Minimising Cost of Poor Quality – CoPQ, is among the commonly identified projects in the Pharma industry. This article covers an Oral Drug project that one of our consultants worked on for a Pharma Client.
Problem Statement: The measured Assay for this product, a Critical Quality Attribute (CQA) in the FP (Finished Product) stage was having repeated issues and the process performance identifier Ppk of this product was coming out to be fairly low (<0.66). Mandate for this project was to identify the real root cause of the issue (since other probable root causes identified earlier and preventive actions had not yielded any results).
We followed Six Sigma Methodology and applied conventional DMAIC steps. As mentioned earlier, we have already integrated Machine Learning and AI modelling as part of the investigation steps.
- Project Charter signed
- CTQ (Critical to Quality) analysis done
- Process Map and Cause and Effect analysis done
- For this project the CTQs are measured at IP stage (in process) and FP (finished product stage)
In the DMAIC’s “Analyze” stage we looked into the previous reports on investigation and Corrective and Preventive Action (CAPA). Reports indicated that the CAPAs had not been effective and issues continued to reoccur.
The Investigation
It was observed that the CTQ is on target at the IP stage but higher in the FP stage, though there are no major process steps in between.
This indicated possible analytical variation. However before proving analytical, we had to Collect the CPP (Critical Process Parameters) data. We ran the exploratory analysis using different ML algorithms. This was the first time we had even operator level data made available to us, so we looked at the possible dimensions of Equipment and Operators. ML model and exploratory analysis did not indicate any potential causes in Manufacturing, which would lead us to suspect the analytical variation.
Data was further collected as per given format and data was analysed. We did extensive experimentation with ML models and analytical variation was validated.
Based on the analysis the variation appeared to be in the standard preparation stage in the Analysis method. This conclusion was presented and it was confirmed by R&D and QC/QA.
As a recommendation a new CAPA was proposed and implemented.
Result: A marked difference in mean value of CTQ was observed. The project was appreciated by the senior production managers and technical and Quality management.
Serendipity
While Water by KF (WKF) was not part of the investigations earlier, the extensive analysis for Analytical variation threw up pointers that directly pointed to the variance seen in WKF measurements. We ran ML models on the collected data on CPP profiles and presented the observations and modelling results to the manufacturing SMEs and our OE counterparts.
Kudos to the Operations Excellence (OE) team to be open minded about the scope of the investigation. These SMEs went to lengths to help us collect the relevant data and corroborate the findings with their observations. In several cases the SMEs would also challenge and help us correct the observations that were unexplained purely looking at the data.
The prediction model for WKF was accepted and was referred for all further working ranges on the equipment.
Lessons in What not to do
We had a sort of false start when our ML Models started showing parameters like Bed Temperature and Outlet Temperature as highly significant in the SHAP Plots. Before it could waste a significant time, the OE Experts and their experience proved very helpful. A Master Black Belt was quick to point out that these are Output Variables and not necessarily Input Variables.
Segregating the KPIVs and KPOVs continues to be one of the first steps in the modelling process.
