NIR Monitoring of a Hot-Melt Extrusion Process

Process analytical technology (PAT) and hot-melt extrusion (HME), commonplace in the food and polymer industries, are becoming increasingly deployed in the pharmaceutical industry. Herein the application of in-line, transmission mode, Fourier-transform near-infrared (FT-NIR) spectroscopy to the HME manufacturing platform for a drug product in development is detailed. NIR spectroscopy and partial least squares (PLS) models were developed for real-time active pharmaceutical ingredient (API) loading (%wt/wt) and surfactant loading predictions. These predictions were used for fault detection, isolation of suspect material, and real-time troubleshooting during HME. Additionally, the NIR/PLS output was used for real-time release of the intermediate drug product.

Extrusion is the process of shaping a material (extrudate) by forcing the material through an orifice (die) using pressure. Although extrusion has been used for decades in the plastics and food industries, it is relatively new to the pharmaceutical industry, with most of the activity occurring within the last decade (1–3). Hot-melt extrusion (HME) is the term usually applied to pharmaceutical applications of extrusion in which temperatures above the glass transition of that being extruded are used. HME is used for the preparation of drug formulations when the active pharmaceutical ingredient (API) has low solubility (Biopharmaceutics Classification System II/IV), assisting in the solubilization in vivo. This is achieved via conversion of a poorly soluble crystalline API to a more soluble amorphous form. The result is a solid dispersion of the amorphous drug substance stabilized within a polymer matrix. This increase in in-vivo solubility yields the potential for enhanced bioavailability in that the API dissolves more rapidly and maintains concentration, allowing greater adsorption potential compared with the parent crystalline form (2). In addition, surfactants or other additives can be included in the extruded formulation to further enhance solubility or otherwise improve the physical properties of the formulation (2). Further advantages of using HME include, but are not limited to, the following: no organic solvent is typically used, there are fewer processing steps, it is easily scalable, and there is the potential of continuous processing with a relatively small equipment footprint (3,4). Furthermore, extrusion enables a number of functions to be performed as a single unit operation (for example, mixing, melting, degassing and densification, and reacting of materials forming) thus increasing manufacturing efficiency. Extrusion enables process flexibility given that the screw and barrel are modular. The locations of venting, material feed, mixing, and heating–cooling zones are flexible to meet various processing objectives. Additionally, the duration of material residence within a given zone can be varied. More specifically, the residence time distribution characterizes the time the material is in the extruder and the extent of mixing.

During HME, two or more materials are added at the front of a screw extruder, and the screw (or screws) convey, compress, and degas the feed stream (or streams) before shearing, heating, and mixing of the feed (refer to Figure 1). The extrusion step for the drug product studied here consists of adding API, a surfactant, and a polymer via respective feeders to the extruder. After the mixture is in a homogeneous and fluid state, the extruder conveys and pressurizes the material so that it flows through the die. The die reduces the barrel configuration that houses the screws to a single circular orifice yielding a hot strand or sheet called extrudate. The extrusion process discussed herein has an additional module; the die adapter. The die adapter provides the NIR interface to the extruder. A chilled roll with a coarse mill, also known as a kibbler, is at the end of the extruder. This equipment is positioned so that the molten extrudate can fall vertically between two rolls where the material will drop below its glass transition temperature and form a brittle glass sheet. This sheet exits onto a conveyor belt that leads to the kibbler where it is broken into smaller flakes of glass to make the material more amenable to the milling process. After extrusion, the material is milled, blended with excipient, lubricated, and compressed into a conventional solid oral dosage form.

Figure 1: Schematic of the hot-melt extrusion process.

Due to the continuous nature of HME, it is often desirable to manufacture large batches of extrudate, with weights reaching thousands of kilograms, and values reaching millions of dollars. Thus, it is imperative that product quality be monitored and ensured throughout the batch; otherwise the entire batch is put at risk. HME is fundamentally a mixing process; ensuring that the product has the correct composition is the most basic and important aspect of product quality. Traditional off-line techniques such as chromatography can be used for composition measurements; however, the nature of the extrudate makes sample preparation difficult (4).

Process analytical technology (PAT), in general, yields numerous manufacturing advantages, but two specific to HME are noted here. PAT, as applied here, eliminates the need for sample preparation and timely off-line analyses. This is advantageous for obvious reasons, but more specifically the introduction of a polymer to the sample matrix can complicate the analyte (in this case API and surfactant) extraction for intermediate drug product quality assessment. Process monitoring better enables continuous manufacturing since any quality attribute deviations are identified in real-time. For example, extrudate that does not meet specifications can be diverted to waste until the real-time quality attribute is within specification as opposed to having to discard the entire batch. Given the necessity for in-line monitoring, near-infrared (NIR) and Raman spectroscopy have been investigated off-line and in-line for API content determination in HME film matrices (4–8). Additionally, the characterization of extrusion processes via in-line ultrasound and UV–vis spectroscopy also have been reported in the literature (7–9).

The implementation of an in-line, transmission mode, Fourier-transform near-infrared (FT-NIR) method and a partial least squares (PLS) model to support the HME manufacturing platform for a drug product in development is discussed. PLS is a commonly used multivariate linear regression technique and was applied to NIR spectral data for real-time API loading (%wt/wt) and surfactant loading predictions. These predictions were used for fault detection, real-time troubleshooting, and real-time release of the intermediate drug product during the hot-melt extrusion processing step.

Experimental Details

A 27-mm twin-screw extruder with a length/diameter ratio of 40:1 was used. The API, surfactant, and polymer were added via independent feeders positioned in zones 1 through 4. Zones 5 through 10 are melt and mixing zones. The extrusion process has two additional temperature-controlled zones where zone 11, the die, reduces the barrel configuration that houses the screws to a single circular orifice. The last zone is the NIR die adapter, which is heated by two cartridge heaters. The temperature set-points in these additional zones are the same as earlier mixing and melting zones.

The NIR system is integrated to the extruder via a custom in-line temperature-controlled die adapter with a fixed optical pathlength. A Thermo Antaris II FT-NIR spectrometer (Thermo Fisher Scientific, Madison, Wisconsin) equipped with two 3-in. transmission probes is used for spectral data acquisition. The single-fiber transmission probes have standard 1/2-20 UNF mounting threads and are screwed into the die adapter so that they are perpendicular to the plane of material. The probes have been designed to be seated flush within the die adapter to minimize optical pathlength fluctuations. More specifically, the probes cannot be over-tightened to alter the fixed optical pathlength provided by the die adapter. Additionally, the probes are designed to be resistant to an exterior operating temperature of 220 °C and to withstand pressures up to 1500 psi. The die adapter for this manufacturing process is typically at 180 °C and yields a pressure of approximately 300 psi.

NIR data acquisition is performed using Thermo RESULT v3 yet is managed via the Siemens nv/sa (Brussels, Belgium) package for process analytical technology, SIPAT v3.1. SIPAT enables instrument control, data analysis, visualization and archiving, and integration of the instrument to an automation system. The spectral range is 4000–10,000 cm^-1 and a resolution of 16 cm^-1 is used. The number of spectral scans averaged is 16. An internal background is acquired every hour throughout data collection.

A typical HME campaign for the drug product studied here involves placebo manufacture immediately followed by active manufacture. Such a campaign can run continuously for several days and yield thousands of kilograms of material. NIR monitoring is used for both placebo and active batch manufacture. Using the aforementioned 16 spectral scans yields a near-IR spectrum, an API loading (%wt/wt) prediction, and a surfactant loading prediction every 8 s. Given the length of these campaigns, a delay is programmed into the method to yield values every minute as to not overload the NIR collector station.

NIR calibration samples were purposely extruded with varying levels of API and surfactant. The target API composition is 20% and the calibration range spans from 0–30.1%. The target surfactant concentration is 10% for placebo and 20% for active and the calibration range spans from 4.8 to 27.0%. Currently, the model includes 20 calibration samples and six independent validation samples. Additionally, extrudate throughput was varied from 20 kg/h to 35 kg/h, within the model, to mimic potential manufacturing rate changes to satisfy demand. Given that calibration and validation samples need to be manufactured rather than prepared in a laboratory, model development, like HME, is a continuous process. With each development campaign, samples are taken for off-line high performance liquid chromatography (HPLC) reference measurements.

The collection of extrudate for off-line HPLC reference measurements involves adjusting the process to yield the varying levels of either API or surfactant and then allowing the process to achieve a steady state. Although the residence time of the material in the extruder is short, a steady state is desired before collecting samples to minimize concentration and throughput fluctuations. The hot-melt extrudate stream is sampled as it exits the die adapter.

Partial least squares models were developed using Umetrics (Umetrics Inc/MKS Instruments, San Jose, California) SIMCA-P+ v11.5 software, but after they were optimized they were transferred to TQ Analyst v8.3.125 software (Thermo Fisher Scientific) to enable real-time data acquisition, analysis, and visualization via SIPAT. A second derivative using a 13-data-point Savitzky-Golay filter with a second-order polynomial was applied to the data. All spectral data were mean-centered. A range of 6302.23–5985.96 cm^-1 and two factors were used for the API loading prediction. These two principal components accounted for 99.6% of the spectral variation. A range of 6302.23–4921.45 cm^-1 and four factors were used for the surfactant loading prediction. These four principal components accounted for 99.5% of the spectral variation.

Results and Discussion

Spectra of polymer, placebo, and active extrudate indicated that NIR in transmission mode sample geometry is feasible for extrusion monitoring of the noted drug product (refer to Figure 2). NIR spectra of the active drug product extrudate, placebo extrudate, and polymer extrudate illustrate amorphous API specificity in the region of 6302.23–5985.96 cm^-1. Given that NIR is sensitive to form changes of the given API; the NIR measurement of extrudate confirms conversion to the amorphous state. Although API specificity is straightforward, surfactant specificity required spectral preprocessing and chemometric analysis.

Figure 2: NIR spectra of the active drug product extrudate (solid), the placebo drug product (dashed), and polymer extrudate (dotted) illustrating absorption specific to amorphous API in the 6302.23–5985.96 cm-1 region.

After ensuring the feasibility of NIR for process monitoring, a development run was planned to initiate the calibration model. This development run began with varying the API and surfactant composition with a target range of ±30% of the active target composition. Extrudate with seven compositions levels were manufactured where the API and surfactant were independently varied. The seven composition levels were manufactured at extruder throughputs of 20 and 25 kg/h. The target active composition was manufactured at higher extruder throughputs of 30 and 35 kg/h. During this development campaign, it was not certain what the target extruder throughput would be, therefore a potential target was bracketed. After the manufacture of calibration–validation samples, routine manufacturing was done as illustrated in Figure 3. Figure 3 was generated after the initial development of the NIR/PLS model to determine if extrudate samples were collected at steady state, if throughput changes affected the prediction values, and to assess the quality of the active batch manufactured after calibration–validation sample manufacture. Detailed inspection of the results, summarized in Figure 3, indicated extrudate samples were collected at steady state. The throughput variations have no observable effect on the prediction results. Additionally, a scatter plot of principal component scores 1 and 2 yielded no observable trends or clustering resulting from throughput variations, indicating the method is robust to the changes explored. Lastly, Figure 3 yields quality assurance for the routine manufacturing portion of the given development campaign.

Figure 3: Loading (%wt/wt) predictions from the NIR/PLS model during a development campaign that resulted in calibration sample manufacturing and routine active drug product manufacturing.

The model was updated further (twice) with samples collected during subsequent manufacturing campaigns. Additional calibration samples provided more surfactant composition levels to bracket the placebo and active batch targets, batch-to-batch variability inclusion, and more validation samples. The model has not significantly improved in regards to prediction error, but the prediction power, or robustness, of the model has increased.

The API loading prediction model yielded a correlation coefficient of 0.9990 and a root mean squared error of calibration (RMSEC) of 0.4%. The surfactant loading prediction model yielded a correlation coefficient of 0.9978 and RMSEC of 0.4%. The calibration model was tested for bias and accuracy with independent validation samples, meaning that they were not included in the calibration set. There was no significant bias between the estimates from the NIR/PLS model and those from the reference off-line HPLC method. The root mean squared error of prediction (RMSEP) values for the API loading and the surfactant loading are 0.3% and 0.6%, respectively.

Process Fault Detection

As with most manufacturing processes, process parameters are visually checked and entered into a batch record. The difficulty in doing this is that there are multiple process parameters to check while on the manufacturing floor. A delicate balance between process monitoring and physically demanding manufacturing support has to be maintained by the manufacturing scientists and operators. Regarding the process monitoring aspect, it is easier to simplify the multivariate nature of such a complicated process into a single data stream that is representative of product quality. The quality critical attribute for this manufacturing step is the material composition and hence the loading has been monitored on the manufacturing floor via the SIPAT operator console. In many circumstances over several manufacturing campaigns, faults have been immediately and easily identified from alarms triggered by the loading predictions exceeding specifications. When this happens, product collection is immediately diverted to waste. Then various process parameters are checked to identify the root cause of the compositional deviation. A representative example of such a fault is given in Figure 4. At approximately 04:59, the API loading (%wt/wt) drifts upward and ultimately exceeds the alarm limits. As soon as the loading trend shifts upward, the molten extrudate stream was diverted to waste that bypasses the chilled roll. Process parameters associated with the material feeders were investigated due to the increase in concentration. In this circumstance, a change in API lots with different bulk densities led to material being fed at a faster rate than the set-point. The material continued to be diverted until the feeder reached steady state around the set-point and the NIR response was closely watched to deem the result of this action. In approximately 11 min, the problem was identified, the suspect extrudate was isolated from quality extrudate, the fault was corrected, and quality product collection was resumed. Without the real-time release testing provided by the NIR/PLS model, at the very least, the drum of extrudate would have been quarantined for further testing. The worst case scenario is that the entire batch would have been lost. One could argue that monitoring the API flow would have ultimately identified the process upset if NIR monitoring was not in place. The issue is much more complex. Even if the upset had been detected by a process parameter, it would not have been known if the upset was of enough significance to put product quality at risk, and if so, for how long. Thus, NIR monitoring was important, in this example and multiple other circumstances, for identifying and facilitating the correction of the process upset while yielding quality assurance.

Figure 4: Example of a fault detected via NIR during hot-melt extrusion and an illustration of how the fault was corrected.

Prediction Model Lifecycle Management

A successful model is dependent upon having a data set that includes anticipated and designed sources of variation and contains chemical or physical insight to the attribute of interest. The model described herein achieves these goals, but will be a work in progress until transferred to routine supply. It is an impractical expectation that all anticipated variations will be captured during drug product development. This is primarily due to the resource restraints given anticipated variation including, but not limited to different lots of raw material, different scales of production, and multiple instruments and processing equipment. Thus, the process analytical technology implementation strategy to support HME includes on-going robustness evaluation and accuracy testing to assess model performance. The model update approach has been, and will continue, adding new calibration samples into the existing calibration sample set and regenerating the model. When routine, commercial production is achieved, model updates will be less frequent and solely dependent upon manufacturing process changes.

Conclusions

A NIR/PLS model was successfully developed to predict API loading and surfactant loading during hot-melt extrusion of a drug product in development. Although verifications and updates are planned as part of the NIR/PLS model lifecycle management, the model is fit for purpose. In other words, the NIR/PLS model evolves as the HME process evolves. The model has been applied to multiple manufacturing campaigns for the purpose of fault detection and real-time release testing of intermediate drug product. The intent of this model is for implementation into routine supply when the drug product is released to market. Given the capacity of the current NIR monitoring platform in place, the next steps are to use the API and surfactant loading prediction values to automate diversion to waste whenever there is an out-of-specification alarm. Presently, diversion to waste is manual. Having NIR monitoring in place during development has facilitated a robust model to process variations and thus will yield process flexibility post approval. This platform has been key at providing process knowledge to improve HME for the given drug product. Therefore, the application provided illustrates how a process analytical method and a manufacturing process are codependent regarding development and optimization.

References

(1) C. Leuner and J. Dressman, Eur. J. Pharm. Biopharm. 50, 47–60 (2000).

(2) M.A. Repka, S. Majumdar, S.K. Battu, R. Srirangam, and S.B. Upadhye, Expert Opin. Drug Deliv. 5(12), 1357–1376 (2008).

(3) M.M. Crowley, F. Zhang, M.A. Repka, S. Thumma, S.B. Upadhye, S.K. Battu, J.W. McGinity, and C. Martin, Drug Dev. Ind. Pharm. 33, 909–926 (2007).

(4) V.S. Tumuluri, S. Prodduturi, M.C. Crowley, S.P. Stodghill, J.W. McGinity, M.A. Repka, and B.A. Avery, Drug Dev. Ind. Pharm. 30(5), 505–511 (2004).

(5) V.S. Tumuluri, M.S. Kemper, I.R. Lewis, S. Prodduturi, S. Majumdar, B. A. Avery, and M. Repka, Int. J. Pharm. 357, 77–84 (2008).

(6) L. Saerens, L. Dierickx, B. Lenain, C. Vervaet, J.P. Remon, and T. De Beer, Eur. J. Pharm. Biopharm. 77, 158–163 (2011).

(7) P.D. Coates, S.E. Barnes, M.G. Sibley, E.C. Brown, H.G.M. Edwards, and I.J. Scowen, Polymer 44, 5937–5949 (2003).

(8) I. Alig, D. Fischer, D. Lellinger, and B. Steinhoff, Macromol. Symp. 230, 51–58 (2005).

(9) Y. Wang, B. Steinhoff, C. Brinkmann, and I. Alig, Polymer 49, 1257–1265 (2008).

Brandye Smith-Goettler is a Senior Analytical Chemist, Colleen M. Gendron is a Project Development Engineer, and Neil MacPhail, Robert F. Meyer, and Joseph X. Phillips are Senior Development Engineers with Merck Sharp and Dohme Corporation in West Point, Pennsylvania. Direct correspondence to: brandyemichelle_smithgoettler@merck.com