ANALYST NOVEMBER 1986 VOL. 111 1225 Techniques for Evaluating Control of Automated Multi-determinant Analytical Instruments by Computer Murray T. Fisher and Julian Lee Applied Biochemistry Division Department of Scientific and Industrial Research Private Bag Palmerston North New Zealand M. Kelly Mara Applied Mathematics Division Department of Scientific and Industrial Research Private Bag Wellington, New Zealand Analytical data collected from calibration standards and other quality control standards using an amino acid analyser (18 ninhydrin products) and a plasma emission spectrometer (20 elements) are logged regularly during analysis periods. Principal components are derived from the data using principal component analysis and those that contain the most variation (usually first three) are plotted on cumulative sum charts (Cusums).Changes in the physico-chemical control of instruments can be clearly demonstrated and out of control events identified. The computer program can be run on any computer with sufficient memory to handle large matrix transformations. The technique is applicable to any instrument measuring several or many variables. Keywords Amino acids; ICP-AES; quality assurance; principal component analysis; Cusums There has recently been a rapid growth in the use of automated instrumentation for multi-determinant analysis. Attention has focused on the importance of quality assurance (QA) associated with these analytical instruments and analysts are complementing subjective assessment based on experience and skill with more objective criteria.The aim of the investigation reported in this paper was to devise and test an objective QA technique for monitoring the control of automated multi-channel analytical instruments. The technique presented involves the collection com-puterised manipulation and interpretation of sets of data derived from the repetitive assay of reference materials or control solutions. In this laboratory both an amino acid analyser and an inductively coupled plasma emission spec-trometer (ICP-AES) are routinely used to assay biological materials. For the amino acid analyser a data set consists of up to twenty ninhydrin positive products from a protein hydrolysate. ICP-AES data sets contain data for more than 20 elements in a sample digest. Ideally a multi-determinant QA technique would need to consider simultaneously all the data in a set and be able to relate that set to previous sets.It should be sufficiently flexible to be applied to sets of control data from both types of multi-channel instruments and be quick to use. Since the introduction of Shewhart charts,l>* their use for process control has become widespread. A more recent control chart is the cumulative sum (Cusum) ~ h a r t . 3 > ~ Con-struction and operating procedures for the charts are de-scribed elsewhere.5 The advantages of the Cusum chart are its strong visual impact and its ability to detect more clearly the subtle changes in the process that may not have emerged in Shewhart charts of the same data. However the properties of these charts have been examined mainly when only one process characteristic is being measured.When several, possibly related characteristics are being observed accurate statistical control becomes more difficult. The simple approach to monitoring multi-channel instrumental variation is to assume that all measured charac-teristics are independent. A Cusum chart may then be constructed for each variable. The process is then considered to be “out of control” if one or more of the charts demonstrates such a condition. There are however several major weaknesses in this approach. Firstly the independence is rarely justified. Secondly even if independence is a suitable assumption the false alarm risks increase dramatically with the number of elements. For example using a multi-channel instrument with ten determinants the use of ten individual charts each with a false alarm risk of 0.05 makes the over-all false alarm risk 1-(0.9510) = 0.4.That is an “out of control” condition will be indicated for 40% of the time when in fact the process is running on target. When the measurements are correlated the calculation of this error rate becomes unduly complex. In addition the simultaneous handling of ten control charts is far too cumbersome. What is therefore required is a more sophisticated multi-variate technique which also possesses the important property of being quick and simple to use. In order to decide which determinant or combination of determinants is contributing most to the variation the use of principal component analysis (PCA) combined with Cusum plotting seems an appropriate technique.The use of principal components (PCs) to monitor production processes has been advocated earlier.68 Traditionally PCA has been used to describe multi-variate data sets and to examine the relation-ships between the original variables. PCA looks for a few linear combinations that can sum-marise the data losing in the process as little information as possible. These linear combinations of the original variables are designed to account for as high a percentage of the variation among the variables with as few PCs as possible. Our procedure consists of generating the PCs from a suitable, stable (equality of variance) set of data and then producing Cusum plots of the scores of the weighting in the first few PCs.The method of PCA transforms the original set of correlated measurements into a mutually orthogonal (independent) set of new uncorrelated variables that are linear combinations of the originals. PCA involves finding the eigenvalues and eigenvectors of the sample correlation or covariance matrix. A detailed description of the method has been given by Jackson.9 Of the several methods9 of assessing how many PCs to retain to describe the data adequately we chose the most commonly used method that of considering the percentage variation accounted for by each PC. gxperimental Instrumentation The amino acid analyser used was an LKB (Cambridge UK) 4150 Alpha using a sodium citrate buffer pH gradient with an 11 X 10-6 m particle size cation-exchange resin.Post-column derivatisation was by reduced ninhydrin reagent. The 570- and 440-nm detector responses were plotted together with pea 1226 ANALYST NOVEMBER 1986 VOL. 111 processing and area integration using a Shimadzu (Kyoto, Japan) CR-2AX Chromatopac connected to a Shimadzu INP R2A instrument. Detailed operating conditions have been given previously.10 The conditions were optimised to give >95% separation between serine and threonine and arginine, the last peak off the column eluted within 95 min. The plasma emission spectrometer used was an Applied Research Laboratories (Sunland CA) 34800 inductively coupled plasma source (ICP) and a vacuum polychromator (Paschen-Rounge system) equipped with a movable primary slit. A standard quartz torch was used with a GMK nebuliser system (Lab-Test) and a Gilson Minipuls I1 peristaltic pump.The analytes were measured at the following lines (nm) A1 I, 308.22; As I 189.04; B I 249.68; Cd 11,226.50; Co 11,228.62; Cr 11 267.72; Cu I 324.75; Fe 11 259.94; Mg 11 279.08; Mn 11 257.61; Mo 11 202.03 Na I 589.60; Ni 11 231.60; Pb 11, 220.35; Se I 203.99; Sn 11 189.98; Sr I 407.77; and Zn I, 213.86. Plasma operating conditions are described else-where.11712 Analyses were carried out under conditions optimised for simultaneous multi-elemental analysis. Materials The calibration standard for the amino acid analyser used was a commercial amino acid hydrolysate mixture containing 17 acids plus ammonium sulphate (No. 20079 Pierce Chemical Co. Rockford IL USA).The in-house reference sample used in this study was prepared as follows. Lipid and carbohydrate were extracted from freeze-dried fish material by the Folch procedure.13 The remaining protein residue was then hydrolysed with 6 M HC1 by an in-house modification of a standard procedure. 14 The in-house reference hydrolysate was assayed with each batch of hydrolysates of biological material. Twenty-two data sets containing assays of up to 15 amino acids (plus ammonia) were recorded and stored in the computer. For the ICP the instrumental control monitoring procedures used were similar to those documented by Botto.15 A 2 M HC1 blank solution was prepared from re-distilled constant-boiling HCl. This solution was used to make a check standard which contained all the elements analysed by the instrument at levels of one hundred times their respective 30 detection limits (approximately equal to ten times the background equivalent concentration).This standard was used routinely to monitor any calibration drift within batch runs. Data Collection and Analysis Amino acid data sets containing results of the analysis of 15 amino acids plus ammonia from the calibration standard and the fish hydrolysate were recorded and stored in separate data files on a VAX 11/780 computer. For unoxidised protein hydrolysates the amino acids cystine and methionine were not reported because of the breakdown of sulphur-containing amino acids during sample hydrolysis. The ICP intensity data of the emission lines measured by assaying the check standard were acquired using the ICP computer.The data were then chronologically logged on the VAX. The measured intensities of the emission lines ratioed to the Ca I1 317.93 nm line were also logged. Data manipulation and the computations of PCs were performed using a computer program which made use of Minitab. A computed value or score for each multivariate data set at each event is derived from the PCs. The Cusums of the scores for each of the main (usually the first three) PCs were plotted by computer against the control sample number (event). These events are in a chronological sequence that may be days hours etc. or batch order. Results and Discussion Amino Acids The PCs derived from the AA data are presented in Table 1. For the calibration standard the first PC derived from the data shows that all amino acids contribute more or less equally to the 52% of the total variation.The opposite sign of the coefficient for arginine means that it varies in a different direction; otherwise the sign is arbitary. This finding of equal contribution to the variation might be expected because the assay of a commercially prepared standard mixture represents the ideal separation in terms of being free of sample matrix effects. For the second PC however mainly histidine lysine, ammonia and arginine contribute to a further 28% of total variation. These four determinants are eluted by buffer three, which has a six-fold greater molarity compared with buffers one and two. This greater molarity results in an eluate stream with a higher absorbance.The associated base line change just prior to the elution of these four and the slight decrease in base-line stability point to possible reasons for these amino acids contributing more to the over-all variation. Further, arginine is the last peak to elute and has the greatest peak area to height ratio or band broadening. For the in-house reference fish material PC analysis of the data reveals that for the first component glutamic and aspartic acids and arginine contribute to 49% of the total variation. Here glutamic acid might be expected to exhibit some variation owing to its peak having a slight trailing base over the proline eluting close behind. Arginine again might be expec-ted to vary for the same reasons but aspartic acid’s contri-bution to the variation is not readily explained in terms of separation.For the second PC (26%) glutamic acid histidine and lysine stand out probably for the above reasons. The third PC (llYo) shows that arginine and glutamic acid are again the main variants. It is worth noting that for both materials most of the variation is accounted for by the first three principal com-ponents. Although the components contributing to the variation were usually the same in both the standard and the reference material aspartic and glutamic acids and arginine give higher weightings in the first PC of the latter. This may indicate matrix effects from the biological material. These differences emphasise the need for using quality assurance reference materials that are similar in nature to the “un-knowns.” The amino acid analyser operator’s log book was also consulted to find possible reasons for the changes in slope of the Cusum of the first PC (Fig.1) coincident with changes in instrumental chemistry. It was found that the bottle of buffer one was almost empty on day 3 and was changed for 4 1 of fresh buffer on day 4 (event 4). The log book also showed that a fresh ninhydrin colour reagent of slightly different absorbance was introduced into the analyser on day 17 (event 17). The second and third PC Cusums showed no significant change in slope. ICP-AES Table 2 gives the total PC variation calculated from the covariance matrix for 17 elements determined by ICP-AES in a standard check sample. This standard is run every 40th sample (hourly).The data shown cover a 6-h period for each of three consecutive days during the analysis of about 750 samples. A fresh normalisation of the stored calibration graphs was made at the beginning of each day using a high and low standard otherwise no further adjustment to the calibra-tion graphs was made. Most of the over-all variance is accounted for by the first PC (71.5%). Three PCs account for 97% of the total variance in the data. The elements apart from Pb and Se contributed more or less equally to the over-all variance in the first PC. The same sign for the vectors of all the elements indicates a positively correlated variation. The examination of the concentration data indicated a downward drift within each day’s run.Selenium contributed markedly to the over-all variance because of higher values obtained on the second day’s run and also due to poorer over-all precision. Indeed Table 1. Amino acids. Weightings for the first three PCs from the correlation coefficients between amino acids determined in both the calibration over 22 consecutive daily runs Principal MPC . . 0.22 0.17 0.29 0.27 0.21 0.25 0.21 0.28 0.25 0.26 0.27 0.25 0.28 component Asp Thr Ser Glu Gly Ala Val Met Ile Leu N-Leu Tyr Phe Calibration standard: 2ndPC . . 0.05 0.04 0.09 0.00 0.07 0.06 0.00 0.00 0.07 0.09 0.05 0.06 0.04 3rdPC . . 0.15 0.03 0.33 0.65 0.06 0.13 -0.1 -0.27 -0.16 -0.09 -0.18 -0.20 -0.46 In-house reference material: 1StPC . . 0.41 0.12 0.12 0.57 0.13 0.22 0.19 -0.04 -0.08 0.25 - 0.05 0.08 2ndPC .. 0.05 0.01 0.02 0.30 0.04 0.03 -0.08 -0.05 -0.02 -0.03 - -0.01 -0.02 3rd PC . . 0.07 0.20 -0.06 -0.6 0.03 -0.04 0.09 -0.09 0.04 -0.03 - -0.01 0.03 Table 2. ICP elements. Weightings for the first two PCs from the correlation coefficients between elements determined in a “check” standard run Principal component A1 As B Cd c o Cr c u Fe Mg Mn Mo Ni Pb 1stPC . . 0.19 0.27 0.22 0.25 0.26 0.17 0.12 0.1 0.26 0.08 0.22 0.26 0.34 2ndPC . . -0.08 0.2 -0.26 0.04 0.03 -0.06 -0.04 -0.27 -0.55 4 - 5 3 -0.01 -0.07 0.04 Table 3. ICP element ratios. Weightings for the first two PCs from the correlation coefficients of emission line intensities ratioed to the intensity containing each of the elements at a concentration designed to give 10 times the background equivalent concentration (10 BEC) Principal component A1 As B Cd Co Cr Cu 1StPC .. 0.33 0.06 0.31 0.02 0.05 0.05 0.32 0.13 0.08 0.11 0.12 0.73 0.08 0.00 2ndPC . . 0.10 0.11 0.38 0.14 0.19 0.07 0.03 0.17 0.15 0.15 0.25 0.46 0.21 0.33 Fe Mg Mn Mo Na Ni P 1228 ANALYST NOVEMBER 1986 VOL. 111 $ 3 0 541 I 1 4 \ 36 30 24 18 12 ---6 -0 --6 -Fig. 1. Amino acids. Cusum plots for the first three principal components of data from daily analyses of a calibration standard containing 16 amino acids plus ammonia. Buffer and reagent changes made at points A and B respectively 1 0.5 0 E 2 -0.5 0 -1.0 -1.5 -2.0 v 2nd PC 1 I 1 I I I I I I 2 4 6 8 10 12 14 16 18 Eventlh Fig. 2. ICP-AES. Cusum plots for the first two principal components of concentration data from the calibration standard (18 elements) run hourly over a 6-h period on three consecutive days.Slope changes indicate daily normalisation of calibration graphs 1 I 0 2 4 6 8 10 12 14 16 18 20 Eventiday Fig. 3. ICP-AES. Cusum plots for the first two principal components of emission intensity data (ratioed to the Ca I1 317.93 nm emission line) of a calibration standard analysed daily at the start of each batch of analyses (1 month period). Arrow indicates new calibration Se and Pb are the least sensitive of the emission lines measured. The second PC indicates a different pattern of variation among the elements. Magnesium and Mn and again Se, contribute the most to the variation. The reasons for this are not readily apparent.The Cusum plots for the first and second PC are shown in Fig. 2. Within each PC the daily drift is well illustrated with an abrupt change in slope indicating the start of each daily run. The use of the Cusum’s associated V-masks indicates that the analysis process is “in control,” however, within each daily run. The marked change in slope of the Cusum at the start of each day’s run reflects the renewed normalisation of the calibration graphs. A different check standard is routinely run before the start of each batch of analyses. The decision to proceed with further analyses depends on the performance of the instrument as reflected in the data from this standard. Several years of emission intensity data have been logged from measurements made on a “check” solution containing all the elements.A pattern has emerged showing changes in emission response in relation to changes in instrumental settings. Consequently a standardised procedure now operates to ensure an optimum response for the analysis of all determinants. Element emission line intensities (mV) are ratioed to the Ca I1 317.92 nm emission line. These ratios are systematically examined for departure from a pre-set value. Criteria similar to that outlined by BottolS are used. The first two PCs and their Cusums are shown for a typical run sequence (approximately 1 month) in Table 3 and Fig. 3, respectively. This particular sequence has been chosen to illustrate the effect of an instrument re-tuning and ensuing re-calibration on the Cusum plot which occurred on day 12 (event 12 Fig.3) and is reflected in the change of slope in the Cusum for the first PC. In the first PC (Table 3) the elements Na Al Cu B and to a lesser extent Sr contribute the most to the over-all variation. This observed pattern may be explained in part by the known response behaviour of the measured elements to various plasma processes.1~19 The emission lines Na I 589.59 nm A1 I 308.22 nm Cu I 324.75 nm B I 249.68 nm and Sr I 407.77 nm show maximum emission intensities outside the “compromise” plasma observation zone. 11 Consequently these elements exhibit responses that are sensitive to slight changes in those instrument parameters which contribute to effective shifts in the plasma stability of this zone viz. changes in the aerosol carrier gas flow-rate.This drift contributes to a greater daily variability in the observed intensities for these elements. For the technique to be routinely usable as a control method it is hoped that the first few (say a maximum of three or four) principal components account for most of the variability. In the data given in this work the variation among the elements is highly correlated and this is reflected in the observation that only two or three PCs are necessary to describe near to or over 90% of the total observed variation. Additionally it is necessary to keep in mind the important question “Does the process seem in control?” In order to answer this it is important that changes that are evident in the Cusum plots on the individual assayed components are also demonstrated in the Cusum charts of the chosen PCs.Conclusion The combined approach of PC analysis and selected automatic Cusum charting has been shown to be effective in monitoring the control of our analytical instruments during multi-determinant assays. This QA technique can be used on any reference data to monitor the control of any multi-determinant instrument and could be a feature of the instrument’s QA software. We thank the following for their help G. F. Filby for data entry Mr. P. Thakurdas for programming and Dr. J. R. Sedcole for helpful discussions ANALYST NOVEMBER 1986 VOL. 111 1229 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. References Shewhart W. A. “Economic Control of the Quality of Manufactured Product,” Van Nostrand New York 1931.Westgard J. O. Barry P. L. and Hunt M. L. Clin. Chem., 1981 27 493. Page E. S. Biometrika 1954 41 100. Ewan W. D. and Kemp K. W. Biometrika 1960 47 363. “Guide to Data Analysis and Quality Control Using Cusum Techniques,” BS 5703 Parts 1-4 British Standards Institution, London 1980. Jackson J . E. and Morris R. H. J. Am. Stat. Assoc. 1957, 52 186. Jackson J. E. Commun. Stat. Theory Methods 1985,14,2657. Woodall W. H. and Ncube M. M. Technometrics 1985 27, 285. Jackson J. E. J. Qual. Technol. 1980 12 201. Fisher M. T. in Richards E. L. Editor “Developments in Food Analysis.” Symposium on Food Chemistry Food Tech-nology Department Massey University New Zealand 1983, Lee J. ICP Inf. Newsl. 1983 8 553. p. 21. 12. 13. 14. 15. 16. 17. 18. 19. Lee J. Sedcole J. R. and Pritchard M. W. Spectrochim. Acta Part B 1985 41 12. Folch J. Lees M. and Sloane S. G. H. J. Biol. Chem. 1957, 226 497. Horwitz W. W. Editor “Official Methods of Analysis of the Association of Official Analytical Chemists,” Thirteenth Edi-tion Third Supplement “Changes in Methods,” AOAC, Philadelphia 1982 p. 496. Botto R. R. Spectrochim. Acta Part B 1984,39 95. Boumans P. W. J. M. and de Boer F. J. Specrrochim. Acta, Part B 1975 30 309. Anderson T. A. Burns D. W. and Parsons M. L., Spectrochim. Acta Part B 1984 39 559. Houk R. S . and Olivares J. A. Spectrochim. Acta Part B , 1984,39 575. Caughlin B. L. and Blades M. W. Spectrochim. Acta Part B 1984 39 1583. Paper A61109 Received April 8th 1986 Accepted June 11 th 198