(424e) Overcoming the Challenges of Noisy Data during the Calibration of NIR As a PAT Technique to Measure Roller Compacted Ribbon Density

 

INTRODUCTION

 

                A number
of research groups have investigated the use of NIR spectroscopy for monitoring
ribbon density [1-4]. These studies have taken advantage of the fact that NIR spectra
are affected by the density of the material being analysed and offer the
potential of measuring/monitoring roller compacted ribbon density in real time [1-3, 5]. In order to develop a calibration equation relating density to the
NIR spectra Gupta [2] fitted a regression
line through all the scanned wavelengths and used the fitted slope of the line.
However, this approach does not take into account issues which arise when there
are multiple scans on multiple samples of the same type of material, and each
scan records simultaneous NIR spectra for several probes. The spectral dataset
can contain erroneous outliers. Data processing in real time is challenging as
sampling error outliers must be either filtered or cleaned from the data set
before further analysis. [6]. Visual inspection
of the data and removal of sampling error outliers though justifiable is not
practical for real time data processing. Existing data cleaning and filtering
models are complex and are essentially off-line operations [6]. A simple solution
is required.

This study aims
to highlight how variability in roller compacted ribbon quality can impact on
NIR spectral measurement and proposes a simple method of data analysis to deal
with this.

 
Materials and Methods

 

Roller
compaction: A Fitzpatrick
CCS220 roller compactor was used for this study. The roll speed was kept
constant at 2 rpm, roll gap at 0.5 mm and the roll pressure was adjusted for
each run to 2, 4, 6, 8, 10, 12, 14 and 16 kN. Blends of microcrystalline
cellulose, lactose monohydrate and magnesium stearate were compacted. Each of
the three blends differed in MCC % moisture content. MCC was stored at 43% for
blend 1, blend 2 at 11% and blend 3 at 75% relative humidity. The variation in
moisture was introduced to assess its impact on the ribbon quality and thus
challenge the NIR technique. Ribbon envelope density was determined by the GeoPyc™
1360.

NIR spectral
collection: Four MultiEye^TM NIR probes
scanned samples at-line in the manner in which they were produced by the roller
compactor i.e. samples were not cut to a specific size. This set up ensured
outliers were detected in a similar manner to which they would be detected
in-line when monitoring the process. The Perkin Elmer Spotlight 400 FT-IR was
also used and the off-line results were compared to the at-line MultiEye™ data.
Samples were cut to a specific size to facilitate off-line Perkin Elmer Spotlight
analysis.

Statistical
data analysis: Three
different calibration methods were developed between spectral slopes and
envelope density and compared; (1) using the entire data set unfiltered, (2) a
visual discard method and finally (3) a 33% Trimmed Mean method.

 

 

 

RESULTS
AND Discussion

 

        Calibration
of the NIR method to ribbon density was challenging due to the presence of
spectral data from broken, split and curved ribbons in the data set
particularly for blend 3 which had the highest moisture content (6.96 % w/w). 
All initial processing of the full data set indicated that some form of data
cleaning was required in order to process the MultiEye™ at-line data set. The
visual discard method was applied to the full data set and was found to be
particularly successful for blends 1 and 2 (Figure 1). After using the
visual discard data cleaning method MulitEye™ data the calibration correlation
was found to be as good as the off-line higher scan resolution Perkin Elmer
spectrometer. Blend 1 off line r=0.92, visual discard r=0.95.

However it is
necessary to replace the visual method of spectra cleaning with a non-subjective
method which was capable of screening for these erroneous probe readings. The trimmed
mean eliminates a specified percentage of the data from both the high and the
low ends of the dataset before evaluating the standard mean of the remaining
data. For this data set the trimmed mean method sets a limit on how data is
cleaned from the data set allowing for the removal of a faulty probe reading
(25% of data) or a poor sample (33% of data). The 33% Trim Mean in place of the
Full Mean reduced the impact of spectral variation or misreads between samples
or probes. The Trim Mean method was found to be at least as successful as the
Visual Discard method at cleaning the data set prior to development of the
calibration equation (Table 1 and Figure 2). The Trim Mean method was preferred
as the decision on data to include in the calibration model as it is not
subjective and simple to apply. Major variation is apparent in the spectral data
for Blend 3 (75%) which renders it unsuitable for prediction and calibration. These
processing variations were attributes to the high moisture of MCC in this
compacted blend.

 

CONCLUSIONS

 

        The 33% Trim Mean method offers a simple and practical solution to
dealing with the NIR spectral challenges when applying this PAT technique to
roller compacted ribbon in-line/at-line. Further development of this method
when used for calibration could investigate the optimal number of ribbon
samples to % data trimmed to achieve a statically robust balance of data
cleaning. Increasing the number of samples used to during calibration may allow
for a greater retention of data in the TrimMean e.g. 5 samples and keeping 3
giving a 20% Trim Mean. The Trim Mean method is therefore no less successful
then other simple calibration methods but offers the advantage of a simple,
easy to follow and non-biased means of removing NIR data collected due to probe
sampling error at-line/in-line.

 

ACKNOWLEDGMENTS

 

        Research is
funded by the Synthesis and Solid State Pharmaceutical Centre (SSPC) under
grant number 12/RC/2275, the Mathematics Applications Consortium for Science
and Industry funded by Science Foundation Ireland Investigator Award 12/IA/1683
and the Pharmaceutical Manufacturing Technology Centre, Ireland. Microcrystalline
cellulose PH102 donated by FMC Corporation. Innopharma Laboratories for use of
the MultiEye^TM NIR Spectrometer.

 

REFERENCES

1.             Acevedo, D., et al., Evaluation
of Three Approaches for Real-Time Monitoring of Roller Compaction with
Near-Infrared Spectroscopy. AAPS PharmSciTech, 2012. 13(3): p. 1005-1012.

2.             Gupta, A., et al., Nondestructive
measurements of the compact strength and the particle-size distribution after
milling of roller compacted powders by near-infrared spectroscopy. J.
Pharm. Sci., 2004. 93(4): p. 1047-1053.

3.             Gupta, A., et al., Influence
of ambient moisture on the compaction behavior of microcrystalline cellulose
powder undergoing uni-axial compression and roller-compaction: A comparative
study using near-infrared spectroscopy. J. Pharm. Sci., 2005. 94(10):
p. 2301-2313.

4.             Lim, H., et al., Assessment
of the critical factors affecting the porosity of roller compacted ribbons and
the feasibility of using NIR chemical imaging to evaluate the porosity
distribution. Int. J. Pharm., 2011. 410(1-2): p. 1-8.

5.             McAuliffe, M., et al., The
Use of PAT and Off-line Methods for Monitoring of Roller Compacted Ribbon and
Granule Properties with a View to Continuous Processing. Organic Process
Research & Development, 2014. 19(1): p. 158-166.

6.             Liu, H., S. Shah, and W. Jiang,
Online outlier detection and data cleaning. Comput. Chem. Eng., 2004. 28(9):
p. 1635-1647.

 

Figure 1
Left panel - Representative spectral data for one piece of ribbon as scanned by
the MultiEye™ NIR probes at-line.  Probe 1 (continuous line), Probe 2 (dashed line),
Probe 3 (dotted line) and Probe 4 (dashed line in lower absorbance region),
each of the 5 scans for each probe is seen as a different color i.e. red,
yellow, blue, green and purple. When applying the visual discard method to this
spectral data set, probe 4 was removed from the data set. Right panel –Representative
spectral data for ribbon samples from the same batch as measured by Perkin
Elmer Spotlight off-line.

Table 1. Calibration models and the
respective correlation
coefficients (r) for each of the three blends as
calculated by the three different data discard methods. ED indicates envelope
density.

Calibration Method	Blend	Calibration model	r
Trim Mean	Blend 1	ED=0.416 + 3584 x Average Spectral Slope	0.96
Visual Discard	Blend 1	ED=0.409 + 3628 x Average Spectral Slope	0.95
Full Mean	Blend 1	ED=0.420 + 3788 x Average Spectral Slope	0.93
Trim Mean	Blend 2	ED=0.319 + 3788 x Average Spectral Slope	0.97
Visual Discard	Blend 2	ED=0.170 + 4257 x Average Spectral Slope	0.88
Full Mean	Blend 2	ED=-0.041 + 5587 x Average Spectral Slope	0.89
Trim Mean	Blend 3	ED=0.465 + 3021 x Average Spectral Slope	0.74
Visual Discard	Blend 3	ED=0.734 + 1515 x Average Spectral Slope	0.92
Full Mean	Blend 3	ED=0.474 + 3115 x Average Spectral Slope	0.74

Figure 2. Regression of NIR spectral
slopes on Envelope Density. Note: symbols 1=Trim Mean samples, 2= Visual
Discard samples and 3 = Full Mean samples. Solid line is fitted line for Trim
Mean samples, dotted line is for Visual Discard samples.