Transfusion Medicine, 2008, 18, 190–196
doi: 10.1111/j.1365-3148.2008.00864.x
ORIGINAL ARTICLE
Use of statistical process control in the production of blood components K. Magnussen,* S. Quere* & P. Winkel† *Clinical Immunological Centre, Copenhagen University Hospital, Herlev, and †The Copenhagen Trial Unit, Centre for Clinical Intervention Research, Copenhagen University Hospital, Copenhagen, Denmark Received 3 November 2007; accepted for publication 11 April 2008
Introduction of statistical process control in the setting of a small blood centre was tested, both on the regular red blood cell production and specifically to test if a difference was seen in the quality of the platelets produced, when a change was made from a relatively large inexperienced occasional component manufacturing staff to an experienced regular manufacturing staff. Production of blood products is a semi-automated process in which the manual steps may be difficult to control. This study was performed in an ongoing effort to improve the control and optimize the quality of the blood components produced and gives an example of how to meet EU legislative requirements in a smallscale production centre. Data included quality control measurements in 363 units of red blood cells, 79 units of platelets produced by an occasional staff with 11
technologists and 79 units of platelets produced by an experienced staff with four technologists. We applied statistical process control to examine if time series of quality control values were in statistical control. Leucocyte count in red blood cells was out of statistical control. Platelet concentration and volume of the platelets produced by the occasional staff were out of control, which was not the case with the experienced staff. Introduction of control charts to a small blood centre has elucidated the difficulties in controlling the blood production and shown the advantage of using experienced regular component manufacturing staff.
In many semi-automated processes like the production of blood components, the manual steps are critical and person dependent and hence difficult to control and standardize. To assess this type of problem, a control chart depicting the quality control (QC) values may be a useful tool to apply (Perrotta et al. 2002; Council of Europe, 2007). The QC data from the blood production have been stored electronically for several years. In the ongoing effort to improve the control and optimize the quality of the blood components produced, control charts were applied to analyse this historical database, and the results were compared with the corresponding logbooks of the production. The study was used to introduce statistical process control to the blood centre and is an example of how
to meet EU legislative requirements in a small-scale production centre. The purpose was to assess if the production processes were stable. If they were not, then to identify possible critical steps in any part of the production and to assess if a need existed for optimizing the production process, perhaps by introducing automation. It was specifically tested if a difference was seen in the quality of the buffy coat produced leucodepleted platelets pools (PLT), when a change was made from a relatively large inexperienced occasional component manufacturing staff to an experienced regular manufacturing staff.
SUMMARY.
Key words: autocorrelation, blood production, control charts, EWMA, EWMAST, statistical process control, quality control.
MATERIALS AND METHODS Red blood cell production
Correspondence: Karin Magnussen, Clinical Immunological Centre, Copenhagen University Hospital Herlev, Herlev Ringvej 75, DK-2730 Herlev, Denmark. Tel.: 145 4488 3765; mobile: 145 6088 3911; fax: 145 4494 4167; e-mail:
[email protected]
190
From December 2001 to June 2004, 363 QC samples were taken from the production of 35 000 red blood cells (RBCs) in additive solution and buffy coat removed. The RBCs were collected in KGR7341 triplet # 2008 The Authors Journal compilation # 2008 British Blood Transfusion Society
Statistical process control and blood production bags and the expressors used were OptiPress II, both from Baxter (Deerfield, IL, USA). Eleven QC results were excluded from this study, six because data were missing and five because the technologists had produced only one or two RBCs. The QC samples were taken weekly from randomly selected products from approximately 1% of the production, according to the current guidelines. The measurements of the leucocyte count (white blood cell, WBC), the haematocrit (HCT) and the haemoglobin concentration (cHb) were obtained using the ADVIA 120Ò from Siemens Medical Solutions Diagnostics (Dublin, Ireland). The results represent samples from individual products.
Platelet production From December 2001 to March 2003 (occasional staff), 79 samples were taken from 79 out of the production of 138 PLTs, produced from four buffy coats. From September 2003 to June 2004 (regular experienced staff), 79 samples were taken from 79 out of the production of 155 PLTs. All PLTs that were produced when staff was available for QC sampling were tested. Leucodepletion was performed using PLX5 filters from Asahi Kasei Medical (Tokyo, Japan), and the tests to monitor the results of the leucodepletion were then performed at a central laboratory by flow cytometry; hence, we do not have the WBC results other than that, the leucocyte count in all tested products was below 08 106. The expressors used were OptiPress II, and the PLTs were stored in the 2410 bags from Baxter. The QC quantities monitored were the platelet concentration (cPlt), the volume of the PLTs and the platelet count per PLT (Pc/PLT). cPlt was measured on ADVIA 120. The identity of the technologist who produced the PLT was obtained from the logbooks and recorded. The occasional staff consisted of 11 technologists with a wide variety of tasks only producing PLTs infrequently. In the second period where the staff of four experienced technologists took over the production of PLT, the centrifugation speed had been changed. Thus, the two production periods differed in terms of the number and experience of the staff as well as the process of production. What can be compared is therefore whether the platelet production is in control or not. During the second period, when the experienced regular component manufacturing staff was producing the PLTs, occasionally a technologist belonging to the inexperienced occasional platelet production staff produced a unit, when none from the experienced staff was available.
191
STATISTICS To assess the stability of the production processes, QC sample results were depicted on control charts. A QC sample may include a single or several values. In the present context, we have used single values from weekly samples of the production of RBC and PLT. A sample function value at a specified sampling time is either the current sample value or some function of the sample values cumulated so far. It is depicted on a control chart showing the function value vs. time of sampling. The control chart depicts a centre line (the expected sample function value) and two control limits calculated as the expected value 3 SD of the expected value. Figure 1 shows the Hb/RBC sample values depicted on a Shewhart X-moving range (X-mr) chart calculated from these sample values. The centre line here is the mean, and the control limits are the mean 1 3 SD and the mean 2 3 SD. If the sample function values (here the actual sample values) stay within the control limits, the process is assumed to be stable (in statistical control). If the process is stable, the quality of the process may be assessed from its mean and standard deviation (SD). Once a control chart of a stable process has been constructed, it may be used to monitor the process in that future sample function values are depicted on the chart to see if they remain within the control limits. For a review of the use of control charts in the health care sector, see Winkel & Zhang (2007). In this study, we have used the control charts to analyse and characterize the processes but not for monitoring purposes. Three types of control charts for individual values were used in combination. The X-mr chart depicts individual values and is sensitive to large abrupt changes (Montgomery, 2000), the exponentially weighted moving average (EWMA) chart depicts an EWMA of the current and all previous values and is sensitive to small but sustained changes in the mean level (Montgomery, 2000), and the EWMAST chart also depicts the EWMA, but it is designed for stationary autocorrelated data in that the SD of the EWMA is adjusted accordingly (Zhang, 1998, 2000; Winkel & Zhang, 2007). Autocorrelation implies that the sample values are statistically dependent. Positive autocorrelation, which is the most common form in biology, may, e.g. appear if a random phenomenon that has an impact on one observation is also having an impact on the next one. Thus, if a random error causes one value to be relatively high (low), the subsequent value also tends to be relatively high (low). The estimate of the process SD (the moving range estimate) used in the X-mr chart and the EWMA chart is biased towards too low a value in the presence of positive autocorrelation (Winkel &
# 2008 The Authors Journal compilation # 2008 British Blood Transfusion Society, Transfusion Medicine, 18, 190–196
192
K. Magnussen et al.
Fig. 1. Control chart of 352 Hb/RBC sample values. The chart depicts the sample values as a function of the sample. Starting from the top of the figure and moving downwards, the graph depicts three lines: the upper control limit, the centre line and the lower control limit. The centre line that is equal to the estimated process mean estimates the expected sample value. The two control limits are equal to the estimated process mean 3 times the estimated process SD. The process SD is estimated using the moving range estimate (Winkel & Zhang, 2007, p. 57). The probability that a sample value will fall outside the control limits is 00027 or equivalently for every 370 (1/00027) sample values only one is expected to fall outside the control limits, provided the process is in statistical control. Here, the chart has been used to study the process. It is noted that the process seems to be in statistical control. The data, however, are autocorrelated (see Statistical Methods), but in this case, the X-mr chart does not detect this. The chart may also be used to monitor the process in that future values may be depicted on the chart to see if they remain within the control limits.
Zhang, 2007). Therefore, in the presence of positive autocorrelation, too many false alarms will occur when either the X-mr chart or especially the EWMA chart is used. The 3 SD control limits were used for all three charts, and in case of the last two, l (a parameter used to calculate the weights) was equal to 02. A 3 SD control limit gives a 00027 probability that a value falls outside the limits if the production is in control. Thus, when the process is in statistical control, one would expect one value for every 370 (1/00027) sample values to fall outside the control limits. The assumed normality of the distributions of the data was assessed using the Shapiro–Wilks W test. Each time series was tested for the presence of significant (P < 005) autocorrelation using the Ljung–Box test (Ljung & Box, 1978). The means of different technologists were compared for significant (P < 005) differences using a one-way ANOVA and the homogeneity of their variances was tested using the CochranÕs C-test and BartlettÕs test (Bartlett, 1937; Cochran, 1941; Conover et al., 1981). If the variances differed significantly between technologists (P value of one or both tests is 0), we have EWMA(t) ¼ X(t) l 1 (1 2 l) EWMA(t 2 1). Using the this equation, we obtain for each sample value a weighted sum of the current sample value and all previous ones, in that the weight decreases exponentially with the age of the sample values. The control limits depict the process mean 3 times the estimated SD of EWMA(t). The calculation of the SD of EWMA(t) depends on whether the sample values are autocorrelated or not (see Statistical Methods). In this case, the Hb/RBC values are autocorrelated (Winkel & Zhang, 2007, p. 87, p. 105). When the EWMA chart, designed for data that are not autocorrelated, was used, the control limits were closer to the centre line and three EWMA values fell outside the control limits. However, because the data were found to be significantly autocorrelated, the EWMAST chart was used. It is noted that now the process seems to be in statistical control because only 1 of 352 values fell outside the control limits. The EWMA(t) values display rather systematic patterns. In the EWMA and EWMAST charts, by contrast to the X-mr chart, this should not give rise to any concern. It is to be expected because by definition the EWMA(t) values are highly dependent.
observations outside the control limits is shown. One outlier is expected to occur at random for every 370 observations. Thus, the quantities logWBC/RBC and HCT are not in statistical control. Therefore, we compared the mean level and the within-technologist variability (SD) between the individual technologists. The mean of logWBC/RBC (P ¼ 0004) and the SDs of the RBC volume and Hb/RBC differed significantly between the technologists. The geometric mean of the WBC content in RBCs produced by each of the technologists ranged from 036 to 116 109 per RBC, which is below the limit of 12 109 per RBC recommended by the European Guidelines for RBC (Council of Europe, 2007). In Table 2, the QC results for the two teams producing PLTs are shown. None of the time series were autocorrelated. The control charts (X-mr and EWMA chart) of the inexperienced occasional staff had between 1 and 3 values outside the control limits in
case of the quantities cPlt and volume. The outliers of volume and cPlt were low values falling below the lower control limit of the X-mr chart. Thus, the processes were not in statistical control. Therefore, we compared the performance of the individual technologists. The mean values of the cPlt and Pc/PLT differed significantly between the occasional staff technologists, whereas the corresponding SDs did not. Also both the SD and the median of the PLT volume differed significantly between the technologists. In general, technologists who produced PLTs low (high) in cPlt and Pc/PLT also produced PLTs with low (high) volumes (Figs 3 and 4). In the PLTs produced by the experienced staff, the volume was out of control in that two values fell outside the lower control limit of the X-mr chart. However, the two PLTs in question turned out to be among the 14 PLTs produced by stand-ins from the occasional staff, who helped out when no one from the experienced staff
# 2008 The Authors Journal compilation # 2008 British Blood Transfusion Society, Transfusion Medicine, 18, 190–196
K. Magnussen et al.
Table 2. QC results from the 79 PLTs produced by the occasional staff and the 79 PLTs produced by the experienced staff, including number of values outside relevant control charts Occasional Experienced Control staff* staff* charts† Mean SD cPlt (109 per L) 717 Volume (mL) 329 Pc/PLT (109) 237
Mean SD
200 820 30.5 310 71 256
X-MR EWMA
190 2 (0)‡ 1 (0) 21.7 1 (2) 3 (0) 66 0 (0) 0 (0)
*The occasional staff consists of 11 technologists with basic training and limited routine in platelet production, whereas the experienced staff consists of four technicians with experience and routine. The production process differs, which means that only the degrees of control can be compared. In a control chart with 3 SD limits the probability that one value is outside if the production is in control is 1/370. †The number of values outside the limits of the control chart is shown. ‡Values in parentheses refer to the experienced staff, and values outside parentheses refer to the occasional staff.
were available. When these values were removed, all processes were now in statistical control. The mean cPlt of the experienced staff is higher than that of the occasional staff, whereas the mean of the volume of the PLTs produced by the experienced staff is lower. The overall effect is that the mean Pc/PLT is significantly (P < 005) higher in PLTs produced by the experienced staff than in PLTs produced by the occasional staff. Whether this is because of difference in the performance of the two teams or the change in
Volume (mL)
410
5
Pc/PLT × 1011
194
4 3 2 1 0 1
2
3
4
5
6
7
8
9
10 11
Technologist Fig. 4. The mean and 95% confidence intervals of the Pc/PLT QC measurements made by each of 11 technologists. The number of measurements made by the technologists 1 through 11 was 11, 15, 9, 11, 5, 2, 7, 12, 3, 2 and 2, respectively. The mean values but not the SDs differed significantly (P < 005) between the technologists.
the centrifugation speed or both cannot be discerned. However, 14 PLTs were produced by technologists from the occasional staff during the second period, and a comparison of the four experienced technologists from the regular staff did not show any significant differences between their mean values. Therefore, we compared the PLTs produced by the two teams during this second period, thus overcoming the problem of the difference in the production process, between the first and the second period. The results of this comparison are shown in Table 3. In spite of the small number of PLTs in the groups, the difference between the occasional and the regular staff was statistically significant, with less variability and a higher platelet content in the PLTs produced by the regular staff.
Table 3. QC results from the second period comparing the results from the technologists from the occasional staff with the results from the experienced staff
390 370 350
Occasional staff (n ¼ 14)*
Experienced staff (n ¼ 65)*
P of the difference
Mean
SD
Mean
SD
Mean/ median SD
253
837
171
NS†
0.0002
330 310 290 1
2
3
4
5
6
7
8
9
10 11
Technologist Fig. 3. The mean and 95% confidence intervals of the volume QC measurements made by each of 11 technologists. The number of measurements made by the technologists 1 through 11 was 11, 15, 9, 11, 5, 2, 7, 12, 3, 2 and 2, respectively. The SD differed significantly (P < 005) between the technologists. A non-parametric test taking this variance heterogeneity into consideration showed that the medians of the technologists also differed significantly.
cPlt 740 (109 per L) Volume 297 (mL) 221 Pc/PLT (109)
30.6
313
18.4
0.03†
0.05
82
263
60
0.03
NS
*During the second period, when platelet production was done by the experienced staff, 14 PLTs were still made by technologists from the occasional staff. No difference in the production process. †The medians were compared using a non-parametric test due to the demonstrated significant differences between the standard deviations.
# 2008 The Authors Journal compilation # 2008 British Blood Transfusion Society, Transfusion Medicine, 18, 190–196
Statistical process control and blood production DISCUSSION The study gives an example of how to meet EU legislative requirements in a small-scale production centre. Lack of statistical control may be caused by donors with extreme blood counts, e.g. suffering from unrecognized anaemia or thrombocytopenia, technologists who make errors either occasionally or systematically, malfunctioning of the instruments used for production of the blood products or errors in sampling or counting method for QC purposes. Inspection of the logbooks of the instruments did not reveal any out of control conditions that would have influenced any of the QC data used in this study. However, the analytical instruments may create a positive autocorrelation between QC results (Winkel & Zhang, 2004). When autocorrelation was taken into account using the EWMAST chart, there were still too many HCT values outside the control limits. We found no explanation for the high HCT in the production process as such, which made us look into the sample technique used for securing QC samples. Here, we found that the mixing of the blood products in connection with sampling could be improved, and since we optimized the mixing in connection with QC sampling, there have been no further HCT values out of control. The WBC values falling outside the control limits of the EWMA chart were actually low values. The values are well within the range of the ADVIA 120, and inspection of the regular controls run on the ADVIA 120, as well as the logbook of the instrument did not give us reason to suspect problems with the counting method. We suggest that a minority of technologists are more careful in their handling of the blood products than the rest, thus preventing leucocytes from migrating into the erythrocyte layer. The analysis of the results from the occasional staff revealed, that the mean level of cPlt, the PLT volume as well as Pc/PLT differed significantly between the technologists. During the second period, where the PLTs were produced by the regular experienced staff, 14 of the 79 PLTs tested were still produced by technologists from the occasional staff, who occasionally took over the production when technologists from the experienced staff were not available. When these 14 values were removed, it was revealed that the process as carried out by the experienced staff was now in statistical control. Furthermore, the PLTs produced by the experienced staff had a higher mean level of the Pc/ PLT when compared with the 14 PLTs produced during the second period by the occasional staff. However, even though the control charts showed no signs of lack of statistical control, the SD of the technologists in the experienced staff actually differed slightly but significantly, demonstrating that statistical control does not
195
imply that the quality of a process cannot be improved. In the PLT material, all values out of control fell below the lower control limit. Furthermore, technologists from the occasional staff, producing units with a low cPlt also produced units with a low PLT volume (Fig. 3 and Fig. 4). This is the pattern that one would expect if the products were not handled with enough care subsequent to the centrifugation step. Physical disturbance may cause the erythrocytes to migrate into the platelet zone, which will cause the OptiPress to stop prematurely when it detects these erythrocytes. As a result, the volume will be reduced and even the cPlt because the cPlt decreases when the distance from the erythrocyte zone increases. Thus, it appears that the handling of the material subsequent to the centrifugation step is quite critical. QC based on control charts also adds the possibility to give feedback to technologists. CONCLUSIONS The results suggest that at least in our hands, the manual steps are difficult to control and standardize. Based on control charts, QC sampling was identified as a critical step for monitoring the RBC production. Once corrected, the RBC production was in control. Platelet production is more sensitive, but good results were obtained using the regular and experienced staff. Still, it is vulnerable with only a few technologists to perform an important task, which made us decide to automate the production of PLTs using the OrbiSac System provided by Gambro BCT (Lakewood, CO, USA). The use of control charts on QC data thus proved to be a useful tool. ACKNOWLEDGMENTS The authors thank leading technologist Lone Nielsen and Nykoebing Falster Blood Centre for excellent collaboration. REFERENCES Bartlett, M.S. (1937) Properties of sufficiency of statistical tests. Proceedings of the Royal Statistical Society, 160, 262–282. Cochran, W.G. (1941) The distribution of the largest of a set of estimated variances as a fraction of their total. Annals of Eugenics, 11, 47–52. Conover, W.J., Johnson, M.E. & Johnson, M.M. (1981) A comparative study of tests for homogeneity of variances, with applications to the outer continental shelf bidding data. Technometrics, 23, 351–361. Council of Europe (2007) Guide to the Preparation, Use and Quality Assurance of Blood Components (13th edn). Council of Europe Publishing, Strasbourg.
# 2008 The Authors Journal compilation # 2008 British Blood Transfusion Society, Transfusion Medicine, 18, 190–196
196
K. Magnussen et al.
Kruskal, W.H. & Wallis, W.A. (1952) Use of ranks in one-criterion variance analysis. Journal of the American Statistical Association, 47, 583–621. Ljung, G.M. & Box, G.E.P. (1978) On a measure of lack of fit in time series models. Biometrika, 65, 297. Montgomery, D.C. (2000) Introduction to Statistical Quality Control. New York. Perrotta, P.L., Ozcan, C., Whitbread, J.A. & Finch, S.J. (2002) Applying a novel statistical process control model to platelet quality monitoring. Transfusion, 42, 1059–1066.
Winkel, P. & Zhang, N.F. (2004) Serial correlation of quality control data – on the use of proper control charts. Scandinavian Journal of Clinical and Laboratory Investigation, 64, 195–203. Winkel, P. & Zhang, N.F. (2007) Statistical Development of Quality in Medicine. Wiley and Sons, Chichester. Zhang, N.F. (1998) A statistical control chart for stationary process data. Technometrics, 40, 24–38. Zhang, N.F. (2000) Statistical control charts for monitoring the mean of a stationary process. Journal of Statistical Computation and Simulation, 66, 249–258.
# 2008 The Authors Journal compilation # 2008 British Blood Transfusion Society, Transfusion Medicine, 18, 190–196
