THE EFFECT OF COMPRESSION FORCE ON THE NEAR-INFRARED SPECTRA OFT ABLET DOSAGE FORMS

The United States Pharmacopoeia/National Formulary (USP/NF) sets the standards and maintains monographs for the evaluation of tablets. These include Official Tests for uniformity of dosage units and disintegration testing, and Unofficial Tests for mechanical strength (hardness, crushing strength) and resistance to abrasion (friability). Current methods of analyzing tablet hardness involve the indirect measurement of the mechanical strength of a tablet through destructive and time-consuming procedures. Near-infrared reflectance spectroscopy (NIRS) is gaining acceptance in the pharmaceutical industry as a non-invasive and non-destructive method for the analysis of finished dosage forms and raw materials. This investigation outlines methods used to evaluate various tablet parameters using NIRS and the achievement of successful predictions of those parameters. NIR models for tablet hardness and density were developed for 15% and 20% hydrochlorothiazide and 2% and 6% chlorphenirarnine maleate in a 0.5% magnesium stearate and rnicrocrystalline cellulose matrix. NIR calibration models for tablet hardness were developed for flat-faced and convex round tablets containing 6% chlorphenirarnine maleate and 0.5% magnesium stearate, with either rnicrocrystalline cellulose or dibasic calcium phosphate dihydrate. Although the NIR response to changing hardness was the same regardless of the drug, separate models were required for tablets of different geometries. Scored tablets also required formulation specific calibrations for NIR hardness determination. Models for upper and lower compression forces were developed for flatfaced round tablets containing 6% chlorphenirarnine maleate and 0.5% magnesium stearate, with either rnicrocrystalline cellulose or dibasic calcium phosphate dihydrate. NIRS prediction of these parameters was at least as precise as the reference hardness test. Calibration of compression forces was successful for rnicrocrystalline cellulose-based tablets, but not for the more variable dibasic calcium phosphate dihydrate systems. The methods described in this investigation may serve as a model for the future acceptance of NIRS as an alternative to current compendia! methods for tablet hardness.

This Dissertation is brought to you for free and open access by DigitalCommons@URI. It has been accepted for inclusion in Open Access Dissertations by an authorized administrator of DigitalCommons@URI. For more information, please contact digitalcommons@etal.uri.edu.

INTRODUCTION
Near-Infrared Reflectance Spectroscopy (NIRS) is becoming a valuable analytical tool in the pharmaceutical industry. NIRS involves the multidisciplinary approaches of the analytical chemist, statistician and computer programmer. The term "chemometrics" refers to the use of knowledge from other disciplines to derive meaningful chemical information from samples of varying complexity. Chemometrics is defined' as the chemical discipline that uses mathematical, statistical and other methods employing formal logic to: a) design or select optimal measurement procedures and experiments, and, b) provide maximum relevant chemical information by analyzing chemical data. Chemometrics has found widespread use in analytical chemistry and is relied upon for the development of NIRS methods. Numerous types of NIRS problems are studied with qualitative and quantitative chemometric methods.
In the NIR region, the radiation can penetrate compacted materials, such as tablets, providing a vast amount of spectral information about the sample. When used in reflectance mode, the NIR light beam is scattered from powder samples after its molecules have absorbed it selectively. The unabsorbed radiation then passes to the several detectors mounted at an angle to the path of the incident rays. The analysis takes approximately 40 seconds per sample. Application of a math treatment, such as second derivative, prepares the raw spectral data for use in a regression and subsequent development of a calibration equation. This type of treatment results in a data file which will yield more information more easily than a raw data file.
Official Standards for the evaluation of tablets are prescribed by the United States Pharmacopoeia 2 (USP) and other compendia and include uniformity of dosage units (weight variation, content uniformity) and disintegration testing. Unofficial tests include those for mechanical strength (hardness, crushing strength) and resistance to abrasion (friability). Tablet hardness or crushing strength is an indication of the mechanical strength of a tablet. Hardness has also been associated with other tablet properties, such as density and porosity'.
Tablet hardness depends on the weight of material and the space between the upper and lower punches at the moment of compression. Inconsistent hardness values are likely to result from variation in these parameters. A tablet should be no harder than necessary for adequate handling and shipping. If the tablet is too hard, it may not disintegrate in the required amount of time. lf it is too soft, it may not withstand the rigors of shipping, handling and dispensing, and may crumble easily. In production, a hardness of four kilograms (or kilopons) is considered to be minimum for a satisfactory tablet 4 • Variation in tablet hardness may be caused by poor mixing, poor flow, unequal length punches, or a variation in the size and distribution of granules being compressed.
Poor mixing of a formulation may lead to poor flow of material, since lubricants and glidants may not have been thoroughly distributed. Poor flow may cause improper filling of the dies, resulting in weight and hardness variation. A slight variation in the lengths of the lower punches can lead to variation in the fill volume and affect tablet hardness.
Finally, if there are too many large granules present, the proportion of large and small particles may change the fill weight in each die' .
Results of mechanical tests give an indication of how the tablet will behave during handling. The crushing strength of a tablet (axial or radial) has been described as a function of the compressional pressure employed during its compaction•. The mechanical strength of a tablet plays an important role in the development and control procedures.
Crushing strength is the most widely used test of mechanical strength. It is defined as the compression force which, when applied diametrically to a tablet, just fractures it7.
The Erweka Hardness Tester, which is commonly used in the pharmaceutical industry, measures horizontal crushing strength by applying a load at 90° to the longest axis. This type of hardness tester is subject to two sources of inherent error: ( 1) the possibility of an incorrect zero and (2)  There is reason to believe that the NIR signal would vary if the compression force used to manufacture tablets was changed. Many drug attributes, such as particle size, density, moisture content and surface area can have a direct effect on the compaction process. Current "wet" methods of analyzing tablets involve the indirect measurement of the mechanical strength of a tablet through destructive and time-consuming procedures.
Many of these tests do not give an entirely accurate indication of how the tablet will behave during handling.
Intuitively, a harder tablet would have a smoother surface, thus changes in hardness (compression force) would be expected to alter the NIR spectra. Presumably, increased compression force causes a harder tablet to be smoother, thus causing less light scattering, leading to greater absorbance and a higher baseline 8 . However, not all of the tablets within a batch or between batches would have the same hardness properties. It is not known to what extent it would be feasible to develop standard procedures to quantify compaction force. The use of NIRS could provide an alternative method of quantifying the compaction properties of a tablet.

HYPOTHESIS TESTED IN THIS Ph.D. PROJECT
There is considerable interest in the area of quantification and prediction of tablet parameters using near-infrared spectroscopy (NIRS) 9 ·w If certain process parameters were altered, a change in the NIR spectra of the sample would be expected. It has been shown that the NIR signal is modified by changes in particle size, morphology and crystallinity. 11 For instance, a very rough surface of many small particles would enhance scattering and absorption of infrared radiation. Thus we can conclude that changes in compression force may have a substantial effect on the intensity of the NIR signal of a drug in a tablet. The hypothesis to be tested in this research is that NIRS can be utilized for the nondestructive determination of tablet hardness in a manufacturing environment, and offers potential as an alternative to traditional methods of hardness testing. Although this project is limited to a small number of drugs and matrices, it is hoped that the general principles will be applicable to many or all drugs. Thus, it is hoped that the project will be of material value in advancing the possible use of NIRS for compendia! standards.  Sci. Ed. 42, p. 194 (1953).

OBJECTIVES
The objectives of this research project were:

1)
To evaluate the utility of Near-Infrared Reflectance Spectroscopy (NIRS) for the measurement of tablet hardness.

2)
To describe the relationship between NIR signal and tablet compaction force or hardness.

3)
To compare traditional "wet" methods of tablet analysis to NIRS methods and calibrate a NIRS instrument to tablet hardness.

4)
To develop a NIRS method of quantifying the compaction properties of a tablet, which might replace or augment existing compendia] tests and demonstrate the potential utility of the technique as an alternative to current methods of tablet hardness testing.

5)
To examine the effect of tablet geometry and excipient matrix on the NlR/compression force relationship.

6)
To evaluate any differences in NIR response due to scoring of a tablet.

7)
To examine the utility of NIRS as a quality control mechanism in determining adherence to batch requirements for hardness.

Introduction
In recent years, near-infrared spectroscopy (NIRS) has become an important analytical technique in industry. It bas been used extensively in the food and agricultural industries for the determination of moisture, protein and starch content in grains'. The use of NIRS to solve pharmaceutical problems is increasing because of technological advances in NIR analytical instrumentation and computer software. Though the spectroscopy itself is not new, its applications and its use of chemometrics are new and innovative.
It is now recognized that NIRS offers significant advantages for a broad range of quantitative applications. NIRS is a rapid analytical technique that uses the diffuse reflectance of a sample at several wavelengths to determine the sample's quantitative composition. Sophisticated software stores calibration equations which correspond to the spectral features of a sample. In the pharmaceutical industry, NIRS has been shown to be useful in determining the percentage of active ingredient as well as in identifying specific tablet formulations . NIRS has potential as both a qualitative and quantitative method in the 9 pharmaceutical industry'. Other pharmaceutical uses include raw material identification 3 as well as monitoring content uniformity in powder mixing operations•.
The NIR region of the infrared spectra was discovered by William Herschel in the early 1800's. It deals with the absorption of electromagnetic radiation in the range of 800 to 2500 nm' . The segment from 1100 to 2500 nm is known as the Herschel region, and is the range most often used in the analysis of pharmaceutical products.
NIR is more often used as a secondary analytical technique than a primary technique.
When used as a primary technique, standards are prepared from reference materials, just as they are for other primary analytical techniques. A library of known spectra is created, then the instrument response is plotted for each sample, yielding a calibration curve.
Sophisticated mathematical techniques are applied to the data via computer software, and the results may be calculated within a few minutes.

Historical Background
Near Infrared Spectroscopy (NIRS) has been used in the food industry for over twenty years to determine the components of feeds and grains. A major stimulus to interest in analytical applications of NIRS has been the success achieved in the analysis of agricultural products. It is anticipated that adoption ofNIRS methods and related technologies will be explosive because they offer the potential for major improvements in quality control, record keeping, and control of product uniformity . However, the requirements for pharmaceutical quality control are more severe than in other fields . Analytical methods are required to be extremely accurate, specific and precise. In addition, since active components are often present in small quantities, the methods must be very sensitive. Absorption in the near infrared region is generally weak, which is an advantage for major components since no sample dilution is needed. However concentrations of minor components are often at or near the detection limit of the instrument' .

Fundamentals And Instrumentation
Theory NIRS deals with the absorption of electromagnetic radiation in the range of 700 to 2500 nm 9 . It is a rapid analytical technique, using the diffuse reflectance of a sample at several wavelengths to determine the sample's composition.
The absorption of infrared radiation is the result of transitions between molecular vibrational and rotational states (twisting, bending). Upon interaction with infrared radiation, portions of the incident radiation are absorbed at specific wavelengths. One of the features of an infrared spectrum is that absorption in a specific region can be correlated to functional groups in the molecule (e.g., fingerprint region 7690-15,380 nm). Multiple vibrations occur simultaneously and produce a complex absorption spectrum that is uniquely characteristic of the functional groups that make up the molecule and of the overall configuration of the molecule.
The NIR region of the spectrum contains overtones and combination bands which are mainly due to hydrogen vibrations (OH, CH, NH). These overtones and combination bands are much weaker than the fundamental vibrations, so the molar absorptivities are between 10 and 1000 times smaller than those of the corresponding infrared bands. Since ordinary glass is transparent in the NIR wavelength range, the optical components of NIR instrumentation don't have to be made of fragile materials. This lack of response by glass as well as quartz enables these materials to be used as transparent containers and also permits the use of optical fibers to transmit the spectra. ' 0 Glass cannot be used in instruments designed for the mid and far IR regions.

Phannaceutical Applications
NIRSysterns Another issue is that of "transferability" of the calibration model, including transferable correlation coefficients, that would be usable on all instruments. A model built on extensive samples and spectra is much more readily transferable than one developed with only a few samples. Although some progress has been made in making calibration transfers between instruments, the situation is far from ideal, and careful monitoring is needed to obtain satisfactory results.

QC and Regulatory
The pharmaceutical industry's interest in NlR technology is in the production of better products at a lower cost, while the regulatory interest is in product control and uniformity and the detection of deviations from the approved formul ations.
Validation of a NIRS method is necessary for acceptance by regulatory bodies. The error of the primary method must be well known. Accuracy, linearity, reproducibility, specificity, sensitivity, and robustness of the method must be demonstrated . The accuracy of the NIR results is obtained by comparison with the reference analytical method.
Specificity of the method can be determined through the use of instrument software which qualifies the sample. Since sample placement is an important source of error, the same sample should be measured, removed and remeasured several times to determine reproduciblity. Robustness of the assay may be tested by varying the operating conditions. solvents, and produced no waste products. This approval is likely to be followed by other computer-based technologies which will rapidly come into use in the pharmaceutical industry. It is anticipated that "adoption of these technologies will be explosive because they offer the potential for major improvements in the control of product uniformity and quality and better record-keeping ... " at a significantly lower cost 12. Canada's pharmaceutical regulatory agency, the Health Protection Branch (HPB), recently (December, 1993)

Moisrure Determination
The classical methods for water determination are based on weight loss by drying  Boehm and Liekmeierl 9 studied the moisture content of solid dyes and organic solvents. They identified three working ranges for quantitative water determination (Table   I). Nonpolar, organic solvents, such as carbon tetrachloride, were analyzed at 2400 nm, the most sensitive band. An indigo dye was analyzed at the 1900 nm wavelength. between an a-hydrogen atom on the lovastatin ester group and an additional a -methyl group in simvastatin. However, the MIRS method was able to distinguish between the two structures, since it is well suited for structural elucidation and identification of compounds.
Lodder and Hieftje26 described a NIRS method of analysis of intact aspirin tablets.
The method involved the use of a double-reflecting aluminum sample holder which preserved the integrity of the tablet during the analysis.
Lodder, et at27 reported a NIRS method for the detection of adulterated nonprescription drugs. This work was triggered by the 1982 incidents28. of potassium cyanide-laced Extra Strength Tylenol® capsules in the Chicago area that resulted in seven deaths The adulterants tested included potassium cyanide, sodium cyanide, ferric oxide, aluminum metal shavings, arsenic trioxide, and sodium fluoride. The detection limit for potassium cyanide was 2.6 mg, or two orders of magnitude less than the lowest reported lethal dose in humans (2.941 mg/kg, or 306 mg for a 70 kg person). One shortcoming of a NIRS method in this situation is that it is not possible to predict what contaminant might be placed in a particular product. The authors' results indicated that a variety of contaminants could be detected in intact capsules by using four wavelengths.
Drennen and Lodder29 developed a non-destructive NlRS assay for determination of the degradation products for intact aspirin tablets. The authors concluded that the salicylic acid formed by hydrolysis of aspirin significantly changed the spectrum of aspirin tablets after exposure to moisture and that this correlation to salicylic acid resulted from salicylic acid formation rather than a correlated process. The mass of salicylic acid formed by hydrolysis in intact aspirin tablets was measured by NIRS with a reported error of 0.04% of the total tablet mass (400 ppm).
Ciurczak30 reported a powder mixing study which utilized NIRS to check for homogeneity. For this study, aspirin and vitamin 812 were the active ingredients.
Aliquots were taken at various times and analyzed via NIRS. A comparison was made between visual matching, spectral matching and principal component analysis. Visual matching provided an approximation, while spectral matching (using computer software) gave somewhat better results. Principal component analysis was the more rigorous method, in that it was able to distinguish between the penultimate and the true final mix.

22
The author suggested that the use of NIRS could save many hours of analysis time in a routine mixing study.
Ciurzcak, et aJ3 I described a method of determination of the mean particle size of pure, granular substances. The method is based on the theories of reflected light, namely the Kubeika-Munk equation, where K is the absorption coefficient and S is the scattering coefficient. Reflectance, R , increases as the mean particle size decreases, while R decreases as the absorptivity increases. Graphs were constructed from log ~ values and used to assess the particle size of pure samples of ascorbic acid, aspirin, and aluminum oxide. The absorbance values for each spectrum at 1658 nm (the major peak) were plotted against the absorbance values at 1784 nm (the baseline), resulting in a linear plot with a correlation coefficient of 0.99999.
The authors concluded that this method could be used as a quality control tool when new materials are received.

Liquid Dosage Forms
Dubois, et al37 reported a method of detennination of five components in a liquid formulation of otic drops. The product contained two active components (phenazone and lidocaine), two solvents (ethanol and glycerol) and one antioxidant (sodium thiosulfate).
Results indicated that the NIRS method was well suited to the quantitation of both of the solvents and one of the active compounds (phenazone). The concentration oflidocaine in the fonnulation (I %) was at the detection limit of the instrument, thus the accuracy of the method was insufficient.
Kumar and Raghunathan38 used NJRS to examine the nature of the water pool fonned in the reverse micellar system, lecithin/nonpolar solvent/water. The three nonpolar solvents used in the study were benzene, carbon tetrachloride and cyclohexane. The NIR spectra indicated the presence of two types of water in the lecithin reverse micellar solutions. One was water-dispersed in the organic phase and the other was watersolubilized in the reverse mi cellar interior. Results revealed that the amount of water present in the organic phase was negligible at all water concentrations in all three solvents.
Grant, et al 39 investigated the quantitative analysis of solutions containing various concentrations of sodium hydroxide, sodium chloride and sodium carbonate using NIRS.
It was observed that the addition of these salts caused changes in the absorption spectrum of water, even though sodium carbonate and sodium chloride do not themselves absorb in the NJR region.

Other Possible Uses
The hydroxyl value is an indicator for the stages of an esterification reaction.
Hansen reported a NIRS method by which the shifting of the hydroxyl value of a reaction could be monitored 40_ This work suggests that it may be possible to monitor the degradation of a reaction, and possibly be useful in stability testing of raw materials.
Tudor, et al 4 1 investigated the use of near-infrared Fourier-transfonn (Ff) Raman spectroscopy in the molecular structural analysis of drugs and biomedical polymers. The authors developed a technique by which the concentration of a drug within a polymer vehicle could be determined over a wide drug concentration range. The Ff-Raman spectrum of diclofenac dispersed in a sodium alginate matrix was monitored, as well as the spectrum of the alginate alone. It was concluded that this method illustrated the potential for quantification of degradation kinetics in certain polymers using Ff Raman infrared spectroscopy.

Conclusion
Recently, there has been a large increase in the amount of research in the near infrared region. NIRS bas been shown to be a valuable tool for a number of important applications. It bas gained official acceptance in the food and agricultural industries, and is now becoming more recognized in the pharmaceutical industry. Specially designed instrumentation for use in the pharmaceutical industry has become more widely available, and is made more powerful by software improvements. In the food industry, NJRS has been used for the qualitative determination of hardness I 5, 16 of wheat kernels. Wheats are classified as hard or soft according to their milling performance. Hard wheats produce larger, more angular particles during the grinding process than soft ones. This angling occurs because hard wheats have cleavage planes associated with the cell walls in the endosperm; the cells come away more cleanly and remain more intact 17. Soft wheats fracture at random, frequently across cell walls, resulting in fragments containing mainly starch. NIRS can be used to discriminate hard and soft wheat kernels through differences in particle size of ground meal . The ground meal from the hard varieties reflects less energy than soft ones ground in the same fashion IS.

Calibration Transfer Between Instruments
In a recent commentary stemming from the 1995American Chemical Society Meeting at the short course on practical NIR (the first NIR short course offered by the A.C.S.), consultant Emil Ciurczak stated that " it is more important for instruments to perform in a reproducible manner than for them to represent absolute values of some mythical standard." l 9 In other words, Ciurczak stressed that it is doesn't matter that a single wavelength of one instrument may be 2 nm off from an absolute standard, as long as it always identifies that point the same way every time. The National Institute of Standards and Technology (NIST) in Gaithersburg, MD has no single standard available for wavelength accuracy in transmission NIRS. A reflectance standard exists but has not yet been widely accepted. The reflectance standard is based on polystyrene, however, and consistency in batches of such polymers has not yet been achieved. At this time, the responsibility of establishing proper calibration is on the manufacturer of the instrument.
Bouveresse and Massart20 described a modified algorithm for standardizing NIR spectrophotometric instruments. The authors used locally weighted regression, which gives more weight to the standardization samples which are in the same spectral intensity range as the samples to be predicted and less weight to the samples farthest from this range. This approach enabled standardization of samples of a different nature.

Standards for Quantitative NIR Analysis
The American Society for Testing and Materials (ASTM) recently published an official document2 l providing a guide to spectroscopists for the multivariate calibration of infrared spectrometers. The scope of the publication, entitled "Standard Practices for Infrared, Multivariate, Quantitative Analysis", includes the use of multivariate 29 calibrations for the determination of physical or chemical characteristics of materials. The practice applies to the near-infrared (780 to 2500 nm) through the mid-infrared (4000 to 400 cm-1) spectral regions. This document is the first official standard for the application of chemometric multivariate analysis to near-IR and IR instruments22.

Validation of NIRS Methods
Validation aspects of a NIRS method are similar to those of other analytical methods. The principal elements of ensuring linearity, accuracy, selectivity and reproducibility of a quantitative method are required. The validation process determines the amount of error due to variation between the values in the population. It is used to check for the existence of a relationship between the calibration set and the validation set.  situation is when this plot shows points scattered evenly along a line that is 45° to the xaxis, thus zero bias. A bias adjustment is required when the value for the bias is more than double the standard error of the bias in the positive or negative direction.

31
The main problem associated with using empirical models is that they are based on correlation rather than causation2 4 . Construction of these models involves finding measurements that simply correlate well with an analyte. The validity of these models depends the ability of the calibration set to accurately represent the samples in the prediction set. A good rule of thumb is to make sure that any type of variation observed in the prediction set also varies in the calibration set over the same range as the variation occurring in the prediction set25. Usually, the complete prediction set is not available at the same time as the creation of the calibration set, and unusual phenomena may be associated with some of the prediction samples.
One source of prediction error is the inherent accuracy and precision of the reference method used. If the reference method produces erroneous analyte values that are consistently high or low, this bias will be reflected in the prediction results. Imprecise (but accurate) reference values may also increase prediction error, in a nonsystematic way. Thus it is very important to minimize the errors in the reference values that are to be used to create an empirical model.
Other sources of prediction error relate to the reproducibility, stability and repeatability of the NIR instrument. Reproducibility (precision) is validated by making repeated measurements of the same sample, removing it between runs. Small changes in conditions may occur due to multiple insertions of a sample onto the instrument. Stability refers to similar changes that may occur over a longer period of time (hours or days).
Repeatability refers to the instrument's ability to generate consistent measurements under the same conditions (without removing the sample from the instrument}, over a relatively short period of time (seconds or minutes). All of these factors must be addressed to assure the validity of the NIR calibration model.

Current NIR News
In 1995, the European Patent Office26 granted a patent to Dr. Paul Aldridge of Pfizer Central Research in Groton, Connecticut for an Apparatus for mixing and detecting on-line homogeneity. This patent involves the use of a NIR fiber-optic probe interfaced on-line with a blender. Sekulic, et a127 recently described the use of this Apparatus for on-line monitoring of powder blend homogeneity . An S.-quart twin-shell V-blender was interfaced with a fiber optic probe at the axis of rotation. Spectra were collected at prescribed intervals, and data analysis was performed by a series of software packages.
Variability in the NIR spectra as a function of time was measured, and it was shown that this variability reached a minimum sooner than traditional blending times suggest.
Official approval of a NIR method as an alternate method for identification and assay of tablets was granted in June 1995 to Glaxo Wellcome in the United Kingdom28.

Conclusions
Although the use of NIRS methods in the pharmaceutical industry is increasing, many scientists are reluctant to accept it as a viable alternative to current testing methods.
The process of developing a calibration and selecting a model is challenging project, as is the validation of the method. Traditionally, pharmaceutical scientists are not trained in chemometric methods, and this remains a stumbling block to the understanding and implementation of NlR technology. Instrument manufacturers and software vendors are aware of this, and now design their products in more user-friendly ways than before. It will be several more years before sufficient data is published to convince the skeptical that NlRS is a usable and extremely useful technology .

20)
Bouveresse, E. and Massart, D. L., Anal. Chem., 67 (8) (3) chlorpheniramine maleate (CTM) 2 % with placebo matrix. Five or six levels of tablet hardness (2 to 12 kg) were used for each formulation. Tablets were evaluated by conventional USP testing methods for weight, hardness, thickness and friability. NlR reflectance analysis was performed on 20 tablets from each batch using a NTRSystems Rapid ContentT>• Analyzer. Tablet evaluations showed hardness variation of 5-10% and weight variation of< l %. NlR analysis of these tablets showed an upward shift in the raw spectra with increasing hardness. Softer tablets had more variable spectra. Principal component analysis correctly (by distance) identified tablets that were two hardness units apart. Results confirmed that there is a difference in raw NlR spectra due to changes in tablet compression force.

2 Introduction
Near-infrared reflectance spectroscopy (NIRS) bas received widespread attention as a nondestructive method for the rapid measurement of the composition of many products 1 ·2.3 . NIRS determines these parameters through the measurement of diffuse reflectance. Diffuse reflectance is light that bas been transmitted through a portion of the sample and emerges from the illuminated surface due to internal light scattering4 . This type of reflectance is affected by the absorbance and light-scattering properties of the product.
Current methods of tablet hardness testing, drug identity and content are destructive in nature and may not always give an accurate representation of the batch being evaluated.
NlRS is a noninvasive and nondestructive method that, in theory , would allow for 100% testing. In this respect, NIRS is attractive from both a quality control and a regulatory perspective.
The purpose of this study was to investigate the feasibility of NIRS for the measurement of tablet hardness.

.3. TabletManufacture
Half-inch round, flat-faced tablets were manufactured using one of the sixteen stations of a Stokes B-2 Rotary Tablet Press. Hydrocblorothiazide and chlorpheniramine maleate were the active components chosen for the formulations, in addition to microcrystal line cellulose(Avicel® PH 102, FMC Corporation) and magnesium stearate. Both are relatively low-dosage drugs and would thus not be expected to interfere with the process of direct compression. The components of each formulation (Table 1) were accurately weighed on a Mettler balance for a batch size of one kilogram. Each blend was mixed for ten minutes in a Turbula mixer, then transferred to a labeled plastic bag to await compressing.
38 Table I. Formulations used in feasibility study. Friability testing was performed using a Roche Friabilator. Twenty tablets from each batch were weighed on a Mettler balance before and after undergoing four minutes (100 revolutions) in the friabilator. The weight difference was calculated and expressed in terms of percent loss due to abrasion or fracture. was used to create color output files for spectral plots. Plots were downloaded to a laser printer in a remote location.

.3.2 Near-Infrared Analysis
NTR reflectance parameters were set at32 scans per sample in the range of 1100 to 2500 nm. A ceramic (Coor' s Standard) reference scan was taken before each set of samples. Single tablet NTR scans were run on 20 samples from each batch of tablets. The sample to be measured was placed directly above the detector surface and centered with the iris. Before positioning each sample, the detector surface was gently cleaned of debris.
Each sample scan took approximately 42 seconds to complete.
Reflectance spectra were collected for the active components (in powder form), the individual excipients used in the blends, and the powdered blends before compaction. The sample to be measured was loosely packed into a 50 ml glass beaker, up to a volume of20 ml. The beaker was tapped lightly three times to level the powder surface. The spout of the beaker was aligned the same way for each sample. The sample was scanned once, then tapped three times on the counter, and rescanned, with the spout shifted 45' from the original alignment A third scan was taken after the beaker was tapped and rotated another 45' . The same beaker was used for all components. The three scans were averaged and overlayed on a plot.
A spectral Ii brary of each " product" (batch) was created through the use of JQ2TM .
Spectral data from each hardness level of HCTZ tablets were entered into the HCTZ product library. Likewise, CTM spectral data were entered into the CTM product library.
Through the software's internal validation program, principal components were calculated.
Identification by correlation and by distance was evaluated. "Correlation" is one of the modes of spectral matching used by the NSAS software. In correlation mode, the products in the spectral library are compared by correlation to see if they match a potential sample for analysis. In the "distance" mode, the program uses the distance between the library records and potential samples. A minimum of three spectra in the library is required for the distance mode to work.

.4.Results o/Tablet Evaluation
Tablet weights were very consistent, with a relative standard deviation of less than 1 % (most batches had a relative standard deviation of 05% or less). Specifications for tablet hardness were set at 10%. Table 2 summarizes the average hardness, weight, thickness and friability for the placebo blend. Relative standard deviations (RSD) for the mean hardness values for the placebo tablets ranged from 4.50% to 11.2%. Table 3 summarizes the average hardness, weight, thickness and friability for HCTZ 15% tablets. The RSD's for hardness ranged from 3.3% to 14.6%, following a trend of increasing variation with a decrease in hardness. The standard error of the Erweka Hardness Tester was calculated to be 0.34 kg.
A summary of average hardness, weight, thickness and friability for CTM 2 % tablets appears in Table 4. The RSD' s for hardness ranged from 3 .18% to 8.49%. This formulation produced tablets with the least amount of deviation from the target hardness.
Overall, the variation (RSD) in tablet hardness ( Figure 4) was observed to be inversely proportional to the hardness value, a trend which was generally reflected in the ftiability data (decrease in hardness results in an increase in friability) . It is logical to expect that a softer tablet may have more variability with respect to values for hardness. 42

Results of NIR Analysis
The raw and second derivative spectra of CTM, Avicel® PHI02 and magnesium stearate are shown in Figures  The NIR absorbance (log l/R) versus wavelength (run) was plotted for each batch of tablets. When the raw spectra from all batches of a blend were overlayed on the same plot, a general upward shift in absorbance was observed in response to an increase in hardness.
Plotting the second derivative spectra at several hardness levels also demonstrated an increase in absorbance at the peak maxima, although less obvious. The raw and second derivative spectra of the placebo blend tablets at five hardness levels are shown in Figures 9 and 10, respectively. Figures 11 and 12 are the raw and second derivative spectra of HCTZ 15% tablets at five levels of hardness. The raw and second derivative spectra of CTM 2% tablets at six hardness levels are shown in Figures 13 and 14.
Observation of the spectra in Figures 12 and 14 demonstrate the smoothing effect of the second derivative math treatment The spectra appear to be nearly superimposed upon each other, except for small changes at absorbance maxima. Linear regression was performed on the spectral data at absorbance maxima Numerous single wavelengths were chosen based on the appearance of an effect from increased tablet hardness. The results of these analyses are summarized in Tables 5 and 6. CTM 2% absorbance at eight single wavelengths was found to be significantly increased with an increase in tablet hardness.

43
HCTZ 15% absorbance at eight (out of ten wavelengths chosen) was also significantly increased in response to an increase in tablet hardness.
A spectral library was created for each of the two blends containing active components, using the second derivative spectra. Each level of hardness represented a separate product, from which the calculation of principal components was generated. The principal component analysis (PCA) for HCTZ 15% resulted in the successful identification of each product by distance. Seven HCTZ 15% samples (47 correct/ 7 incorrect) were incorrectly identified when the library was validated by correlation. The correlation method also identified four out of ten conflicting pairs of products in this library. This information is useful for future updating of the library , since there may be a better set of samples for use in this spectral library.

Conclusions
The results of this feasibility study indicate that there is a change in the NIR signal as a function of tablet hardness. Three tablet formulations were manufactured of varying compression forces and analyzed by NIRS. An increase in tablet hardness resulted in a consistent upward shift in NIR absorbance. Tablet samples of different hardness levels were successfully differentiated by principal component analysis. NIRS has the potential as an altemati ve method of tablet hardness testing.

Acknowledgments
The Rapid Content Analyzer was loaned by Perstorp Analytical/NIRSystems, Silver Spring, MD.
I would also like to thank Mr. Rajeev Jain for his assistance with the tablet manufacturing and hardness testing portions of this study.    .

.2 Evaluation ofTablets
Twenty tablets from each batch were evaluated for weight, hardness, thickness and friability. The USP tolerances for weight variatioo3 allow a percentage difference of 7.5 for an average tablet weight of 130 to 324 mg. Tablets weighing more than 324 mg may differ by no more than 5%. The USP requires that twenty tablets be individually weighed and their average weights calculated. The weights of no more than two of the tablets may differ from the average weight by more than the prescribed percentage. No single tablet weight may differ by more than double that percentage. In this study, the target weight was 324 mg+/-5.0o/o. Tablet thickness and friability were measured according to the protocol described in Manuscript II .
Tablet hardness was measured using the Erweka Hardness Tester. Hardness testing was performed on twenty tablets from each batch after all non-destructive physical tests were completed. This order of testing allowed direct correlation of data to a specific  ( l): where r = radius of 0.5 cm.
The RSD of tbe density values were tabulated and compared with the RSD of the laboratory hardness values. The calculated density values were entered into the NSAS™ computer files for the corresponding NIR spectra. For practical purposes, surface porosity was disregarded, i.e., it was assumed that the tablets were flat.
The standard error (standard deviation) was calculated for each laboratory (reference) method. The overall and single product values for tbe reference standard errors were used for comparison to the NIR standard errors.

.3.3 Near-Infrared Spectroscopic Analysis
A NIRSystems Rapid Content Analyzer® Model 5000 was used for the analyses of tablet samples. This instrument and corresponding setup were described in Manuscript II.
NIR reflectance parameters were set al 32 scans per sample in the range of 1100 lo 2500 nm. A ceramic (Coor's Standard) reference scan was taken before each set of samples.
Single tablet NIR scans were run on 20 samples from each batch of tablets. The sample to be measured was placed directly above the detector surface and centered witb the iris.
Before positioning each sample, the detector surface was gently cleaned of debris. Each sample scan took approximately 42 seconds to complete. For each sample tablet, the lab hardness value for the corresponding NTR spectra was entered into the computer as the constituent value for hardness. The NIR spectral data were mathematically transformed to their second derivative spectra using a segment of 20 and a gap of 0. The segment size refers to the number of wavelengths the computer averages into one data point to improve the signal to noise ratio. Gap size is the distance in nanometers between wavelength segments. These two parameters will vary according to the math treatment in use.
Of the twenty spectra collected per batch, thirteen spectra per batch were selected for inclusion in the calibration set. To test for bias in the data, several calibration sets were created for each formulation , either by random computer selection or simply using the first thirteen spectra. The remaining seven spectra were used to create a validation sample set.
Each of the HCTZ calibration sets contained a total of 65 samples ( 13 x  Validation of each model was performed by applying it to a set of validation (or prediction) samples to test the model's predictive ability. These predicted values were then statistically compared to laboratory hardness values measured for these samples and checked for agreement of the model with the reference method. The standard error of prediction (SEP), also known as the RMS error, was also calculated.

J .3 .4 Effect of Sample Position on NIR Spectra
A validation study was undertaken to evaluate the effect of sample position on the resulting NlR spectra. A comparison was made between spectra taken from (A) one tablet in ten positions without the use of the iris, which is normally used to center the tablet before analysis, (B) ten tablets in one position, both sides, using iris, and (C) one tablet scanned 20 times using iris.
In Part A, one tablet was scanned in ten positions on the instrument's detector surface without the use of the iris. Samples 2 through 10 were scanned with an edge of the tablet touching the approximate center area of the detector surface. Figure  In Part B, one tablet was scanned twenty times in the center position. The tablet was scanned, removed, then repositioned between scans using the iris to center the tablet on the detector surface.
In Part C, twenty tablets were scanned once using the iris to position the sample in the center of the detector surface.

Results
The tablets used in this study were manufactured by direct compression. This is the simplest tablet production method and has several advantages over wet granulation methods. It is an economical method, since few processing steps are involved in comparison to wet granulation. No moisture is involved in the preparation of the blends for direct compression, thus the tablets made from this process tend to be more stable than those produced by wet granulations. However there are several d. isadvantages to direct compression methods. There are relatively few crystalline substances that may be directly compressed. Also, many products contain a low effective dose of the active drug, which may be difficult to adequately distribute throughout the tablet matrix. Another disadvantage of direct compression is that a limited amount of active ingredient may be incorporated into the matrix (usually no more than 30% ). This is a major limitation when formulating high dose products.
A relationship between NIR signal and tablet hardness or compaction force was established for four tablet formulations containing active drugs, and for one placebo formulation . The same excipient matrix was used for the entire study.

Results of Physical Testing of Tablets
Results of tablet evaluations for hydrochlorothiazide 15% (HCTZ), chlorpheniramine maleate 2% (CTM) and placebo blends were previously reported in Manuscript 11. Tables 2 and 3 summarize the results of physical testing of HCTZ 20 % and CTM 6 %,respectively. Tablet weights for all batches were very consistent, with a relative standard deviation of less than 0.6 %. Friability was observed to increase in response to a decrease in tablet hardness. The properties of all batches fell within acceptable performance guidelines.
The mean values for tablet hardness were generally lower than the target hardness value for each batch (Table 4). An exception to this occurred with the placebo blend, where batches in the 4, 6, and 8 kg range were higher than the target values. One other  Overall, the relative standard deviation (RSD) for lab hardness values ranged from 3.23 to 13.24%. For each formulation , the RSD was found to increase as the average hardness decreased. The RSD for hardness was plotted against the age of the tablet, relative to the time that hardness data was collected ( Figure 2). No relationship was found between RSD for hardness and the age of the sample. However, when the age of the product was plotted ( Figure 3) against the percentage difference from the target hardness, there was a linear relationship (r2= 0.586). In Figure 4, the data was plotted and regression was performed by product. No correlation between age and percentage difference from the target hardness was found for the single products HCTZ 20o/o (r2= 0.047) or CTM 2% (r2= 0.00). Io Figure 5, we find similar behavior between both HCTZ products and CTM 6%. The overall r2 for both HCTZ products was 0.82, and 0.83 for CTM 6% alone. Similar slope values resulted when both HCTZ products (slope= 0.446) were compared with the CTM 6% (slope= 0.531) formulation. Although inconsistent between products, this evidence suggests that tablet hardness may change over time, and hardness data should be collected immediately after tablet manufacture for the most accurate results.  When all of the HCTZ data was plotted together as in Figure 11 , the r2 dropped to 0.729. Figure 12 is a plot of density and hardness values for all four formulations. When from all four formulations were plotted together, the r2 was 0.45. As a linear relationship was demonstrated between density and tablet hardness for these products, I concluded that it would be feasible to develop NIR calibration models for density as well as hardness.

.4.2 Results of NIR Spectral Analysis
The raw spectra of CTM 6% and HCTZ 20% are illustrated in Figures 13 and 14.
The corresponding second derivative spectra are displayed in Figures 15 and 16. The overall shapes of the spectra are similar between the two formulations, due to the fact that the major portion of each formulation was composed of the same matrix. It may be observed from the spectra in Figures 15 and 16 that the baseline at 1100 nm starts out quite close for all hardness levels, then begins to diverge after about 1500 nm.
The raw and second derivative spectra ofCTM 2% and HCTZ 15% were given in Manuscript II. The second derivative spectra of all four products at a hardness level of 2 kilograms is shown in Figure 17. The spectral differences between the formulations are more obvious in this plot. Each spectrum represents the average spectra of twenty single tablets.  .c < HCTZ 20% Tabs (h=2 to 12 k g)

82
. 694 .552 "' I Wavelength (nm) Figure 17. Second derivative NIR spectra of HCTZ 15%, HCTZ 20%, CTM 2% and CTM 6% tablets at a hardness of approximately 2 kg. Figure 18 illustrates the effect of various sampling positions on the NTR spectra of one tablet The resulting spectra cover nearly a four-fold range in absorbance. Without the use of the iris to position the sample, even the spectra of the two replicates at 12:00 were shifted from one another. The sample scanned in the center position was closest to the actual baseline. The remaining samples were grossly shifted upward in relation to the center sample. Samples positioned at 180° and 225° were nearly superimposed, as were samples at 45" and 270°. It is interesting to note that the second derivative treatment ( Figure 19) reduced the baseline offset of the samples so that they were almost superimposed upon each other. The region from 1476 nm to 1594 nm contained a significant amount of variation, as is evident from the plot The standard deviation spectrum of these samples appears in Figure 20, where the highest amount of variation appears around 1500 nm. It is likely that a loss of spectral information occurs when the sample is not reproducibly and accurately positioned.
There was no significant difference between the raw spectral plots of one tablet scanned twenty times ( Figure 21) versus twenty tablets scanned once ( Figure 22). These spectra appear to be one solid line of varying thickness at specific wavelength regions. The iris was used on both sets of samples, thus the shapes of the spectra are the same. Figure   23 and Figure 24 further illustrate the similarity between these two sets of samples. These are the standard deviation spectra of the respective sample sets. When superimposed, they differ slightly in absorbance, but not in "peak" position. This portion of the study demonstrates the precision of the NIR method. Multiple scans of the same sample yielded nearly the same results as a single scan of several samples. Overall, using the iris to center the tablet dramatically reduced the error due to sample positioning.
One Tablet in 10 Positions Wavelength (nm) Figure 24. Standard deviation spectra of twenty tablets scanned once.

.4.4 Results of Calibration Development
Once a suitable set of calibration models was developed, they were applied to a set of validation samples that were part of the original population but not included in the calibration set. The choice of the best calibration models was made by comparing statistical parameters that were calculated by the NSAS™ software. First, since a NIRS method cannot be more sensitive than its primary analytical method6, it is important that the The validation process involved the calculation of various statistical parameters that measure goodness of tit. These parameters (defined in Appendix I) include the ratio of bias to standard error of the bias (bias/SEB) and the ratio of standard deviation of the difference to root mean square (SDD/RMS). The RMS is close to the standard error of prediction (is not n-1); it is the non-bias corrected standard error. Standard deviation of the difference (SDD) is the bias-corrected standard error. The desirable ratio between these parameters is 0.9 to 1.0. A rule of thumb for goodness of fit using these parameters is that bias/SEB should equal no more than 3.0, and the ratio SDD/RMS should be close to 1.0.
It is desirable to have the bias and the SEB close to zero. It is also important that the value for slope adjustment be close to l.0 as well , since this indicates linearity. Note that it is unacceptable to change the slope9, as this indicates that the model does not fit the sample population.
When Therefore, eight factors were chosen as the maximum number allowed in the regression.
The "mixed" calibrations required more factors, since they included the addition of drug concentration as a variable.
As the current study was extensive in the development of various calibration models, only the summary results will appear in the text of this manuscript The reader is directed to Appendix 2 for more detailed calibration results, which include wavelength and calibration coefficient information for each formulation.

Results of HCTZ Calibrations
Tables 6 and 7 summarize the calibration of HCTZ 15% tablets and HCTZ 20% tablets, respectively. Different MLR wavelengths were selected by the computer for calibrating each HCTZ formulation. Although it might be expected that the calibration wavelengths would be the same for a given drug, it must be reiterated that the equations are developed through chemometric methods using the wavelengths having the greatest  ( 107 variation in absorbance. It was possible to preselect the regression wavelengths, but in this experiment, the resulting predictions were not generally as good. Calibrations that were designed for Herz 15% did not fit data sets from HCTZ 20% samples. Table 8 contains a series of Herz 15% calibrations at 2076 nm. The calibration sets used to generate these models differed in the manner of sample selection and exclusion of outliers. (An outlier is a data point that falls well outside the main population.) The first model (hct l Scal) was generated from a sample set that was not randomly selected. The first thirteen spectra from each of the five hardness levels were selected for the calibration sample set. The next two models were generated from a calibration set that was randomly selected. These two models differ in the number of wavelengths used to perform the regression. The addition of another wavelength to the model resulted in an improved SEE and correlation coefficient. The next set of three models in Table 8

Results ofCTM Calibrations
Calibration models for CTM 2% and CTM 6% are summarized in Tables 10 and   11, respectively. Numerous equations were developed at multiple wavelengths for each formulation. The computer selected different wavelengths for the two different drug concentrations. In general , the addition of wavelengths to the calibration resulted in a higher correlation coefficient and a better SEE. However, better predictions (validations) resulted from the models that used fewer wavelengths in the multiple regression of both CTM 2% and CTM 6% tablets. PLS performed somewhat better than MLR.
Validation of CTM 2% models resulted in higher (ten-fold) bias values than the CTM 6% predictions. This may be a reflection of the higher variability in the laboratory hardness data.
All of the MLR models that were developed for CTM 6% tablets successfully predicted their corresponding validation sets. It is evident from Tables 10 and 11 that the model with the best SEE may not yield the best prediction results. The best models consisted of three or four terms, but the same equations did not yield the best validation results. Overall, PLS performed slightly better than MLR for CTM 6% samples.
Equations developed for combined CTM 2% and 6% samples followed the same rule as the combined HCTZ formulations (Table 12). The computer selected wavelengths for the combined formulations that were different from those forCTM 2% or CTM 6%.

.4.4.3 Results of"Mixed" Calibrations
"Mixed" calibrations were developed by performing regression on combined data from the four tablet formulations (Table 13). In the process of developing these calibration equations, the software defined a significant number of data points as outliers. The majority of the outlying points originated from the CTM 2% data. One PLS calibration was developed using the data from all four formulations (Table 14). This model only marginally fit both sets ofHCTZ validation data and the CTM 6% data. None of the developed models fit the CTM 2% data. The best performance was obtained from the Ill where the computer chooses a wavelength that is unrelated lo the constituent of interest.
The instrument evaluates the overall spectral variation-changing the hardness in a specific formulation has an overall effect on the spectra, which may vary between formulations.
Although the present study did not find a universal calibration equation for hardness, it was found that the NIR signal responded in the same way to a change in Absorbance changes at these wavelengths suggest changes in moisture content due to changes in tablet compaction forces. Since all four formulations were composed of the same excipient matrix, similarities in the general peak shape would be expected.

Results of Placebo CalibraJions
The results of MLR and PLS calibrations and validations for the placebo tablets are summarized in Table 15. Several multiple term models were developed using MLR. The best MLR model utilized three wavelengths and resulted in multiple correlation coefficient constructed from only four levels of hardness. Better performance could likely be achieved by including more data in the models.

Results of Density/NJR Calibration
Calibrations were also developed for the four formulations using the calculated tablet density versus NIR signal. The models themselves were much better than those developed using hardness data. One factor that might be expected to contribute to this improvement is the degree of variability in the laboratory measurements. Relative standard deviations (fable 16) in laboratory hardness measurements were 3 to 13% compared with errors in density of only 0.2 to l.0%. In evaluating the laboratory hardness data, a trend was observed in the percentage standard deviation: as hardness was decreased, percentage variation increased. This effect was not observed with the calculated tablet density data.
In order to evaluate the performance of the NIR determination of density, the NIR values were compared to the reference SE for density, i.e., the laboratory values. The overall reference SE for density was 6.67 x l Q-3 (giml). The calculation of the SE for density is dependent on the precision of the weight and thickness measurements. These measurements are generally very good, due to the quality of the tablet press tooling and the accuracy of the balance.
Numerous one and two-term density calibration models were developed for each of the four products and are summarized in Tables 17 to 20  Overall, PLS models also performed better in the validation process. Although the laboratory density values were less variable than the laboratory hardness values, calibrations produced from hardness data were more rugged than the density models. and were better at predicting hardness. Jn other words. the models based on hardness contained built-in variability which contributed to the good prediction ability of the hardness calibrations.

Conclusions
This work presents a viable and non-destructive alternative to hardness testing of tablets. A method was developed which offers the potential of 100% quality control testing for tablet hardness. There is a correlation between the hardness or compression force of a tablet and its NIR spectra. As tablet hardness increased, an upward shift in the raw NIR spectra was observed. This relationship was modeled by the development of formulation specific calibration equations for the determination of hardness via NIRS . The NIR method of hardness testing did not suffer from subjective differences in reading the results.
Because the method is non-destructive. the samples can be further tested or even packaged for sale after NIR testing. The use of the iris was important to maintain accurate and reproducible sampling technique.
Equations based on tablet density produced statistically improved models for NIR density determination. Predictions based on tablet density had slightly lower multiple correlation coefficients than predictions based on tablet hardness.
In applying multivariate regression techniques. we are assuming that there is an equation that will best fit all the data in a set. We also assume that a perfect fit of any model to all the data cannot be made because of random errors in the data. Thus. we end up with a list of potential equations that fit our criteria (none of them a perfect fit It may not be possible, or desirable to develop a single, global equation for the evaluation of hardness. Since hardness is a physical property for which there is no single analytical wavelength, PLS may be a more reliable approach to calibration. PLS models the entire spectra, not just data at specific wavelengths. The goal in developing a global calibration is •· ... to cover as broad a range of samples as possible while maintaining acceptable accuracy" .1 1 Unique equations that are developed for a particular product can secondarily act to identify or qualify the product. In the agricultural industry. a global calibration is one that is designed to analyze 90 to 95% of samples of a given product 1 2. Specific calibrations based on a small range of samples typically perform bener than general calibrations, provided the samples to be analyzed are represented in the calibration set.

6 Considerations for future work
The calibrations developed in this study covered a hardness range of 2 to 12 kg. Jn a manufacturing sening, a more realistic range of acceptable hardness values would be+/-20%, at most. In this situation, a slightly different approach to calibration would be required. Rather than manufacturing tablets covering a broad range of hardness values.
spectral data from several lots of a product would be collected over a period of time. An acceptable range of hardness would be identified for the product. and a calibration would be developed using the collected data.
Jn order to fully characterize the potential of NIRS as an alternative to conventional hardness testing. several tablet matrices should be evaluated. A comparison of tablets produced by direct compression versus wet granulation would be useful in the detennination of the extent of NIRS utility for hardness testing.
In the current study. second derivative data was used to develop the calibration models. Derivative spectroscopy is a powerful technique for magnifying the fine structure of spectral curves. The result is an enhancement of structure that is offset by a decrease in ( the signal-to-noise (SIN) ratio. The major advantage is an increase in resolution. which may be very useful for resolving bands that are too close to be resolved in their absorption spectrum. In the case of a physical property, such as tablet hardness, the use of the second derivative spectra may not be desirable since taking derivatives minimizes non-chemical composition effects, such as particle size. Since we are interested in a physical property. it seems logical to include such influences by using the untreated (raw) spectra.

Acknowledgments
The author gratefully acknowledges Perstorp Analytical/NIRSystems, Silver

. Abstract
The purpose of this study was to evaluate the effect of matrix and geometry on the detennination of tablet hardness via near-infrared reflectance spectroscopy (NIRS). A secondary objective was to evaluate any differences in NIR response due to scoring of a tablet. Flat-faced and convex (scored on one side) tablets were manufactured at five level s of compression force, using two excipient matrices. Blend# I consisted of chlorpheniramine maleate6%, magnesium stearate 0.5% and microcrystalline cellulose.
Multiple linear regression and partial least squares were used to develop calibration models using NIR spectral data and tablet hardness. Differences in these models enabled comparisons of matrix and geometry, as well as scoring. NIR absorbance data at several individual wavelengths were subject to linear regression and one-way analysis of variance in order to assess the potential of hardness prediction at one wavelength.
Formulation specific calibration models were developed for two tablet matrices and two geometries. NIRS calibration models successfully predicted tablet hardness for both matrices and both geometries. Absorbance (log I /R) values were higher for convex tablets than flat tablets. Scored tablets produced slightly more variable results than nonscored tablets . Tablets containing di basic calcium phosphate di hydrate produced more variable hardness and spectral results than those containing microcrystalline cellulose. Models developed for one fonnulation could not be used to predict hardness in other fonnulations.
Models developed for flat tablets could not be used to predict hardness for convex tablets.
The previously established hardness/NIR relationship was valid for other formulations. The CTM fortbeAvicel®/CTM blend was screened (20 mesh) prior to mixing.
The target weight for the CTM/Emcompress® /magnesium stearate tablets was 800 mg. The target weight for the CTM/Avicel®/magnesium stearate tablets was 540 mg.

132
After adjusting the tablet press for correct target weight, the hardness level was adjusted to achieve five different levels for each blend.
Target hardnesses of 2, 4, 6, 9, and 12 kg (as monitored by a Heberlein, vectortype Hardness Tester) were used for each combination of blend and geometry for a total of 20 different batches of tablets. The tablet press was adjusted to the desired hardness level, beginning with the lowest level , and monitored for constant tablet hardness. After achieving a constant hardness level, the next ten tablets were collected from each batch. in order, and stored in labeled Whirl-top® plastic bags. The instrumented tablet press recorded compression force data for the first six tablets of the ten collected from each batch.
Upper and lower compression forces (in kN, or kilonewtons) and ejection force data were recorded for each batch. This compression data was correlated to the first six reserved tablet samples, for later calibration with NIR data. The tablet press was allowed to continue compressing until approximately 200 tablets per batch were manufactured. The same process was repeated for the next hardness level.
Each batch of tablets was labeled according to geometry, matrix and hardness (low to high) level. Flat-faced Avicel®/CTM tablets were labeled AF! through AF5. Convex Avicel®/CTM tablets were labeled AC! through ACS. Flat-faced Emcompress®/CTM were labeled EFI through EF5, and convex Emcompress®'CTM tablets were labeled EC I through EC5.

2 Near-Infrared Spectroscopic Analysis
A Perstorp Analytical/NIRSystems Rapid Content Analyzer® Model 5000 was used for the analyses of tablet samples. This instrument and corresponding setup was described in Manuscript II.
NIR reflectance parameters were set at 32 scans per sample in the range of 1100 to 2500 nm . A ceramic (Coor' s Standard) reference scan was taken before each set of samples. Near-infrared reflectance measurements were made on the ten reserved tablets from each batch. The sample to be measured was placed directly above the detector surface and centered with the iris. Before positioning each sample, the detector surface was gently cleaned of debris. Flat faced (FF) tablets were scanned once on each side. Standard round convex tablets (SRC) were measured twice, alternately , on each side, and given spectral sample names corresponding to score and replicate number.

J .3.3 Tablet Evaluation
After completion of the NIR scans, the ten reserved tablets from each batch were subjected to weight, thickness and hardness testing using a Vector Systems-3 hardness tester, which was interfaced with a Mettler balance, Model AM.SO. The tablets were manually added to the hardness tester without regard to the orientation of the scoring on the SRC tablets. The order in which the instrument performed these tests was thickness, weight, and then tablet hardness.
The laboratory hardness and compression data were analyzed using Mini tab®

.3.4 Effect of Sample Position ojSRC Tablets on the NIR Spectra
Reproducibility of sample position is an important factor in NlR analysis. Flat faced, round tablets can be reproducibly positioned on the detector surface (this experiment was described in Manuscript Ill ) using the iris. However, one might suspect that SRC tablets may be subject to more variation in position since the point of contact between the tablet and the detector surface is so much smaller than that of a flat faced tablet. Scoring of 134 Emcompress®/CTM 6%/magnesium stearate SRC tablets (6 kg hardness) were chosen for this portion of the study. One tablet was placed in the center sampling position using the iris and scanned ten times on each side, removing and flipping the tablet over between scans. First the nonscored side of the tablet was scanned, and then the scored side was scanned. During the analysis of the scored side, an attempt was made to align the score at the same point for each scan.
Next, one tablet was placed in the center position using the iris. and scanned ten times on each side without removing it from the detector surface. The nonscored side was run first, followed by the scored side. In the third phase of the study , ten tablets were scanned once on each side in the center position , using the iris.
Finally, one tablet was scanned in ten positions on the detector surface. The methods for this study were previously described and illustrated in Manuscript III ( Figure I in section 1.3.4). Beginning with the center position, the tablet was moved to the 12:00 (twelve o' clock) position , then rotated to a position 45• from the previous spot. At each rotation , one edge of the tablet was touching the center position . The 12:00 position was repeated, as position number ten. The resulting spectra were entered into spectral libraries, where the mean and standard deviation spectra were extracted for each condition. Standard deviation spectra were plotted and compared. A separate calibration set was needed to test the second and third constituents, upper and lower compression force. This was because compression data was available for only six tablets from each batch. The same sample selection process was followed . except four spectra per batch were selected by the computer from the samples having associated compression data, and the remaining two spectra from each batch were used in the validation (prediction) set.

Near-Infrared Calibrarion of Hardness and Compression Force
Each subset of spectral data was converted to the second derivative. Standard multiple linear regression (MLR) and partial least squares regression (PLS) were performed on each calibration set. One to three wavelengths were used for each MLR model. PLS regression was limited to eight factors and the wavelength range was 1100 to 2500 nm .
Equations were developed from second derivative and "raw" (untreated) spectral data.
137 Validation sets for hardness were created using the remaining four spectra from each batch of tablets (for a total of 20 validation samples per formulation). Validation of the models was performed using the Percent Predict function ofNSAsr~. Equations that were developed for one calibration set were applied to the corresponding validation set (of the same math treatment), as well as validation sets from other blends, to test their fit.
The subsets of scored and nonscored data were compared chemometrically to detect spectral differences. Mean and standard deviation spectra for each subset were calculated using the JQ2"' function of NSAS. Standard deviation spectra of raw and second derivative data were compared to evaluate differences between scored and nonscored sides of the samples. Calibration equations for hardness were also developed for the scored and nonscored data, as described above.

1.4.l Resuirs ofTabler Evaiuarion
The results of the evaluation of physical tablet parameters for Avicel®/CTM FF and SRC tablets are summarized in Tables 3 and 4. Emcompress® /CTM FF and SRC tablet data are summarized in Tables 5 and 6. All hardness, thickness and weight evaluations were completed within one week of the date of tablet manufacture. Good reproducibility was achieved in all batches for tablet weight and thickness. Avicel® /CTM tablet weights varied from 0.54% to 1.84% (relative standard deviation , RSD) and thickness varied from 0.59% to 1.42%. Emcompress® /CTM tablet weights varied from 0.30% to 0.49% and thickness varied from 1.07% to 2.10%.
It was generally noted in all formulations that an increase in tablet hardness was associated with an increase in variability (RSD Hardness data from two batches of the Emcompress®/CTM SRC tablets were found to have equal means. Hardness level 4 (9.41 ± 0.97 kg) was deterrnined to be not significantly different (p== 0.05) from hardness level 5 (9.86 ± 1.29 kg). In contrast to this finding , the corresponding upper and lower compression force data did not follow the same pattern. The five levels of upper and lower compression force were found to be significantly different (p== 0.05) from one another in these batches.
A comparison of average hardness data was made between FF and SRC tablets at each hardness level for both forrnulations (  A comparison of average hardness means was also made between formulations of the same geometry and hardness level ( Table 9). The purpose of these comparisons was to evaluate the spectral differences between excipients (matrices) at one hardness level. Mean hardness values for AC and EC were found to be equal at hardness levels I, 2 and 3.
Mean hardness for AF at level 5 was found to be equal to EF at level 5 (p= 0.13).
Avicel®/CTM upper (UC) and lower (LC) compression force data is summarized in Table 10. Variability (% RSD) in compression data was generally from 2 to 5.7%. UC and LC data for Emcompress®/CTM tablets are summarized in Table l l.
In order to achieve the desired levels of tablet hardness, significantly greater compression forces were required to compress tablets containing Emcompress®. UC forces required for Avicel®/CTM batches ranged from l.8 to 5.1 kN , while those for Emcompress®/CTM tablets ranged from 6.7 to 22.7 kN. This was partially due to the fact that Emcompress® has a higher bulk density than Avicel®. Particle size distributions al so differ between the two excipients. Emcompress® bulk tapped density was reported to be 0.99 g/ml (per the manufacturer's certificate of analysis). Its density is 2.89 g/cm3. The average particle size is 120 to 150 µm. Emcompress® is a crystalline solid or powder, consisting of granules of which over 95% are less than 425 µm and less than I 5% are under 75 µm. It is not known to be hygroscopic5. Compaction of dibasic calcium phosphate takes place primarily by brittle fracture. Due to its abrasive nature, it is important to include a lubricant in the tablet formulation .
Avicel® PH 102 (microcrystalline cellulose) is a crystalline powder composed of porous particles. Its bulk density is 0.3 g/ml3 and its density is 1.55 g/cm3. Typical mean particle size is 20 to 200 µm. Particle size distribution consists of less than 8% of particles greater than 250 µm and over 45% of particles greater than 75µm . The moisture content of Avicel® is less than 5.0% but it is known to be hygroscopic6.

.4.2 Results of NIR Specrral Analysis
The effect of blending Emcompress® with CTM and magnesium stearate is demonstrated in Figure 1. This plot compares the raw NTR spectrum of Emcompress® powder (alone) with that of the blended formulation . The spectra have similar shapes except for the two small peaks between 2100 and 2300 nm on the blend spectrum.
In Figure 2, the raw spectrum of the Emcompress®/CTM/magnesium stearate blend is plotted with the average spectrum of the FF tablets (n= 10) at each of five hardness levels. (Note: unless otherwise indicated. all spectral plots represent the average of I 0 tablets). This plot demonstrates the effect of varying the compression force on the blend.
As the hardness (compression force) was increased. the absorbance value also increased.
The spectrum of the uncompressed powder blend shifted slightly between the spectra of the second and third hardness levels.   The difference between the flat and SRC Emcompress®/CTM tablets is shown in Figure 4. Here, the high and low hardness level tablets of each geometry type were plotted together. The two upper spectra compare flat to SRC tablets at hardness level 5. At hardness level 5, the mean hardness value for the SRC tablets was 9.86 ± \.29 kg, while that of the FF tablets was 145 ± 0.85 kg. Thus, due to the large difference in the hardness data at this level, a direct comparison of the spectra based on hardness at this level is not possible. The plot at hardness level 5 shows a higher absorbance for the FF tablets in comparison to the SRC tablets. At hardness level I, absorbance values are higher for the SRC tablets (hardness means were equal at level I}. Observation of the pattern of the spectra in Figure 4 reveals striking similarities between FF and SRC tablets. The chemical composition of the sample is responsible for the unique spectrum of the formulation. We can see that tablet hardness and geometry do not affect this uniqueness; instead, these physical changes are responsible for the drifting baseline and ''peak" shifts in the spectra.   Table 8 will be useful for reviewing Figures 7 through 11. As discussed earlier, the mean values for hardness for Emcompress®/ CTM tablets were equal for hardness levels I, 2, 3 and 4, and so direct comparisons could be made about the effect of tablet geometry. In each of these four plots, the absorbance of the SRC tablets was higher in comparison to the FF tablets. It is also noted that the FF and SRC spectra partially overlap in the 19<Xl nm region (water band). It would be expected to find the same concentration of water in the two batches, since they originated from the same blend. However, the maximum absorbance at about 1930 nm was higher in the SRC tablets for all but hardness level 3 spectra. In Figure 11. a direct comparison could not be made since the mean hardness values at level 5 were not equivalent. Figure 12 is the raw spectrum of Avicel® /CTM/magnesium stearate powder blend plotted with that of the FF Avicel® /CTM/magnesium stearate tablets (n= 10) at five hardness levels. In the region of 1100 nm to approximately 2100 nm, the absorbance of the powder blend spectrum is much lower than the absorbance of the FF compressed tablets. From 2100 nm to 2500 nm , the powder blend spectrum falls much closer to the tablet spectrum. In Figure 13 , a similar comparison is made between the powder blend and the compressed SRC tablets at five hardness levels. Again, the greatest amount of difference is seen from 1100 nm to about 2100 nm , where the powder blend spectrum is closer to the tablet spectrum.
In contrast to the observation of the deviation of powder blends from the tablet spectrum (deviation on the left side of the plots), the tablet spectra behaved in the opposite fashion. The tablet response to increasing compression force was greatest on the right side of the plots. In each of the plots comparing tablet spectra, the raw spectra were often indistinguishable in the 1100 nm to 1500 nm region. Increases in absorbance due to increasing compression force were not noted until 1500 nm.    The raw spectra of FF and SRC Avicel® /CTMJmagnesium stearate tablets are compared in Figure 14. The spectra of five hardness levels of SRC tablets were plotted against the spectra of five hardness levels of FF tablets. In the region from 1100 nm to 1500 run, each set of samples was practically overlayed. From approximately 1475 nm. the effects of increasing compression force are manifested in the spectra. The maximum absorbance of the SRC tablets was greater than that of the FF tablets. Figure 15 is a second derivative plot comparing the FF and SRC tablets at the highest and lowest hardness levels. The absorbance of the Avicel®/CTM SRC tablets was slightly greater than that of the FF tablets. The difference between A vicel®/CTM FF and SRC tablets was small in comparison to that seen with the Emcompress®/CTM tablets.
This expanded view (in the region from 1414 nm to 1646 run) of the derivitized spectrum further illustrates the effect of increasing hardness on the NIR spectrum. As previousl y mentioned, derivative treatment of spectral data enhances spectral features, but may decrease the signal to noise (SIN) ratio.

.4.3 Results of Sample Position Study
Emcompress®/CTM 6%/magnesium stearate SRC tablets (6 kg hardness) were used to evaluate the effect of various sampling positions on the NIR spectra. Examination of the raw spectra of each condition does not initially appear to reveal much information. Wavelength (nm) Figure 17. Raw NIR spectra of Erncompress®/CTM nonscored tablets I tablet scanned I 0 times with replacement versus not moving it between scans. Figure 18 compares the reflectance spectra (raw data) of one tablet scanned ten times on each side without moving it. The upper spectrum is the average of ten scans of the scored side of the tablet. The lowermost spectrum is the average of ten scans of the nonscored side of the tablet. Observation of the plot shows that the spectra were overlayed from 1100 to about 1450 nm, then began to diverge. Beyond 1900 nm, the difference between the two spectra was at its maximum. This plot demonstrates slightly higher absorbance values for the scored tablet data.
The standard deviation spectrum is a useful tool for extracting more information from the samples. When comparing second derivative data from the scored and nonscored sides of the tablet scanned ten times without moving it, there was so little difference that the computer could not create a standard deviation plot for either of these conditions.
Obviously, ten replicates of this type of sample (without moving it) would not be needed in a real calibration. Figure 19 compares the standard deviation spectrum (second derivative) of a scored versus nonscored tablet scanned ten times. There appeared to be more variation in the water band for the scored sides of the tablets; the rest of the nonscored spectra are in the same range as the scored plot. Figure 20 compares the raw spectra of ten scans of a scored side of a tablet versus ten scans of a nonscored side of a tablet. The spectrum of the scored side had a slightly higher absorbance than that of the nonscored side. Figure 21 compares the standard deviation spectrum of scored versus nonscored sides of ten tablets scanned once. Figure 22 displays the second derivative transformation of the same data. Again, the second derivative treatment reduced much of the apparent baseline shift caused presumably by sample placement. There was greater variability in the scored data. Figure 23 illustrates the difference between the raw spectra often scored tablets scanned once and one tablet scanned ten times. The average absorbance spectrum of ten .... scans of the same tablet was lower than that of ten tablets scanned once each. Figure 24 is the second derivative of the standard deviation spectrum comparing one scored tablet scanned ten times versus ten scored tablets scanned once each. The standard deviation spectrum of the scored tablets under these sampling conditions are essentially the same. Figure 25 shows the raw spectra of nonscored tablets, comparing one tablet scanned ten times versus ten tablets scanned once. The difference was not obvious to the naked eye. Figure 26 displays the standard deviation spectra of these conditions. Note the vastly greater difference in variation in the spectrum representing the average often tablets scanned once (upper spectrum). A second derivative plot of the same samples ( Figure 27) shows two very different looking spectra. Overall, there was greater variability in the average spectrum of ten tablets scanned once each. There is one region of overlap of the data in the 1900 to 1-930 nm range, which is due to water content.
The raw spectrum of one tablet scanned in ten positions on the detector surface is shown in Figure 28. The absorbances at 1100 nm ("origin" of the plot) range from -0.102 to 0.41. demonstrating a five-fold shift in response to a change in sample position. The naked eye cannot discern any real differences in the overall shape of the spectra. It is noteworthy that the spectra of the two replicates scanned in the 12:00 position (above the center) were quite close to each other. Also near the 12:00 spectra were the 135° and 315° samples. The spectra of the 45°, 180°, and 270° samples were close to each other, with the 225° sample not far above it. The sample from the center position was the lowermost spectrum in the plot, beginning at an absorbance of -0. I 02. The 90° sample originated at approximately 0.130.
The second derivative spectra of the tablet in ten positions are shown in Figure 29.
Much of the baseline shift was removed by the derivitization of the spectra. The spectra appear to be completely overlayed except for the region from about 1450 nm to 1600 nm. Wavelength (nm) Figure 30. En larged view of seco nd derivative NIR spectra of one nonscored tablet sca nned ten times.
absorbance maxima at 1513 nm, 1540 nm , and 1573 nm. It is not known why this region was so highly affected by sample position in comparison to the rest of the spectrum.
The raw spectra in Figure 28 followed a similar panem of presentation to those in Figure 4 , where increases in compression force caused the spectra to shift upwards . When sample position was changed, the spectra also shifted, although not in a regular, predictable manner. Changes in sample position are known to result in this apparent "particle size" effect. The difference is that the second derivative of data from Figure 4 were not completely overlayed at the absorbance maxima, as they were in Figure 29. This indicates that the apparent "particle size" effect from changing hardness could not be entirely removed by derivitizing the spectra.
In summary, the results of the sample position study demonstrate that variability in the spectral data may be introduced through slight changes in sample placement. This variability was greater in scored than in nonscored tablet data. Much of the variation could be reduced by using a second derivative math treatment. The center position was the most reproducible sampling position due to the ability to use the iris.

.4.4 Single wavelength regression and analysis of variance
During NIR analysis of these products, some degree of consistency was observed in the spectral response to increasing tablet hardness. For a given formulation , regular Second derivative absorbance values at the specified wavelengths were tabulated for each product (n=IO) at all five hardness levels. Linear regression and one-way analysis of variance were applied (Minitab®) to the hardness and spectral data to assess their relationship. Cricket Graph was used to plot the data at each wavelength. Since NIR is being used to predict tablet hardness from absorbance data, hardness becomes the dependent variable (y). Regression data are summarized in Table 12. A linear relationship was found to exist for each set of data; in most cases r2 was better than 0.92.
Graphic representation of the data provides useful information for comparative purposes. Data for flat versus SRC Emcompress®/CTM tablets at 1430, 1898, and 1926 nm are plotted in Figures 31 through 33   To further investigate the si ngle wavelength regression theory, NSAS was used to develop NIR calibrations for each product at each of the aforementioned wavelengths.
Models were developed for flat, scored, nonscored and mixed (scored/nonscored) tablet surfaces. It was surprising to discover that the results of these calibrations were quite good. Table 13 summarizes the regression coefficients for Avicel®/CTM tablets. R2 values ranged from 0.951 to 0.990. Calibration and prediction results for the    (Table 12) with NSAS results (Tables 13 and 15) reveals slight differences in regression coefficients. The primary reason for the differences was that the data sets for NSAS calibration were constructed from six spectra rather than ten (Minitab®) in order to facilitate the development of validation sets.
It should be reiterated that tablet hardness is a physical property, and thus a single wavelength cannot be assigned exclusively to hardness. The absorbance at 1926 nm is associated with water content, yet each formulation demonstrated a regular increase in absorbance at 1926 nm in response to increased hardness. Since this pattern of increased absorbance occurred at several wavelengths for each product, it may be concluded that one or more of these wavelengths could be used to calibrate tablet hardness for a particular formulation. Absorbance of interfering substances would have to be evaluated before choosing a single wavelength for this application (e.g., 1926 nm may not be a reliable single wavelength).

Calibration ojTabJet Hardness
The yardstick for comparison of calibration equations is the set of statistics that results from the application of the equation to a validation set. The statistical parameters used to evaluate calibration models were described in Manuscript Ill.
The standard error (SE) of the reference method must be known before undertaking a calibration experiment. The NIR calibration is only as good as the reference values from which it was generated. ln addition to the aforementioned selection criteria, the SEE of the calibration must be compared to that of the reference method. If the calibration SEE is better than the reference SE, the model may be overfilled and may not accurately predict unknown samples. For comparative purposes, the ratio of bias/standard error of the bias There were no outstanding features to note concerning calibration ofEmcompress® based tablets in comparison to Avicel® based tablets. The process was the same, and similar patterns of behavior were noted for both matrices.

.4.5.J Hardness Calibration o/SRC Avicel®!CTMTablets
General calibration equations were developed for SRC tablets using several different calibration, or training, sets. All four replicates of NIR tablet scans were averaged (AC data set) and subject to random sample selection for the calibration set. MLR and PLS regression were used to develop equations using raw and second derivative spectral data.
Improvement of the MLR models was obtained by regressing on more than one wavelength. ln some cases, the first regression wavelength. was preselected based on spectral information, or to provide a basis of comparison to the other groups of calibrations (e.g. , scored, nonscored, mixed). To summarize the calibration process for AC data sets, all of the AC calibrations were able to predict AC validation sets with a good degree of accuracy. None of the calibrations based on raw data were able to sufficiently predict AN or AS data. One second derivative PLS calibration was able to predict an AN (r2= 0.994) as well as an AS (r2=0.995) validation set. Two second derivative MLR equations adequately fit both AN and AS data from one validation set. ln general , a higher bias/SEB value resulted when an model was used to predict an AN or AS model. There did not appear to be a difference in performance between PLS models developed from raw data versus second derivative data.
MLR models were slightly better when second derivative data was used. AN calibrations using second derivative data were only marginally successful at fitting AS or AC validation sets. One equation fit the AC data, and two others were marginally acceptable. When applying AN calibrations to scored data sets (AS), two out of seven equations produced a fair prediction. Performance of the models appeared to be the same between raw and second derivative data for both PLS and MLR models.
Calibrations based on spectral data from scored (AS) tablets followed the same pattern as the nonscored (AN) tablets: equations developed specifically for AS data were able to fit AS validation sets but not AN or AC validation sets. Calibration and validation results are summarized in Table 19. Performance of raw data PLS models was slightly better than second derivative. Raw and second derivative MLR models were about equal in performance. None of the AF equations were able to predict AC data sets. There did not appear to be a significant difference in calibration performance using raw data versus second derivative data for either MLR or PLS models. Table 21 summarizes the results of calibrations developed for spectral data consisting of mixed AC and AF samples (ACAF). Lo general , it was possible to model the combined sets of Avicel® tablet data. PLS calibrations were better than MLR for these samples, and raw data produced slightly better results than second derivative data.

Hardness Calibration of Mixed FF & SRC Avice/®/CTMTab/ers
The calibrations performed best when applied to validation sets of mixed samples. When applied to sets of flat, scored or nonscored tablet data, a large bias value resulted.
Better calibrations resulted from data sets that excluded AF hardness levels 2 and 4.
These two sample sets were very close to the " adjacent" sample sets in spectral appearance.
These findings provide valuable information about the suitability of the calibration sample set.
The effect of increasing the number of regression wavelengths can be observed in the first three MLR second derivative calibrations of    Flat vs SRC Avicel / CTM (ami xpls2 ) 4 Two MLR calibrations were created for scored tablets using raw data. The wavelengths used were the same as those in the nonscored data, with the same results. The two-term equation was better than the one term equation and SEP values were in the same range as the reference SE.
One PLS calibration was developed for scored tablets using second derivative data.
The SEE was 0 .854 and the SEP was 0.567 for scored tablets. This model predicted nonscored and mixed (scored with nonscored) nearly as well as it predicted scored data.
This result indicates that the PLS model is not precise enough to discriminate the different data sets. 1.4.5.5 Hardness Calibration of FF Emcompress®ICTMTablets The second section of Table 25 contains MLR models developed from second derivative (2d) data. These were based on one, three, and two terms, respectively. The three-term model produced the best SEE (0.583 kg) and SEP (0.440 kg).
PLS models follow the MLR results in Table 25. The raw data PLS model uti lized one factor and produced acceptable results (SEE= 0.700 kg, SEP= 0366 kg). The second derivative models varied in the number of factors, which were one and three, respectively. The three-factor model had a better SEE (0.680 kg) than the one-factor model (SEE= 0.800) and a better SEP (from 0.473 to 0.415 kg). It appeared that PLS and MLR models performed equally well, and second derivative models were slightly better than raw data models (bias/SEB ratios were lower for second derivative models).

Hardness Calibration of Mixed FF & SRC Emcompress®/CFMTablets
Numerous calibration equations were generated for Emcompress®/CTM tablets of mixed geometry (ECEF). Results of hardness calibration and validation for ECEF data are summarized in Table 26. Raw data and second derivative data performed equally well for ECEF data sets. All of the MLR and PLS models adequately fit validation sets of mixed EC and EF spectra. The addition of a second or more regression wavelengths improved the models. This point is illustrated by calibration files ecef-p3, ecef-p4, ecef-c5 and ecef-c7 in

Calibration of Compression Force
Upper and lower punch compression forces were calibrated against NIR data for flat faced tablets using both raw and second derivative data. As with hardness calibrations, the addition of a second regression term tended to improve a MLR model. Tablet matrix appeared to be a factor in the success of the calibrations, as it was possible in this study to calibrateAvicet® /CTM but not Emcompress® /CTM tablets.

1.4.6.J Compression Calibration of FF Avicel®!CTMTab/ets
The upper compression force (UC) calibration and validation results for A vice!® /CTM tablets are summarized in Table Tl derivative data, and in the hundreds for raw data models. Two PLS models were developed for each type of data, using the full spectral range for one (1100 to 2500 nm) and a restricted range for the other model . Reducing the wavelength range improved the calibration statistics but did not improve prediction performance. Both of the raw data calibrations made adequate predictions of the validation set. Comparisons of raw data versus second derivative models were inconclusive, since there were only two models for each data type. Due to the small sample size (n=20) and high amount of variability, the calibration of Em compress® /CTM flat tablets was deemed unsatisfactory. The process might be improved by using a larger calibration set and excluding the EFl tablet data. It bas been shown in other studies 7 that a greater matrix variability requires that many more samples be included in order to have a valid calibration. It is evident from the Avicel® /CTM results that compression force can be modeled using NIRS , however improper selection of the calibration set can produce inadequate results.

7 Conclusions
These results describe the utility of the hardness/NIR relationships for different formulations. The NIR/hardness relationship established in Manuscript II still holds true for other matrices: an increase in tablet compression force resulted in an increased NIR absorbance. There is a matrix effect involved in the NIR calibration of hardness and compression force. Calibration models must be formulation (matrix) specific.
Tablet geometry also bas an effect on the NIR calibration of hardness. CaHbration models designed specifically for one geometry type did not fit another shape of tablet surface. Raw data models bad smaller regression coefficients than second derivative models, suggesting an increase in the amount of noise in the second derivative models.
Mixed geometry models gave variable results, supporting the assertion that calibration models should contain samples of homogeneous composition.
The NIR spectra of scored sides of tablets contained more variability between replicates (higher standard deviation) than the nonscored sides of tablets. This study did not find a statistical difference in NIR response to hardness due to scoring of Erncompress®/CTM tablets, possibly due to the higher variability in reference and spectral data versus tbeAvicel®/CTM data. There was a significant difference in NIR response due to scoring of the Avicel® /CTM tablets. This portion of the study suggests that separate calibrations should be developed for scored and nonscored tablets.
This document described the determination of several tablet properties and subsequent caHbration using NIRS. Further work may lead to an explanation of the mechanism behind changing NIR absorbance in response to increased tablet hardness.
Specific studies may be undertaken to determine which features of the tablet surface are most responsible for the change in NIR response to increased compression force.

Addendum to Manuscript IV
The following observations were made with regard to the results of Manuscripts Ill and IV.

Data from two tablet presses
There was a difference in the NIR spectra between two batches of CTM 6% tablets manufactured with two different tablet presses. One batch was produced at the University of Rhode Island (URI), using one station of a sixteen-station Stokes Rotary Tablet Press.
The second batch was made at Pfizer Central Research, using a Korsch six-station tablet press. The same raw materials were used in both blends. The same NlR spectrophotometer was used to obtain the reflectance spectra of all of the samples. When the spectra were plotted together, there was one region (Figure I a) where the spectra were different; the plot was essentially the same for the remainder of the spectra.
This finding has implications in the validation process of a NIR hardness testing method. The spectral variation may be due to inhomogeneity of the blend or differences in the concentration of one or more of the components. As the URJ press was not instrumented, upper and lower punch data were not available for comparison to the Pfizer press data. In any case, a good NIR calibration set for this product would include this type of variation if a company produced these tablets at several manufacturing sites.

Data from two hardness testers
The overall standard error of the Pfizer hardness tester was 0.50 kg, as compared to 032 kg for the Erweka hardness tester at URI . This difference may be attributed to the number and type of samples available for calculation. The error calculation for the Erweka Hardness Tester was based on sampling 20 flat faced, nonscored tablets from 20 batches of 300. All of these samples were of the same excipient matrix, so the calculation was based on the testing of 400 similar samples.
The calculation of the overall standard error of the hardness tester at Pfizer was based on a combination of flat faced and standard round convex tablets (with one scored side) and two different excipient matrices. Ten tablets from each batch of200 tablets were included in the calculation, for a total of 200 samples tested. Hardness testing of the flat tablets resulted in a much lower (0.41 kg) standard error than convex tablets (0.60 kg).
This difference should be considered when calibrating hardness with NIRS. The standard error for all batches containing Avicel® was 0.52 kg, while that for Emcompress® was 0.49 kg. This finding was unexpected, since NIR spectral data for the Avicel® -containing tablets appeared to be less variable than the Emcompress® tablet data.

SUMMARY OF CONCLUSIONS I)
Near-infrared (NIR) spectrophotometric methods are becoming increasingly utilized by the pharmaceutical industry. However, the appearance of NIR methods in the pharmaceutical literature has been sparse, and indicates the reluctance of the industry to accept the technology.

2)
Although the potential of NIR to predict tablet parameters has been suggested, the published literature appears to have little data in this area. The present work appears to be the first study to evaluate the practical utility of the NIR/compression force relationship.

3)
Feasibility studies using chlorpheniramine/Avicel and bydrochlorothiazide/Avicel tablets indicated that a relatively simple relationship may exist between NIR absorbance and tablet hardness. An increase in tablet compression force resulted in a consistent increase in NIR absorbance.

4)
Chlorphenirarnine/Avicel and hydrochlorothiazide/Avicel tablets of increasing hardness levels were successfully differentiated by principal component analysis. This result indicates the potential use of NIR as a useful and non-destructive, quality control mechanism. The NIR signal responded in the same way to a change in hardness, regardless of the drug.

5)
A series of formulation specific equations was developed by calibrating tablet hardness data against NIR reflectance response (absorbance) for each formulation. The results of NIRS hardness prediction were at least as precise as the reference hardness test (SE= 0.32).

6)
Calibration models developed using the calculated tablet density were better than those from hardness data. This was probably due to the accuracy of weight and thickness measurements and the high quality tablet press tooling. However, the predictive ability (SEP) of the hardness models was better than the density models. These results indicate that the hardness models contained built in variability and were thus more rugged.
Conversely, the density models may have been overfitted.

7)
Separate calibration models are required for scored versus nonscored tablets, regardless of the tablet matrix. Mixed models consisting of scored and nonscored tablets adequately predicted mixed validation sets, but better predictions resulted when the calibration and validation sets contained only scored or nonscored tablets.

8)
NlR spectra are different for tablets of different excipient matrices, thus separate calibration models are required.

9)
Tablets composed of the same excipient matrix required slightly different calibration equations when the geometry was changed. Mixed models, composed of both nat and convex tablets, satisfactorily predicted mixed validation sets, but better predictions resulted when the calibration and validation sets contained tablets of only one type of geometry.

10)
It may be possible to predict tablet hardness using a NlR model based on only one wavelength. Good linearity and SEP values were obtained from simple linear regression equations from one or two wavelengths.
NIR calibrations for tablet hardness performed best when formulation specific models were constructed. Ruggedness was improved when all expected types of variability were included in the model. Linearity and SEE were improved when the calibration and validation sets contained a single type of product.

12)
The key factor in calibration development is the selection of the calibration set. It was illustrated in this work that careful selection of representative samples is imperative to the successful performance of the calibration model.

13)
This project demonstrated that it is possible to calibrate various tablet parameters with NIR absorbance data and achieve successful predictions of those parameters. NIR models for tablet hardness, density , upper and lower compression force data were developed using several different tablet matrices.

Prediction Stability Coefficient
The average difference between the calculated and reported results.
The standard error expressing the confidence interval of the reported bias.
The bias corrected estimate of random errors.
A non-bias corrected estimate of random errors. RMS will be equal to the SOD when the bias is zero.
Factor by which the slope and bias terms are multiplied to adjust the existing equation to fit the current data. 1.00 = no adjustment.
Estimate of the error on the computed slope.
Adjustment to be made to the intercept term of the calibration when a slope adjustment is made. 0 is equivalent to no adjustment.
Estimate of the error in the computed values using a slope and bias corrected equation.
Correlation between the calculated and reported values using the slope and bias corrected equation.
The best estimate of the achievable standard error of performance using the available data from each group.
Ratio of the achievable SEP to the actual root mean square deviation (SEP). Thus, the achievable is more often less than the actual SEP. Hardness calibration results for chlorpheniramine maleate (CTM), hydrochlorothiazide (HCTZ) and placebo tablets in Manuscript Ill.  The following tables contain density calibration coefficients and other data from Manuscript Ill.