OPERACIÓN DEL SISTEMA EN MODO DE FALLA DE SENSORES

Squares Regression (PLSR) of Near Infrared

and Reference Analysis Measurements

4 .1 Introduction

In Chapter 3, MSPC procedures were applied to PC scores of NIR spectral

measurements of blends and tablets. The ability of this ‘model-free’ approach to process control and monitoring was determined by comparison of MSPC results with reference analytical measurements. In this chapter, the multivariate methods known as partial least squares regression (PLSR) and multiblock* PLSR are examined. These methods maximise the covariance between NIR spectral measurements and reference analytical measurements and therefore produce latent vectors which are most closely related to the reference analytical values. The PLS scores produced may be monitored in the same manner as described in Chapter 3 for PC scores.

In Sections 4.3, singleblock^ PLS models of blends and tablets respectively, are

subjected to MSPC. The predictive abilities of these models are compared with those of Chapter 3.

Section 4.4 models the entire process by multiblock PLSR. Conclusions regarding these methods of process monitoring are discussed in Section 4.6.

multiblock data sets are a collection of three-way data sets of process data at each process stage. ' singleblock data sets are three-way data sets of process data of one process stage.

4. 2 Near Infrared And Reference Analysis Data Sets Used

In this study, the NIR spectral and reference analysis data sets (Section 3.4) used were those of Section 3.6.3 for multiway PC A of the process for the two strengths of tablet (Tables 3.4 and 3.5).

The number of batches of blends and tablets used in the manufacture of the lower strength tablet was therefore 39 (Table 3.6). For the higher strength tablet process, 41 batches of blends and their corresponding tablets were examined (Table 3.6).

With a few batches produced, some reference analysis data were missing.

4. 2 .1 Data Analysis And Pre-treatment

The spectral data were analysed using code programmed in Matlab 5.2 Scientific and Technical Programming Language (The Mathworks Inc., Natick, MA, USA). A number of different pre-treatments of blend and tablet data were examined. The data pre­

treatments examined were those which were previously shown to produce the best multivariate principal components analysis process models for such data (Section 3.11). These were:

1. Raw spectral data (absorbance/transmission); 2. SNV-DT;

3. S g2dll.

These 2 pre-treatments were applied to blend absorbance data, and absorbance and transmission data for each of the two strengths of tablet. This provided 18 data sets (including multiway data sets of blend and appended tablet absorbance and transmission data) to be studied via projection to latent structures and multiblock projection to latent structures.

4. 3 Statistical Quality Control of Pharmaceutical Blends And Tailets by Single

Block PLSR

The method of PLS summarises the important variability in both the process (NIR spectral data) (X) and the final product quality data (Y) (Morud, 1996). This procedure projects the information in the high-dimensional data spaces (X, V) dowi onto low­ dimensional spaces defined by a small number of latent variables. The NR (X) and final product quality (Y) data sets were first mean-centred and scaled to mit variance and then decomposed according to equations (1.7.18) and (1.7.19) respectively.

The number of PLS components required to extract the information fromX and Y was judged to be 6 components for all models. This number of components vas selected as it modelled a considerable amount of the Y data and was also the numbeiof

components used with most principal components analysis model in Chapter 3. An advantage of this algorithm is its ability to handle missing data.

4. 3. 1 Singleblock PLS Model Variability

Singleblock PLS models (Wangen and Kowalski, 1988) were created forraw and pre­ treated spectral data sets (SNV DT, 11 point quadratic Savitzky-Golay snoothed 2"^ derivative) for the blends used to produce lower strength tablets {n = 39 batches) and for combined absorbance and transmission measurements of the lower strength tablets {n = 39 batches). The models were calculated using the average spectrum of tlose recorded for each batch. The average spectrum of each batch was used instead of several

measurements to eliminate systematic variability within measurements of the same batch introduced by particle size effects and differences in scatter from tie surfaces of the glass vials and tablets and also because average measurements tend tc follow a normal distribution.

Single block PLS models were also created for raw and pre-treated NIR data sets of blends used to produce higher strength tablets (« = 41 batches) and of combined NIR absorbance and transmission measurements of higher strength tablets (« = 41 batches) using the average spectrum of each batch. This produced a total of 12 models of blends and tablets for each strength and for raw and pre-treated data sets which were monitored subsequently.

For each of the singleblock PLS models 6 components were considered to be an appropriate rank for the models. The decision to use this size of model was based on previous experience of multivariate PCA projection of these data and because this rank explained most variability within the NIR data sets and significant amounts of

variability within the certificate of analysis data sets (Appendix C: Tables C l to C4). With raw data sets for blend and tablet models, the high amount of variance explained, typically above 99.6%, is due to multiplicative scatter within the data sets which accounts for most variability in the spectra. The amount of variance accounted for by these models of the certificate of analysis data was therefore not surprisingly lower: 37 and 53% for lower strength blend and tablet models respectively and 27 and 29% for higher strength blend and tablet models respectively. The lower variability accounted for in the certificate of analysis data probably arises from the fact that these data do not contain any particle size information.

With the SNV DT singleblock PLS models, slightly lower variability in the NIR data sets is accounted for by the models (Appendix C: Tables C l to C4) due to scatter correction. The 11 point Savitzky-Golay 2"^ derivative transformation effectively removed much of the multiple scatter information from the NIR data sets and produced single block PLS which accounted for similar amounts of variability for both the NIR and certificate of analysis data sets (Appendix C: Tables C l to C4).

4. 3. 2 Quantitative Calibration o f Individual Certificate o f Analysis Blend and Tablet Variables by Partial Least Squares Regression

Multivariate projection methods for monitoring process operating performance of multivariate processes have been shown to work well where all process data and product quality data are monitored (MacGregor et al, 1994). This is because the variability within and between process and product quality variables is required to model the process (MacGregor et al, 1994). In this study, the ability to produce quantitative models of the individual certificate of analysis variables was investigated. Only low amounts of variability could be modelled for some of the individual blend and tablet variables and was not considered accurate enough for future prediction of

individual reference analysis variables (Appendix C: Tables C7 & C8). The low variance within each variable is likely to be the reason for this.

4. 3. 3 Singleblock PLS Loadings

The loadings of the single block PLS models appeared to represent physical and chemical information. However, their precise interpretation was difficult, especially with the combined tablet absorbance and transmission data.

4. 3. 4 Q Statistic Monitoring o f Unusual Batches

The performance of this control chart was determined for each model type and for the different data pre-treatments by comparison of results above the 99% significance level with the results from the certificate of analysis reference data {i.e. significant Q values and corresponding unusual reference analysis values). In particular, the ability of this chart to identify anomalies in the process in batches preceding those which produced lower quality product and failed reference laboratory tests, was examined as this would be useful for detecting trends in the process over time. PLS models for each process

stage (blend and tablet) were created for the different pre-treated data sets in a recursive fashion, with batches whose Q statistic exceeded the 99% significance level excluded from the model. Each PLS model had 6 components

Singleblock PLS Models

With the lower strength tablet data set, raw data showed batches 22 to 26 as having processed unusually at the blend stage (Appendix C: Table C9). Of these batches,

batches 23 (BN5039), 24 (BN5039 re-blend) were re-blends of a blend which was found to exhibit excessive moisture deviation throughout (moisture deviation for batch 23 = 5.94%, limit: < 5%). With the SNV DT and Savitzky-Golay 2"^ derivative blend data sets, batch 25 (BN5040) was found to have processed unusually despite normal

certificate of analysis data (Appendix C: Table C9). With the SNV DT data set, batches 34 (BN5065), 35 (BN5067) and 36 (BN5075) were found to have processed unusually at the blend stage (Appendix C: Table C9) but had reference laboratory results within limits. The PLS models for lower strength tablets showed some agreement with this result: batches 36 and 37 (BN5076) were outside the 99% limit with the SNV DT and Savitzky-Golay 2"^ derivative models; batch 32 (BN5062) was outside the control limit with raw and SNV DT models and batch 33 (BN5064) was outside the limit for raw data (Appendix C: Table CIO). The certificate of analysis results showed that batch 33 produced tablets which were friable (1 mg). Reference data for batches 34 and 35 were missing. For batch 37 the drug substance content was found to be low (4.83 mg/tablet, range: 4.85 to 5.15 mg/tablet). With the Savitzky-Golay tablet model, batches 26 (BN5041), 27 (BN5042) and 28 (BN5055) were found to have processed unusually (Appendix C: Table CIO). Their certificate of analysis data revealed that batches 26 and 27 showed friability of 1 mg and 2 mg respectively.

Clearly, these control charts were able to detect unusual process behaviour at the blend

and tablet stages with consistency between results for the two stages. Raw data and SNV DT data sets were able to detect process anomalies at the blend stage which were not detected by the reference analytical data.

With the higher strength tablet data set, batches 35 (BN5071), 36 (BN5073), 37

(BN5074), 38 (BN5078), and 39 (BN5079) were all found to exceed the 99% limit with the SNV DT blend data set (Appendix C: Table C l 1). The certificate of analysis data for the blends of these batches were all within limits. However, the reference analytical data showed that batches 35, 37 and 38 produced tablets which exhibited friability of 1 mg, 4 mg and 1 mg respectively. In addition, batches 35, 38 and 39 also had average tablet thicknesses which were above the limit of 4.6 mm (all 4.62 mm). Batches 32 (BN5068), 35 and 38 were outside the 99% limit for the Q statistic with the Savitzky- Golay 2"^ derivative blend data (Appendix C: Table C l 1) and batches 33 (BN5069), 34 (BN5070) and 39 were outside the 99% limit with the blend raw data (Appendix C: Table C l 1). Batches 32 and 33 did not produce tablets which were friable, but the average tablet thicknesses for these batches were 4.62 mm and 4.61 mm respectively - outside the limit. Batch 34 produced tablets which showed longer than normal

disintegration time (17 seconds) and had an average tablet thickness of 4.66 mm, which is above the limit. The tablet Q statistic control charts which performed best were those for raw and Savitzky-Golay 2"^ derivative data. These identified batches 35 and 38 and batches 35, 37 and 39 as unusual respectively (Appendix C: Table C l2). With the Savitzky-Golay 2"^ derivative tablet PLS model, batches 27 (BN5053) and 28 (BN5054) were identified as unusual (Appendix C: Table C l2). These batches were both friable (4 mg and 1 mg respectively) and batch 28 had an average tablet thickness of 4.63 mm. With the higher strength tablet raw data PLS model, batches 28 (BN5054), 29 (BN5058), 30 (BN5059) and 31 (BN5060) were identified as unusual (Appendix C: Table C12). Batches 29 to 31 had average tablet thicknesses above the limit of 4.6 mm.

The Q statistic control charts for the higher strength tablet process were also able to identify batches and groups of consecutive batches which differed from the normal data set at both the blend and tablet stages. The manufacture of these batches ultimately produced tablets of lower quality. This deviation from normal process operating performance was identifiable from blend data despite showing no unusual reference analytical results at that stage. Overall, the SNV DT data set showed the best

performance for the higher strength tablet process of the pre-treatments tested.

4. 3. 5 M SPC o f Singleblock PLS Models o f Blends A n d Tablets

The PLS scores of the single block models corresponding to the latent vectors of the NIR data were used for MSPC monitoring. Estimation of the control phase 1 batches was performed by a Monte Carlo simulation. This involved random selection of the scores of 8 or 9 batches and construction of 99% Hotelling’s f^ control ellipses as described in Chapter 3. The Hotelling’s distance was then measured for the scores of the remaining batches from this ellipse and batches which had significant values were recorded. This process was repeated 200 times to produce a frequency bar chart which showed the frequency that any batch had been found to be significantly different from the control group. Batches which had a frequency greater than zero were not used for estimation of the process variance-covariance matrix and process mean vector. All batches were then monitored in control phase 2 using those batches which were deemed to be in-control {i.e. Monte Carlo bar chart frequency = 0) as the control phase 1 group.

Hotelling^s Control Phase 1 For Blends A nd Tablets

With raw data for the blends of lower strength tablet process and higher strength tablet process batches, the Monte Carlo search was unable to identify any unusual batches (Appendix C: Tables C17 and CIS). Most batches were therefore considered by the

algorithm to be in-control {n - 39 batches of lower strength tablet process blends, « = 36 batches of higher strength tablet process blends). For both processes, examination of the scores revealed that they were evenly divided into two clusters. This was found to be due to a difference in offset in the original blend absorbance data (Fig. 4.1), probably arising from different particle size distributions and porosity. With the PLS models produced from raw tablet data, these distinct clusters were not observed on the score plots, with some batches identified as unusual at control phase 1. The results of the tablet models produced from raw data for both strengths of tablet showed agreement with results from the Q statistic control charts (Appendix C: Tables C19 & C20), hence raw data was not considered useful for monitoring the blends and spectral scatter correction was considered necessary. With the SNV DT transformation, 17 and 31 batches were used in control phase 1 for the lower strength tablet process and higher strength tablet process blends respectively (Appendix C: Tables C17 & CIS). With the Savitzky-Golay 2"^ derivative data, 29 and 32 batches were used in control phase 1 for the lower strength tablet process and higher strength tablet process blends (Appendix C: Tables C l7 & CIS). PLS models for the two strengths of tablet used between 17 and 2S batches for the lower strength tablet process tablets (Appendix C: Table C l9) and used between 19 and 34 batches for the higher strength tablet process tablets (Appendix C: Table C20).

Hotelling*s Control Phase 2 For Blends A nd Tablets

Scatter correction was considered a useful pre-treatment of the blend absorbance data. With the SNV DT scatter correction of the lower strength tablet process data set, a number of batches, some consecutive in batch number, were found to have significant Hotelling’s 7^ values (Appendix C: Table C17). These were compared with results of

0.3 -0.1 -0.2 1200 1400 1600 1800 2000 2200 2400 Wavelength/nm 0.3 n -0.1 -0.2 1200 1400 1600 1800 2000 2200 2400 Wavelength/nm

Fig. 4.1. Blends batch mean absorbance spectra for: A) lower strength tablet process {n = 39 spectra) and B) higher strength tablet process (n = 41 spectra), showing spectra for each process divided into two classes with different offsets (batches 1 to 18 for lower strength process blends have lowest offsets, batches: 1 to 17 ; 19 to 20; and 23 to 25 for higher strength process blends have lowest offsets).

the lower strength tablet process SNV DT lower strength tablet process data set. Batches 25 (BN5040), 26 (BN5041) and 27 (BN5042) and batches 33 (BN5064) and 35

(BN5067) were found to have significant values at both the blend and tablet stage (Appendix C: Tables C17 & C19). Batches 26 and 27 produced tablets with friability of

1 mg and 2 mg; batch 33 produced tablets with friability of 1 mg (tablet reference analysis data for batch 35 was not recorded). The Savitzky-Golay 2"^ derivative transformation did not detect these batches as unusual from their blends, however batches 33, 34 (BN5065) and 35 were detected as unusual from the tablet PLS model (Appendix C: Table C l9) (tablet reference analysis data for batch 34 was not recorded). For the lower strength tablet process, the SNV DT transformation was considered to be the most appropriate pre-treatment of those tested for blend and tablet data.

With the higher strength tablet process data sets, the SNV DT transformation detected batches 23 (BN5049)and 24 (BN5050) as unusual at both the blend and tablet stages. Batch 25 (BN5051) was detected as unusual at the blend stage. The blends were not found to have unusual reference analysis values, however they produced tablets with average thicknesses of 4.62 mm, above the limit of 4.6 mm. The SNV DT and Savitzky- Golay tablet models detected batches 34 (BN5070) and 38 (BN5078) and 39 (BN5079) as unusual (Appendix C: Table C20). Batch 34 produced tablets with unusually long disintegration time (17 seconds) and which had an average thickness greater than the maximum limit (average thickness = 4.66 mm). Batches 38 and 39 had average tablet thicknesses above the maximum limit (4.62 mm for both batches) and batch 38 produced tablets with 1 mg friability. With the Savitzky-Golay blend model, batch 38 could be detected as unusual despite having apparently normal reference analysis results (Appendix C: Table C l8). Both the SNV DT and Savitzky-Golay 2"^ derivative pre­ treatments produced useful blend and tablet models.

4. 4 Statistical Quality Control of The E ntire Process by M ultiblock PLSR

4. 4 .1 Multiblock Partial Least Squares Model Generation

Multiblock PLS has been proposed as an alternative projection method to single block PLS for situations with large numbers of variables that can be divided into distinct process sections (X blocks) (MacGregor et al, 1994; Wangen and Kowalski, 1988). The data sets used in this study may be considered as two process X blocks: blend stage, X I (blend spectra) and tablet stage, X2 (combined tablet absorbance and transmission data). The final product quality data, Y, are the blend and tablet certificate of analysis

reference data combined (14 variables). MacGregor (1994) states that multiblock projection methods allow for easier interpretation of process data because smaller meaningful blocks can be individually monitored as may the relationship between these blocks.

The multiblock PLS algorithms used in this study were variations of those of Wold et al (1987) and Wangen and Kowalski (1988). This algorithm leads to a set of orthogonal loading vectors ( w / a , a = 1 , 2 , . . . ) and orthogonal latent vectors ( r / a , a = 1 , 2 , . . . ) for

each block X/. The X/ blocks are then represented in terms of their leading A PLS components as:

= (4.4.1)

a=\

X 2 = '^ t 2 ^ p 2 l + E 2 (4.4.2)

a=\

This enables monitoring and construction of diagnostic plots for each block separately, as previously described for singleblock PLS. This algorithm is also able to effectively handle missing data. An overall monitoring space for the process may be obtained by

using projections in the latent vector space {tCa, a= 1,2, ...) of the consensus matrix T formed by collecting the latent vectors from the individual blocks. The score vectors of this consensus matrix are no longer orthogonal, however it has been shown that where blocking of the process variables has been done in a meaningful fashion, these vectors should continue to define the same plane as the latent vectors obtained by single block PLS, and provide essentially the same predictions of Y:

Y = ± t c , g l (4.4.3)

a=]

A check on whether the blocking has been done well is to compare predictions of Y obtained from the singleblock and multiblock algorithms for the same number of dimensions (A). These should be comparable.

4, 4. 2 Multiblock PLS Model Variability

The multiblock PLS models were found to account for similar amounts of variability within each process stage NIR data set and for the certificate of analysis data as was