Overview: A Pharmaceutical Manufacturing Situation
A pharmaceutical company recently implemented Six Sigma within its manufacturing and development divisions to improve the quality of existing manufacturing processes and to develop new manufacturing processes right the first time. After reviewing the manufacturing balanced scorecard, the Vice President of Manufacturing was concerned about the relatively high number of lots of an established product that did not meet release specification requirements.
This product has been marketed for more than 20 years, and the Vice President is aware that significant knowledge about the product manufacturing process has been lost over time. Therefore, the Vice President asks a process engineer, also a Six Sigma green belt, to analyze and document the process using Six Sigma and to identify ways to dramatically improve process capability.
Define the Quality Characteristic
Many pharmaceutical manufacturing processes are lot-based, where large quantities of raw material pass from one process step to the next. Each step processes and transforms the input materials incrementally. These steps lead to the production of the final product, in this case, a tablet.
The process under review is the secondary manufacturing process, a five-step process responsible for turning the active pharmaceutical ingredient (API) into the tablet. The steps are:
1. A sizing operation, which grades the incoming API
2. A milling operation, which aims to deliver the API with uniform particle size
3. A blend operation, which mixes the milled API with a number of inactive ingredients with the goal of delivering a uniformly blended material
4. A compression operation, which aims to produce tablets with target hardness characteristics
5. A coating operation that delivers the final product—a coated tablet
After completing the process, sample tablets from the lot are tested to characterize the lot against a number of criteria, one of which is the percentage of the sample of tablets dissolved after 60 minutes.
The first step in defining the variables impacting the quality characteristic we want to impact, which is percentage tablet dissolved after 60 minutes. The process engineer baselines current performance by examining the process capability of the key quality characteristic for the last 12 months’ production. The key quality characteristic is the percentage of the tablet dissolved after 60 minutes, with a lower specification limit for product release of 70%.
Capability analysis shows that approximately 16% of the lots produced over the last 12 months fail to attain the lower release specification, with a resulting capability index (Cpk) of 0.27 and equivalent sigma quality level of 2.3. Having verified the need for improvement, the engineer decides to review and document the current tablet manufacturing process.
Measure the Results
The potential causes of slow dissolution are the input materials and processing parameters at each step of the process. The process engineer indicates, on the process map, the key inputs that might influence dissolution. Data on dissolution and the potential causes identified in the process map for lots manufactured over the last 12 months were collected using a combination of manual data input and imports from Microsoft Excel and database files. The resulting merged data matrix consists of a row for each lot and a column for each potential cause or quality characteristic.
Because dissolution is a destructive measurement, conventional gauge R&R could not be applied to assess the sources of measurement variation relative to the process window.
Analyze, Part 1: Visualize and Explore Relationships
The starting point for determining the actual relationship between each potential cause and dissolution is visualization using interactive graphs.
The process engineer brushes the data points in the histogram of dissolution with a value below 70% and then examines the distribution of each process factor for obvious relationships between one or more process factors and for those lots that fail release specifications.
In Figure 4, failing lots are identified in dark green across all process factor distributions, and the process engineer looks to see if a setting of one or more factors leads to no failures. Although some process factors are associated with reduced failure rate; e.g., medium API particle size, mill times exceeding 25 minutes, small screen size, and low spray rate. Tighter control of one factor does not remove all lot failures.
The potential effect of each process factor on dissolution is then investigated in more detail to verify that several factors must be operating together to cause lot failures.
For the process factors measured on a discrete scale such as API Particle Size (large, medium, small), box plots of the effect of each process factor against dissolution display with the individual data points superimposed in red for lots that pass release specification and in blue for lots that fail release specification.
The comparison circle to the right in each graph indicates the levels of the process factor that result in statistically significant differences in the value of mean dissolution. Mean dissolution for screen size 5 is significantly lower than mean dissolution for screen sizes 3 and 4. This implies that screen size 5 should be avoided. However, failing lots continue to occur to a lesser extent at lower screen sizes; this verifies the earlier conclusion that tighter control of any single factor does not remove all defects.
For process factors, such as Mill Time, that are measured on a continuous scale, a scatter plot matrix of the effect of each process factor against dissolution displays. Successful lots are identified in red and failing lots identified in blue. The top part of the display is a matrix of the linear correlation coefficients; the middle part of the display is a scatter plot matrix, and the bottom part of the display is a color map of correlations.
The scatter plot matrix shows that many variables have a weak correlation with dissolution; Mill Time and Spray Rate appear to have a stronger relationship with dissolution.
Analyze, Part 2: Generate Hypotheses
Although many factors seem to have a weak relationship with dissolution, tighter control of one factor alone is unlikely to result in a dramatic improvement in process capability for dissolution. From the view of achieving a Six Sigma process, the engineer “can’t see the wood for the trees.” Therefore, the engineer selects two hypothesis-generation techniques that look at how the potential causes operate collectively to influence dissolution. The two JMP techniques appropriate to the problem are partition analysis and stepwise regression.
Hypothesis Generation Technique: Partition Analysis
The process engineer first considers a partition analysis to see if it is possible to group good lots from bad lots according to the values of one or more of the potential causes. The main graphical display (Figure 7) colors passing lots in red and failing lots in blue. Each time the engineer clicks on the split button, the lots subgroup in a way that one subgroup contains a greater proportion of failing lots and the other subgroup a greater proportion of passing lots.
The first split creates a subgroup with screen size 3 or 4 with relatively fewer failing lots than screen size 5. The second split further subdivides the screen size 5 group according to whether Mill Time is < 11 or = 11; again, the = 11 subgroup has proportionately fewer failures. The third split further sub-divides the screen size 3 or 4 subgroup according to whether spray rate is < 404 or = 404. The < 404 sub-group has no failures. The engineer could continue making splits that might identify other combinations of the process factors that result in no failures.
The tree on the bottom right in Figure 7 summarizes the four subgroups created by the splits. Scrolling down, we have a leaf report for each node of the tree; the last leaf shows that the subgroup defined by screen size 4 or 3 and Spray Rate < 404 contains 47 lots—all of which pass the lower specification for product release.
The process engineer now looks at a capability analysis of these 47 lots. The analysis confirms no rejected lots and a Cpk of 0.55 or the equivalent of a 3.2 sigma process. The tighter controls on screen size and spray rate proposed by partitioning result in a modest improvement that might remove lot failures. However, these tighter controls result in a process that is operating on the edge since several lots have dissolution values too close to the lower specification limit for comfort. Certainly, this change does not result in the dramatic improvement in process capability requested by the Vice President of Manufacturing.
However, this change does represent an incremental improvement in the process, and this process improvement should be standardized while other improvement activities continue. Thus, a process change request is initiated to ensure screen size of 3 or 4 and spray rate < 404 in future production in an effort to remove potential future failures from the process while further improvements are investigated. A risk assessment within the process change request states that independent validation of the change is not required as the change is within the bounds of the currently validated process.
Hypothesis Generation Technique: Stepwise Regression
The process engineer now considers an alternative hypothesis generation technique called stepwise regression. This identifies those process factors that have the statistically strongest relationships with dissolution.
Stepwise regression identifies API Particle Size, Mill Time, Screen Size, Lactose Supplier, Blend Time, Blend Speed, Compressor, Coating Viscosity, and Spray Rate as process factors that might collectively cause poor process capability. These factors are considered in more detail in the Improve phase.
The process factors identified by stepwise regression are progressed into a multiple regression model. The model confirms that all but two factors—API Particle Size and Lactose Supplier—significantly contribute to the tablet dissolution issue. The relationship between dissolution and each process factor included in the multiple regression model is illustrated in Figure 10. The figure also shows the setting of each process factor that maximizes mean dissolution. In summary, Mill Time, Screen Size, Blend Time, Blend Speed, Compressor, Coating Viscosity, and Spray Rate are identified as being the collective set of the most likely causes of poor process capability for dissolution.
To investigate the best-case process capability that can result from controlling these factors at their “best” settings, the process engineer uses the simulation option to simulate the process under these conditions superimposed with the level of uncontrolled variability estimated by multiple regression.
The proposed optimum settings of the key process factors result in a predicted Cpk exceeding 2.3. This is equivalent to 8.4 sigma process—substantially better than a Six Sigma process and consistent with the goal of delivering a dramatic improvement in process capability.
A process change request is initiated, and a control plan is defined to run the process at the tighter controls implied by the proposed optimums. A risk assessment within the process change request states that independent validation of the change is not required since the change is within the bounds of the currently validated process.
A control plan includes a review of lots after two months when approximately 20 lots will have been processed under the new control plan.
The improvement is verified with an estimated short-term Cpk of 2.1 and long-term Cpk of 1.9 (equivalent to a 7.7 sigma process in the short term and a 7.2 sigma process in the long term).
Using JMP for each step in the Six Sigma DMAIC process, the team identified the cause of unacceptable tablet manufacturing yields, simulated a proposed solution, and implemented the improvement. The improvement represents more than a $2M annual saving by eliminating lost batches due to dissolution failure with significant upside through predictability and consistency of supply.