The year 2023 marks the 40th anniversary of the commercialization of inductively coupled plasma mass spectrometry (ICP-MS). Today, there are close to 2000 ICP-MS systems installed worldwide every year, representing around $400M in annual sales. These instruments are used for a wide variety of applications, from routine, high-throughput multielement analysis to more complex tasks, such as trace element speciation studies with high-performance liquid chromatography and monitoring of nanoparticles. My column this month will look at the evolving ICP-MS marketplace, highlighting current application trends (compared to other analytical spectroscopy techniques), and where the technique might be heading 5–10 years from now.

At the 1983 Pittsburgh Conference, SCIEX introduced the ELAN 250 quadrupole-based inductively coupled plasma (ICP) mass spectrometer (Figure 1). It would be another 12 months before its joint venture with PerkinElmer, Inc. was announced, but the event was the start of the meteoric rise of ICP-mass spectrometry (ICP-MS) as the dominant multielement technique used for ultra-trace elemental analysis, and the beginning of a competitive race which would see a number of other vendors and other mass spectrometer technologies come and go over the next 40 years.

Today, there are close to 2000 ICP-MS systems installed worldwide every year, representing around $400M in annual sales, performing a wide variety of applications, from routine, high-throughput multielement analysis to more complex tasks, such as trace element speciation studies with high-performance liquid chromatography and monitoring of nanoparticles (1). As more and more laboratories invest in the technique, the list of applications is getting significantly larger and more diverse.

In previous publications, I have given detailed information about the various ICP-MS application sectors being carried out by the user community (2). However, 40 years after the commercialization of the technique, it would be almost impossible to capture all of them. The most common ones include environmental monitoring, geochemical, metallurgical, petrochemical, petroleum, food, clinical, toxicology, semiconductor, industrial, energy, agricultural, nuclear, pharmaceutical, and cannabis, but every year it seems that a new market has woken up and realized the full potential of its capabilities. So, to celebrate the 40th anniversary of the technique, I have already begun working on the 4th edition of my ICP-MS textbook, “Practical Guide to ICP-MS: A Tutorial for Beginners,” which will be available in the early summer of 2023.

I cannot believe it has been almost 20 years since I published the first edition of the book in 2004, and 8 years since the 3rd edition was launched in 2014. But I have not been idle, as I have written two additional books since, one on elemental impurities in pharmaceuticals published in 2018, and the other on heavy metals in cannabis and hemp in 2021. What was originally intended as a series of ICP-MS tutorials on the basic principles for Spectroscopy Magazine in 2002 (3) quickly grew into a textbook focusing on the practical side of the technique. With over 6000 copies sold, including a Chinese (Mandarin) edition, I’m very honored that the book has gained the reputation as being the reference book of choice for novices and beginners to the technique all over the world. Sales have exceeded my wildest expectations.

However, this edition will be a little different, because, based on reader feedback, I have decided to also include fundamentals and applications of other complementary atomic spectroscopy (AS) techniques, including inductively coupled plasma optical emission spectroscopy (ICP-OES), flame and graphite furnace atomic absorption spectroscopy (FAAS/GFAAS), atomic fluorescence (AF), microwave induced plasma atomic emission spectrometry (MIP-AES), X-ray fluorescence (XRF), laser induced breakdown spectrometry (LIBS) and laser ablation, laser ionization time-of-flight mass spectrometry (LALI-TOF-MS). Moreover, to capture the staggering flexibility and capability of ICP-MS, the new book will also include a chapter dedicated to the continually evolving application landscape of the technique compared to other AS techniques. So, to give readers of Spectroscopy a flavor of the new book, particularly the ICP-MS marketplace, my AP column this month will highlight current application trends and where the technique might be heading 5–10 years from now.

As a result of the widespread use and acceptability of ICP-MS, the cost of commercial instrumentation has dramatically fallen over the past 40 years. When the technique was first introduced, $250,000 was a fairly typical amount to spend, whereas today one can purchase a quadrupole or TOF system for under $150,000. Although it can cost a great deal more to invest in magnetic sector technology or a “triple quadrupole” collision/reaction cell instrument, most laboratories that are looking to invest in the technique should be able to justify the purchase of an instrument without price being a major concern. One of the benefits of this kind of price erosion is that, slowly but surely, the AA and ICP-OES user community are being attracted to ICP-MS, and as a result, the technique is being used in more and more diverse application areas. Figure 2 shows the approximate percentage breakdown of the major market segments being addressed by ICP-MS on a worldwide basis over the past 40 years (2).

Three points should be emphasized here. First, these data can be significantly different on a geographical or regional basis, because of factors such as a country’s commitment (or lack of it) to environmental concerns or the size of a region’s electronics or nuclear industry, for example. Furthermore, many laboratories carry out more than one type of application, and as a result can be represented in more than one market segment. Finally, the research market segment has been listed as a separate category to show the instruments that are being used in an academic environment or for non-routine applications. However, many universities, federal organizations, or corporate R&D groups might be using their instrumentation for research purposes in a particular application segment. For these reasons, these data should only be considered an approximation for comparison purposes.

It’s fair to say that ultra-trace detection capability, combined with increased sample throughput, have been the main incentives for analytical chemists to invest in ICP-MS. The technique is now considered a mature tool that is being used for the routine, high-throughput analysis of complex samples with relatively seamless sample preparation procedures. In the early days, it was used by academic institutions and large corporate R&D laboratories. Nowadays, ICP-MS instrumentation is also being used by contract laboratories in the routine analysis of environmental, pharmaceutical, and biomonitoring applications, among others. This has been made possible by greater automation on the sample preparation side and seamless method development tools, which have resulted in far less operator expertise required to run them. It is also worth pointing out that stricter regulations are requiring lower and lower detection limits (DLs), particularly in the semiconductor, toxicology, and food market segments. In addition, for the past five years, the rapid growth in the cannabis industry has meant cultivators and processors have to ensure the safety of consumer products by meeting strict state-based limits for contaminants like heavy metals. So, the drive toward lower detection limits is just one factor, whereas the other is the analytical needs and demands of the customer. For example, the electronics industry is faced with the enormous challenge of “chasing zero,” so manufacturers must measure the lowest level of elemental impurities in process chemicals and electronic devices to ensure the highest performance of their products.

As ICP-MS has become more applicable in a wide variety of application areas, it has become more readily accessible to startup laboratories. Part of this is the development of turnkey methods for food, environmental, biomonitoring, geochemical, pharmaceutical, and cannabis applications for inexperienced and novice users, which will be critical for the growth of this technique in these market segments.

Important considerations include ease of use, seamless software, low maintenance, and the ability to handle different matrices, which give the analyst confidence in the results being generated and ensure they get the correct answer the first time and every time. The user community expects these qualities, particularly in high-throughput laboratories. They want to run the instrument for extended periods with minimum maintenance and short downtime while simultaneously running multiple samples, many containing high concentrations of mineral acids. It is important to emphasize that this kind of sampling environment is not really instrument friendly. When samples are being run that contain >10% mineral acid or >10% dissolved solids, it will take its toll on the instrument’s sample introduction components. Therefore, the instrument must be rugged enough to handle this kind of sample matrix, without contributing significantly to the routine maintenance of the system, while at the same time delivering high quality data, day in and day out.

Therefore, ruggedness is critically important when designing an ICP-MS instrument. However, this is just one side of the analytical challenge. As mentioned previously, the semiconductor industry is the one application area where this kind of ruggedness is expected, but the lowest possible detection limits are also required. Just 15 years ago, 10 ppt was the guideline for elemental impurities in semiconductor process chemicals; today, the industry is requiring 1–2 ppt. This puts unique demands on the design of any instrument used for this kind of analysis.

Consequently, with the push for lower and lower detection capabilities, it is becoming a major challenge not only to develop instrumentation that has high analyte sensitivity and extremely low background noise, but also to have an ultra-clean laboratory and sample preparation environment that sets the stage for the detection of elements at such low levels. Both factors are absolutely critical. The trend towards ease of use has meant that operators have less time to be involved with the analytical procedure, and therefore are often not versed in the nuances of working in the ultra-trace environment and won’t necessarily recognize issues which require user intervention, should they arise. For that reason, instruments have to offer maintenance reminders that mitigate potential problems before they actually happen.

The increasing demand for state-of-the-art ICP-MS equipment that offer practical solutions has been mainly driven by regulatory requirements. Whatever the industry, there’s a huge need for the detection of lower and lower impurities of inorganic pollutants like heavy metals. For example, many regulations require the testing of soil for toxic metals before any crops can be grown in the soil. In addition, industrial effluent must be regulated to know what is being discharged into waste streams. This is particularly the case with nanomaterials, which are finding their way into the environment. Experts have not extensively studied the long-term impact of inorganic nanoparticles on plants and animals, so going forward, industry clearly must have a much better understanding of this issue.

Another example is the mining of minerals such as titanium dioxide, which is used as a whitening agent and silicon dioxide, used for its drying and anti-caking properties in the manufacture of pharmaceutical excipients, fillers and tablet binders. There are new regulations that require manufacturers to measure 24 elemental impurities at the microgram (µg) level in their products. If they do not comply with these permitted daily exposure (PDE) limits, they could be liable for significant fines or even the closure of their manufacturing plant. So, there are clearly many roles in which metals play a critical role in the world we live in, and, as a result, the quantification and regulation of these metals and metalloids is required. Hence the importance of using very sensitive techniques like ICP-MS as well as performance and productivity enhancing accessories to help us carry out those extremely low measurements.

There are two major trends to address those needs. One is at the high end, where the semiconductor and the biomonitoring markets are chasing after the lowest detection limits, being able to analyze as many inorganic analytes as possible. This is where you see significant research being carried out to develop offerings with the lowest detection limits, maximum interference reduction, and optimum flexibility—in other words, the highest-performance instruments. A separate trend involves the routine analysis of food, environmental, pharmaceutical, and cannabis related samples, for which ICP-MS sensitivity is already sufficient to meet their requirements. For these laboratories, the trend is toward improved robustness, less routine maintenance, more seamless integration with automated sampling solutions and ease of use, which is clearly a very different application segment from ultra-low background levels and detection capability.

Factors Impacting ICP-MS Instrumental Development?

Since many academic groups are not developing hardware intended for commercialization these days, the trend in these institutions seems to be focused more on application development. This is in contrast to the 1970s and 1980s, when, for example, the fundamentals of ICP-MS research were carried out predominantly in academia. However, more recently, we have seen a shift toward ICP-MS development being carried out mostly by manufacturers, and much of the cutting-edge application work being carried out or supported by academia. A good example of this is the expansion of ICP-MS into the field of nanomaterials, and specifically detecting nanoparticles in environmental and biological samples over the past four to five years. Here there have been several excellent collaborations occurring between universities and manufacturers. In general, there is an increased interest from universities to expand into niche application areas that are not necessarily being addressed by vendors or other users such as single particle and single cell analysis to monitor the metal content within single cells to better understand the mechanism of how metals such as platinum play a role in cancer treatment research.

Another example is in the pharmaceutical industry, where 

 Q3D Guidelines have opened up new markets for ICP-MS through recent regulations for elemental impurities testing in drug products and dietary supplements. An extension of this application is the testing of cannabis and cannabinoid consumer products for heavy-metal contaminants known as the Big Four—lead (Pb), cadmium (Cd), arsenic (As), and mercury (Hg)—using regulations initiated by the pharmaceutical industry. However, unlike the manufacture of drugs, there is no real tracking of the sources of heavy metals in cannabis, so it is not uncommon to see product recalls because of elevated levels of elemental contaminants. As a result of this, it has become critical to monitor the cannabis plants and products for a wider panel of elemental contaminants than the Big Four and to include others that are worthy of consideration using ICP-MS. Again, this is clear indication that some application sectors are being stimulated by the demands of the industry and supported and driven by the manufacturers, as opposed to coming from university-directed research.

Instrument vendors are firmly committed to the ICP-MS marketplace, with sales growing at approximately 6% per year (compared to about 3% for ICP-OES). If we look at the main market segments, such as semiconductor, biomonitoring, environmental, geological, industrial, food, pharmaceutical, and cannabis, it is clear that demand for ICP-MS is still growing rapidly. This growth is achieved mainly by collaborating with key market-driven customers to ensure that they are addressing their future needs, and using this feedback to help them design innovative products, whether related to application development, hardware improvements, or software enhancements. The aim of this article has therefore been to give the reader a snapshot of the ICP-MS application landscape, and a better understanding of why it is still the fastest growing trace element technique available today. If there is one common theme that runs through many of these applications, it is the unparalleled detection limits the technique has to offer, for both the total and speciated form of the element. When this is combined with its rapid multielement characteristics, isotopic measurement capability, freedom from interferences, and ease of use, it is clear that ICP-MS will continue to dominate the application landscape above all other AS techniques.

(1) Global ICP-MS System Market Research Report 2022 (Industry Research, Pune, India). 

https://www.industryresearch.biz/global-icp-ms-system-market-21299553



(2) Practical Guide to ICP-MS: A Tutorial for Beginners; (CRC Press, Boca Raton, FL, 3rd ed., 2014). 

https://www.routledge.com/Practical-Guide-to-ICPMS-A-Tutorial-for-BeginnersThird-Edition/Thomas/p/book/9781466555433#



(3) Beginners Guide to ICP-MS, Spectroscopy Magazine Tutorial Series, 2002, 

http://www.scientificsolutions1.com/Beginners%20guide%20to%20ICP-MS.pdf

Articles

Columns: Atomic Perspectives

The use of ICP-MS is constantly expanding into an ever-wider variety of applications. We assess the current landscape and where the technique is going in the future.

40 Years Old and Still Solving Problems: Evolution of the ICP-MS Application Landscape

Until now, our discussion of polymers has been limited to functional groups containing carbon, hydrogen, and oxygen. In this column, we begin covering polymers that have functional groups containing nitrogen as well. We review the structures and bonding characteristics of the nitrogen atom, and then begin our examination of the spectra of organic nitrogen polymers with a very important class of polymers: polyamides.

In the last column, we concluded our study of the infrared (IR) spectroscopy of carbonyl-containing polymers by discussing acrylates (1). Traditionally, the chemical element nitrogen is considered part of organic chemistry (2). In this column and several following it, we turn our attention to polymers that contain organic nitrogen compounds.

The nitrogen atom has an atomic number of seven and contains five outer shell electrons. The drive of chemical elements, of course, is to get their outer shells filled. In most organic compounds, nitrogen tries to fill its outer shell by forming three chemical bonds to other atoms, be they hydrogens, carbons, or whatever. Carbon and nitrogen can form single, double, or triple bonds, as seen in Figure 1.

Note for the C-N bond seen on the left in Figure 1 that the bond angles are approximately 120°, and there are three atoms bonded to the nitrogen. For the C=N bond seen in the middle of Figure 1, the bond angles are approximately 90° (2), and there are two atoms attached to the nitrogen. Lastly, for the C≡N bond seen on the right in Figure 1, the bond angles are approximately 180° and there is one atom (carbon) attached to the nitrogen. Nitrogen has an electronegativity of three, which means it tends to hog the electrons in any bonds it forms. Carbon and hydrogen have electronegativities of 2.5 and 2.1, respectively (2), so C-N and N-H bonds are covalent.

Since we need to study nitrogen containing polymers, is there an IR signature for the presence of nitrogen in a sample? This question was answered previously when we first studied nitrogen containing functional groups (3). Briefly, C-N stretching vibration peaks are common but weak because they have a small value of dμ/dx, which is the change in dipole moment with respect to bond length during a vibration. Recall (4) that the intensity of the IR peaks depends upon (dμ/dx)

 amongst other things. The small value of (dμ/dx)

 means C-N stretching peaks are weak and hard to see. Also, they show up in the middle of the fingerprint region from 1400 to 1000 cm

 (all peak positions and ranges will be in cm

 units going forward even if not so noted) where they are lost in the level of detail there. Thus, C-N stretches are not good group wavenumbers and are not useful for determining if a sample contains nitrogen.

C=N stretching peaks are even more problematic because this functional group is not stable and rarely seen. Lastly, the C≡N or nitrile group has a strong and unique peak around 2200 as we have seen (5). This peak is an excellent group wavenumber, but not all nitrogen containing polymers have a C≡N bond, so this peak cannot be used as a diagnostic marker for the presence of nitrogen in a polymer. This means we have struck out with trying to use carbon-nitrogen stretching vibrations to solve our problem. As it turns out, N-H stretching and bending peaks are our best bet for figuring out whether there is nitrogen in a polymer (3).


The first type of nitrogen containing compounds we will discuss are polyamides. These polymers contain an amide group in their backbone. The structure and parts of the amide functional group are seen in Figure 2. There are three types of amides: primary; secondary; and tertiary. These amides are named depending upon the number of C-N bonds present in the functional group as you can see in Figure 3. Note that as the number of C-N bonds goes up, the number of N-H bonds goes down. Thus, primary amides have one C-N bond and two N-H bonds; secondary amides have two C-N bonds and one N-H bond; and tertiary amides have three C-N bonds and no N-H bonds. Polyamides can be divided into roughly three groups. Nylons, which you have probably heard of, is another name for polyamides (6). Aromatic polyamides, where there is an aromatic ring in the polymer backbone, are also known as aramids (7). Lastly, an important group of polyamides are proteins, which I am sure you have heard of. Now, biochemists hate it when I say this, but from the point of view of an organic chemist, proteins are polymers made from amino acid monomers where each repeat unit is linked to the next via an amide bond. We will discuss the spectroscopy of proteins in a later column.

As we have studied previously (8), all amides have their C=O stretching peaks between 1680 and 1630 because all of them are conjugated. Because most polyamides contain secondary amide linkages, that will be our sole focus in the study of polyamides. Table I lists the group wavenumbers for secondary amides.

The number of N-H stretching peaks for amides equals the number of N-H bonds. Thus, for secondary amides there is only one N-H stretching peak which falls from 3370 to 3170 as stated in Table I. This peak is at 3301 and is labeled A in the spectrum of nylon 6,6 seen in Figure 4.

Note that the N-H stretching peak in Figure 4 falls in the same range as O-H stretches (9). However, N-H stretches are weaker than O-H stretches because dμ/dx is greater for O-H than N-H bonds. Also, N-H stretches are narrower than O-H stretches because hydrogen bonding is weaker for N-H bonds than O-H bonds (10). The C=O stretch for nylon 6,6 falls at 1641 as expected from Table I, and is labeled C in Figure 4. Note that it has an intense peak next to it labeled D at 1542, which is assigned as the N-H in plane bend (10). In the many years I have been writing these columns, we have said very little about in-plane bending vibrations because they are often times weak and show up in the crowded fingerprint region where they are hard to see. However, the N-H stretch of secondary amides are unusually intense, making them excellent group wavenumbers. Also, the N-H stretch of secondary amides is one of the few sharp, intense peaks we see in the 1600 to 1500 range in IR spectroscopy (10).

Note that in Figure 4 the C=O stretch and N-H in-plane bend fall at about 1640 and 1540 respectively, and that they are the two biggest peaks in the spectrum. In my experience, if I see the spectrum of a polymeric sample with a pair of intense peaks near 1640 and 1540, my first thought is nylon.

The C-N stretch for nylon 6,6 is at 1274 and is labeled E in Figure 4. As I mentioned above, C-N stretches are weak and peak E in the figure is a perfect example of this. Notice there are a lot of other peaks around it. If this fingerprint region were any busier we might not see the C-N stretching peak at all. Lastly, the N-H wag for nylon 6,6 falls at 691 and is labeled F in Figure 4. This peak is small but is noticeably broadened because of hydrogen bonding (10).

In the nomenclature for nylons, the numbers refer to the number of carbon atoms in the monomer(s) used to make the polymer. For example, for nylon 6,6, each repeat unit contains six carbon atoms. There is a variant of nylon 6,6 called nylon 6. The structures of these two polymers are seen in Figure 5. 

Note that for nylon 6,6 the order of functional groups is C=O, C=O, N-H, and N-H, whereas for nylon 6, it is C=O, N-H, C=O, and N-H. This is a small but real chemical difference. Now, both nylon 6,6 and nylon 6 are important commercially, and both these polymers are made into fibers for clothing and carpets. It would be useful if IR spectroscopy could tell the difference between the two. Figure 5 shows the spectra of these two polymers in part of the fingerprint region from 1350 to 1050. Note that the C-N stretch of nylon 6,6 is at 1274, whereas for nylon 6, it is at 1262. Also, nylon 6 has a peak at 1171 that nylon 6,6 does not have, and nylon 6,6 has a peak at 1145 that nylon 6 does not have. The point is that even though these two polymers have very similar chemical structures, they have spectra that can be distinguished from each other. This has been important in my own work when I once was tasked with developing a method for distinguishing nylon 6,6 from nylon 6 so these materials could be sorted and recycled separately. Infrared spectroscopy did the trick (11). This then is an example of the utility of IR spectroscopy to distinguish similar materials from each other.

We reviewed the topic of nitrogen containing functional groups, and then looked at the spectrum of a polyamide, nylon 6,6, in detail. It shows all the classic peaks of a secondary amide, and the 1640/1540 intense pair of peaks from the C=O stretch and N-H in-plane bend, which are a strong sign that a polymeric material is a nylon. We wrapped up by showing how infrared spectroscopy can easily distinguish between two chemically similar and economically important polymers, nylon 6,6 and nylon 6.

(1) Smith, B. C. Infrared Spectroscopy of Polymers X: Polyacrylates, 

https://doi.org/10.56530/spectroscopy.mi9381w4

(3) Smith, B. C. Organic Nitrogen Compounds, Part I: Introduction. 

(4) Smith, B. C. Why Spectral Interpretation Needs to Be Taught. 

(5) Smith, B. C. Organic Nitrogen Compounds IV: Nitriles. 

(6) Wikipedia, Nylon (accessed February 2023). 

(7) Wikipedia, Aramid (accessed February 2023). 

(8) Smith, B. C. Organic Nitrogen Compounds VI: Introduction to Amides. 

(9) Smith, B. C. The C-O Bond, Part I: Introduction and the Infrared Spectroscopy of Alcohols. 

Infrared Spectral Interpretation: A Systematic Approach

Columns

We study the spectra of organic nitrogen polymers with a particular focus on polyamides.

Infrared Spectroscopy of Polymers, XI: Introduction to Organic Nitrogen Polymers

Offering the high spectral resolution of conventional Raman spectroscopy combined with reduced acquisition time, multiplex coherent anti-Stokes Raman scattering (MCARS) microspectroscopy using sub-nanosecond laser pulses has been accepted as a mature and straightforward technology for label-free bioimaging. In a recently published paper, Philippe Leproux and associates have introduced the combination of the MCARS imaging technique with unsupervised data analysis based on multivariate curve resolution (MCR). Leproux spoke to 

 about combining the techniques, as well as discussing the potential of the results.

Your paper (1) states that coherent Raman imaging has been extensively applied to live-cell imaging in the last two decades, allowing to probe the intracellular lipid, protein, nucleic acid, and water content with a high-acquisition rate and sensitivity. What advantages does this particular technique offer compared to other label-free spectroscopic procedures?

It combines a high spatial resolution (~300 nm laterally) and a high acquisition speed, making it compatible with the observation of biological specimens like living cells or fresh tissues. This is a major asset for a chemically-selective (spectroscopic) imaging technique.

Does label-free spectroscopy produce better results for bioimaging than labeled methods, in terms of spatial resolution, and definition of composition of cells? What are the biggest advantages of label-free methods?

Confocal and multiphoton microscopy approaches are very powerful for investigating biological cells with subcellular resolution. They are supported by a mature and reliable technology, and the myriad of available fluorescent molecules allows one to collect loads of structural and functional cellular information. However, the use of labeled samples implies that those samples were somehow altered (chemically or genetically) and that the observation be restricted to “what” was labeled (for example, organelle or protein). These drawbacks disappear with label-free methods, not to mention that sample preparation is drastically simplified.

You and your co-authors introduce a new methodology for hyperspectral cell imaging and segmentation, based on a simple, unsupervised workflow without any spectrum-to-spectrum phase retrieval computation, utilizing multiplex coherent anti-Stokes Raman scattering (MCARS) microspectroscopy. Would you give our readers a basic explanation of this statement and share with us what inspired the development of this methodology?

MCARS is a microscopy technique based on spectral measurement, which is mature today (2), but generally requires a complex subsequent data processing step under the supervision of an expert. Yet it provides datasets that, in principle, should be computed by means of chemometric tools. In light of this, we focused our efforts on developing an algorithm to perform the unsupervised (“automated”) analysis of MCARS data. By combining MCARS technology with such automated data processing, we could introduce a simple method for label-free cell imaging and segmentation, that we intend to be accessible to a broad range of users, especially from the biomedical field.

What is the meaning of hyperspectral cell imaging versus basic cell imaging methods?

“Hyperspectral” means that the collected datasets contain a high number of spectral channels (one thousand, typically). Comparatively, “basic” fluorescence imaging uses one spectral channel per label or fluorophore, but spectral measurement is also possible (in this case, the word “basic” should be removed).

While, as you state, chemometric methods such as multivariate curve resolution (MCR) are widely used in spectroscopy-based data analysis, to date, MCR has never been applied to the analysis of MCARS hyperspectral datasets. What is the advantage of applying MCR, and why do you think MCR has not been applied and published before your work?

The main benefits are the unsupervised analysis, and the fact that output results are directly readable and interpretable by biologists, chemists, or physicists. Before our work, few studies reported on the use of other chemometric methods for processing CARS data in the context of cell imaging. We believe that we could successfully apply MCR owing to the quality of our datasets, including the high spectral resolution.

What was the most surprising result in your work?

Rigorously, the assumptions behind MCR should make the method non-compatible with MCARS data. Nevertheless, we could demonstrate the validity of our work from the numerical point of view, and we are currently working on the theoretical aspects. As a precaution, we consider that our approach is qualitatively valid at the moment and that further investigations would confirm quantitative imaging. Anyhow, at the very beginning of this work, we were fairly surprised by the quality of the results obtained, in terms of both spectral unmixing and chemically-selective imaging!

Were there any memorable limitations or challenges that you encountered in your work?

The study of 3D tissue samples is highly challenging because of the quantity of data to be processed. In that case, the required (random-access) memory–typically a few terabytes–can be a limitation, and subsampling strategies need to be implemented.

Another challenge was to establish and keep an interdisciplinary approach in our work within a broad team, including cell biologists, computer scientists, and physicists. This type of approach, although in the era of time, is nevertheless sometimes tricky to handle (technical language, practices, objectives, etc.).

Can you please summarize the feedback that you have received from other researchers regarding this work?

Actually, the feedback primarily came from biologists who would like to apply our method to their research field. This was expected following the scope of the journal in which we published our work (1). However, researchers from other fields, such as materials science and earth science, have also expressed their interest in our work.

What discoveries do you hope to realize using this method of cell imaging, and what do you think its impact will be in cell research and diagnostics?

We are now starting to use this label-free method to study the biological mechanisms at the interface between living cells and calcium phosphate ceramic scaffolds, in the context of bone tissue engineering. We conduct this work in the frame of Σ-LIM laboratory of excellence (3). Furthermore, we are collaborating with our Japanese partners (the research group of Prof. Hideaki Kano at Kyushu University) on the subject of label-free brain imaging, with the expectation of building a whole brain atlas with single-cell resolution (4).

What are the next steps in this research?

First, we will maintain our efforts to develop biomedical applications of MCARS technology, as illustrated above. Moreover, in terms of data processing methodology, the next step is to perform multivariate curve resolution by using deep learning approaches, given the large amount of data available and with the aim of accurately modeling the physical phenomenon of vibrational signal generation. In that respect, we are already evaluating autoencoders as candidate artificial neural networks (ANNs) for such a purpose.

(1) Boildieu, D.; Guerenne-Del Ben, T.; Duponchel, L.; Sol, V.; Petit, J-M.; Champion, É.; Kano, H.; Helbert, D.; Magnaudeix, A.; Leproux, P.; Carré, P. Coherent Anti-Stokes Raman Scattering Cell Imaging and Segmentation with Unsupervised Data Analysis. 

(2) Kano, H.; Leproux, P. Multiplex CARS Microscopy in the « Long-Pulse » Regime : Where Are They Now? Proc. SPIE 11973, 

Advanced Chemical Microscopy for Life Sciences and Transactional Medicine 2022

(3) https://www.unilim.fr/labex_sigmalim/?lang=en

(4) Murakami, Y.; Miyazaki, S.; Boildieu, D.; Rajaofara, Z.; Helbert, D.; Magnaudeix, A. ; Carré, P.; Leproux, P.; Honjoh, S.; Kano, H. Toward Whole Brain Label-Free Molecular Imaging with Single-Cell Resolution Sing Ultra-Broadband Multiplex CARS Microspectroscopy. SPIE 11973, 

Combining multiplex coherent anti-Stokes Raman scattering (MCARS) microscopy with automated data processing enables a simple and more accessible method for label-free bioimaging.

Coherent Anti-Stokes Raman Scattering Cell Imaging and Segmentation with Unsupervised Data Analysis

The past year was a unique and challenging one for spectroscopists. With talk about a global recession—and several countries already reporting economic recessions—greater uncertainty in the job market was on the mind of many. For this 2023 edition of our annual salary and employment survey, we gathered information about average salaries and compared those numbers to previous years. We also addressed broad trends in the global economy, asking respondents to offer their thoughts on whether they think a recession is already impacting their country, and how they are responding to current economic conditions. The results reveal that anxiety about the global economy is on the minds of many spectroscopists. We present these results in the text and graphics below.

As in previous years, we analyzed all reported salaries that were between US $15,000 and $250,000. Based on those data, the average salary for spectroscopists came in at $96,683, representing an approximate 3% decrease from last year ($99,740). However, this salary decrease reflects the respondent profile, as more younger workers responded to the 2023 survey (where 34.7% of respondents had fewer than 10 years of experience) than in 2022 (with 15.5% of respondents in that bracket).

Work Environment and Seeking Alternative Employment 

Most spectroscopists (63%) reported that their current work environments are better than last year. Despite this improvement, a majority (59%) indicated an interest in seeking better employment opportunities beyond their current situation.

When citing motives for pursuing alternative employment, the top two reasons were seeking a higher salary (19%) and dissatisfaction with their employer (9%). On the other end, for respondents who are not seeking to leave their current jobs, the most-cited reasons were their salary (18%) and a convenient work location (16%).

We also asked respondents to comment on the job market for spectroscopists. Respondents were mostly positive, with approximately 73% categorizing the job market as either “excellent” or “good.”

With a global economic recession either already happening or looming, we also asked respondents to comment on their job security. A sizable minority (43%) consider their jobs to be just as stable now as they were last year. About 31% said their jobs are more secure now than last year, but nearly 26% said their jobs are less secure. Just under 26% expressed anxiety about an economic recession, and some 37% said that losing their job would be their biggest concern during a recession. Approximately 69% said that their country is already in an economic recession.

 salary survey was available online from November 22, 2022, to January 12, 2023, and received a total of 153 responses from spectroscopists at various stages in their careers. Of the 153 respondents, 25.2% have 21–35 years of spectroscopic experience; 19.7% have 5–9 years of experience; 15% have fewer than 5 years of experience in the field; 13.6% have between 36–40 years; 12.9% have 10–15 years of experience; 10.9% have 16–20 years in the field; and 4.1% have 41 years or more of experience in the field.

The gender breakdown was 28.6% female and 65.3% male; the other 6.1% preferred not to answer.

Half of our respondents reported that they work in industry, 32% work in academia or at an academic institution, 10% work in government or a nationally funded laboratory, and 1% said they work for the military. Respondents came from 16 different countries: Belgium, China, Croatia, Germany, Greece, India, Pakistan, Poland, Russia, South Korea, Sweden, Switzerland, Tunisia, Turkey, the United Kingdom, and the United States.

Special Feature

Spectroscopists offer their views on their salaries and the current job market, weighing in on topics such as job security, economic recession, and skill sets of new workers.

The 2023 Spectroscopy Employment and Salary Survey: Is a Recession Looming—Or Even Already Here?

 refers to an assortment of statistical tools developed to handle situations in which more than one variable is involved. MVA is indispensable for data interpretation and for extraction of meaningful data, especially from fast acquisition instruments and spectral imaging techniques. This article reviews trends in the application of MVA to pharmaceutical manufacturing and control. The MVA models most commonly used in drug analysis are compared. The potential of MVA to resolve analytical challenges, such as overcoming matrix effects, extracting reliable data from dynamic matrices, clustering data into meaningful groups, removing noise from analytical response, resolving spectral overlaps, and providing simultaneous analysis of multiple components, are tackled with examples. Industrial applications of MVA capabilities are described, with special emphasis on process analytical technology (PAT) and how MVA can aid in process understanding and control. A scheme for selecting an MVA model according to the available data and the required information is proposed.

Modern pharmaceutical manufacturing involves the integration of analytical chemistry into critical steps of the manufacturing process for real-time monitoring and control. This integration is essential, especially in continuous manufacturing schemes, to ensure uniform product quality and process robustness, and to reduce manufacturing cost. The continuous monitoring required in a continuous manufacturing environment is challenging for conventional analytical methods. To resolve such challenges, instruments that allow fast and simultaneous capture of multiple data points are preferred, such as near infrared (NIR), Fourier transform infrared (FT-IR), and Raman spectroscopy; mass spectrometry (MS) and UV-visible (UV-vis) photodiode array detectors that capture multiple wavelengths are also commonly recruited. To interpret the complicated analytical signal of these tools, multivariate regression analysis (MVA) is necessary.

MVA refers to an assortment of statistical tools developed to handle situations in which more than one variable is involved (1–5). MVA can be applied to resolve numerous analytical challenges in the pharmaceutical industry, such as eliminating sample extraction steps during process monitoring, determining the drug substance in an unknown solution that is constantly changing or in a dynamic matrix, minimizing the impact of instrumental fluctuations, and enabling the simultaneous determination of more than one component in a certain matrix.

Selecting the appropriate MVA model has always been challenging task (2,3,6). In this review, the MVA models most commonly used in drug analysis are discussed and compared. A scheme for selecting a suitable MVA model is proposed (Figure 1). The available input data and the type of information required determine whether a supervised or unsupervised model, and which algorithm, should be used. Supervised models are those in which the data include input variables and an output variable, and an algorithm is applied to learn the relationship between the variables and train the model to predict the output when new data are provided. Supervised models can be used for either regression or clustering. Examples of supervised models are multiple linear regressions (MLR), classical least squares (CLS), partial least squares (PLS), artificial neural networks (ANNs), and locally weighted regression (LWR). Unsupervised models are used when the available data are only the “X” input variables. They can be used for clustering and association. Examples of unsupervised models are principal component analysis (PCA) and hierarchical cluster analysis (HCA) (4,5).

CLS and MLR are used for discrete measurements, such as absorbance at certain wavelengths. CLS can determine multiple analytes even in the presence of interfering matrix components, but all spectrally active ingredients should be fully known and modeled during calibration (7). MLR is best applicable for determining a single analyte in the presence of specific matrix components. For MLR, calibration can be done by varying the single analyze concentration in the sample matrix, which allows test samples to be directly analyzed without extraction (8). In both methods, irrelevant spectral information may affect the model prediction accuracy.

In more complicated models where the analytical signal is a full range that may contain both relevant and irrelevant information, PLS can be applied. PLS finds the latent variables. These latent variables relate to the largest covariance between the signal and response matrices (9–11). In this context, PLS is considered another type of supervised learning. It’s worth mentioning that CLS is considered the first generation of the PLS method (12). ANN (13) and LWR (14) can work well in nonlinear models.

The unsupervised multivariate methods are PCA and HCA. In both techniques, a correlation matrix (or sometimes a covariant matrix) is usually computed between the different input 

–variables (factors). In PCA particularly, the number of generated principal components is equivalent to the number of input variables. Importantly, the variation in the studied and investigated data and scale ranges is usually normalized to an important statistical parameter, the unit variance (1/SD), prior to analysis (where SD is standard deviation). The correlation matrix is utilized to determine the data principal components, which are actually the eigenvectors of the matrix. Accordingly, the computed eigenvectors are ranked according to their eigenvalues (scalar values) and consequently, the eigenvector (which is actually a geometrical direction) with the largest eigenvalue is selected as the main principal component, followed by the vector having the second eigenvalue in magnitude, and so on (15).

Similarly, HCA depends on the correlation matrix to generate its specific dendrogram (

Both techniques help in clustering the data into meaningful groups for pattern recognition applications (12).

When information is scarce—a situation usually encountered in the early stages of screening—unsupervised methods, such as HCA and PCA, are preferred. In the later stages of screening, supervised methods such as MLR, PLS, and ANN are often used to predict specific sample properties while still using the former methods to detect similarities and dissimilarities in the data (16). A comparison of MVA techniques discussed in this paper, in terms of advantages and limitations, is presented in Table I. Examples of situations where MVA is analytically useful are discussed.

Eliminating Sample Extraction and Preparation 

Eliminating sample extraction steps is necessary in many situations, such as for online, inline, or atline analysis. This would imply recruiting MVA for extracting relevant information from the captured data where the analytical signal contains information from the analyte of interest as well as other interfering materials.

Tablet monitoring using hyperspectral imaging (HSI) is an example of such an application (17). HSI is a nondestructive and fast analysis option that gives information about tablet component homogeneity and distribution as well as the presence of impurities (18–20). It works by simultaneous acquisition of spatial and spectroscopic data from a sample and gathering them into an HSI data cube where the 

-axes contain the spatial information while the 

-axis contains the spectral information. Collected data are then processed to correct baseline drifts, spectral overlap, and multiplicative scattering; then, the region of interest (ROI) that best represents the analyte of interest is chosen with the aid of MVA models such as PCA. Following the identification of the ROI, a quantitative model can be generated either in univariate or multivariate mode.

Another application of MVA in chemical imaging is in the detection of counterfeit pharmaceuticals. The following techniques were applied for the detection of counterfeit Viagra tablets: Raman microscopy imaging for data acquisition, PCA for determination of the ROI, and K–nearest neighbor (KNN) for data classification. KNN is considered a basic classification technique. It is a supervised learning model that has great application in classification and pattern recognition. KNN functions by determining the Euclidian distances (the square root of the sum of the square differences) between the test sample and the standards available in the training set. It determines the 

 number of points that is closest to the test data. The test is identified based on the group to which the majority of the 

Another example of the use of MVA is seen in HSI of the UV region from 225 to 400 nm for the determination of ibuprofen, acetylsalicylic acid, and paracetamol in tablets. The collected data were deconvoluted by PCA (22).

Monitoring the content and spatial distribution of salicylic acid in film tablets using FT-Raman mapping coupled to multivariate curve resolution (MCR) has also been described. The MCR assumes that the spectral response is a linear combination of the spectra of the pure components in the system. The model decomposes the instrumental response to the individual spectra and gives distribution maps of the components in the spectral image. In this study, pure spectra of salicylic acid and the other excipients were individually collected. FT-Raman mapping spectra were collected as 16 sample points from the flattened tablet surfaces over an area of 3000 × 3000 μm with a step size of 1000 × 1000 μm

. The ROI was selected based on the characteristic peak for salicylic acid and a salicylic acid distribution map was generated in which the active pharmaceutical ingredient (API) was located and the API concentration is represented by color intensity. Quantitation of the API content was performed by averaging the Raman intensity at 1635 cm

 (the characteristic peak for salicylic acid) as collected from the 16 data points (23).

The solid state transitions of the API and excipients can be assessed by HSI. An example is the assessment of piroxicam and lactose dehydration on the surface of a tablet using NIR-based HSI using MCR–alternating least squares (ALS) and parallel factor analysis (PARAFAC) (24).

Another example is the application of fiber-array–based Raman HSI for the simultaneous, chemically selective monitoring of particle size and shape of pharmaceutical ingredients—acetylsalicylic acid, acetaminophen and caffeine—in analgesic tablets (Figure 2) (25).

Monitoring blending and analyzing blend components is another analytical challenge that must be performed without extraction in continuous manufacturing. An example is the atline analysis of API in powder blends using FT–NIR where the analytical signal was captured in the range of 4000 to 12,500 cm

. PLS was used to determine the latent variables that represent the covariance between spectral signal and the amount of API in sample (26). Another example is the determination of hydrochlorothiazide in powder blends by the use of NIR adjunct to PLS. The NIR spectra were recorded for powdered samples in the range of 100 to 2500 nm from which the region 640 to 1780 nm was chosen for the determination. A calibration curve was computed on laboratory-prepared powder blend samples in which additives were included. The procedures were accurately applied to test samples with no sample extraction or preparation (27).

Reducing the consumption of organic solvents is an important benefit of eliminating sample extraction. An example of this concept is the use of thermogravimetric analysis (TGA) for the direct quantitation of ascorbic acid in an effervescent dosage form. In the reported method, a calibration set was prepared using a series of different ascorbic acid concentrations in the presence of additives. The analytical response used was TGA measured at two temperatures for MLR modeling. A relationship was constructed between the ascorbic acid concentration and the TGA measurement at the two selected temperatures in the presence of additives. The method was validated and successfully applied to the determination of ascorbic acid in the dosage form. The method enabled direct analysis of ascorbic acid with no sample pretreatment steps, which reduces the volume of solvents and waste compared to other analytical procedures (28).

Determination of Drug Substance in an Unknown, Constantly Changing, or Dynamic Matrix

The components of a sample change frequently and rapidly during drug manufacturing. Continuous monitoring implies fast analysis to cope with the product evolution and minimizing the effect of variability in the sample components.

NIR is very valuable in this respect. It offers a very fast spectral acquisition method that can be applied for consistent analysis of blend uniformity or monitoring of the dissolution profile using MVA (such as PLS) for data interpretation (26,29). Similarly MVA can be applied to the monitoring of extraction processes (30) and completion of chemical reactions (31).

Minimizing the Impact of Instrumental Fluctuation, Inconsistency, Noise, Irrelevant Data, and Response Overlap 

Proper application of an MVA model can correct for instrumental fluctuation and noise by extracting only relevant information. It can also model (qualitatively or quantitatively) spectral data containing overlapped signals. This capability enables the use for previously abandoned analytical tools for quantitative applications.

In this context, X-ray diffraction (XRD) is the most popular tool for polymorphic state assessment. It is a noncontact, nondestructive inspection technique; however, its quantitative application is hindered by the high probability of error because of sample heterogeneity or improper crystal orientation on sampling (28,32). Such obstacles can be overcome by using MVA. A very good example was presented by Otsuka and colleagues where PCA was used to extract four PCs from the XRD measurement that best represent the hydrate content in theophyllin powder. The method was successfully applied to the determination of theophylline monohydrate in theophylline anhydride bulk substance (33). Other examples are presented in other studies (33,34). Another example is the use of FT-IR for estimation of crystalline purity of cimetidin using PCA for data refinement (35).

In this context, it is very important to mention prediction-augmented CLS (PACLS), which is a CLS modification that adds possible spectral variation to the CLS calibration model. Such modification allows identification and removal of spectral variation resulting from a spectrally active component or instrument fluctuation and drifts that were not introduced during model development. The model enables method transfer between different instruments (36,37).

Selecting the Derivative of a Therapeutic Molecule with the Highest Loading on a Nanocarrier

The differences in results between two natural molecules for the treatment of Alzheimer’s disease (curcumin and bisdemethoxycurcumin) were investigated. The difference between these two molecules in terms of their behavior in the biopharmaceutical formulation (loading on poly [lactic-coglycolic acid] [PLGA] nanoparticles) and their therapeutic levels (interaction with certain receptors) were attributed to differences in their main constitutional, electronic, and physicochemical descriptors (38,39). Five or more hypothesized molecules were introduced and their corresponding descriptors were calculated (40). Because the number of descriptors was actually six (more than two dimensions), a multivariate statistics method, PCA, was used to analyze the results and cluster the investigated molecules, according to their descriptors, into meaningful groups (41,42). Interestingly, the docking behavior of the clustered hypothesized molecules with curcumin was very close to that of the parent compound on all of the investigated macromolecules and receptors. Thus, PCA helped address the analytical and manufacturing challenge of determining which of the of the chemical derivatives achieved the highest load on the investigated nanocarrier (the PLGA nanoparticles).

Simultaneous Determination of More than One Component in a Matrix 

MVA is useful for resolving spectra of interfering analytes—extracting selective information from unselective data. CLS is especially important in resolving spectral overlaps where during the model development the individual spectrum of each component is input into the system to determine its absorptivity. An example is the simultaneous determination of the spectrally overlapping lamivudine and stavudine (43). Similarly, trimethoprim and sulfamethoxasole were simultaneously determined in an animal bolus preparation without an initial separation step despite their overlapping spectra (44). Another important MVA model for multicomponent analysis is PLS. An example is the simultaneous determination of spectrally overlapping analytes using NIR coupled to PLS without sample pretreatment (45,46).

Handling Nonlinearity in an Analytical Method

LWR and ANN are commonly used to resolve nonlinearity in measurement. In LWR models, a large database of samples is used containing variable amounts of the analyte of interest. On the analysis of a new sample, a higher weight is given to database samples close to the analyzed one to create a modified calibration curve. Adjunct to LWR, PLS can be used to extract the most relevant region in the spectra that best represent the analyte of interest; this reduces spectral noise and irrelevant data in the spectra. An example is the analysis of ibuprofen and magnesium stearate during monitoring of cleaning procedures (47). Another example is the use of LWR– PLS to handle nonlinearity in NIR measurements during determination of API content in granules for tableting (48).

Another application is the simultaneous determination of rifampicin and isoniazide by continuous–flow chemiluminescence (CL) using N–bromo succinic acid (NBS) as a reagent and ANN for data deconvolution (49). Rifampicin degrades in alkaline medium, affecting its reaction with NBS and the resulting CL intensity, but this is not the case for isoniazide. This variation in the reaction kinetics of the two drugs that can be harnessed for their simultaneous analysis, provided proper data management. Meanwhile, both compounds react with each other, resulting in CL nonlinearity. For their simultaneous analysis, alkaline solutions of both drugs were pumped into the CL system and merged with NBS for the reaction to take place. The CL emission was measured over 300 s with a 1 s interval. PCA was used to select the PC that represented the concentration of both components and the PC information was used as an input into the ANN model to improve its performance. For the ANN model, optimum performance was obtained when the input nodes were 20, the iterations were 6000, which represents 20 samples each measured at 300 s (20 x 300), the hidden nodes were 22, and the output was 2 to represent the two active ingredients (49).

Extracting Useful Information from Complex and Overlapping Spectra of Commonly Used PAT Tools 

Process analytical technology (PAT) is a system for designing, analyzing, and controlling manufacturing through timely measurements of critical quality attributes (CQA) and critical performance attributes (CPA) of raw and in-process materials and processes with the goal of ensuring final product quality (50). In other words, it is the integration of analytical chemistry into critical steps of a pharmaceutical process for real-time monitoring and control. This integration is essential, especially in continuous manufacturing schemes, to ensure uniform product quality as well as process robustness, and to reduce manufacturing cost.

To achieve real-time monitoring and control, PAT tools are recruited. These are process analyzers (sensors) enabling real-time or nearly real-time analysis. A proper PAT tool should be nondestructive, noninvasive, not require sample withdrawal or contact, and provide fast data acquisition. Ideally handheld instruments and those connected to a probe (inline accessory) are preferable (50).

MVA is used with PAT tools to extract useful information from overlapping and complex spectra, analyze multiple components, generate robust data in a continuously changing and dynamic matrix, and handle signal nonlinearity and correlating CPPs affecting the process. The combination of MVA with PAT tools enables adequate process understanding and control. Examples of PAT tools are FT-IR, NIR, and Raman spectroscopy and other whole-spectrum acquisition instruments. Table II includes examples of such applications.

The use of an immersion probe attenuated total reflectance FT-IR (ATR-FT-IR) instrument was presented to monitor the production of sulphathiazole and assess its polymorphic state. The methods were capable of quantifying sulfathiazole product in the range of 0–30 g and predicting the predominant sulphathiazole polymorphs. In both applications, PLS was used to process data (51). Similarly, probe-based ATR-FT-IR spectroscopy was used to monitor a Curtius rearrangement reaction (52). Other examples of recruiting FT-IR for inline monitoring are the flow FT-IR cells suggested by Ley and coworkers (53–56).

NIR is a superior PAT tool. It is the most widely used tool for PAT because it allows fast data acquisition through transparent glass, plastic containers, or (most commonly) through a probe, with acceptable accuracy. NIR radiation can penetrate deep into the sample, allowing extraction of chemical and physical information (57,58). The analysis is nondestructive and thus will not affect final product yield and also will not require a break in the production line for sampling (59). Both NIR imaging and spectroscopy are described in literature. However, 2D imaging data are inadequate for monitoring deep and dynamic events such as blend uniformity. In such cases, the use of NIR probes will be superior, provided proper placement for full data acquisition (60).

Immersion-probe based NIR has been described for inline monitoring of amino acid concentration during hydrolysis of the traditional Chinese medicine Cornu Bubali using PLS for data processing (61). Similarly, in an other example, NIR was used to monitor total solid, paeoniflorin, and benzoic acid during red peony extraction using a PLS regression model (62). In another case, analysis of the chromatographic purification of crude heparin was described, employing NIR monitoring augmented with PLS to select the variable (signal) representing the heparin content (63).

In another case, a probe-based NIR system placed at the tablet feed frame was used to monitor API and excipients during tablet preparation. The spectral data were processed by PLS to give information about powder uniformity (64). A similar application was the real-time determination of acetaminophen in a powder blend and tablets during tablet production (65). Likewise, an inline NIR spectroscopic method for continuous blend potency determination in the feed frame of a tablet press was reported using sodium saccharine as a model active ingredient. In this method the NIR probe was implemented in the feed frame of the rotary tablet press. A PLS model was used for data analysis (66).

NIR was found to be a successful PAT tool for inline monitoring of a tablet coating using PLS. A fixed-dose tablet of glimepride and metformin was used as an example, where metformin constitutes the tablet core and glimepride exists in the tablet coating. Inline NIR determination of glimepride enabled confirmation of complete coating and identification of the appropriate coating time, which further improved the pharmaceutical process. The process is illustrated in Figure 3 (67).

The use of FT and dispersive NIR instruments to handle continuous process changes such as variation in powder flow rate have been compared (68). In this ex- periment a reflectance NIR probe was connected to FT and dispersive instruments. PLS was used for data processing in both cases. The FT-NIR system showed greater sensitivity and selectivity than the dispersive instrument (68).

Raman spectroscopy is another nondestructive optical tool, with molecular fingerprinting capabilities that enables its application in process monitoring. It has been described for inline Knoevenagel condensation reaction monitoring for entacapone formation using PCA as the associated MVA model. The combination of Raman spectroscopy and PCA enabled faster identification of a reaction end point (69). Determination of minerals in whole tablets has also been described, where Raman spectroscopy was coupled to PLS (70). This technique has also been used for the assessment of dapivirine in vaginal rings (71).

Light–induced fluorescence (LIF) is another emerging PAT tool, and its high sensitivity enables its use for monitoring APIs at very low concentrations. In one example, LIF was mounted using a fiber optic probe into a V–blender. Adjunct to PLS, LIF was able to monitor blend uniformity and accurately determine API content (57).

Handling Process Variability and a Constantly Changing, Dynamic Matrix

Changes in process variables, such as powder rheology, flow rate, and process temperature, may affect beam scattering and thus pose a spectral challenges for PAT tools (72,73). Also, physical properties of measured samples, such as particle size and shape and sample density and compactness may affect the analytical performance of PAT tools (74).

MVA can be used to correct the influence of process variables on accuracy of analytical data (Table III) (74,75). An example is handling the effect of temperature on the accuracy of acquired spectra. This correction can be achieved by augmenting a calibration model to handle temperature-induced variations, as seen in temperature-augmented CLS models (76) and an augmented PLS model (77). An alternative approach is the MVA-assisted selection of a temperature-insensitive variable in the analytical response (78).

Similar to correcting the effect of powder segregation on the accuracy of NIR spectra, inline collection was performed of 500 spectra of a melt extrude, from which 300 were chosen by the PCA model to represent the API content in the melt (79).

Optical path changes may include a change in particle size and sample density, resulting in multiplicative light scattering that in turn affects spectral data reliability. In one example changes in the optical path were corrected by optical pathlength estimation and correction algorithms. The model was applied to correcting NIR spectrum in the presence of variable particles that induced beam scattering (80). Examples also can be found in another study (81).

Beside MVA, proper placement of PAT tools has been suggested as a solution (72). Other researchers have developed statistical models to correct for such signal effects (82). The use of two calibration models has been suggested. The first is a coarse model working over a wide concentration and parameter range throughout a dynamic process (such as blending). The second is a “local” model with fine concentration prediction for a specific stage of blending (58). A recent study applying Raman spectroscopy to inline monitoring of mammalian cell culture production proposed optimizing the Raman acquisition time and including multiple batches in calibrating the model. This was done because acquisition time and batch-related variabilities may affect the integrity of the spectral results (83).

Enhancing Process Understanding and Control

Process understanding helps identify critical points and aspects of a manufacturing process, reducing efforts in process development and scale-up. Figure 4 illustrates a PAT application that reduced the criticality of numerous parameters in a blending process (82). Previously, all the CPPs were considered critical to achieve adequate manufacturing. On applying PAT, an NIR probe was mounted to monitor blend uniformity (the CQA). Mixing speed was identified as the CPP that needed to be adjusted to achieve the desired CQA (82).

In another case, PAT enabled understanding the CPPs affecting production of transdermal glimepride liposomal films. Plackett–Burman screening design was applied to quantify the impact of each process variable on the final CQA of the product. MLR was used to correlate CPPs to the CQA (59).

An NIR probe mounted inline was employed to understand CPPs affecting the hot melt extrusion (HME) of vegetable calcium stearate as a matrix carrier and paracetamol as the API using PLS for NIR data interpretation. The results demonstrated that the screw design was the predominant factor affecting uniformity of the preparation, followed by the effect of multiple feeders; the screw speed was shown to not be influential (79).

Process control is another benefit associated with MVA. A clear example of process control is given in the example of a HME process as part of a continuous solid dosage form production process (84). Extrusion is forcing molten material through a die under pressure. The extrudate leaves the die at high temperature and meets chilled rolls, causing its temperature to fall below its glass-transition temperature, leading the extrudate to form brittle glass sheets that are then ground down into smaller flakes for milling followed by blending, addition of excipients, lubrication, and compression into a final product. The HME process combines multiple steps (mixing, melting, degassing, and densification).

In this example FT-NIR spectroscopy was used to monitor form change, assay, and uniformity of the mixed substance (all CQAs) where the CPPs included temperature, mixing, and feed rate. If CQAs are not met, immediate automated discard of out-of-specification material occurs. Then the necessary CPPs are adjusted. This is controlled through a MVA model. If the same process had only monitored by final product inspection, all subsequent stages after the HME would have been completed, requiring that an entire batch of defective final product be discarded (84).

In another example, PAT was applied to evaluate and control a twin-screw HME process for continuous production of tamper–resistant tablets (85). NIR probes were used to monitor content uniformity assisted by PLS for data interpretation. Other CQAs were also monitored, as shown in Table IV. These data were used to control the process by changing CPPs such as barrel temperature or by diverting defective samples from the production line (85). Other important PAT chemometric models such as multivariate image analysis are described in detail in the literature (86).

A plethora of MVA algorithms exist. They encompass linear as well as nonlinear analytical models. Collectively, they represent a solution to many analytical challenges that encountered during pharmaceutical manufacturing. The application of MVA can reduce the cost and environmental hazards of sample preparation and organic solvent usage. A proposed scheme for selecting the optimal MVA model is provided (Figure 1).

MVA offers reduction of sample pretreatment and screening steps and allows the use of nondestructive analytical tools. It enables prediction and determination in nonlinear relationships, clustering into meaningful groups, accurate determination in multi-component samples, removal of instrumental noise, simultaneous determination of many analytes, and processing of overlapping spectra. Additionally, MVA can relate different CQAs with different CPPs and material parameters and allow adequate process understanding and real-time process adjustment and control. These attributes and capabilities make MVA a perfect match for PAT tools for the development and monitoring of pharmaceutical production processes.

(1) Bro, R. Multivariate Calibration: What Is in Chemometrics for the Analytical Chemist? 

(2) Roggo, Y.; Chalus, P.; Maurer, L.; Lema–Martinez, C.; Edmond, A.; Jent, N. A Review of Near Infrared Spectroscopy and Chemometrics in Pharmaceutical Technologies. 

(3) Wold, S.; Sjostrom, M.; Eriksson, L. PLSRegression: A Basic Tool of Chemometrics. Chemometrics and Intelligent Laboratory Systems. 

(4) Rajalahti, T.; Kvalheim, O.M. Multivariate Data Analysis in Pharmaceutics: A Tutorial Review. 

Quality by Design in Practice, Multivariate Analysis in the Pharmaceutical Industr

Compendium of Chemical Terminology. Vol. 1669

; Blackwell Scientific Publications, 1997.

(7) Franke, J.E. Inverse Least Squares and Classical Least Squares Methods for Quantitative Vibrational Spectroscopy, in 

(8) Rawlings, J.O.; Pantula, S.G.; Dickey, D.A. 

Applied Regression Analysis: A Research Tool

; Springer Science & Business Media, 2001.

(9) Wang, S.; Xiang, B.; Su, Y.; Tang, Q. Direct Determination of Dichlorvos in Water by Partial Least Square-Discriminant Analysis. 

(10) Mirghani, M.E.S.; Che Man, Y.B.; Jinap, S.; Baharin, B.S.; Bakar, J. A New Method for Determining Aflatoins in Groundnut and Groundnut Cake Using Fourier Transform infrared Spectroscopy with Attenuated Total Reflectance. 

(11) Abdi, H.; Chin, W.W.; Vinzi, V.E.; Russolillo, G.; Trinchera, L. 

New Perspectives in Partial Least Squares and Related Methods

; Springer Science and Business Media, 2016.

(12) Hathout, R.M.; Metwally, A.A.; Woodman, T.J.; Hardy, J.G. Prediction of Drug Loading in the Gelatin Matrix Using Computational Methods. 

(13) J. Schmidhuber, Deep Learning in Neural Networks: An Overview. 

(14) Perez–Guaita, D.; Kuligowski, J.; Quint.s, G.; Garrigues, S.; de la Guardia, M. Modified Locally Weighted—Partial Least Squares Regression Improving Clinical Predictions from Infrared Spectra of Human Serum Samples. 

(15) Hathout, R.M. Using Principal Component Analysis in Studying the Transdermal Delivery of a Lipophilic Drug from Soft Nano-Colloidal Carriers to Develop a Quantitative Composition Effect Permeability Relationship. 

(16) Botker, J.; J. Rantanen, J. Multivariate Analysis Supporting Pharmaceutical Research, in 

Multivariate Analysis in the Pharmaceutical Industry

(17) Kandpal, L.M.; Cho, B.–K.; Tewari, J.; Gopinathan, N. Raman Spectral Imaging Technique for API Detection in Pharmaceutical Microtablets. 

(18) Fauteux–Lefebvre, C.; Lavoie, F.B.; Colbert, M.J.; Guay, J.M.; Gosselin, R. Determining the Number of Components for Multivariate Curve Resolution: Case Study Using Raman Mapping of Pharmaceutical Tablets. 

(19) Li, C.; Zhang, D.; Slipchenko, M.N.; Cheng, J.–X. Mid-Infrared Photothermal Imaging of Active Pharmaceutical Ingredients at Submicrometer Spatial Resolution. 

(20) L. Coic, P.–Y. Sacre, A. Dispas, A.K. Sakira, M. Fillet, R.D. Marini, P. Hubert and E. Ziemons, Evaluation of the analytical performances of two Raman handheld spectrophotometers for pharmaceutical solid dosage form quantitation. 

(21) Sacre, P.–Y.; Deconinck, E.; Saerens, L.; De Beer, T.; Courselle, P.; Vancauwenberghe, R.; Chiap, P.; Crommen, J.; De Beer, J.O. Detection of Counterfeit Viagra. by Raman Microspectroscopy Imaging and Multivariate Analysis, 

In this review, we show a wide range of examples of the expanding use of multivariate analysis (MVA) in pharmaceutical manufacturing and control. MVA is being used to resolve numerous analytical challenges, such as overcoming matrix effects, extracting reliable data from dynamic matrices, and more.

Resolving Analytical Challenges in Pharmaceutical Process Monitoring Using Multivariate Analysis Methods: Applications in Process Understanding, Control, and Improvement

March 2023

40 Years Old and Still Solving Problems: Evolution of the ICP-MS Application Landscape

Infrared Spectroscopy of Polymers, XI: Introduction to Organic Nitrogen Polymers

Coherent Anti-Stokes Raman Scattering Cell Imaging and Segmentation with Unsupervised Data Analysis

The 2023 Spectroscopy Employment and Salary Survey: Is a Recession Looming—Or Even Already Here?

Resolving Analytical Challenges in Pharmaceutical Process Monitoring Using Multivariate Analysis Methods: Applications in Process Understanding, Control, and Improvement

Efficient Discrimination of Some Moss Species by Fourier Transform Infrared Spectroscopy and Chemometrics