mimicked additional API peaks in feeding due to sticking and later dislodging of API from feeder outlet. The step disturbances resemble transient offset from target feed rate that takes place during feeder refill. The key quality attributes were content and content uniformity in the blend and the tablets. These were measured with NIR and Raman spectroscopy. Content uniformity was assessed against pharmacopoeia requirements. The disturbance data was used to calculate residence time distributions (RTD) of the unit operations. These enabled insight into the roles of different process units on the back-mixing capability of the line and estimations of magnitude and length of a disturbance to result in tablets not fulfilling the pharmacopoeia requirements. Similar studies focusing on the robustness of a full continuous direct compression line using commercial formulation were difficult to find in the literature. A few studies with a simulation oriented approach have been presented (Engisch and Muzzio, 2015; García-Muñoz et al., 2018). There is a great number of papers that focus on the relationship between process parameters or material properties and RTD of unit operations with the implications that RTD has on the quality of the product. However, studies that are based on fully integrated lines, including all piping and interfacing, are less frequent.

2. Materials and methods

2.1. Formulations

The formulations in this study were selected from those used in the previous studies in the same line (Karttunen et al., 2019b; Lakio et al., 2017). The materials, their respective grades and formulation compositions are shown in Table 1. Both formulations included paracetamol as a model API with cohesive nature, two fillers, a disintegrant and a lubricant. The primary formulation (formulation A) provided the most robust and consistent tablet quality in previous studies. For the current study this formulation was slightly improved by pre-blending the standard (std) API with silicon dioxide (SiO2) to enhance the flow. The std API and SiO2 were blended with a high shear blender (Lödige MGTL 5/15, Lödige Maschinenbau GmbH, Paderborn, Germany) in 2.5 kg batches (99.5 % API, 0.5 % SiO2) at 320 rpm for 10 min. The study design also included a modified formulation in which the std API was changed to a larger granulated grade (formulation B). In all runs, MCC and L-HPC were used as a premix blended with the same continuous blender and the same feeder type as was used in the actual runs (mixing speed 1200 rpm, MCC feed rate 20 kg/h, L-HPC feed rate 3.2 kg/h).

Table 1: Formulation compositions and material suppliers Material (abbreviation) , trade

name/grade Supplier

Amount (% w/w), Formulation A

Amount (% w/w), Formulation B Paracetamol, standard (std API) Chemtronica HK Company

limited, Hong Kong, China 22 -

Paracetamol, granulated (gran API)

Mallinckrodt

Pharmaceuticals, St. Louis, MO, USA

- 22

Mannitol, Pearlitol 200SD Roquette, Lestrem, France 46.9 47 Silicon Dioxide, Fumed (SiO2),

CAB-O-SIL M-5P

Cabot GmbH, Rheinfelden,

Germany 0.1 -

Microcrystalline cellulose

(MCC), Avicel PH200 FMC BioPolymer, Ireland 25

Low-substituted hydroxypropyl cellulose (L-HPC), LH-B1

Shin-Etsu Chemical Co. Ltd.,

Tokyo, Japan 4

Sodium stearyl fumarate, PRUV Moehs, Barcelona, Spain 2

2.2. Continuous Direct Compression (CDC) line and PAT sensors

The top-down CDC line of the study consisted of feeding, blending and tableting units (Figure 1A).

The feeders and the blender were adjusted and monitored by an in-house LabVIEW-based software (LabVIEW 12.0, National Instruments, Austin TX, USA). Formulation A of the study was manufactured at a total powder flow rate of 5 and 10 kg/h, and formulation B at 5 kg/h only. The

CDC line has been previously used in same configuration in several studies (Ervasti et al., 2015; Lakio et al., 2017) and is therefore only summarized briefly.

Figure 1: (A) The top-down CDC line and near-infrared (NIR) sensing tools, (B) Powder flow chute and NIR probe after blender exit, (C) Near-infrared chemical imaging (NIR-CI) set-up after tablet press

Raw materials were fed with loss-in-weight feeders (Coperion K-Tron, Niederlenz, Switzerland) into a continuous blender (Hosokawa, Modulomix, Micron, Doetinchem, The Netherlands). The blender was operated at 500 rpm in all runs. Mannitol and MCC/L-HPC pre-blend were fed with a K-ML-D5- KT20 feeder, API with K-CL-SFS-KT20 and PRUV with K-CL-SFS-MT12. The mass flow (kg/h) of each feeder was acquired once per second by the monitoring software. The blend was transported by gravity from the blender into the feed tube (diameter 5 cm) of the tablet press (PTKPR1000, PTK CO., Ltd, Incheon, Korea). A chute was placed right after the blender to acquire NIR spectra through a sapphire window using an in-line NIR spectrometer (SentroPAT FO, Sentronic, Dresden, Germany) with an optical fiber connected probe (DR LS, Sentronic, Dresden, Germany) mounted on the chute (Figure 1B). The target tablet weight was 300 mg, compacted at 28 (5 kg/h) or 56 (10 kg/h) rpm turret speed using 10 pairs of 10 mm diameter flat punches in the press. Before the start of tableting, approximately 500 g of powder was collected into the tablet press feed frame and feed tube. The level of material in the feed tube was monitored visually during the runs and maintained by adjusting the tablet fill depth slightly, as needed. The feed frame was operated at 150 % of turret speed. After compression, the tablets slid down to a conveyor belt that transported the tablets through an NIR-CI spectrometer (EVK Helios G2-320 Class, EVK DI Kerschhaggl GmbH, Graz, Austria) (Figure 1C).

2.3. Calibration runs and reference analysis

One calibration run was performed for each formulation-flow rate pair to produce reference samples for PLS model building with NIR, NIR-CI and Raman to predict API content in the blend and tablets. Calibration samples (blend and tablets) were manufactured for concentration levels 12, 17, 20, 22, 24, 27 and 32 % (w/w). The center point concentration 22 % (w/w) was included in triplicate.

Different concentration levels were achieved by changing the set points of the feeders while maintaining the same total flow rate. Blend calibration samples (n=1) were collected from the outlet of the NIR chute and tablet samples (n=6) were collected after the tablet press. The tablets were then separately analyzed with the NIR-CI equipment, and the same tablets were used for transmission Raman calibration. The reference method for both the blends and tablets was UV/Vis spectroscopy. Calibration samples were dissolved in phosphate buffer (pH 5.8), filtered, diluted and measured with a UV/Vis spectrometer (Shimadzu UV-1800, Shimadzu Suzhou Instruments Wfg. Co.

Ltd., China) at a wavelength of 242 nm.

2.4. In-line NIR measurements and model building

NIR spectra from the blend were collected from the wavelength range between 1100 and 2200 nm using 20 ms integration time and averaging of 40 spectra, using the SentroSuite 3.4 software (Sentronic, Dresden, Germany). These settings were considered to be the best compromise between decent attenuation of noise and sufficiently frequent sampling rate. With these settings, one sample was recorded roughly every 1.2 s. The wavelength range between 1160 and 2160 nm was used to build a PLS model for API concentration range between 12 and 32 % (w/w), using Simca (v14.0, Umetrics AB, Umeå, Sweden). However, there was still a substantial amount of noise between spectra. Therefore, a moving average of three consecutive spectra was used for the model and predictions. In the final model, the spectra were pre-treated with 1^st derivative window 31 cubic fitting followed by standard normal variate (SNV) transformation followed by mean centering. In the calibration, each reference sample was matched with the spectra measured during sample collection with additional two spectra before and after (i.e. each sample was matched with five consecutive spectra). Since only one sample per concentration level was available, prediction error was quantified with root-mean-square error of cross-validation (RMSECV). RMSECV in API content prediction for formulation A at 5 kg/h was 0.79 (% w/w) with an R² of 0.99. The corresponding values were RMSECV 0.45 (% w/w) and R² 1.00 for formulation A at 10 kg/h, and RMSECV 0.64 (% w/w) and R² 0.99 for formulation B at 5 kg/h. Each model was fitted with 4 PLS components. It is noteworthy that the scale of scrutiny in the NIR measurement for blend is different than for NIR-CI and Raman measurements. The two latter measurements were conducted for single tablets but each NIR result provides a measurement for a mass which corresponds to several tablets. The scale of scrutiny for blend was intentionally different than for tablets. The blend NIR measurement provides only the possibility to monitor the trends seen in content at this stage with reasonable smoothening of the data as described above.

NIR-CI spectra from tablets were acquired with an EVK Helios push-broom camera, model G2 class 320. A detailed description of the system is presented in Wahl et al. (2017) and only a short summary is given here. The spectra were collected from the wavelength range 989 – 1660 nm (220 pixel resolution in wavelength) with a frame rate of 334 Hz as raw 12 bit data stream. A 35 mm objective lens was pointing down, covering the central part of a conveyer belt (312 pixel resolution covering 180 mm). All post processing of data was performed in Matlab 2016b (MathWorks, Natick, USA) with in-house scripts. A PCA model and thresholding was used to generate input for image recognition of round objects of tablet sizes against background. The raw spectra of recognized tablets were extracted and stored in databases for enhanced data handling. The spectra were corrected for black and white reference, converted to absorbance spectra and treated with SNV normalization. Binary masks were used to extract the same amount of 297 spectral pixels per tablet, to sample tablets uniform and with the same amount of edge pixels and a mean spectrum calculated. A PLS model was built using the calibration tablets with concentration levels from 12 to 32 %. A separate PLS model was created for each formulation-flow rate pair. The root-mean-square error of prediction (RMSEP) for formulation A at 5 kg/h was 0.68 (% w/w) with an R² of 0.96. The corresponding values were RMSEP 0.39 (% w/w) and R² 0.99 for formulation A at 10 kg/h, and RMSEP 1.09 (% w/w) and R² 0.93 for formulation B. Each model was fitted with 5 PLS components.

2.5. Transmission Raman measurements

Transmission Raman measurements were performed with a Cobalt Light System TRS100 (Agilent, Santa Clara, United States). All tablets were measured for 0.3 sec, using 200 accumulations with a laser power of 500 mW and a 4 mm laser spot size. Multivariate analysis of spectra was done using PLS Toolbox for Matlab (Eigenvector, Manson WA, United States). A PLS calibration model with 162 samples, which included variation of seven API concentrations (12 to 32 %), two particle sizes and two flow rates, was developed for prediction of API. Automatic weighted least squares, with order two, was applied on the raw spectra for baseline correction as well as mean centering. The performance of the calibration model was evaluated using the root mean square error of prediction (RMSEP) and the bias of an independent test set, with tablets that were not included in the calibration model. The RMSEP of the model was 0.95 (% w/w) with an R² of 0.96. A total number of 252 tablets selected from the disturbance runs were predicted with the PLS model. The tablets were selected around both the resulted deviations and baseline (based on in-line NIR-CI data), to confirm the events seen in in-line data.

2.6. Disturbance runs

The API feeding was subjected to two types of intentional disturbances: a pulse disturbance and a step disturbance. In the pulse disturbance runs, the pulse sizes (g) were related to API feed rate per second, and were roughly 300, 600, 900 and 1800 %. The sizes of pulses and theoretical increases in content are summarized in Table 2. The pulses were manually poured into the blender inlet as instantaneous pulses. The size range of pulses was selected so that the events could potentially take place in actual manufacturing. In the step disturbance runs the step size was 30 % increase or decrease in the API feeder feed rate and the steps had a length of 10, 30 and 60 s. The length of the steps was chosen so that the first step represents a short transient disturbance which would normally be corrected by the feeder after detection and does not last longer than 10 s. The longer steps were chosen to represent the lengths which would be encountered especially during feeder refills. In addition to these three steps, the API feed rate was increased with 30% for the last 12 min of the step disturbance runs to allow for an RTD measurement with the step change method. Tablet samples for Raman measurements were collected every 10 s at the end of the conveyor belt. The sampling started in each run at the same time as the first disturbance was initiated and sampling lasted until the end of the run.

Table 2: Sizes of pulse disturbances and theoretical increases in content Size in relation to

normal API feed (%)

Size of pulse (g) at 5 kg/h

Size of pulse (g) at 10 kg/h

Theoretical concentration (% w/w)

300 0.9 1.8 52.5

600 1.8 3.6 65.9

900 2.7 5.4 73.4

1800 5.4 10.8 83.9

2.7. Tablet data evaluations

Tablet content results from the NIR-CI were used to test uniformity of dosage units (UDU) in all runs against the pharmacopoeia requirements in both Ph.Eur and USP (European Pharmacopoeia, 2018a;

U. S. Pharmacopoeia, 2018). The newer Ph.Eur chapter 2.9.47 ‘Demonstration of uniformity of dosage units using large sample sizes’ (European Pharmacopoeia, 2018b) was used in the test against Ph.Eur and the acceptance table approach proposed by Bergum et al (2014) in the test against USP. In addition, a requirement for single units (within ± 25 % of target) was applied according to USP <905> UDU chapter.

2.8. Residence time distribution estimation and modeling

Residence time distributions (RTD) were calculated for each formulation-flow rate pair using both pulse disturbances and the last (12 min) step data. The RTD, noted as E(t), from a pulse input is calculated by normalizing the concentration profile of the pulse by the area underneath the curve according to equation 1. The last step (12min step) in the step runs (see supplementary) has the shape of a cumulative distribution function F(t), and has to be first normalized to go from 0 to 1.

Then, the RTD can be derived from the F(t) function according to equation 2.

𝐸(𝑡) = ^𝐶(𝑡)

∫ 𝐶(𝑡)₀^∞ (1)

𝐸(𝑡) = ^{𝑑 𝐹(𝑡)}_𝑑𝑡 (2)

From E(t), mean residence time (MRT or τ) and mean centered variance (MCV or στ2) can be calculated. Mean residence time is the first moment of the RTD that quantifies the center of the distribution, i.e. how long on average the particles reside inside the unit in question. Mean centered variance can be used to quantify the back-mixing efficiency of the unit in question. τ can be calculated according to equation 3. For the calculation of the MCV, one has to first calculate variance, σ², of the RTD according to equation 4. MCV, στ2, can then be calculated according to equation 5.

𝜏 = ∫ 𝑡𝐸(𝑡)𝑑𝑡₀^∞ (3)

𝜎² = ∫ (𝑡 − 𝜏)₀^{∞ 2}𝐸(𝑡)𝑑𝑡 (4)

𝜎_𝜏² = ^𝜎_𝜏2² (5)

In this study, tanks in series (TIS), tanks in series plus plug flow reactor (TIS+PFR) and axial dispersion models were all tested to fit the raw data (blend or tablet NIR content prediction). All models yielded practically the same root mean square error in all cases, and the differences in predicted residence times were minimal. The TIS model was chosen to model the RTDs in this study. RTDs were fitted with TIS according to equation 6. In equation 6, n is the number of tanks and τ is the mean residence time. Values for n were estimated by a non-least squares method using the trust region reflective

algorithm in Python (Python 3.6.5, Anaconda, Inc., USA). The variance, σ², was calculated according to equation 7, with the fitted values from equation 6. When the residence time distribution is known, the output concentration of the system can be calculated for different input concentrations, e.g. disturbances, according to equation 8.

𝐸(𝑡) =^𝑛(

𝑛𝑡 𝜏)^𝑛−1

𝜏(𝑛−1)! 𝑒𝑥𝑝 (−^𝑡𝑛_𝜏) (6)

𝑛 =_𝜎^𝜏2₂ (7)

𝐶_𝑜𝑢𝑡(𝑡) = 𝐶_𝑖𝑛(𝑡) ∗ 𝐸(𝑡) (8)

When multiple process units are linked together, the MRT and variance of the whole system are the sum of the unit MRT and variance. Since the content was measured after blending and at the end of the line, the MRT and variance of the tablet press were calculated using equations 9 and 10.

𝜏𝑡𝑎𝑏𝑙𝑒𝑡 𝑝𝑟𝑒𝑠𝑠 = 𝜏_{𝑙𝑖𝑛𝑒}− 𝜏_{𝑏𝑙𝑒𝑛𝑑𝑒𝑟} (9) 𝜎𝑡𝑎𝑏𝑙𝑒𝑡 𝑝𝑟𝑒𝑠𝑠2 = 𝜎_{𝑙𝑖𝑛𝑒}² − 𝜎_{𝑏𝑙𝑒𝑛𝑑𝑒𝑟}² (10)

3. Results

As described in materials and methods, the pulse and step disturbances were deliberately introduced in the feed stream entering the blender. The two main units in the top-down CDC line downstream of the feeders are the blender and the tablet press. Thus, the smoothening of the disturbances was expected to mainly take place in these units. In this section, the responses of the CDC line to the two different types of disturbances are first inspected separately for each type of disturbance. Then residence time distribution (RTD) data generated from these disturbances are examined and limits for allowable disturbances are calculated from the RTD data. In what follows, the API concentration displayed for feeding is theoretical; actual feeder data is shown in the supplementary data. In addition, the NIR and NIR-CI results shown in figures were low-pass filtered with a 30 mHz cut-off frequency in Matlab (Matlab 2017b, The MathWorks, Natick, USA) to better visualize the trends in API content. The tabulated maximum deviations are calculated from individual tablet results.

3.1. Pulse disturbance runs

The predicted API contents for the blend and the tablets in the pulse disturbance runs are illustrated in Figure 2 and the maximum API content deviations in tablets (% w/w) from preceding baseline concentration are compiled in Table 3. The time-axis in Figure 2 is truncated from the beginning.

The data shown start 5 min before the second pulse for all runs. At this point, the process seemed to be in steady-state. In addition, the first pulse was relatively small in all runs and did not induce a

well detectable deviation in tablets. The full runs are shown in supplementary data. As shown in Figure 2, formulation A at 10 kg/h showed the clearest response to the disturbances both in blend and in tablets. This should be partially attributed to this run exhibiting the smallest baseline variation in predicted tablet contents, and smallest error of prediction for NIR-CI. In addition, this run resulted also in the shortest mean residence time (see chapter 3.3) for tablet press which should lead to lower ability to smoothen disturbances. At the blend stage the flow inside the chute was stable at this higher flow rate which contributed to a more consistent sample presentation over the NIR probe in comparison to 5 kg/h runs. However, all the runs showed a lot of variation in API content at blend stage between the intentional disturbances. This is most likely a result of the NIR interface (see Figure 1) which may not be the optimal set-up to produce a robust sample presentation to the probe. The interface also most likely allows sifting segregation to take place. This is seen especially for formulation B due to large particle size difference with the API and major filler. This should, at least partially, explain a higher baseline variation in tablet contents for formulation B. Due to this baseline variation, the pulses were not detected as well from the data as with formulation A. The tablets analyzed with UV/Vis for the NIR-CI model building also exhibited higher variation between tablets with formulation B than with formulation A. This provides assurance that it is not just a result of a poorer prediction model. The results are in line with our previous papers showing that the variation in the API content between tablets during steady processing was higher in formulation B than in formulation A(Karttunen et al., 2019b; Lakio et al., 2017).

In addition to the NIR-CI predictions, selected tablets (3 tablets per selected time point) were also analyzed with transmission Raman spectroscopy. The results from the transmission Raman analysis were similar to the NIR-CI predictions although there was a bias between the two methods for formulation A. The NIR-CI predictions of the API content were consistently lower than the Raman predictions. For formulation B, no bias was observed but the Raman predictions exhibited a larger variation than for formulation A, although the averages of Raman predictions followed similar changes in API content as NIR-CI predictions. More detailed comparisons between NIR-CI and Raman are provided in supplementary data. In addition to the bias for formulation A, the NIR-CI predictions exhibited consistently higher variation than the Raman predictions. Raman is assumed to provide more accurate results since it was performed for static samples and in transmission mode (i.e. scanning through the tablet and averaging multiple scans). It has also been postulated that Raman predictions can provide as reliable results as traditional wet chemistry (Bonawi-tan and Williams, 2004). However, the focus in this paper is on the NIR-CI data, since this technique was applied in-line, providing larger coverage of tablets for the entire run, and as compared above the deviations were followed with both techniques.

The resulting magnitudes of change in tablet content due to induced disturbances (Table 3) were quite similar for formulation A. For formulation B there was a lot of variability between tablets in general which led to larger deviations and smaller differences between the induced disturbances.

Overall, the resulting deviations in tablets for formulation A were well within the ± 25 % limit for individual dosage units as required by USP <905> UDU test. It is also worth noting that the largest, 1800%, pulses caused the content after the blender to exceed this limit (Figure 2), but these deviations were smoothened out in the tablets which implies that the tableting step also possesses back-mixing capability which led to further smoothening of the disturbances. With formulation B there were some individual tablets that deviated more than + 25 % from target which means that

the USP UDU test criteria was not met with this formulation. Out of specification tablets occurred during both normal processing and disturbances. There is a slight discrepancy in the results for the 600 % and 900 % pulses for formulation A at 10 kg/h. The resulting change in content for the 600 % pulse was larger than for the 900 % pulse. This could, at least partially, be a result of slight overfeeding of API around the time of 600% pulse (see actual feeder data in supplementary).

However, the difference between the values is small and within the error of prediction for NIR-CI.

Other than this, results for formulation A follow a logical trend that the higher the disturbance, the higher is the deviation. With formulation B there was large variation between tablets which made it hard to identify the deviations from content predictions of single tablets. In the smoothened trends in Figure 2 the deviations due to disturbances seem to be roughly the same size as for formulation A. In general, the results suggest that the integrated CDC line as a whole was robust to tolerate the short pulse disturbances.

The results for the UDU testing against pharmacopoeia limits are compiled in Table 4. Although formulation B would already fail the USP UDU test due to deviations of individual tablets, the acceptance criteria for the general variability was still inspected. The run with formulation B did not pass these limits even if the disturbances were excluded from the data. This further verifies the larger baseline variability in API content compared to formulation A. A part of this larger variation can be a result of the poorer NIR-CI model, but the Raman results showed increased variation for formulation B as well (Figure 2). There is a good chance that sifting segregation takes place in the NIR chute between blender and tablet press. Segregation might also take place during the drop from the NIR chute to the tablet press feed tube. Segregation can, at least partially, explain the larger baseline variability in tablets for formulation B. Both runs with formulation A passed the UDU test according to both Ph.Eur and USP requirements.

Table 3: Maximum measured deviation in tablet content (% w/w) from the preceding baseline due to the pulse disturbances. The theoretical concentration (% w/w) in feeding during the pulses is shown in brackets below the pulse size. Nominal concentration is 22 % (w/w).

Formulation- flow rate

Pulse size (%)

600 (65.9 %)

900 (73.4 %)

1800 (83.9%)

Formulation A at 5 kg/h 0.0 2.2 3.4

Formulation A at 10 kg/h 2.7 2.2 2.9

Formulation B at 5 kg/h 4.0 4.4 4.4

Table 4: Results for uniformity of dosage unit testing for pulse runs¹ against pharmacopoeia limits Formulation-flow rate Ph.Eur (AV-value)* USP (Acceptance table)**

Incl. deviations Excl. deviations Incl. deviations Excl. deviations

Formulation A at 5 kg/h 8.2 7.7 Pass Pass

Formulation A at 10 kg/h 7.9 7.6 Pass Pass

Formulation B at 5 kg/h 19.9 19.1 Fail Fail

1Calculated at stable process conditions, i.e. start-up excluded

*Ph.Eur 2.9.47Uniformity of dosage units testing with large n, AV-value requirement <15

**According to recommendation by Bergum et al (2014)

Figure 2: API content in the powder blend and the tablets during the pulse disturbance runs. (A) Formulation A at 5 kg/h, (B) Formulation A at 10kg/h, (C) Formulation B at 5 kg/h.

3.2. Step disturbance runs

The predicted API contents in the step disturbance runs for all formulation-flow rate pairs are illustrated in Figure 3 and the maximum API content deviations in tablets (% w/w) from preceding baseline are compiled in Table 5. In Figure 3, the x-axis is truncated from the beginning. The data shown start 5 min before the first step for all runs. At this point the process seemed to be in steady state. The last 12min step for the calculation of RTD is also omitted from the figure. The full runs are shown in supplementary data. According to maximum deviations in Table 5, formulation B exhibited the largest deviations due to the disturbances. However, the magnitude of the deviations is also increased due to the large variation between tablets. For formulation A, the higher flow rate led to larger deviations for 30 s and 60 s disturbances. For the shortest step there was no real difference between flow rates. According to the trends in Figure 3, formulation A exhibits the largest and clearest deviations due to the disturbances, especially for the 60s step. The reasons for this were already discussed in the previous chapter, and are thus not repeated here. There is one discrepancy in the tablet content results; for formulation A at 5 kg/h the 10 s step leads to a larger deviation in tablet content than the 30 s step although the deviations in content at blend stage were as expected.

However, no clear explanation for the discrepancy was found. In blend API content, the disturbances were clearly seen for all formulation-flow rate pairs. However, there is again a lot of variation in the blend content predictions due to the same issue as described for pulse runs.

It seems clear that the ability of the CDC line to tolerate long deviations decreases with increase in flow rate. The mass hold-up in the blender increased only slightly (roughly from 40 to 55 g) when the flow rate was doubled and the mass hold-up in the tablet press remained the same as the feed frame and hopper were filled to same extent in both production rates. This leads to significantly shorter residence time in the process (see chapter 3.3), which subsequently leads to poorer ability of the line to dampen the disturbances. This can be relevant at large volume commercial production if there is further decrease in robustness when production rates are elevated from the rates tested in the current study.

Overall, the resulting deviations in tablets for formulation A were well within the ± 25 % limit for individual dosage units, as required by USP <905> UDU test. Theoretically, if the content during

steady processing is exactly in target, the largest individual deviations in tablets could be 5.5 % (w/w), with the drug load in question. The expectation was that at least some of the individual tablets would have exceeded the limit, especially with formulation B. For formulation B, the individual tablet contents were very generally close to the upper (+ 25 %) limit, although within the limits. The maximum deviations in tablets for the last step would have exceeded the limit if that step would have been an increase in content instead of a decrease since the tablet contents were generally a bit too high. Again, the transmission Raman results verified the NIR-CI predictions, although the same bias between the methods for formulation A existed, and NIR-CI predictions for formulation B contained more noise. The results for the UDU testing against pharmacopoeia limits are compiled in Table 6. The run with formulation B did not pass the pharmacopoeia requirements but both runs with formulation A did. Again, the variation between tablets for formulation B led to failure to meet the UDU requirements.

Table 5: Maximum measured deviations (% w/w) from preceding baseline in tablets due to the step disturbances

Formulation- flow rate

Step length (s)

10 30 60

Formulation A at 5 kg/h 2.1 1.8 2.2

Formulation A at 10 kg/h 1.9 2.5 3.7

Formulation B at 5 kg/h 3.9 4.1 5.3

Table 6: Results for uniformity of dosage unit testing for step runs¹ against pharmacopoeia limits Formulation-flow rate Ph.Eur (AV-value)* USP (Acceptance table)**

Incl. deviations Excl. deviations Incl. deviations Excl. deviations

Formulation A at 5 kg/h 11.2 10.2 Pass Pass

Formulation A at 10 kg/h 9.5 8.1 Pass Pass

Formulation B at 5 kg/h 17 17 Fail Fail

1Calculated at stable process conditions, i.e. very beginning and last 12 minutes excluded

*Ph.Eur 2.9.47Uniformity of dosage units testing with large n, AV-value requirement <15

**According to recommendation by Bergum et al (2014)

Figure 3: API content in the powder blend and the tablets during the step disturbance runs. (A) Formulation A at 5 kg/h, (B) Formulation A at 10kg/h, (C) Formulation B at 5 kg/h.

3.3. Residence time distributions and back-mixing effect of the equipment

Residence time distributions (RTD) for the blender (Figure 4) and the entire line (Figure 5) were calculated using both pulse and step disturbance results. Mean residence times (MRT), mean centered variances (MCV) and the number of tanks in the TIS model fit are compiled in Table 7. The RTDs in blender were very short which is typical for this blender, as presented also in previous

papers for other formulations using the same blender and range of processing rates (Karttunen et al., 2019a; Martinetz et al., 2018; Rehrl et al., 2018). In comparison, the RTDs of other continuous blenders from e.g. Glatt and GEA are considerably longer (Osorio and Muzzio, 2016; Van Snick et al., 2017b, 2017a). MCV is a value that can be used to compare units for their back-mixing efficiency, i.e. smoothening ability (Vanarase and Muzzio, 2011). The higher the MCV, the higher is the amount of back-mixing. The MCV values (from pulse data) for the blender with formulation A, were in the same range as for other continuous blenders, and indicate that the blender contributes to the smoothening ability of the CDC line as expected (Van Snick et al., 2017a; Vanarase and Muzzio, 2011). However, the results for the blender with formulation B indicate less back-mixing than with formulation A. Some of the latest continuous blenders have higher MCV values than the ones exhibited by our blender (Osorio and Muzzio, 2016), indicating even better smoothening ability that increases the tolerance towards disturbances. It is noteworthy, that the MCV results for formulation A were rather different between pulse and step results whereas formulation B yielded rather comparable results from both pulse and step data. In the current study, the variance was retrieved from the TIS model fit using mean residence time and number of tanks, n. As n was rather large in the fit for formulation A in the step runs, the calculated variance was very small. This can be a result of a change in flow properties of the formulation due to the step which has influenced the measurement.

The MCV values of the tablet press, as a result of back-mixing in feed frame (and hopper), were in most cases even larger than those of the blender. This indicates that the tablet press contributes strongly in the smoothening of the disturbances. The main purpose of the feed frame in tablet press is to provide consistent filling of the dies but it also clearly provides additional back-mixing. As the RTDs of the tablet press are considerably wider than those of the blender, the role of the tablet press in the smoothening of the disturbances is equal to or more important than the role of the blender. This is especially important for the smoothening of disturbances that last longer than the mean residence time of the blender. It is noteworthy that there was some variation in the RTDs of tablet press (Table 7). This can be due to a change in fill level in the tablet press feed tube or due to variability in mixing inside the tablet press. Thus, a confidence interval that considers this variation, and baseline variation in tablets, is essential when making decisions on tracking and rejection of material based on RTDs. There was also a considerable difference between the RTD values calculated from pulse input and step input. Two hypotheses for the reason of the difference can be postulated. Firstly, the step change might have influenced the flow inside the process, which should not happen during an RTD measurement. It has been shown that using tracer with similar properties as the nominal formulation is critical for measuring the RTD of the bulk flow and not just the tracer (Escotet-Espinoza et al., 2019). Secondly, there could be some dead zones in the feed frame or the feed tube where the powders do not mix as efficiently. This would mean that a short pulse of tracer does not fully mix into the entire volume inside the tablet press. Since the tracer in this case was the API of the formulation, the very small portions that do mix into these zones were not seen as the tail of the distribution. This should lead to much shorter mean residence time than from a step change test, as shown in the results. In fact, mean bulk residence times calculated by dividing the mass hold-up with flow rate of the line were very similar to the mean residence times calculated from step change data. The RTD curves drawn from raw data (without TIS fitting) also show some similarity with the compartment model RTD for mixer with dead zones (Levenspiel, 1999).

Figure 4: RTD curves for the blender in pulse and step runs. The curves are from tanks in series fit to raw data.

Figure 5: RTD for the CDC line in pulse and step runs. The curves are from tanks in series fit to raw data.

Table 7: Mean residence times, mean centered variances and number of tanks used for fitting in tanks in series model for blender, tablet press and line. Values in brackets indicate one standard deviation for pulse runs were RTDs were successfully fitted to several pulses.

Formulation/flow rate MRT (s) MCV n MRT (s) MCV n

Pulse runs Step runs

Blender

Formulation A at 5 kg/h 17.5 0.20 5 13.8 0.07 14

Formulation A at 10 kg/h 19.4 (1.3) 0.26 (0.08) 4 (1.1) 8.6 0.03 33 Formulation B at 5 kg/h 16.9 (2.5) 0.15 (0.02) 7 (0.9) 19.0 0.33 3

Tablet press

Formulation A at 5 kg/h 146 0.33 3 241 0.41 2

Formulation A at 10 kg/h 94 (9) 0.25 (0.02) 4 (0.3) 143 0.47 2 Formulation B at 5 kg/h 135 (12) 0.32 (0.01) 3 (0.09) 206 0.33 3

Line

Formulation A at 5 kg/h 164 0.26 4 255 0.37 3

Formulation A at 10 kg/h 114 (10) 0.18 (0.01) 6 (0.4) 151 0.42 2 Formulation B at 5 kg/h 152 (9) 0.25 (0.00) 4 (0.1) 225 0.28 4 MRT = Mean Residence Time, MCV = Mean Centered Variance, n = Number of tanks in series

The RTD results (pulse runs) were used to calculate the limits for the allowable magnitude and length of disturbance that do not cause more than ±25 % deviation in single tablets. Maximum deviations due to different perturbation intensity and duration were calculated according to the convolution integral (equation 8). Similar calculations were also performed for the blend to illustrate the contribution of the tablet press in the smoothening of the deviations. The results are presented as funnel plots in Figure 6. In this figure, the white area represents the situation where both the blend and tablets are within the ±25 % from label claim limit. The light grey area represents the situation where the tablets are within the ±25 % limit but the blend has gone out of specification. The dark grey area represents the situation where also the tablets are out of specification. The limits of the areas are the 95% confidence intervals of the calculated maximum deviations, and consider the variation in RTD parameters (MRT and n) as well as baseline variation in blend and tablets at steady state. The figure depicts well the effect of tablet press on the total back-mixing capability of the line

(light grey area). The magnitude and length of the allowable deviations would be quite limited if the tablet press would act as a plug flow reactor without back-mixing. In that case, the white are would represent the allowable disturbances for the full line. Due to the contribution of tablet press, the disturbances that last longer than the RTD of the blender can also be smoothened.

The pulse disturbances in this study ranged roughly from 230 % to 380 % of label claim in the scale of Figure 6 and were all well smoothened out in tablets. The funnel plots show that the magnitude of short (0-1 s) perturbations that would cause the individual tablets to deviate more than 25 % is extremely large. Putting this into general context, the largest pulses tested in the current study were roughly 5 g (5 kg/h) or 10 g (10 kg/h) which were both very large quantities in physical size. Larger lumps, that would first stick and then dislodge from the feeder, should not be expected. Thus, the CDC line can tolerate short perturbations well even if they are large. Since the smoothening ability is not markedly lowered due to increase in production rate, it can be postulated that short pulse disturbances should not pose a large threat at higher production rates either.

The step disturbances in this study meant a perturbation intensity of about 122% from label claim in feeding. Looking at the funnel plot for formulation A at 5 kg/h, even longer than 60 s perturbations of this magnitude would mean that individual tablets stay inside the ±25 % limit. For formulation A at 10 kg/h and formulation B at 5 kg/h, the situation is not as good but the maximum deviations in single tablets are still within the 95% confidence interval limit for the steps tested. However, deviation in tablets due to the longest step (60 s) is almost at the ±25 % limit. In general, a typical reason for long deviations of this magnitude is the refill of feeders. Muzzio and Engisch (2015) studied different refill scenarios with K-Tron feeders leading to refill times of less than 60 s with API feed rate deviations up to 30 - 40 % (equal to about 120-135% perturbation intensity in Figure 6).

With the current line, these kinds of disturbances would be smoothened out if the duration is less than 90 s for formulation A at 5 kg/h. For formulation A at 10 kg/h and formulation B at 5 kg/h, an increase of 30 - 40 % in API feed rate should not last more than 50 to 60 s to ensure, that the tablets will not exceed the ±25 % limit. The robustness towards long perturbations could be an issue at commercial high-volume production (e.g. 50-60 kg/h) if the robustness of the process is further aggravated due to increases in production rate. At certain cases, the amount of back-mixing, especially in the blender, can be affected through changes in blender speed. However, it was outside the scope of this study to test if the back-mixing in the blender and in the tablet press could be increased through changes in process parameters.

Figure 6: The effects of the magnitude and the length of the disturbance on the resulting blend and tablet content. Inside the white area, the blend and tablets are within the ±25% from label claim limit (with 95% confidence interval). In the light grey area, the tablets are within the ±25%

limit (with 95 % confidence interval) while blend is out of specification. In the dark grey area both the blend and the tablets are out of specification. (A) Formulation A at 5 kg/h, (B) Formulation A at 10 kg/h, (C) Formulation B at 5 kg/h.

4. Discussion

Out of the two formulations used in the study, the primary formulation (formulation A) with API in small particle size showed the best results. The content uniformity was clearly better using this formulation and the back-mixing ability of the CDC line was higher than using formulation B. The results for the effects of magnitude and length of perturbation were in line with a previously reported outcome for another CDC line (García-Muñoz et al., 2018). The system tolerates even large perturbations of short duration without severe increases in tablet content. However, the magnitude of allowable perturbations, i.e. disturbances that do not cause single tablets to deviate more than

±25 % from target, decreases with an increase in perturbation time. A higher flow rate leads to shorter residence times and subsequently lower back-mixing capability. The flow rates used in the current study were in the range used in development studies, i.e. in the low end. The decrease in the back-mixing ability is a relevant issue to consider, while the process is operated at higher production rates in commercial manufacturing as the allowable duration for disturbances can be much more limited.

In our case, the tablet press contributed strongly to the overall back-mixing capability of the line. In another CDC line there were much smaller differences in the content of blend and tablets after a pulse disturbance (Engisch and Muzzio, 2016). On the other hand, the back-mixing capability was much larger in a CDC line that consisted of both a larger blender and a larger tablet press than in our line (García-Muñoz et al., 2018). Tablet presses with long and broad RTD, i.e. presses with back- mixing capability, are thus beneficial considering the robustness of the line. The mass hold-up in our line was rather small (350-400 g at the end of the run) and the RTD was narrower than e.g. in other presses used in Karttunen et al. (2019a). Thus, the deviations witnessed could have been smaller if another tablet press was used. The feed frame has a major role considering back-mixing inside the tablet press. It has been shown previously that a high feed frame speed can increase the width of RTD, i.e. lead to better back-mixing (Dülle et al., 2018), without affecting the mean residence time.

Thus, if the feed frame speed is not critical for other tablet qualities, e.g. hardness, feed frame speeds can possibly be used to optimize the robustness of the line towards disturbances. However, this seems formulation and equipment dependent since there is also evidence that feed frame speed can affect the mean residence time (Mendez et al., 2010).

In general, the back-mixing capability of a continuous line is the sum of all unit operations that include at least some back-mixing. Moving downstream from the feeders, the resulting deviation decreases unit by unit and this can be used to allow small disturbances at the beginning of the line (Almaya et al., 2017; Engisch and Muzzio, 2016; Karttunen et al., 2019a; Kruisz et al., 2017). How much the deviation decreases depends on the line, parameters, formulation and disturbance. One solution to enhance the back-mixing capabilities in blending would be to use a continuous blender with individually controllable mass hold-up and mixing speed which should increase smoothening abilities (Toson et al., 2018). This type of individual adjustment could increase the flexibility in designing a process with adequate back-mixing but also the capability for rapid changes in set points if that is needed. In addition, a double mixing and integrated direct compression, such as presented by Taipale-Kovalainen et al (2019) and Van Snick et al (2017b), should also increase the ability of mixing units to smoothen out disturbances. Naturally, mitigation of the deviations in feeding should enhance the control over the line and ensure a more consistent end-product. Control systems

utilizing feed-back control have previously been presented to provide more stable feeding (Nagy et al., 2017; Singh et al., 2014).

5. Conclusions

Overall, the CDC line in question was notably capable in smoothening the disturbances investigated in the present study. The line would have been able to smoothen out even larger disturbances, especially with the primary formulation at lower flow rate, without the API content of individual tablets deviating more than ± 25 % from target. The line was efficient in smoothening short-term deviations with all formulations and flow rates. For long-term deviations, the primary formulation at low flow rate was the most robust setting towards disturbances. With an increase in production rate or a change to a formulation with larger API size, the smoothening capability towards long disturbances became more limited. However, whether or not the pharmacopoeia limits were passed was dependent on the general performance of the formulations and not how the individual disturbances were smoothened. The modified formulation with an API with a large particle size showed increased baseline variability in tablet content which resulted in failing the pharmacopoeia even if the disturbances were excluded from the data. Based on the RTD results, both the blender and the tablet press contributed to the smoothening of the disturbances, but the tablet press had a larger effect on the overall back-mixing capability of the line. This is especially relevant for the smoothening of the long perturbations that last longer than the mean residence time of the blender.

Acknowledgements

This work is a part of the European Consortium for Continuous Pharmaceutical Manufacturing sponsored by member companies, universities and institutions (ref: www.eccpm.com). In particular, this was funded through the Austrian COMET Program by the Austrian Federal Ministry of Transport, Innovation and Technology (BMVIT), the Austrian Federal Ministry of Economy, Family and Youth (BMWFJ) and by the State of Styria (Styrian Funding Agency SFG). The authors thank the PROMIS Centre consortium, funded by Tekes ERDF and North Savo Centre for Economic Development, Transport and the Environment for providing research facilities.

References

Alam, M.A., Shi, Z., Drennen, J.K., Anderson, C.A., 2017. In-line monitoring and optimization of powder flow in a simulated continuous process using transmission near infrared spectroscopy.

Int. J. Pharm. 526, 199–208. https://doi.org/10.1016/j.ijpharm.2017.04.054

Allison, G., Cain, Y.T., Cooney, C., Garcia, T., Bizjak, T.G., Holte, O., Jagota, N., Komas, B., Korakianiti, E., Kourti, D., Madurawe, R., Morefield, E., Montgomery, F., Nasr, M., Randolph, W., Robert, J.L., Rudd, D., Zezza, D., 2015. Regulatory and quality considerations for continuous

manufacturing May 20-21, 2014 continuous manufacturing symposium. J. Pharm. Sci.

https://doi.org/10.1002/jps.24324

Almaya, A., De Belder, L., Meyer, R., Nagapudi, K., Lin, H.R.H., Leavesley, I., Jayanth, J., Bajwa, G., DiNunzio, J., Tantuccio, A., Blackwood, D., Abebe, A., 2017. Control Strategies for Drug Product Continuous Direct Compression—State of Control, Product Collection Strategies, and Startup/Shutdown Operations for the Production of Clinical Trial Materials and Commercial Products. J. Pharm. Sci. 106, 930–943. https://doi.org/10.1016/j.xphs.2016.12.014

Bergum, J., Parks, T., Prescott, J., Tejwani, R., Clark, J., Brown, W., Muzzio, F., Patel, S., Hoiberg, C., 2014. Assessment of Blend and Content Uniformity. Technical Discussion of Sampling Plans and Application of ASTM E2709/E2810. J. Pharm. Innov. 10, 84–97.

https://doi.org/10.1007/s12247-014-9208-z

Bonawi-tan, W., Williams, J.A.S., 2004. Online Quality Control with Raman Spectroscopy in Pharmaceutical Tablet Manufacturing. J. Manuf. Syst. 23, 299–308.

Brennan, Z., 2016. FDA Allows First Switch From Batch to Continuous Manufacturing for HIV Drug.

Regul. Aff. Prof. Soc. https://www.raps.org/regulatory-focus%E2%84%A2/news- articles/2016/4/fda-allows-first-switch-from-batch-to-continuous-manufacturing-for-hiv-drug (accessed 17 September 2019).

Dave, V.S., Shahin, H.I., Youngren-Ortiz, S.R., Chougule, M.B., Haware, R. V., 2017. Emerging technologies for the non-invasive characterization of physical-mechanical properties of tablets.

Int. J. Pharm. 532, 299–312. https://doi.org/10.1016/j.ijpharm.2017.09.009

De Beer, T.R.M., Bodson, C., Dejaegher, B., Walczak, B., Vercruysse, P., Burggraeve, A., Lemos, A., Delattre, L., Heyden, Y. Vander, Remon, J.P., Vervaet, C., Baeyens, W.R.G., 2008. Raman spectroscopy as a process analytical technology (PAT) tool for the in-line monitoring and understanding of a powder blending process. J. Pharm. Biomed. Anal. 48, 772–779.

https://doi.org/10.1016/j.jpba.2008.07.023

De Leersnyder, F., Peeters, E., Djalabi, H., Vanhoorne, V., Van Snick, B., Hong, K., Hammond, S., Liu, A.Y., Ziemons, E., Vervaet, C., De Beer, T., 2018. Development and validation of an in-line NIR spectroscopic method for continuous blend potency determination in the feed frame of a

tablet press. J. Pharm. Biomed. Anal. 151, 274–283.

https://doi.org/10.1016/j.jpba.2018.01.032

DiProspero, D., 2018. Continuous OSD Manufacturing – A Product & Patient Perspective. ISPE.

https://ispe.org/pharmaceutical-engineering/ispeak/continuous-osd-manufacturing-product- patient-perspective (accessed 17 September 2019)

Dülle, M., Özcoban, H., Leopold, C.S., 2018. Investigations on the residence time distribution of a three-chamber feed frame with special focus on its geometric and parametric setups. Powder Technol. 331, 276–285. https://doi.org/10.1016/j.powtec.2018.03.019

Engisch, W., Muzzio, F., 2016. Using Residence Time Distributions (RTDs) to Address the Traceability of Raw Materials in Continuous Pharmaceutical Manufacturing. J. Pharm. Innov. 11, 64–81.

https://doi.org/10.1007/s12247-015-9238-1

Engisch, W.E., Muzzio, F.J., 2015. Feedrate deviations caused by hopper refill of loss-in-weight feeders. Powder Technol. 283, 389–400. https://doi.org/10.1016/j.powtec.2015.06.001

Ervasti, T., Simonaho, S.P., Ketolainen, J., Forsberg, P., Fransson, M., Wikström, H., Folestad, S., Lakio, S., Tajarobi, P., Abrahmsén-Alami, S., 2015. Continuous manufacturing of extended release tablets via powder mixing and direct compression. Int. J. Pharm. 495, 290–301.

https://doi.org/10.1016/j.ijpharm.2015.08.077

Escotet-Espinoza, M.S., Moghtadernejad, S., Oka, S., Wang, Y., Roman-Ospino, A., Schäfer, E., Cappuyns, P., Van Assche, I., Futran, M., Ierapetritou, M., Muzzio, F., 2019. Effect of tracer material properties on the residence time distribution (RTD) of continuous powder blending operations. Part I of II: Experimental evaluation. Powder Technol. 342, 744–763.

https://doi.org/10.1016/j.powtec.2018.10.040

Esmonde-White, K.A., Cuellar, M., Uerpmann, C., Lenain, B., Lewis, I.R., 2017. Raman spectroscopy as a process analytical technology for pharmaceutical manufacturing and bioprocessing. Anal.

Bioanal. Chem. 409, 637–649. https://doi.org/10.1007/s00216-016-9824-1

European Pharmacopoeia, 2018a. 2.9.40. Uniformity of dosage units, 9th ed. EDQM Council of Europe, France.

European Pharmacopoeia, 2018b. 2.9.47. Demonstration of uniformity of dosage units using large sample sizes, 9th ed. EDQM Council of Europe, France.

FDA, 2016. Questions and answers on Current Manufacturing Practices - Production and Process controls. https://www.fda.gov/drugs/guidances-drugs/questions-and-answers-current-good- manufacturing-practices-production-and-process-controls#15 (accessed 17 September 2019) FDA, 2004. Guidance for Industry: PAT — A Framework for Innovative Pharmaceutical Development,

Manufacturing, and Quality Assurance. https://www.fda.gov/media/71012/download (accessed 17 September 2019)

García-Muñoz, S., Butterbaugh, A., Leavesley, I., Manley, L.F., Slade, D., Bermingham, S., 2018. A flowsheet model for the development of a continuous process for pharmaceutical tablets: An industrial perspective. AIChE J. 64, 511–525. https://doi.org/10.1002/aic.15967

Garcia, T., Bergum, J., Prescott, J., Tejwani, R., Parks, T., Clark, J., Brown, W., Muzzio, F., Patel, S., Hoiberg, C., 2014. Recommendations for the Assessment of Blend and Content Uniformity:

Modifications to Withdrawn FDA Draft Stratified Sampling Guidance. J. Pharm. Innov. 10, 76–

83. https://doi.org/10.1007/s12247-014-9207-0

Goodwin, D.J., van den Ban, S., Denham, M., Barylski, I., 2018. Real time release testing of tablet content and content uniformity. Int. J. Pharm. 537, 183–192.

https://doi.org/10.1016/j.ijpharm.2017.12.011

Järvinen, K., Hoehe, W., Järvinen, M., Poutiainen, S., Juuti, M., Borchert, S., 2013. In-line monitoring of the drug content of powder mixtures and tablets by near-infrared spectroscopy during the continuous direct compression tableting process. Eur. J. Pharm. Sci.

https://doi.org/10.1016/j.ejps.2012.12.032

Karande, A.D., Heng, P.W.S., Liew, C.V., 2010. In-line quantification of micronized drug and excipients in tablets by near infrared (NIR) spectroscopy: Real time monitoring of tabletting process. Int. J. Pharm. 396, 63–74. https://doi.org/10.1016/j.ijpharm.2010.06.011

Karttunen, A.-P., Hörmann, T.R., De Leersnyder, F., Ketolainen, J., De Beer, T., Hsiao, W.-K., Korhonen, O., 2019a. Measurement of residence time distributions and material tracking on

three continuous manufacturing lines. Int. J. Pharm. 563, 184–197.

https://doi.org/10.1016/j.ijpharm.2019.03.058

Karttunen, A.-P., Wikström, H., Tajarobi, P., Fransson, M., Sparén, A., Marucci, M., Ketolainen, J., Folestad, S., Korhonen, O., Abrahmsén-Alami, S., 2019b. Comparison between integrated continuous direct compression line and batch processing - The effect of raw material properties. J. Pharm. Sci. 133, 40–53. https://doi.org/10.1016/j.ejps.2019.03.001

Khorasani, M., Amigo, J.M., Sun, C.C., Bertelsen, P., Rantanen, J., 2015. Near-infrared chemical imaging (NIR-CI) as a process monitoring solution for a production line of roll compaction and tableting. Eur. J. Pharm. Biopharm. 93, 293–302. https://doi.org/10.1016/j.ejpb.2015.04.008 Kruisz, J., Rehrl, J., Sacher, S., Aigner, I., Horn, M., Khinast, J.G., Engineering, Pharmaceutical,

Engineering, Particle, 2017. RTD modeling of a continuous dry granulation process for process control and materials diversion. Int. J. Pharamceutics 528, 334–344.

https://doi.org/https://doi.org/10.1016/j.ijpharm.2017.06.001

Lakio, S., Ervasti, T., Tajarobi, P., Wikström, H., Fransson, M., Karttunen, A.-P., Ketolainen, J., Folestad, S., Abrahmsén-Alami, S., Korhonen, O., 2017. Provoking an end-to-end continuous direct compression line with raw materials prone to segregation. Eur. J. Pharm. Sci. 109, 514–

524. https://doi.org/10.1016/j.ejps.2017.09.018

Lee, S.L., O’Connor, T.F., Yang, X., Cruz, C.N., Chatterjee, S., Madurawe, R.D., Moore, C.M.V., Yu, L.X., Woodcock, J., 2015. Modernizing Pharmaceutical Manufacturing: from Batch to Continuous Production. J. Pharm. Innov. 10, 191–199. https://doi.org/10.1007/s12247-015-9215-8

Levenspiel, O., 1999. Compartment Models, in: Chemical Reaction Engineering. John Wiley & Sons, New York, pp. 283–292.

Martinetz, M.C., Karttunen, A.P., Sacher, S., Wahl, P., Ketolainen, J., Khinast, J.G., Korhonen, O., 2018. RTD-based material tracking in a fully-continuous dry granulation tableting line. Int. J.

Pharm. 547, 469–479. https://doi.org/10.1016/j.ijpharm.2018.06.011

Martínez, L., Peinado, A., Liesum, L., Betz, G., 2013. Use of near-infrared spectroscopy to quantify drug content on a continuous blending process: Influence of mass flow and rotation speed variations. Eur. J. Pharm. Biopharm. 84, 606–615. https://doi.org/10.1016/j.ejpb.2013.01.016 Mendez, R., Muzzio, F., Velazquez, C., 2010. Study of the effects of feed frames on powder blend

properties during the filling of tablet press dies. Powder Technol. 200, 105–116.

https://doi.org/10.1016/j.powtec.2010.02.010

Nagy, B., Farkas, A., Gyürkés, M., Komaromy-Hiller, S., Démuth, B., Szabó, B., Nusser, D., Borbás, E., Marosi, G., Nagy, Z.K., 2017. In-line Raman spectroscopic monitoring and feedback control of a continuous twin-screw pharmaceutical powder blending and tableting process. Int. J. Pharm.

530, 21–29. https://doi.org/10.1016/j.ijpharm.2017.07.041

Nasr, M.M., Krumme, M., Matsuda, Y., Trout, B.L., Badman, C., Mascia, S., Cooney, C.L., Jensen, K.D., Florence, A., Johnston, C., Konstantinov, K., Lee, S.L., 2017. Regulatory Perspectives on Continuous Pharmaceutical Manufacturing: Moving From Theory to Practice. September 26- 27, 2016, International Symposium on the Continuous Manufacturing of Pharmaceuticals. J.

Pharm. Sci. 106, 3199–3206. https://doi.org/10.1016/j.xphs.2017.06.015

Osorio, J.G., Muzzio, F.J., 2016. Effects of processing parameters and blade patterns on continuous