Continuous downstream processing of milled electrospun fibers to tablets monitored by near-infrared and Raman spectroscopy

Electrospinning is a technology for manufacture of nanoand micro-sized fibers, which can enhance the dissolution properties of poorly water-soluble drugs. Tableting of electrospun fibers have been demonstrated in several studies, however, continuous manufacturing of tablets have not been realized yet. This research presents the first integrated continuous processing of milled drug-loaded electrospun materials to tablet form supplemented by process analytical tools for monitoring the active pharmaceutical ingredient (API) content. Electrospun fibers of an amorphous solid dispersion (ASD) of itraconazole and poly(vinylpyrrolidone-co-vinyl acetate) were produced using high speed electrospinning and afterwards milled. The milled fibers with an average fiber diameter of 1.6 ± 0.9 μm were continuously fed with a vibratory feeder into a twin-screw blender, which was integrated with a tableting machine to prepare tablets with ~ 10 kN compression force. The blend of fibers and excipients leaving the continuous blender was characterized with a bulk density of 0.43 g/cm 3 and proved to be suitable for direct tablet compression. The ASD content, and thus the API content was determined in-line before tableting and at-line after tableting using nearinfrared and Raman spectroscopy. The prepared tablets fulfilled the USP <905> content uniformity requirement based on the API content of ten randomly selected tablets. This work highlights that combining the advantages of electrospinning (e.g. less solvent, fast and gentle drying, low energy consumption, and amorphous products with high specific surface area) and the continuous technologies opens a new and effective way in the field of manufacturing of the poorly water-soluble APIs.


Introduction
A large portion (up to 75%) of small molecule drug candidates in today's development portfolio has poor water solubility and is one of the main challenges in the pharmaceutical industry (Williams et al., 2013). Amorphous solid dispersions (ASDs) are a common technology and effective way to enhance the dissolution properties of poorly water-soluble molecules (Jermain et al., 2018;Pandi et al., 2020). Several techniques are known to prepare ASDs among which spray drying and hot-melt extrusion are the most typically used (Jermain et al., 2018;Pandi et al., 2020). However, electrospinning is also a very gentle and promising continuous method since it is capable of drying at ambient temperature and operates at atmospheric pressure (Yu et al., 2018). Furthermore, the energy consumption and the solvent needs during ES are favorable compared to the other solvent-based ASD preparation methods (Drosou et al., 2017;Kang et al., 2020;Levit et al., 2018;Sóti et al., 2015;Vass et al., 2019a).
Although the mentioned advantages make the technique suitable for pharmaceutical applications, there are still some key challenges that need to be addressed.
Besides increasing the dissolution of the potential drug molecules, the development of the technologies and integration of continuous manufacturing (CM) processes is a hot topic nowadays in the pharmaceutical industry too. The marketed products prepared by CM and the increasing number of publications about CM show that continuous technologies get more and more attention in the pharmaceutical field (Burcham et al., 2018;Lawrence and Kopcha, 2017). Nonetheless, there are two FDA-approved CM medicines contain spray-dried samples, which indicate the opportunities in the simultaneous application of ASDs and CM . Although spray drying is a widely used method with high productivity, only a few publications have been published relating to the CM of spray-dried materials to tablets (Adali et al., 2020;Hu et al., 2011;Vanhoorne et al., 2016). From the industry point of view, the two FDA-approved spray-dried material-loaded CM medicines confirm that the topic has great industrial relevance. Using CM in the case of other ASDs products might also be effective with respect to the practical application since a high volume of ASD-loaded products are manufactured year by year and the cost of production or the energy consumption can be reduced in this way (Burcham et al., 2018;Pandi et al., 2020). However, it is worth keeping in mind that the handling of products with poor flowability is a general challenge for tablet development and it is even more difficult during the CM of spray-dried samples or electrospun materials with low bulk densities (Al-Zoubi et al., 2021). Therefore, overcoming the flowability problems of critical materials such as active pharmaceutical ingredients (APIs) or ASDs is a crucial part of continuous process development (Besenhard et al., 2016;Chattoraj and Sun, 2018;Pingali et al., 2009;Szabó et al., 2019).
In general, one of the most critical parts of continuous tablet preparation is the feeding because it has a great influence on the homogeneity and the content uniformity of the blend and later the tablets (Blackshields and Crean, 2018). Consequently, the feedability of materials with poor flow properties needs to be improved with different formulation techniques. In the case of APIs or poorly flowable excipients, the most common method is coating the given material with silica, which improves the flowability and thereby the performance of the feeding (Escotet-Espinoza et al., 2020;Kunnath et al., 2018;Mullarney et al., 2011). Besides, the feedability enhancement can also be done through equipment design, for instance via application of vibratory feeders since the danger of clogging, and electrostatic charging of the particles is less than in the case of the twin-screw feeders (Besenhard et al., 2016). However, not only accurate feeding but also effective blending is important to prepare appropriate tablets continuously. The homogeneity of the blend largely depends on the process parameters of the continuous blending thus the determination of the influencing factors is a significant element of the process development (Liu et al., 2018;Vanarase et al., 2013).
Furthermore, introducing the process analytical technology (PAT) principles, the drug content needs to be monitored during the continuous blending to produce good quality products. Besides, the use of in-line analytical methods may contribute to the feedback control of the processes, which is a critical part of CM systems . To accomplish quick, real-time determination of the API content, different non-destructive in-line analytical tools must be applied (De Beer et al., 2011). A widely used method is the near-infrared (NIR) spectroscopy, which has been already used successfully in many continuous blending processes (Colón et al., 2014;Koller et al., 2011;Vanarase et al., 2010). Similar results can be achieved by Raman spectroscopy since the organic API molecules have strong signals in the Raman spectrum Vergote et al., 2004). The real-time monitoring of the homogeneity and drug content has an important role in the process development because the effect of different parameters can be followed continuously (Martínez et al., 2013).
CM of electrospun material containing tablets is a rarely researched area because the industrial relevant scaled-up production of API-loaded fibers has not yet spread. The commonly used single needle electrospinning apparatus has very low productivity, circa 0.01-1 g/h (Vass et al., 2019c); therefore, it of limited use for ASDs where production rates of double digit kg/h are typically required (Jermain et al., 2018;Pandi et al., 2020). To achieve the necessary production rate, different electrospinning principles and equipment designs have been researched to increase the efficiency of the technology (Vass et al., 2019c). For instance, the multi-jet electrospinning utilizes more needles (Kumar et al., 2010), the free surface electrospinning methods are based on the curvature formation of the solution surface (Ahmed et al., 2020;Jiang and Qin, 2014), the alternating current electrospinning replaces the common used direct current techniques Kessick et al., 2004), and high speed electrospinning (HSES) takes the advantage of both electrostatic forces and centrifugal forces . The latter can be very promising from the pharmaceutical application point of view since even a 450 g/h production rate is achievable and the technology is considered scale-able  and is similar to rotary spray dryers, which are scale-able to large capacity (Masters, 1985).
The second main question relates to the conversion of the electrospun fibers into a final dosage form. Although preparing orally dissolving webs via electrospinning is an easy and possible solution for pharmaceutical usage Celebioglu and Uyar, 2019;Sipos et al., 2019), the manufacture of the most commonly used dosage forms of capsules and tablets using electrospun fibers is more complicated due to their physical properties such as low bulk density or the fibrous structure (Démuth et al., 2017). The tableting of the pure electrospun products seemed to be possible in lab-scale (Hamori et al., 2016) but an industrially relevant process requires the application of the appropriate downstream steps and the suitable excipients (Démuth et al., 2016;Vass et al., 2019b).
Besides, coupling continuous processing steps to the electrospinning is also challenging since the electrospun products with low bulk density can cause difficulties during the feeding, blending and tableting . Increasing the flow properties of fluffy fibers via milling is the key point for preparing tablets (Vass et al., 2019b).
The main goals of the current research were to investigate the continuous downstream processing of milled electrospun fibers to tablet form for the first time and to develop in-line analytical methods for determining the ASD content, and thus the API content in real-time.
An integrated system consisting of continuous feeding, blending and tableting was tested for the continuous manufacture of electrospun materials. Furthermore, the comparison of two non-destructive, in-line analytical tools, namely NIR and Raman spectroscopy, was also performed to investigate their applicability during continuous downstream processing of electrospun samples. The development of a fully continuous manufacturing line using electrospinning coupled with the appropriate PAT tools might be very promising in the pharmaceutical industry for effective manufacture of ASDs.

High speed electrospinning (HSES)
Electrospinning of the ITR-loaded fibers was accomplished using HSES equipment, which was combined with a cyclone (Figure 1) (Vass et al., 2019a). A fan provided a constant 120 m 3 /h gas flow rate in the system during the whole production period. The key element of the apparatus is the stainless steel, round-shaped spinneret with 36 orifices (d = 500 μm) on its edge. This spinneret is connected to a pneumatic air-bearing turbine to reach a high rotation speed. The rotation speed of the spinneret was set to 40000 rpm. The preparation of fibrous product was carried out at 25 °C  1°C and 45 ± 5% relative humidity, while the applied voltage was fixed at 40 kV. The solid material concentration of the investigated composition was 0.375 g/ml (consisted of 40% ITR and 60% PVPVA64) and the solids were dissolved in the mixture of dichloromethane:ethanol (volume ratio 2:1) . The solution containing the drug and the polymer was fed with a built-in peristaltic pump (Watson−Marlow Fluid Technology Group, Budapest, Hungary) with a flow rate of 1000 mL/h.

Milling of the electrospun product
A QUICKmill Lab multifunctional milling apparatus (Quick2000 Ltd., Tiszavasvári, Hungary) was used to reduce the fiber length to increase the bulk density and thereby to obtain a powder from the fibrous material with better flow properties (Section 3.1.). The equipment can be operated both in oscillating and conical milling mode depending on the applied grinding head. During this work, the oscillating mode was chosen due to its higher capacity and less material loss. A sieve with holes of 2.0 mm was used for the milling of the prepared electrospun sample. The milling rate was adjusted to 200 cycle/min, which resulted in circa 200 g/h milling capacity in the case of the investigated ITR-loaded fibrous system.

Characterization of the electrospun product
The basic characterization of the ITR-loaded electrospun material after milling was performed using differential scanning calorimetry, X-ray powder diffraction, in vitro dissolution testing, scanning electron microscopy, and laser diffraction measurements, as in our previous studies (Démuth et al., 2017;Démuth et al., 2018;. Nonmilled fibers were investigated only by scanning electron microscopy since the non-milled samples cannot be examined by laser diffraction and the results of all the other measurements did not show any differences between the milled and non-milled samples according to prior experiences. Besides, the amount of the residual solvents is also a crucial factor during the application of solvent-based methods, which can examine well with headspace gas chromatography . The method in the case of ITR-loaded fibers was similar than in one of our prior studies (Nagy et al., 2012). Our previous researches connected to the same composition showed that 700 ppm of ethanol and 300 ppm of dichloromethane was found in the ITR-loaded electrospun samples after two days of storage at normal circumstances (open air), which values were far under the regulation limits (5000 ppm and 600 ppm for ethanol and dichloromethane, respectively). During this work, the electrospun fibers were used in further experiments only few days later thus residual solvents determination was not performed.
The prepared sample proved to be amorphous ( Figure S1 and S2) and showed good dissolution ( Figure S3). Since the basic characterization of the same composition was detailed in our previous studies, only the results of the newest in vitro dissolution tests, the scanning electron microscopy records, the particle size analysis and the bulk-tapped density test are showed in this paper. Therefore, only the above-mentioned methods are detailed in the following subsections.

In vitro dissolution tests
In vitro dissolution studies were carried out on a Pharmatest PTWS 600 dissolution tester (Pharma Test Apparatebau AG, Hainburg, Germany). The pure electrospun samples and the reference crystalline API were examined by the combination of the USP I and USP II method (tapped-basket method) . The fibrous materials were milled and then the powders were weighted into the dissolution baskets. The dissolution of tablets was measured with the common USP II (paddle) method. The drug content was 50 mg in each case. The applied dissolution media was 900 mL of 0.1 N HCl set at a constant temperature of (37 ± 0.5) °C. The stirrer speed by the tapped-basket method and by the paddle method were 50 and 100 rpm, respectively. An Agilent 8453 UV-Vis spectrophotometer (Hewlett-Packard, Palo Alto, USA) was connected to the dissolution tester through flow cells to on-line measure the concentration of the dissolved ITR. Based on a preliminarily built calibration at a wavelength of 254 nm (from 1 to 50 mg/L), the concentration of the dissolved ITR was calculated in real-time during the measurements. Each different sample was investigated in triplicate.

Scanning electron microscopy
A JEOL JSM 6380LA (JEOL, Tokyo, Japan) type scanning electron microscope (SEM) was utilized for examining the morphology and size of the ITR-loaded electrospun material. The investigated specimen was fixed with conductive double-sided carbon adhesive tape. To avoid electrostatic charging, the electrospun sample was sputtered with gold in the next step of sample preparation. The SEM measurement was accomplished in a high vacuum.
The applied working distance was 15 mm while the accelerating voltage was 15 kV.
Calculation of the average fiber diameter was performed by a randomized diameter determination method (Balogh et al., 2015).

Particle size analysis
The particle size distribution of the electrospun sample after milling was determined by a Malvern Mastersizer 2000 type laser diffractometer (Malvern Instruments Ltd., Worchestershire, UK). Before the measurement, a background recording was adjusted for 45 seconds. Then a vibratory feeder added the powder into the equipment with 75% intensity of the vibrational amplitude. The measurement took 60 seconds while the applied pressure was 1.5 bar. The measured d (0.5) values were described as the 50% cumulative undersize of the volumetric distribution, which was used to characterize the particle size.

Bulk and tapped density test
Furthermore, the bulk and tapped densities of the milled electrospun material were also investigated using an ERWEKA SVM12 (Heusenstamm, Germany) type tapped density tester. The flow property of the powder was determined based on the calculated Hausner ratio and the Carr index (Carr, 1965;Hausner, 1967).

Continuous blending and tableting
The targeted API content was 50 mg per 600 mg tablet, which was equal to 125 mg 40% ITR-loaded electrospun material ( muth et a 0 ). The experimental set-up of continuous blending and tableting consisted of two feeders, a twin-screw blender, a conveyor belt and a tableting machine ( Figure 2). The milled ITR-loaded fibers were fed into the blender with a LABORETTE 24 vibratory feeder (Fritsch GmbH, Idar-Oberstein, Germany). During the experiments, a Vshaped channel was applied to direct the powder into the blender. The feeding rate could be controlled by adjusting the vibration amplitude. Since gravimetric feeding could not be performed with this feeder, a pre-calibration of the vibration amplitude and the feeding rate was accomplished before each experiment to set the feeding rate to the targeted API content.
The pre-blend of all other excipients (including the lubricant) was fed with a Brabender DDW-MD0-MT type (Brabender Technologie, Duisburg, Germany) twin-screw feeder in gravimetric mode. The adjusted feeding rate was 750 g/h to achieve appropriate API content in the blend.
A TS16 QuickExtruder® (Quick 2000 Ltd., Tiszavasvári, Hungary) multifunctional equipment was applied for continuous twin-screw blending. The diameter of the screws was 16 mm, while the L/D ratio was 25. The rotation speed of the screws was set to 70 rpm during the continuous blending of the electrospun sample and the excipients.
After the continuous blending, a conveyor belt carried the existing powder mixture from the blender to the tableting machine. The tablets were compressed on a Dott Bonapace

NIR spectroscopy
A Bruker MPA Multi Purpose Fourier-transformed Near Infrared (FT-NIR) Analyzer (Bruker OPTIK GmbH, Germany) with a high intensity NIR source (Tungsten) and PbS detector was utilized as one of the PAT tools for monitoring the continuous blending and investigation of the prepared tablets. The spectra were collect in reflection mode while the investigated spectral range was set to 4000-10000 cm -1 . The resolution of 8 cm -1 was applied during the measurements and 16 scans were accumulated with double-sided, forwardbackward acquisition mode. 5 spectra per blend and 6 measurements per 3 tablets (both sides of each tablet) were used for calibration. seconds with two scans during the calibration and measuring of blends and tablets.

Multivariate data analysis
Each calibration sample consisted of the mixture of the excipients and the electrospun sample in different concentrations. The composition of the excipients-loaded powder was chosen based on a prior study and it is summarized in  thus in the calibration spectra, the spectral signs of the prepared ASD were characteristics and not of the pure API. Therefore, the calibration models were built on the ASD content and not on the API content. However, the API content of the electrospun fibers was also determined by UV-Vis measurements before the calibration. The number of latent variables was chosen by minimizing the root mean square error of cross-validation (RMSE CV ). Interval PLS method proved to be suitable in all cases for variable selection, where the number of maximal latent variables was changed based on the pretreated PLS model. All of the NIR spectra were preprocessed using Savitzky-Golay first derivative (a second-order polynomic function was fitted, while the number of points in the filter was fifteen, and only included data were applied). Multiplicative signal correction (MSC) using the mean spectra as reference and mean centering were also applied in the further steps of NIR spectra pretreatment. The Raman spectra were smoothed at first with Savitzky-Golay smoothing method (a second-order polynomic function was used, while the number of points in the windows was fifteen, and only included data were applied). Then all Raman spectra were baseline corrected using Automatic Whittaker Filter baseline correction with an asymmetry parameter p = 0.001 and a smoothing parameter λ = 0 5 . The Raman spectra of the blends were normalized to unit length, which is a widely used weighted normalization method. The intensity of the raw spectra showed greater differences (see Figure S5) thus this form of normalization, where the larger values were weighted more heavily in the scaling, seemed to be suitable for the spectra of calibration powders. The intensity values of the Raman spectra of the different concentration tablets were closer to each other; therefore, these raw spectra were normalized to unit area, which proved to be appropriate for these data ( Figure S5). In the final step of the pre-processing, mean centering was utilized in the case of the Raman spectra as well. For cross-validation, 6-fold and 7-fold venetian blind cross-validation were used, leaving out one sample per concentration levels at each calculation step methods , in the case of the blends and tablets, respectively. This validation found to be suitable since the replicate measurements were ordered randomly. Besides, external validation was also performed in the case of powders, which confirmed the applicability of the built models.
The models were compared by the coefficient of determination for calibration, crossvalidation, and prediction (R 2 C , R 2 CV, R 2 P ), and the root mean square error of calibration, cross-validation, and prediction (RMSE C , RMSE CV , RMSE P ). The performance of the built models was also characterized by the limit of detection (LoD) and limit of quantification (LoQ), which were calculated by equations 1 and 2 (de Carvalho Rocha et al., 2012).
Furthermore, the limit of Hotelling T 2 was calculated in the case of the selected models to determine an acceptance limit for the real-time experiment (equation 3) . (1) In equations 1-2 σ indicates the standard deviation (SD) of the predicted y-values for each x-value; S denotes the slope calculated from the measured and predicted concentrations.
The calculation of these values are detailed in the supplementary materials (Eq. S1 and S2). In equation 3 σ HotellingT 2 means the standard deviation of the Hotelling T 2 values and Hotelling T 2 MAX expresses the largest Hotelling T 2 value of the given model.

Determination of content uniformity
The content uniformity in the tablets was measured with UV-Vis spectroscopy, which was applied as a reference analytical method for verifying the results of the NIR and Raman spectroscopy measurements. The same wavelength and calibration were used as for the in vitro dissolution tests to determine the ITR concentrations in the tablets. The tablets were dissolved in 2 L of 0.1 N HCl dissolution media and stirred with a magnetic stirrer for 2 days.
The solutions were fi tered through a 0 45 μm fi ter before the UV-Vis measurements.

Characterization of tablets
Friability was measured on PharmaTest PTF 20E (Hainburg, Germany) type friability tester after 100 rounds on 10 tablets. Tablet breaking force was determined on a Schleuniger 4 M (Thun, Switzerland) type hardness tester with 10 tablets. A Sartorius MA40 (Göttingen, Germany) moisture analyzer was used for measuring the moisture content of 10 ground tablets. Loss in drying was determined at 105°C for 10 minutes. The thickness of the tablets was determined with a Pro-Max Electronic Digital Caliper (NSK, Tokyo, Japan).

Preparation of the electrospun and milled samples
The development of an integrated continuous formulation system requires the synchronization of each processing step with respect to capacity. During this research, the production rate of the electrospinning experiments was chosen to fit the next milling step, while the further steps (feedings, blending, tableting) were adapted to the production rate of the milled fibers to see the potential of a possible fully continuous line (from the electrospinning to the tableting). The feeding rate of the solution was set to 1000 mL/h, which seemed to be suitable to prepare adequate quality fibers (grindable, dry product with small fiber diameter). The yield of 78% was reached with the applied process parameters, which resulted in ~ 200 g/h production rate for the solid fibrous material. During longer productions, the yield of the electrospinning might be further enhanced. On the other hand, the material loss was observed on the wall of the drying chamber thus further optimization of the formulation and the equipment (e.g. additional air knives in the HSES equipment) could further increase the yield. The basic characterization showed similarities to the results of our previous articles (data not shown), which means that electrospun fibers containing an ASD of 40% ITR and 60% PVPVA64 with good dissolution properties (90% released within 10 min) were prepared using the aforementioned settings .
Milling of the fibrous products may result macroscopically near round-shaped particles with enhanced flowability therefore, it is important to choose well grindable APIpolymer compositions for electrospinning. The ITR-PVPVA64 system was easy to grind with an oscillating milling machine right after the ES because the powder did not stick into the hole of the sieve and the material loss was less than 5%. The successful milling right after the HSES assumes that the fibers dried enough during the continuous fiber collection by cyclone.
The average diameter of the ITR-loaded electrospun fibers was 1.6 ± 0.9 µm and the fibrous structure remained after the oscillating milling (Figure 3). The macroscopic characteristic of the milled product was investigated with laser diffraction and the results are shown in Figure 4 and Table 2. The observed multimodal particle size distribution with an average diameter of 12.6 ± 1.0 µm can be explained by the formation of different sized agglomerates after the milling. The high standard deviations and relative standard deviations suggest inhomogeneity in the macroscopic particle size after the milling, which can be a critical factor during the downstream processing of the milled fibers.
The SEM images also presented that the milled fibers became entangled and form smaller and larger agglomerates or bundles (Figure 2b).  Although the size distribution of these agglomerates showed higher deviations, the flowability of the milled fibers found to be good according to the bulk-tapped density test (Table 3) thus the sample seemed to be suitable for further downstream processing steps.
However, the milled fibrous material still has a low bulk density (~ 0.13 g/cm 3 ) and is prone to electrostatic charge, which can cause difficulties during the formulation (e.g. the sample could stick to the wall of the feeder, blender, and tableting machine and lead to weight variations during the processes). For this reason, the application of excipients with good flow properties such as large particle size microcrystalline cellulose or mannitol (Démuth et al., 2017), and effective blending is indispensable to continuously produce tablets. Furthermore, the investigated electrospun system with low bulk density revealed that the applicability of the powders cannot be predicted based on on y the Hausner ratio and Carr's index. Wider powder characterization can give more information about the materials with respect to the formulation processes (Van Snick et al., 2018).

Effect of the lubricant
One of the crucial parts of the method development was the preparation of the calibration samples. The first important question, had to be considered, was how to add the lubricant. A possible answer is the feeding of the lubricant directly before the tableting since the over-lubrication can be avoided in this way. However, continuous feeding of the lubricants with poor flowability may be challenging, therefore a pre-blend with the other excipients was more suitable from the CM point of view. For this reason, the effect of the sodium-stearyl-fumarate (SSF) on the dissolution was investigated first to see if it could be added to the mixture of the excipients before the continuous blending with the electrospun material. To test the impact of the lubricant, tablets were prepared in small quantities with batch homogenization method. The results showed that the tablets, where the SSF was added to the powder mixture with the other excipients and electrospun sample and homogenized for 30 minutes showed similarities with the dissolution of the tablets, where the SSF was added only after the homogenization of the blend for 30 minutes ( Figure 5). Although the applied homogenization times during the batch blending processes were much higher than the residence times during a CM, the dissolution did not deteriorate. Therefore, it can be stated that the lubricant does not mean a problem in the case of the examined composition and processing steps, thus the SSF was added to the calibration samples as well. The physical mixture contained the crystalline ITR, the polymer and the tableting excipients in the same concentrations as the ASD-loaded tablets.

Chemometric model building
To reduce the time and the material consumption of the calibration, 2.4 g blends were prepared at each concentration and measured off-line mode without any special sample preparation and destruction. Then the NIR and Raman spectra were pre-processed to find the most reliable regression models. Several variable selection methods were also tried for choosing the appropriate spectral regions that carry the most information. The raw and the pre-processed NIR and Raman spectra of the powders can be seen in Figure S4 and S5, respectively. The off-line performance characteristics of the selected models seemed to be satisfying according to the R 2 , RMSE and bias values (Table 4). Although the LoD and LoQ values proved to be lower for the NIR spectra-based model, the Raman spectroscopy also seemed to be appropriate to predict accurately the concentrations around the target value. The main goal of this work was to develop in-line applicable models to the continuous blending thus off-line performance parameters of the models needed to be supplemented with other important indicators. Before the real in-line tests, the validation of the selected models was accomplished with 7, 13 and 22 w/w% ASD-loaded samples, which were made in the same way as the calibration samples but not used during the model building. The validation blends were measured five times and the concentrations were predicted by the selected models. The predicted values were subtracted from the known concentrations and the obtained residuals can be found in Figure 6. The residuals of the repeated measurements were below 5 w/w%, which indicates acceptable predicting performance. The higher SDs in the case of the 13% ASD-loaded sample can refer to the inhomogeneity of the validation sample and not the error of the models since the SDs of the other two investigated concentrations proved the be suitable. However, it can be stated that the Raman spectra-based model can predict the concentrations more accurately since the SDs of the repeated measurements were lower. Furthermore, the averaged results of the Raman-based model showed fewer deviations from the known concentration of the compositions; therefore, it could be more promising in the in-line applicability point of view than the NIR spectra-based model (Table 5). It is important to note that the calculated deviations agree with the LoD and LoQ values of the models since the highest differences were observed at the 7% ASD-loaded sample, which concentration is below the LoQ values of 7.48 and 14.93 in the case of NIR and Raman spectra-based models, respectively. Furthermore, the R 2 P and the RMSE P values were also calculated to see the goodness of the models (Table 5). These indicators suggest that the Raman spectra-based model can be more accurate to predict in-line the ASD content during a real-time continuous blending process. For outlier detection, the critical limit of Hotelling T 2 values was determined, which were 26.68 and 25.06 in the case of NIR and Raman spectrabased models, respectively. The Hotelling T 2 values of the validation samples were far under the calculated limits therefore both models seemed to be suitable for in-line application at these wider acceptance limits.  electrospun materials seemed to be a proper solution. The application of excipients composites provide several advantages and has industrial relevance as well since some powder mixtures for CM are already on the market (Pharma, 2016). In this way, fewer feeders need to be used for continuous blending processes, which decreases the possibility of weight variations due to feeding errors. The blend of the excipients was characterized with excellent flowability and 0.45 g/cm 3 bulk density thus it proved to be perfect for a CM process.
Besides, the feedability of the powder was appropriate and well adjustable using a twin-screw gravimetric feeder. Although the feeding of the milled electrospun materials is more challenging due to the low bulk densities, a vibratory feeder was suitable for handling the where the rotation speed was changed from 50 rpm that found to be an appropriate setting to transport the mixture of excipients. The finally selected rotation speed was 70 rpm since this was the lowest adjustment, which proved to be suitable for efficient transport of the incoming pre-blend and electrospun fibers together. The bulk density of the outgoing blend was 0.43 g/cm 3 and its flowability proved to be appropriate for the tableting. At higher rotational speed, the powders were not able to totally fill the screws, which resulted in inhomogeneous blend.

Real-time monitoring of continuous blending
To investigate the homogeneity of the outgoing blend in the case of the set rotation speed the developed NIR and Raman spectra-based models were applied in a continuous experiment ( Figure 8). The aim of the continuous blending process was two-fold: to check the in-line applicability of the built PLS models, and to prepare homogenous electrospun ASDloaded blend continuously, which was never accomplished before according to the best knowledge of the authors. At the beginning of the process, only the excipients were fed into the blender. Both the NIR and Raman spectra-based models calculated values around 0% but the NIR spectroscopy showed some outlier based on the Hotelling T 2 values, especially at the first few measured points. Since the powder layer on the conveyor belt was thinner when the excipients were fed, these outliers suggest that the NIR spectroscopy was more sensitive to the thickness of the powder flowing under the probe. For this reason, it is worth paying a special attention to the sample thickness and because of it to the appropriate selection of the screw speed during the continuous blending processes . After the feeding of the electrospun material, the measured ASD content of the blend increased continuously, which was measured well with both applied spectroscopic methods. After emptying the vibratory feeder, the ASD content decreased to the starting point according to the NIR and Raman measurements, which also confirmed the efficiency of the built models. Black symbols indicate the outliers from the models according to the calculated Hotelling T 2 limits.
Grey background indicates that period when the feeding set up and the system reached a steady-state.
As expected based on the higher prediction performance values of the NIR validation, the NIR spectra-based model showed more outliers based on the Hotelling T 2 values, measured higher concentrations, and resulted in higher SD and relative standard deviation (RSD) in steady-state (Table 6). In contrast, the Raman monitoring showed more reliable values and the ASD concentrations were closer to the target concentration (20.8 w/w%) during the steady-state. It is worth mentioning that the presented work is a proof-of-concept relating to the continuous manufacturing of electrospun fibers to tablet forms. Therefore, the variability was high from the industrial point of view but standard deviations can be decreased when fully integrated lines are constructed. To reach the target concentration more accurately, the feed-back control can be applied based on the Raman spectra  or a gravimetric vibratory feeder could be used to adjust the feeding rate of the ASD. Table 6 The averaged ASD and API content in steady-state.
NIR spectra-based model Raman spectra-based model ASD content (w/w%)

Characterization of continuously manufactured tablets
After reaching the steady-state, tablets were prepared and analyzed from the continuously blended powder (Figure 9). The obtained average hardness of ten examined tablets was 95 N while the loss in drying was 3.34% according to the moisture analysis. The thickness of the prepared tablets were between 4.38 and 4.42 mm. Besides, the friability was 0.67%, which is under the 1.00% regulatory limit determined for uncoated tablets. The measured values of the basic characterization methods met the usual pharmaceutical requirements and well correlated the previous results of the tablets containing ITR-loaded fibers (Démuth et al., 2017). For further investigations, ten tablets were selected randomly with an average tablet weight of 578.2 mg and measured off-line by NIR and Raman spectroscopy. The evaluation of the spectra was performed using the built models from the calibration tablets (Table 7). The raw and the pre-processed NIR and Raman spectra of the tablets can be seen in Figure S4 and S5, respectively. The off-line performance parameters showed similarities to the models of the blends, and the LoD and LoQ values were also below the target concentration. The Hotelling T 2 values of the investigated tablets were well below the acceptance limit, as the highest values were 73.54 and 6.42 for the NIR and Raman measurements, respectively.
Although the tablets were measured off-line, the low Hotelling T 2 values and the higher acceptance limits seemed to be promising thus it would be worth testing the models in-line. The API content of the selected tablets was measured by a reference UV-Vis method as well, and the calculated API content results from the three different measurements (NIR, Raman, UV-Vis) are depicted in Figure 10. The ITR content of each tablet was in a narrow range around the targeted 50 mg dose and the deviation of the average values from the target value was less than 5% in all cases (Table 8), which also proved the feasibility of the presented CM setup. The API concentration calculated by the three different measurements showed similarities, and the average ITR content of the ten tablets was comparable in all cases (Table 8).  The Acceptance Value (AV) of the tablets calculated by Raman and UV-Vis methods were below the L1 = 15.0 acceptance limit, which means that the tablets passed the USP <905> content uniformity test (Pharmacopeia, 2007.). Furthermore, the ITR contents of the individual tablets were between 37.5 and 62.5 mg, i.e. between 75.0 and 125.0% of the label claim (the 100.0% target content was 50 mg in this case). The NIR spectra-based model resulted in higher AV due to the higher RSD values but the predicted API contents were similar to the results of the other two methods. Besides, the NIR and Raman spectra-based models were compared according to the error of the prediction, where the UV-Vis results were used as known API content. The RMSEP values were 2.96 mg and 2.52 mg for the NIR and Raman spectra-based calculations, which proved to comply with the USP limits. The UV-

Conclusions
Continuous manufacturing of tablets containing milled ITR-loaded electrospun fibers was successfully achieved through a multi-step system including electrospinning, milling, feeding, blending and tableting processing steps. The synchronization of the different parts of the investigated experimental setup was an important aspect of the work. The adjusted feeding rate during the electrospinning resulted in ~ 200 g/h ASD productivity and this value was the starting point of the continuous blending as well. For the examined ITR-PVPVA64 formulation, the results showed that it was possible to produce 600 mg tablets continuously.
Further developments could enable a fully continuous line integrating electrospinning with milling to the continuous tableting process. In that case, the application of the presented production rates could result in circa 38.400 tablets/day in the case of the 600 mg tablets with 50 mg API content. Moreover, further scale-up is also achievable , which can satisfy the requirements of the pharmaceutical industry.
Besides, real-time monitoring of the ASD content proved to be achievable using NIR and Raman spectroscopy and spectra-based PLS regression models. The off-line performance parameters of the models, the validation and the in-line experiments revealed that the Raman spectroscopy-based models are more robust and accurate than the NIR spectra-based models.
The UV-Vis measurements, used as a reference analytical method, confirmed that appropriate homogeneity was achieved in the final dosage form, which was well measurable with both NIR and Raman spectroscopy with 2.96 mg and 2.52 mg RMSE P values, respectively.