Process analytical technology for continuous manufacturing tableting processing: A case study

Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

Contents lists available at ScienceDirect

Journal of Pharmaceutical and Biomedical Analysis journal homepage: www.elsevier.com/locate/jpba

Process analytical technology for continuous manufacturing tableting processing: A case study Victoria Pauli a,b,1 , Yves Roggo a,∗,1 , Laurent Pellegatti a , Nhat Quang Nguyen Trung a , Frantz Elbaz a , Simon Ensslin a , Peter Kleinebudde b , Markus Krumme a a b

Novartis Pharma AG, Continuous Manufacturing (CM) Unit, CH-4002 Basel, Switzerland Institute of Pharmaceutics and Biopharmaceutics, Heinrich Heine University, Universitaetsstr. 1, 40225 Dusseldorf, Germany

a r t i c l e

i n f o

Article history: Received 10 July 2018 Received in revised form 31 August 2018 Accepted 4 September 2018 Available online 13 September 2018 Keywords: Continuous manufacturing Tableting Content uniformity Blend uniformity Near-infrared spectroscopy Process analytical technology

a b s t r a c t The use of Near Infrared Spectroscopy (NIRS) as a fast and non-destructive technique was employed for the control and monitoring of the tableting step during a continuous manufacturing process. Two NIRS methods were optimized in order to in-line control the blend uniformity in the tablet feed frame and the API concentration of freshly pressed tablets prior the ejection. The novelty of this work first lies in the acquisition speed of NIR spectra reaching up to 70,000 tablets/h. Partial Least Square (PLS) regression was used as chemometric tool for the computation that resulted in excellent predictive calibration results. A coefficient of correlation (r) value of 0.99 was obtained for both probes. The root mean square error of calibration (RMSEC) and the root mean square error of prediction (RMSEP) were respectively 1.8% and 1.8% for active content in the tablet feeder and 2.2% and 2.3% for the tablet content. In addition, calibration performance and robustness of the methods were evaluated. Moreover several qualitative methods were proposed to monitor the tableting process in different stages of development (single wavelength, Principal Component Analysis, and Independent Component Analysis). In early phase development, the requirement/quality of the input material is not established yet; hence the use of a qualitative approach allows to confirm the suitability of the PAT methodology for in-process material monitoring & control. Later, the qualitative approach constitutes the foundation for the quantitative approach when input materials are fixed and larger production size occurs. The proposed strategy is a performant PAT tool for continuous manufacturing and a step forward to real time release. © 2018 Elsevier B.V. All rights reserved.

1. Introduction Continuous Manufacturing (CM) of pharmaceutical drug products is a new approach within the pharmaceutical industry, opposing traditional batch manufacturing processes based on its potential to increase manufacturing flexibility and efficiency. In CM, all process units are directly connected to each other. Starting material is continuously charged into the first process unit at the beginning of the line, while final product is simultaneously discharged at the end. In combination with suitable process analyzers, this manufacturing approach allows to constantly monitor the processed material and if required, to control critical process parameters and thus critical quality attributes of in process mate-

∗ Corresponding author at: Novartis Pharma AG, WSJ-27.4.021.01, Continuous Manufacturing (CM) Unit, CH-4002 Basel, Switzerland. E-mail address: [email protected] (Y. Roggo). 1 The first two authors listed contributed equally to this work. https://doi.org/10.1016/j.jpba.2018.09.016 0731-7085/© 2018 Elsevier B.V. All rights reserved.

rial and final products. In a continuous manufacturing process, and contrary to a traditional batch production process, the tablet press runtime is not limited by the batch size; tableting of granules can be proceeded as long as the upstream unit operations are supplying materials to the press in a controlled process state. A controlled process state can be ensured by implementing suitable in-line and/or at-line PAT-tools (PAT = process analytical technologies) that deliver real-time information about the process state and/or product quality, together with associated control loops to adapt critical process parameters accordingly [1–3]. Near Infrared Spectroscopy (NIRS) has become a popular qualitative and quantitative PAT-tool in the pharmaceutical industry, as it is a safe, fast, and non-destructive method which does not require sample preparation. The integration of NIRS into pharmaceutical manufacturing was initially endorsed by the US Food and Drug Administration (FDA) as part of their PAT initiative in 2004, and further emphasized by the International Conference on Harmonization (ICH), by publishing a guideline on NIRS method validation [2,4]. The guideline was then picked up by the FDA and

102

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

Fig. 1. Installation of the Near Infrared probes in the tablet press. (A) Sentro PAT FO for granules blend uniformity in the tablet press feed frame, with light source (1) and probe head (2) inserted into the tablet press feed frame (3). (B) VisioNIR for content uniformity of tablets after compression with the probe head (4) installed between the upper punch (5) and tablet being ejected by lower punch (6).

the European Medicines Agency (EMA), as well as the United States Pharmacopeia Convention (USP) and the European Pharmacopoeia (Ph. Eur.), who developed their own guidelines based on the initial ICH document [5–10]. Nowadays, NIRS is frequently applied for the identification of raw materials, in-line monitoring of process steps like blending, granulation, and drying by compound quantification, as well as for process troubleshooting [11–16]. Another well described application is the quantification of API content in tablets to demonstrate content uniformity [17–20]. However, most reported methods are merely suitable for offline or at-line sample measurement, as long sample handling times and spectral acquisition times of available spectrometers allow only a limited percentage of produced tablets to be tested [17–20]. This approach can be sufficient for process control, where blending and granulation are done according to a validated range and content uniformity for batch-release is demonstrated via offline analysis of finished goods according to Ph.Eur. Chapter 2.9.40 – Uniformity of Dosage Units [21]. However, when blending and granulation are done continuously, variations in process parameters or starting material characteristics over the course of production may cause fluctuations in granules content uniformity and therefore in the tablets API-content, water content, or tablet hardness. These variations might go undetected with the common approach of sampling and analyzing only a limited number of randomly selected tablets. Alternatively, calibration-based approaches for 100% tabletingprocess control through chemometric compression parameter measurements could be implemented [22–24]. These methods however are mainly limited to control the tablet weight variations, while lacking the ability to measure the actual API content of each tablet [22–24]. Thorough quantitative in-line control of solid dosage forms via NIRS at full production speeds is still limited up to this date, due to limitations in measurement speed. One available instrument that can perform 100% quantitative testing of API content in tablets at production speeds of up to 125.000 tablets/hour is the VisioNIR LS (Uhlman, VisioTec GmbH, Laupheim, Germany), as reported by Järvinen at al. in 2013. The authors described 100% control of powder mixtures and tablet drug content during a continuous mixing and direct compression process of a three-ingredient powder blend at varying tableting speed and compaction force in a technical research environment [25]. Based on this paper, further work on establishing NIRS for 100% in-line real-time control of tablet API content was conducted on a GMP-qualified, continuous from powder-to-tablets manufacturing process with a six-ingredient formulation similar to a commercial

one. In detail, granules at varying API-contents from a twin-screw wet granulation process were dried in a continuous fluid bed dryer and subsequently transferred to a tablet press for compression while granules and tablets content uniformity was monitored with two independent NIRS instruments. The first instrument was installed in the tablet press feed frame, to ensure the absence of segregation issues that could occur after milling and transfer. The second instrument was installed in the tablet press to measure every tablet right before being ejected from the press. A NIRS validation strategy was proposed in order to create a robust calibration where focus was put on selection of suitable calibration and validation samples, assessment of different preprocessing techniques, and evaluation of different chemometric options for process monitoring. Additionally, uniformity of dosage units using large samples sizes based on NIR predictions from tablets being ejected from the press was demonstrated according to Ph.Eur. 9.5 chapter 2.9.47 [32]. The development of an efficient ejection system and corresponding control loops is currently ongoing and not discussed in this manuscript. 2. Materials and methods 2.1. Materials Diclofenac Sodium (Acros Organics, Geel, Belgium) was used as Active Pharmaceutical Ingredient (API) in all performed trials. The standard formulation contained 25% Diclofenac Sodium, 5% Sodium Starch-Glycolate, 5% Sodium Stearyl Fumarate, 4% Hypromellose (Cellulose-HP-M 603), 12.2% Calcium Hydrogen-Phosphate Anhydrous and 48.8% Microcrystalline Cellulose PH102 (all excipients supplied by Novartis Pharma AG, Stein, Switzerland). To generate granules at varying API content, formulations containing 70–130 % of the original drug content were prepared, by adapting the amounts of Hypromellose and Calcium Hydrogen-Phosphate Anhydrous accordingly, while keeping their ratio constant at 80:20. Purified water was used as granulation liquid at a Liquid/Solid-ratio of 0.3. The targeted tablet weight was 240 mg with a targeted label claim (LC) of 60 mg API/tablet (LC = 100%). 2.2. Solid dosage forms 2.2.1. Preparation of powder blends For preparation of powder blends for continuous granulation and tableting, all ingredients were weighted into a drum and blended two times for 10 min at 11 rpm with a Pharma Telescope

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

Blender PTM 300 (LB Bohle GmbH, Ennigerloh, Deutschland) in batch sizes of 5–50 kg, depending on the trial. Between the two blending steps, the powder was sieved through a 1.25 mm-mesh sieve to break agglomerates. 2.2.2. Twin-screw wet granulation Continuous wet-granulation was performed on a Thermo Fisher Pharma 16 Twin Screw Granulator (TSG) (Thermo Fisher Scientific, Karlsruhe, Germany) with screw diameter (D) of 16 mm and screw length of 53 ¼ x D. Screw configuration from inlet to outlet of the barrel was as follows: 2 D Feed Screw Elements (FSE), 2 D Long Helix Feed Screws, 22 D FSE, 2 ¼ D 30◦ Mixing Element, 22 D FSE, 3 D Distributive Feed Screw Elements. The powder blend was fed into the barrel through the first feeding port by a loss-in-weight powder feeder (K-Tron T20, Coperion K-Tron GmbH, Niederlenz, Switzerland). Granulation liquid at room temperature was fed through the second port by a custom made dispensing pump system (based on Watson Marlow, Zollikon, Switzerland). Granulation settings were as follows: barrel temperature 35 ◦ C, granulator screw speed 500 rpm, solid feed rate 4 kg/h, liquid feed rate 1.2 kg/h. 2.2.3. Continuous fluid-bed drying Continuous fluid-bed drying of wet granules was performed on a Glatt GPCG 2 CM fluid-bed dryer (Glatt GmbH, Binzen, Germany), directly connected to the Twin-Screw Granulator. Drying settings were as follows: 80 ◦ C drying temperature, 140 m3 /h air flow rate, dryer rotation speed 17 rph. 2.2.4. Sieving and tableting Dried granules were sieved through a 1.25 mm sieve (Oscillowitt-Lab type MF-lab, Frewitt, Granges-Paccot, Switzerland) directly before compaction in a rotary tablet press (Calibration 1200i, FETTE Compacting, Schwarzenbek, Germany), with round punches of 10 mm diameter at tableting speeds between 17.000 and 70.000 tablets/hour. 2.3. Analytical methods 2.3.1. Near-infrared spectroscopy ® A SentroPAT FO instrument (Sentronic , Dresden, Germany) configured with an immersion probe (SentroProbe DR LS NIR) was used for the monitoring of dried and sieved granules in the tablet press feeder. In-line spectra for calibration and validation were continuously measured in reflection mode using 60 scans of 0.011 s, a 2 nm resolution, and a spectral range of 1150–2200 nm in dynamic mode (i.e. tablet press in function). Offline spectra for calibration were measured with the same settings in static mode. A VisioNIR LS instrument (VisioTec GmbH, Laupheim, Germany) was used for control of tablet content uniformity. In-line spectra of tablets for calibration and validation were measured in reflection mode in the tablet press using one scan of 0.004 s, an 8 cm−1 resolution, and a spectral range of 4000-10000 cm−1 . Offline spectra for calibration were measured at the same settings in static mode outside of the tablet press. The distance between tablets and the probe head is fixed at 3 mm. In-line spectra were collected at three different tableting speeds (17000, 30,000 and 70,000 tablets per hours). With both instruments, a background spectrum was acquired every hour for off-line samples and before the campaign for in-line acquisition. The instruments and the installation of the two NIR probes are presented in Fig. 1. 2.3.2. Reference analysis by HPLC Diclofenac Sodium content of collected samples was quantified by HPLC with an Agilent 1260 analytical system (Agilent Tech-

103

nologies, Santa Clara, CA, USA), equipped with a 250 x 4.6 mm YMC Pack ODS-AM Column (YMC CO. LTD., Kyoto, Japan). A mixture of Methanol and 0.8 g NaH2PO4*H2O in water at pH = 2.5 (adjusted with H3PO4 85%) was used as mobile phase, with gradient mixing of 35/65 at 0–10 minutes, 65/35 at 11–30 minutes, and 35/65 at 31–35 minutes, at a flowrate of 1.5 ml/min. Column temperature was 40 ◦ C, detection wavelength was 254 nm. The sample solvent contained 320 ml tri-sodium citrate di-hydrate 1% and 680 ml Ethanol 90%. For sample preparation, 80 mg of dried granules or ground tablet were dissolved ad 100 ml solvent and filtered through a 1 ␮m glass filter before injection of 40 ␮l for analysis. Diclofenac Sodium 99.9% in solvent was used as reference standard. 2.3.3. Calibration and validation datasets For blend uniformity assessment in the tablet feed frame, offline spectra and in-line spectra of granules containing 70%, 80%, 90%, 95%, 100%, 105%, 110%, 120%, and 130% of the label claim API content were acquired. Offline spectra were taken with the probe directly submerged into a granule sample at three different positions and five spectra per position, in-line spectra were continuously collected directly in the tablet feed frame during processing. The whole dataset containing 441 offline- and 1809 in-line spectra, was split into two independent calibration and validation sets (221 offline- and 905 in-line spectra for calibration and 220 offline and 904 in-line spectra for validation). For tablet content uniformity assessment, offline and in-line spectra of tablets containing 70%, 80%, 90%, 95%, 100%, 105%, 110%, 120%, and 130% of the label claim API content were acquired. Offline spectra were taken in static mode from 10 tablets per label claim, in-line spectra were taken directly in the tablet press at tableting speeds between 17.000 and 70.000 tablets/hour. The full dataset containing 4500 offline and 6750 in-line spectra, was split into two independent calibration and validation sets, each containing 50% of the overall data. During calibration development it was ensured that calibration and validation spectra were acquired on different days with different batches of powder blends. Three different NIR calibrations were created for blend uniformity and tablet content uniformity, each: one with only off-line spectra, one with only in-line spectra and a third one combining in-line and off-line data. Reference analysis by HPLC was only performed for off-line analysis of tablets (i.e. 10 tablets per label claim). The gravimetric method was used as reference for the calibration model when the spectra have been acquired inline as proposed by Nagy and coworkers [35]. For each of the calibrations, different types of preprocessing were evaluated. Details are explained in the results section. 2.4. Chemometric methods 2.4.1. Data preprocessing Variables such as particle size, dried granules water content, and NIR probe position have an impact on the spectra. Standard normal variate (SNV) combined with first derivative (gap of 5) were applied on the raw data for spectral adjustment and reduction of noise [26]. The selection of the preprocessing is discussed as part of the calibration optimization. 2.4.2. Qualitative methods 2.4.2.1. Principal component analysis (PCA). PCA is an unsupervised method used to highlight similarities and differences in a set of observations using linearly uncorrelated variables called Principal Components (PCs). This procedure permits a visualization of the repartition of the dataset and in this case the spectra. PCA was applied in order to have a first overview of the sample distribution and for qualitative process monitoring [27].

104

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

Fig. 2. Near Infrared spectra after Standard Normal Variate (SNV) normalization, colored by API content (70–130 % LC). (A) Granules in the tablet feed frame: A-1: full spectral range, A-2: zoom main API peak at 1670 nm. (B) Tablets: B-1: full spectral range, B-2: zoom main API peak.

2.4.2.2. Independent component analysis (ICA). ICA is special case of blind source separation that performs a decomposition of a multivariate signal (i.e. spectra) into independent non-Gaussian components (also called latent variables or sources) with the assumption that the components are statistically independent. In contrast to PCA where original signals cannot be recovered from a multivariate one, ICA can apply information on statistical independence to recover the original sources. There are numerous algorithms available that do ICA, all differing in the way independence is defined. In the presented work it was defined by maximizing non-Gaussianity according to Hyvarinnen et al [28]. Further details on the mathematics behind ICA can be found in literature [29,30]. 2.4.3. Quantitative method – partial least squares regression (PLS) Partial least squares (PLS) regression was carried out on the preprocessed NIRS spectra with a full cross-validation using a nonlinear iterative PLS-algorithm. PLS is the most used chemometric algorithm for regression calibrations [31]. 2.4.4. Software and statistical computation 2.4.4.1. Blend uniformity. The “Sentro Suite” package(version 2) ® (Sentronic , Dresden, Germany) was used for spectra acquisition in the tablet feed frame. SIMCA software (version 13.3, Umetrics/Sartorius, Umea, Sweden) was used to compute the calibration used during in-line acquisition. 2.4.4.2. Tablet content. The “NovaPAC/NovaMath” package ® (Prozess Technologie Inc. , St. Louis, Missouri, USA) was used for spectra acquisition of tablets in the tablet press.

®

Unscrambler version 10.5 (CAMOs Software AS, Oslo, Norway) was used for the preprocessing, the PLS computation for in-line prediction. Hierarchical Calibration development module (version1.5 - CAMOs Software AS, Oslo, Norway) was used to create the in-line calibration. 2.4.4.3. Offline data evaluation. R Software (version 3.3.3, The R Foundation for Statistical Computing) was used for computation with RStudio (Version 0.99.878 – RStudio, Inc.) as interface. The following libraries were used for this publication: Library signal (version 0.7–6) for preprocessing, library PLS (version 2.6.0) for Partial Least Square regression, Library HyperSpec (version 0.99) for spectral importation, FactoMineR (version 1.36), Factoextra (version 1.0.5) and Factoshiny (version 1.0.5) for Principal Component Analysis and library ICA for Independent Component Analysis (version 1.0-1) 2.5. Calibration assessment Calibrations were assessed via linearity, range, and robustness, according to current guidelines on validation of NIR methods. To test linearity and range, calibration and validation samples were plotted against values obtained from reference analytics as recommended by EMA. Results were evaluated by considering the range, the coefficient of correlation (R2 ), the intercept, the slope, and the bias. R2 greater or equal to 0.95 should be observed while the slope and intercept are expected to be as close as possible to 1, and 0, respectively [7]. Accuracy was demonstrated by evaluating root mean square error of calibration (RMSEC), of cross-validation (RMSECV), and of prediction (RMSEP) and a graphical evaluation of residues from the validation samples to determine their minimum

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

105

Fig. 4. PCA plot of SNV-pretreated spectra at three different tableting speeds demonstrates the impact of the acquisition speed on tablets NIR spectra: Score plot PC1 vs PC2, colored by label claim 70–130 %.

needs to be taken to design the calibration methods in a robust way that is not biased by the tableting speed. 3.2. Quantitative determination of API content

Fig. 3. Principal Component Analysis (PCA) of SNV treated datasets. (A) Granules in the tablet feed frame: Score plot PC1 vs PC2, colored by label claim 70–130 %. (B) Tablets: Score plot PC1 vs PC2, colored by label claim 70–130 %, acquired at moderate tableting speeds of 17.000 and 30.000 tablets/h.

and maximum [7]. To demonstrate robustness, common variations during routine analyses should not affect the ability of the method and should be evaluated [4,7–9]. In this paper the robustness of tablet content uniformity towards the acquisition speed (i.e. tableting speed) was evaluated.

3. Results and discussion

3.2.1. Strategy for content uniformity of tablets 3.2.1.1. Initial offline calibration. In the first part of the study methods were calibrated with spectra acquired offline by measuring samples of the different label claims in static mode. Calibration was done against two different reference values for API content, to evaluate their performance: first, the target concentrations of the individual formulations compressed (w/w) and second, the samples’ HPLC-references. Excellent predictive performances were achieved by PLScalibration for both reference values, as shown in Fig. 5 A and B, respectively. R2 of 0.989 and 0.977 indicated a good overall linearity of both calibrations, the low intercept values and bias of nearly zero confirmed the absence of constant systematic errors. SEP errors of 1.695% and 2.575% indicated accurate correlation between NIRS and granules’ API target concentration and HPLCreference values, respectively. In summary, calibration was feasible with both reference values. Consequently, it was decided to use the target concentration for all following calibrations in this paper, since sampling is simplified by this approach.

3.1. Overview of spectral data Fig. 2 displays the NIR spectra of calibration samples for granules and tablets after SNV normalization between 1100 and 2200 nm. The main absorption of Diclofenac Sodium is observed at 1670 nm, caused by the first overtone of −CH. Absorption increases as the samples API content increases from 70% to 130% label claim. Additionally, the API’s −OH peak at 1930 nm is detected. However, since the region between 1900–2000 nm is also highly correlated to water absorption, this peak is not suitable for API content quantification [32]. PCA was performed on SNV-transformed granule- and tabletdatasets, as seen in Fig. 3 A and B, respectively, to get a better overview of acquired spectra. The first PC reflects the change in API content. Spectra in Fig. 3 B were acquired at two moderate tableting speeds (17.000 and 38.000 tablets/h). Here, no difference in spectra is detected. After the inclusion of spectra collected at high tableting speeds (70.000 tablets/h) to the PCA analysis, two clusters of spectra that can be correlated to the different speed settings were identified (see Fig. 4). NIR spectra from granules and tablets provide suitable information on the materials API content and should allow a qualitative monitoring of the process. Nevertheless, the tableting speed seems to have an influence on spectral acquisition. Therefore, special care

3.2.1.2. Apply offline calibration to in-line data for calibrationevaluation and optimization. The offline-calibration was applied to in-line measured spectra. SEP increased from 1.695% (SNV offline calibration validated with offline spectra) to 7.366%, when the SNVoffline calibration was applied to in-line spectra. As a result, bias increased from close to zero to 8.981%. In addition, it was observed that bias varied depending on the tableting speed (i.e., spectral acquisition speed, see Fig. 6). Based on these observations it was decided that the offline calibration has to be optimized before it can be applied to in-line data. By comparing the effects of different preprocessing techniques (namely SNV and SNV in combination with 1st Derivate), the different available calibration data sets (offline, in-line, and combined), and two different spectral ranges (1094–2103.5 nm and 1094–1983.5 nm), the optimal calibration was selected. Details are listed in Table 1. The initially observed in-line prediction error was reduced in three steps. First, the offline calibration dataset was extended with in-line data. The addition of in-line data did not have a significant impact on SEC (1.722% offline vs. 1.741% combined), but SEP of SNV- preprocessed in-line spectra was reduced to 1.559% (from 7.366%, see Table 1A-1). Second, the spectral range was decreased from 1094 to 2103.5 nm to 1094–1983.5 nm to decrease the spec-

106

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

Fig. 5. (A) Calibration against target concentration: A-1: Selection of number of components, A-2: Calibration plot (API in %) SEC = 1.722, A-3: Validation plot (API in %) SEP = 1.695, A-4: Residuals (predicted-measured). (B) Calibration against HPLC reference: B-1: Selection of number of components, B-2: Calibration plot (API in %) SEC = 2.596, B-3: Validation plot (API in %) SEP = 2.575, B-4: Residuals (predicted-measured).

Table 1 : (A) Calibration optimization with SNV-processed spectra of three calibration data sets: offline spectra, in- line spectra acquired at three different speeds (17.000, 38.000 and 70.000 tablet/h) and fusion of the offline and in-line spectra. A-1: Full spectral range calibrated (1094 to 2103.5 nm). A-2: Reduced spectral range calibrated (1094 to 1983.5 nm). (B) Calibration optimization with SNV and 1st Derivative processed spectra. B-1: Full spectral range calibrated (1094 to 2103.5 nm). B-2: Reduced spectral range calibrated.(1094 to 1983.5 nm) A-1 Calibration Dataset

Offline (4500 spectra)

Online (3 speeds, 6750 spectra)

Fusion offine and online (11250 spectra)

PCs SEC SEP (with offline spectra) SEP (with online Spectra) SEP (with fusion Spectra)

7 1.722 1.695 7.366 5.805

7 1.554 5.868 1.514 3.892

7 1.741 1.923 1.559 1.714

Calibration Dataset

Offline (4500 spectra)

Online (3 speeds, 6750 spectra)

Fusion offine and online (11250 spectra)

PCs SEC SEP (with offline spectra) SEP (with online Spectra) SEP (with fusion Spectra)

7 1.682 1.650 6.635 5.244

7 1.476 6.879 1.424 4.488

7 1.637 1.810 1.448 1.602

Calibration Dataset

Offline (4500 spectra)

Online (3 speeds, 6750 spectra)

Fusion offine and online (11250 spectra)

PCs SEC SEP (with offline spectra) SEP (with online Spectra) SEP (with fusion Spectra)

7 2.299 2.243 2.681 2.515

7 1.822 3.158 1.755 2.416

7 2.080 2.401 1.761 2.041

Calibration Dataset

Offline (4500 spectra)

Online (3 speeds, 6750 spectra)

Fusion offine and online (1125011,250 spectra)

PCs SEC SEP (with offline spectra) SEP (with online Spectra) SEP (with fusion Spectra)

7 1.924 1.911 2.057 2.000

7 1.723 3.105 1.634 2.336

7 1.868 2.115 1.608 1.827

A-2

B-1

B-2

tral noise, causing SEC and SEP of the fused calibration to decrease slightly when applied to offline, in-line and fused spectra. Third, SNV-preprocessing was combined with 1st Derivative, with the aim

to reduce the tableting-speed dependent systematic error. Similar SEPs were achieved between both preprocessing methods (1.448% for SNV vs. 1.608% for SNV&1st Derivative, see Table 1A-2 and B-2,

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

107

tablet press is high 20000–70 000 tablets /h. At this speed a single tablet cannot be isolated in order to perform HPLC and retrieve the inline predicted value. The gravimetric approach was proposed by [35]. In a laboratory, it has been shown that the two calibrations against HPLC or against gravimetric method provide very similar results (Fig. 5). The gravimetric approach is useful, especially in early phase of the development (e.g. the reference HPLC is not always fully validated) or to support the validation of qualitative or a semi-quantitative methods. 3.3. Process monitoring: comparison of qualitative and quantitative approaches

Fig. 6. Prediction of in-line and offline data with the offline-calibration (see Fig. 5A). Higher residuals were computed with in-line spectra.

respectively). However, the combination of SNV plus 1st Derivative preprocessing, demonstrated to successfully minimize the impact of the process speed on the NIR spectra. In summary, the combination of in-line and offline spectra, SNV plus 1st Derivative preprocessing and a decreased spectral range could improve the accuracy and robustness of calibration the most, with a SEC = 1.868% and SEP in-line = 1.608% (see Table 1 and Fig. 7 for details). 3.2.2. Optimization for blend uniformity (tablet feed frame) Similar optimization was performed with the spectra for granules’ blend uniformity. SEP increases when applying in-line data to a SNV-processed calibration made exclusively from offline data (from 1. 591% with offline data to 4.815% with in-line data). By fusing in-line and offline data during calibration, a significant reduction of SEP to 2.606% was achieved. By combining SNV and first derivative, SEP was further decreased to 2.240% (applied to in-line data). In conclusion, the combination of SNV and 1st Derivative reduced the bias of the sampling mode (offline vs. in-line) on prediction errors, and increased the calibrations accuracy and robustness. Results of the initial calibration and the final, optimized calibration are compared in Fig. 8 and Table 2. In summary, the observed linear relationship between NIRS and reference values was confirmed through a quantitative calibration. Combination of in-line and offline calibration data, preprocessed by SNV and 1st Derivative, achieved an accurate and robust prediction of API blend uniformity of granules in the tablet press feed frame. The gravimetric method was used as reference method due to the sampling challenge during real production. The speed of the

In the second part of this project, different qualitative and quantitative solutions for NIRS-process monitoring of granules’ blend uniformity and tablets content uniformity were compared. To do so, granules varying in API content over time were compressed in the tablet press and in-line NIRS spectra were recorded in the feed frame and from final tablets. In detail, tableting was started with granules containing 100% of the label claim (LC) API content. In this step, two different tableting speeds were applied (38.000 and 70.000 tablets/h). After about five minutes, granules containing 130% LC were added to the tablet press hopper for approximately five minutes, before switching to granules containing only 70% LC granules (tablet speed remained at 38.000 tablets/h during these changes). Tablets were compressed continuously throughout this experiment without emptying the hopper between the different granule types. Qualitative and Quantitative methods were compared only with spectra obtained from tablets with the VisioNIR instrument. In a second step, quantitative results from both NIRS systems were compared. 3.3.1. Qualitative process monitoring Qualitative analysis of NIRS spectra allows process monitoring without vast prior knowledge of the process and materials. Possible solutions are single wavelength monitoring (SWM) and Principal Component Analysis (PCA). In SWM, the absorption of the main Diclofenac Sodium peak at 1670 nm was monitored, in PCA the scores of SNV and 1st Derivative preprocessed spectra were observed over time. Results are depicted in Fig. 9 A. For SWM, the use of derivative reduces the effect of the speed and the qualitative method provides accurate information on API-content trending (see Fig. 9 A-1). PCA analysis depicts the same qualitative results on process trends, but results are influenced by the speed of the tablet press, despite of the 1st derivative (see Fig. 9 A-2). Generally, multivariate methods are known to improve a methods robustness, by considering numerous spectral variations. In this case however, PCA was not effective to eliminate the effect of speed. Consequently,

Fig. 7. Prediction of in-line and offline data with the optimized calibration computed from offline and in-line data (preprocessed by SNV and 1st Derivative, spectral range reduced to 1094–1983.5 nm). A-1 Measured vs predicted calibration set (API content in %). A-2 Measured vs predicted validation set (API content in %). A-3 Residuals from offline and in-line data. No Bias observed with in-line data.

108

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

Fig. 8. Comparison of offline and fused (offline and in-line) calibration and validation results for blend uniformity in tablet feed frame. (A) Initial offline calibration: A-1: calibration with offline spectra (SNV preprocessed), A-2- validation with offline and in-line spectra, A-3- residuals with offline and in-line spectra. (B) Optimized calibration: B-1: calibration with offline and in-line spectra (SNV and 1st Derivative), B-2: validation with offline and in-line spectra, B-3: residual with offline and in-line spectra. Table 2 : Calibration optimization for API content in granules’ blend uniformity in tablet feed frame. A-1 SNV preprocessing. A-2: SNV & 1st Derivative preprocessing. A-1 Calibration Dataset

Offline (221 spectra)

In-line (905 spectra)

Fusion offine and in-line (1126 spectra)

PCs SEC SEP offline spectra SEP online spectra SEP fusion spectra

7 1.624 1.591 4.815 4.375

7 2.292 10.507 2.260 5.073

7 2.739 3.172 2.606 2.726

Calibration Dataset

Offline (221 spectra)

In-line (905 spectra)

Fusion offine and in-line (1126 spectra)

PCs SEC SEP offline spectra SEP online spectra SEP fusion spectra

7 1.631 1.521 5.271 4.775

7 2.177 4.072 2.042 2.571

7 2.308 2.057 2.240 2.205

A-2

Independent Component Analysis (ICA) was applied [28–30]. ICA allows the separation of sources of variation, and thus allows to separate the effect of the API content from the effect of tableting speed, as demonstrated in Fig. 9 B, making it the more robust method for qualitative process control. In conclusion, ICA proved to be an adequate solution for qualitative API content monitoring, when no quantitative calibrations are available. This might be especially useful during early stages of development. 3.3.2. Quantitative process monitoring The optimized PLSR calibration from Fig. 7 was used for quantitative process monitoring of tablets content uniformity in real-time. Results are summarized in Fig. 10. The predicted API

content at 100% LC was accurate and robust to the change in tableting speed. Uniformity of dosage units was further demonstrated in accordance to Ph.Eur. 9.5, chapter 2.9.47 - Demonstration of uniformity of dosage units using large samples sizes. The average NIRS predicted API content of 4500 100% LC-tablets was X¯ = 98.29%, with standard deviation s = 2.11%. According to the monograph, this results in an acceptance value (AV) of 5.07, which is far below the allowed limit of 15. Furthermore, the number of individual dosage units outside of the defined content range 74% - 123% was zero, while the allowed limit was 42 [33]. Details on the calculation and the results are shown in Fig. 10A-2. The subsequent changes to 130% and 70% of LC were clearly indicated by the NIRS method, including the transient phases, where

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

109

Fig. 9. Qualitative approaches to monitor tablet content uniformity. A-1: Single wavelength monitoring with SNV and 1st Derivative preprocessed spectra. A-2: PCA scores of SNV and 1st Derivative preprocessed spectra. PCA scores are influenced by API content and speed, despite 1st derivative preprocessing of spectra. (B) Independent Component Analysis (ICA) allows the separation of sources of variations API content (B-1) and tableting speed (B-2).

Fig. 10. A-1: Quantitative monitoring of tablets content uniformity throughout the experiment by optimized PLSR calibration (Predictions in % of label claim). A-2 Demonstration of uniformity of dosage units using large samples sizes according to Ph.Eur. 9.5, chapter 2.9.47 [33]. Both, the acceptance value (AV) and the maximum allowed number of individual dosage units outside of the recommended range were met.

tablets of slowly increasing and decreasing API contents were compressed, due to mixing of different types of granules in the tablet hopper (see Fig. 10A-1). In summary, quantitative monitoring of tablet content uniformity allows more accurate process control than qualitative approaches, but time-consuming sampling and reference analytics for calibration are commonly required during method development. Consequently, this method is more suitable for late stage and commercial manufacturing processes.

3.3.3. Redundant quantitative process monitoring of blend and content uniformity In the last part of the project, quantitative predictions from the two different NIR probes (tablet press feed-frame and final tablets) were compared. With this redundant setup, tight monitoring of API content over time is facilitated, which can be beneficial for real® time-release approaches in the future. Since the VisioNIR LS solely scans a tablets surface for content uniformity prediction, the NIRS in the feed-frame adds a second layer of security to the process

110

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111

Fig. 11. Quantitative Monitoring with the two NIR probes for redundant process control (x axis: time in minute – y axis prediction in %).

control strategy, as it demonstrates the granules blend uniformity shortly before tablet compression. Results of the comparison are depicted in Fig. 11. The observed time-shift between the two graphs represents the materials residence time in the press. The variation in API content in the feed-frame, especially during the transition between granules of different label claims, reflects the mixing phase inside the feed frame. Due to the press design, excessive granules that are scraped off from the die are looped back into the fill-shoe. Consequently, the residence time of granules with 100% LC is prolonged by the loop, which is visible as the tablets API content still exhibits a slight positive slope over time after the initial, steep transient phase is finished. The partial recycle loop inside the feed frame manifests itself in the negative peaks that indicate a recycle loop period of about 1 min. The same is observed after switching from 130% to 70%. Consequently, the application of two NIR probes also allows thorough studies on the materials residence time and backward-mixing in the feed-frame, which is of high importance for continuous manufacturing processes and needs to be studied in more detail in the future. As well it serves as characterization tool to optimize material flow in the feed frame by modifications of the feed frame design with the goal of improving the transient behavior of the feed frame in order to eliminate partial dead-time characteristics in the material flow. 4. Conclusion The results presented in this study demonstrate that the tableting step in a continuous manufacturing process can be precisely monitored through the combination of two near infrared probes. The first probe analyzes the dried granules in the tablet press feed frame and verifies their uniformity in API content, to avoid segregation issues that could occur after milling and transfer. The second probe allowed 100%, accurate and robust monitoring of tablet content uniformity in real-time at varying tableting speeds up to 70,000 tablets/h. Different approaches for the development and optimization of calibration methods were presented and evaluated. To minimize the effect of tableting speed the choice of spectral preprocessing methods was of high importance. Also, selecting the right data for calibration played a huge role for the calibrations accuracy, linearity and robustness. In the presented example, SNV and 1st derivative were required to reduce the impact of process speed on spectral

acquisition and in-line and offline spectra had to be included into the calibration and validation data set for the best accuracy and robustness. Qualitative approaches that allow spectral data analysis without prior time-consuming calibration, can be used during early development when few samples are available. Especially Independent Component Analysis proved to be a useful qualitative multivariate analysis tool, as it allowed to separate the sources of spectral variation between API content and tableting speed for analysis. Quantitative calibration approaches require reference analytics of samples and more effort for method validation. In turn, the results are more accurate and can be applied for precise process monitoring and control, especially in late phase and commercial production, once the final formulation and dosage form has been selected. The study demonstrated that the speed of the tablet press is critical and spectra acquired at different speed as to be included in the calibration set. In addition to API content, NIRS allows accurate prediction of granules or tablets water content (LOD) after appropriate calibration. Since LOD can impact the final products shelf life, it is considered another critical quality attribute that needs to be monitored closely [34]. The implementation of NIRS calibrations on LOD is currently ongoing, to further improve the process control system. In the future, the combination of the two NIRS probes with chemometric monitoring of compression parameters will be evaluated. This combination promises 100% real-time control of both, content uniformity and tablet weight variation, giving a clear picture of every single tablet that is produced [22–24]. With this knowledge, single tablets that do not meet their predefined quality characteristics can be ejected precisely and feedback control loops can be implemented to conduct countermeasures as soon as a drift in the process was detected. The development of such an ejection system and the implementation of corresponding control loops is currently ongoing. Likewise, tablets exiting the press that demonstrated to be within quality specification are eligible for real-time release. Yields are increased and thus economical losses and quality risks decreased.

References [1] V. Pauli, et al., Methodology for a variable rate control strategy development in continuous manufacturing applied to twin-screw wet-granulation and continuous fluid-bed drying, J. Pharm. Innov. (2018) 1–14. [2] Guidance for Industry, PAT - A Framework for Innovative Pharmaceutical Development, Manufacturing and Quality Assurance (Pharmaceutical cGMPs), U.S. Department of Health and Human Services, Food and Drug Administration, Center for Drug Evaluation and Research (CDER), Center for Veterinary Medicine (CVM), Office of Regulatory Affairs (ORA), 2004. [3] G. Allison, et al., Regulatory and quality considerations for continuous manufacturing May 20–21, 2014 continuous manufacturing symposium, J. Pharm. Sci. 104 (3) (2015) 803–812. [4] I.C.H. Harmonized, Tripartite Guideline: Validation of Analytical Procedures: Text and Methodology: Q2(R1), Available from:, 2005 https://www.ich.org/ fileadmin/Public Web Site/ICH Products/Guidelines/Quality/Q8 R1/Step4/Q8 R2 Guideline.pdf. [5] Analytical Procedures and Methods Validation for Drugs and Biologics, Food and Drug Administration (FDA) - Guidance for Industry, 2015. [6] Development and Submission of Near Infrared Analytical Procedures Guidance for Industry (Draft Guidance), United States Department of Health and Human Services Food and Drug Administration, Center for Drug Evaluation and Research (CDER), 2015. [7] Guideline on the Use of Near Infrared Spectroscopy by the Pharmaceutical Industry and the Data RequireMents for New Submissions and Variations EMEA/ CHMP/ CVMP/ QWP/ 17760/ 2009 Rev2, European Medicines Agency, 2014. [8] Chapter 2.2.40. Near-infrared spectroscopy, in: European Pharmacopoeia 9.5, 2018. [9] Chapter 5.21. chemometric methods applied to analytical data, in: European Pharmacopoeia 9.5, 2018. [10] General Chapter <1119> Near-Infrared Spectroscopy in United States Pharmacopeial Convention (USP38), 2015. [11] Y. Roggo, et al., A review of near infrared spectroscopy and chemometrics in pharmaceutical technologies, J Pharm. Biomed. Anal. 44 (3) (2007) 683–700.

V. Pauli et al. / Journal of Pharmaceutical and Biomedical Analysis 162 (2019) 101–111 [12] C. De Bleye, et al., Critical review of near-infrared spectroscopy methods validations in pharmaceutical applications, J. Pharm. Biomed. Anal. 69 (2012) 125–132. [13] O. Berntsson, et al., Quantitative in-line monitoring of powder blending by near infrared reflection spectroscopy, Powder Technol. 123 (2) (2002) 185–193. [14] M. Blanco, R. Gozalez Bano, E. Bertran, Monitoring powder blending in pharmaceutical processes by use of near infrared spectroscopy, Talanta 56 (1) (2002) 203–212. [15] A. Burggraeve, et al., Development of a fluid bed granulation process control strategy based on real-time process and product measurements, Talanta 100 (2012) 293–302. [16] Y. Hayashi, T. Sato, M. Otsuka, Real-time monitoring of the drying of extruded granules in a fluidised bed using near infrared spectroscopy and kinetic evaluation of the drying process, J. Near Infrared Spectrosc. 21 (2) (2013) 107–115. [17] P. Chalus, et al., Near-infrared determination of active substance content in intact low-dosage tablets, Talanta 66 (5) (2005) 1294–1302. [18] D.J. Goodwin, et al., Real time release testing of tablet content and content uniformity, Int. J. Pharm. 537 (1-2) (2018) 183–192. [19] L.M. Kandpal, et al., Quality assessment of pharmaceutical tablet samples using fourier transform near infrared spectroscopy and multivariate analysis, Infrared Phys. Technol. 85 (2017) 300–306. [20] M. Dyrby, et al., Chemometric quantitation of the active substance (containing C≡N) in a pharmaceutical tablet using near-infrared (NIR) transmittance and NIR FT-Raman spectra, Appl. Spectrosc. 56 (5) (2002) 579–585. [21] Chapter 2.9.40. - Uniformity of dosage units, in: European Pharmacopeia 9.5, 2018. [22] A. Bhaskar, F.N. Barros, R. Singh, Development and implementation of an advanced model predictive control system into continuous pharmaceutical tablet compaction process, Int. J. Pharm. 534 (1-2) (2017) 159–178. [23] Y. Hattori, M. Otsuka, Modeling of feed-forward control using the partial least squares regression method in the tablet compression process, Int. J. Pharm. 524 (1-2) (2017) 407–413.

111

[24] S. Matero, et al., Towards better process understanding: chemometrics and multivariate measurements in manufacturing of solid dosage forms, J. Pharm. Sci. 102 (5) (2013) 1385–1403. [25] K. Jarvinen, et al., 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 48 (4-5) (2013) 680–688. [26] Å. Rinnan, Fvd. Berg, S.B. Engelsen, Review of the most common pre-processing techniques for near-infrared spectra, TrAC Trends Anal. Chem. 28 (10) (2009) 1201–1222. [27] M. Clavaud, et al., Global regression model for moisture content determination using near-infrared spectroscopy, Eur. J. Pharm. Biopharm. 119 (2017) 343–352. [28] Aapo Hyvärinen and Erkki Oja, Independent component analysis: algorithms and applications, Neural Netw. 13 (4-5) (2000) 411–430. [29] Aapo Hyvärinen, Juha Karhunen, Erkki Oja, Independent Component Analysis, 1 ed., John Wiley & Sons, New York, 2001. [30] James V. Stone, Independent Component Analysis: a Tutorial Introduction, MIT Press, Cambridge, Massachusetts, 2004. [31] S. Wold, M. Sjöström, L. Eriksson, PLS-regression: a basic tool of chemometrics, Chemom. Intell. Lab. Syst. 58 (2) (2001) 109–130. [32] F.J.S. Nieuwmeyer, et al., Granule characterization during fluid bed drying by development of a near infrared method to determine Water content and median granule size, Pharm. Res. 24 (10) (2007) 1854–1861. [33] Chapter 2.9.47. - Demonstration of Uniformity of Dosage Units Using Large Sample Sizes, European Pharmacopeia 9.5, 2018. [34] T. Monteyne, et al., Continuous melt granulation: influence of process and formulation parameters upon granule and tablet properties, Eur. J. Pharm. Biopharm. 107 (2016) 249–262. [35] B. Nagy, et al., Spectroscopic characterization of tablet properties in a continuous powder blending and tableting process, Eur. J. Pharm. Sci. 123 (2018) 10–19.