Implementing process analytical technology (PAT) became a goal in the pharmaceutical industry 15 years ago, with the publication of a guidance document from the U.S. Food and Drug Administration (1). Recently, the industry’s drive to use continuous manufacturing—instead of batch processing—has given new impetus to PAT applications. Steve Hammond, a PAT consultant who recently retired from Pfizer, has spent four decades focused on process-based analytics, using techniques like near-infrared (NIR) spectroscopy. He recently spoke to us about this work, including the current status of PAT implementation in the pharmaceutical industry and the challenges involved, illustrating his points with examples from a recent implementation at Pfizer.
The FDA released its guidance for industry on process analytical technology (PAT) 15 years ago. How would you describe the current state of implementation of PAT at this point?
I was a founding member of the FDA team that wrote that FDA Guidance. PAT is now established, and many applications have been registered. In recent years, PAT applications have received a boost, from the development of continuous processing for solid dosage forms. New spectrometer types, with better performance and robustness, have helped with the deployment of PAT. Much better software platforms have also increased the probability of success for PAT applications.
There is a drive in the pharmaceutical industry toward continuous manufacturing. How does the shift toward continuous manufacturing change the way one implements PAT? What new challenges does it present?
In the continuous manufacturing scenario, there is a need to continuously monitor the performance of each step in the manufacturing stream, to ensure equipment is performing within expected tolerances. This is particularly important for the feeder mixer units. See Figure 1.
This has led to development of better PAT interfaces that ensure PAT probes perform in an optimum way in a dynamic environment. That means probes are collecting good quality spectra, and that fresh material is presented to the probe window to ensure good discrete sampling of the process step. It is one of the obvious developments with continuous processing that the engineering of sensors into the process has been given particular attention and investment. See Figures 2–5.
Another important point is the recognition of the need for software platforms that can simultaneously take sensor data from multiple sources, and turn data into information that a control system can use. In a continuous manufacturing system, PAT and the data handling software platform have to be engineered into the process and be a part of the process. Validation of a continuous manufacturing system is focused on the control system, which, in turn, is based on timely process measurements. Those measurements can be made with PAT sensors, other sensors like load cells used in mass balance models, or, hybrid models, with a combination of equipment sensors and PAT devices.
Why is NIR, particularly in diffuse reflectance mode using fiber optic probes, a key analytical method for PAT in continuous pharmaceutical manufacturing— compared to other techniques, such as Raman or infrared spectroscopy?
When using NIR in diffuse reflectance mode, it is easy to fashion fiber optic probes that can be engineered into the processing equipment used for continuous processing.
The wavelengths of NIR radiation are easily transported through fiber optics cables, and sensitive detectors are available for NIR wavelengths. The availability of these sensitive detectors is due, in part, to the telecommunication industry, which uses the NIR region for telephone signals. For example, the Viavi spectrometers are a spinout of JDSU, a telecom company. The JDSU spectrometers were originally designed to test telephone signals transported through fiber optic cables.
Raman instruments have slower scanning speeds and a lower signal-to-noise (S/N) ratio than NIR. Raman spectroscopy tends to be more useful for liquid media. Fiber optics for Raman spectroscopy are good, as Raman tends to use the 1–2 μm wavelength light that NIR also does.
Fiber optics for mid-IR exist, but are low capability and very expensive compared to those used with NIR. Mid-IR fibers tend to be only a few meters in length, which reduces their viability for use in a manufacturing plant, although applications in reactor monitoring are very successful.
Modern diode array NIR spectrometers are capable of very fast data acquisition, while maintaining good selectivity and accuracy. For example, the Sentronic NIR probe system can scan at a rate of one acquisition scan in 20 ms, so several acquisitions (scans) can be integrated over 0.5–1.0 s to obtain high quality, low noise spectra. Mid-IR and Raman are not capable of this fast data acquisition at the same S/N.
Mid-IR is very difficult to interface with powders, because the intensity of mid-IR bands is about 100 times greater than in NIR. This means the mid-IR light intensity returning from the surface of a sample to a detector is very low, as most of the midIR radiation is absorbed by the sample material. The only solution to this with solid samples is to dilute them, traditionally in sodium chloride or Nujol. That is totally impractical in a processing environment. Mid-IR works well in a liquid environment where ATR crystals enable undiluted samples to be scanned effectively.
You have worked on deploying PAT at five different steps in the manufacture of pharmaceutical tablets, where four of the steps used NIR (Figure 1). What types of challenges were presented by those different steps? Why was NIR not the technique chosen for the other step?
The one application that did not use NIR used a focused beam reflectance measurement (FBRM), which is a probe that can measure the physical size of particles. This is not a chemical composition measurement, which is what NIR is generally used for. FBRM can directly estimate the particle size distribution of powders. The direct FBRM measurements can be used to calibrate an indirect NIR method for particle size. See Figure 4.
The four NIR applications had challenges related to sampling. The mixer we used had high throughput and little retained mass, so material passed through the mixer in seconds. That meant the measurement systems had to be fast to monitor a fast-throughput process. A measurement that took less than 1 s was imperative to provide timely data on the function of the mixer and the feeders placing ingredients into the mixer.
Wet powders exiting a granulator are sticky. So development of ways of making sure the sticky material did not blind a probe window, and good sampling could be achieved, were the secrets of success of all the NIR applications. In addition, the process engineers warned the PAT people of dire consequences if the probe interfaces blocked the flow of material in the processing equipment.
The devices in Figures 2–5 illustrate how these challenges were met. Devices were installed that captured the powder to the correct mass and density in front of a probe to ensure good quality spectra, and wiped the window, but also ensured material moved on down chutes.
Powder exiting a dryer or a mill tends to be in the form of a dispersed cloud. The device shown in Figure 3 answered this need. The spoon device captures the powder into good mass and density to be scanned in an optimum way for the NIR probe.
Developing systems to optimize the spectroscopy of the tablet feed frame, where product release decisions are made, was critical. Figuring out how to engineer a probe into the tablet feed system, and not have the moving parts interfere with the spectroscopy, required some engineering modifications to the tablet press feeder system. Historically, it has been very difficult to have engineers agree to modifications to equipment to accommodate sensors. Continuous manufacturing changed all that, because the equipment could not be operated without continuous monitoring of performance. The results of our designs are shown in Figures 2–5.
How do you ensure that your NIR probes are collecting spectra on representative samples?
The FDA has indicated that, for blend uniformity analysis, the effective sample size should be comparable to a unit dose. How did you ensure and document that?
The solution was to engineer devices that could ensure unit dose sampling. The combination of mechanical movement of powder married with a fast scanning spectrometer can achieve unit dose sampling.
Experiments showed a typical fiber optic probe scanning a typical pharmaceutical powder interrogates about 3 mg of powder each time it scans. In other words, only 3 mg of powder contributes to a single scan. Typical tablet weights are 100–300 mg, so to obtain a tablet-weight spectrum, the fiber optic probe is set to scan 50 times, and the 50 scans are integrated to produce the “spectrum” of a virtual 150 mg tablet that is used in the calculation of a concentration result. See Figure 7.
How do you minimize interferences like specular reflectance and scattering when making measurements?
The spider wheel fingers were modified to avoid the spectrometer scanning bare metal. Either 2-mm deep notches were cut in the fingers, or the back of the fingers was coated with white PTFE, similar to the reference material. See Figures 2 and 6.
What was your process or protocol for developing your chemometric models for this work? What do you think is the FDA’s approach toward chemometric methods, and what could be done to improve procedures using chemometrics?
Pfizer does have standard protocols for development of models and their validation. These are based on the many textbooks and publications on this subject. One important point about calibration development using a continuous manufacturing rig is that the loss-in-weight feeders can be programmed to transition across the target concentration, from minus 10% to plus 10% of the nominal, thus providing calibration samples that can be sampled and analyzed using conventional methods.
When applying data preprocessing, SIMCA, MLR, PCA, and PLS models, is there a recommended document or set of documents that gives guidance on precisely how those models are developed, validated, and maintained? Does the FDA give guidance on this process?
Again, Pfizer does have standard protocols for development of models and their validation. These are based on the many text books and publications on this subject. The FDA will not provide guidance on these processes, mainly because there are so many possible ways to do them that are all “right.” As computer systems develop, more possibilities for producing models become available. The development of artificial intelligence systems will almost certainly produce the calibration algorithms of the future—artificial neural networks (ANNs), for example. FDA will expect the company to submit the protocols used, and a detailed explanation of what was done, as well as a scientific explanation that validates that the system works. For new algorithms, or data treatments, the FDA has set up a team, called the “Emerging Technologies Team” to examine these new systems and provide feedback in a neutral environment. That team can be approached to examine new technologies, and will provide feedback, but that feedback does not in any way give official FDA approval to any technology.
- US Food and Drug Administration, Guidance for Industry: PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance (FDA, Rockville, MD, 2004). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/pat-framework-innovative-pharmaceutical-developmentmanufacturing-and-quality-assurance
Steve Hammond is an independent consultant, working in the area of process-based analytics. His work includes the development of PAT interfaces using systems from ExpoPharma (www.expopharma.ie). In May 2018, he retired from Pfizer, where he spent 39 years as the senior director or team leader of the Process Analytical Sciences Group, in Peapack, New Jersey. During his career with Pfizer, he authored or co-authored many scientific publications on the pharmaceutical applications of process analytical technologies, and contributed many oral presentations on this subject as an invited speaker. In his most recent work there, Hammond took responsibility for developing a suite of process analytical technology sensors and their interfaces for the monitoring and control of continuous manufacturing streams, for both solid dose forms and continuous flow reactors. Hammond is a graduate of the Royal Institute of Chemistry in London, and has a master’s degree in analytical chemistry from the University of Kent, UK. He was appointed to the Royal Society on Chemistry in 1981, and held the status of Chartered Chemist in the UK.