Three Steps To A Better Continued Process Verification With Machine Learning
Jan 2, 2020
Pharmaceutical companies typically use periodically collected samples and traditional statistical methods for Continued Process Verification (CPV) – partly due to limited computing power. As a result, you get old information and low accuracy as well as a nonscalable approach for massive data volumes.
Today, however, computing power is no longer an issue. As a result, pharmaceutical companies can use Machine Learning for accurate CPV. This allows continuous learning from the changes in bioprocesses, and the approach scales up as the amount of data grows.
This article explains how you can implement better CPV with Machine Learning in 3-steps.
Why is lifecycle-long regulation critical in pharmaceuticals?
The lifecycle thinking in regulations – especially in FDA’s Stage 3, Continued Process Verification – combines product and process development, qualification of the commercial manufacturing process, and maintenance of the process in a state of control throughout entire commercial production. That is, the time when you, the manufacturer, own the product.
In comparison to other manufacturing sectors, pharmaceuticals have some fundamental differences that make continuous process verification more critical and painful.
As an example, bioprocesses change gradually over time. All modifications in the process – equipment wear, maintenance breaks, changes in the medium or bioreactor, or switches of suppliers and subcontractors – add up over time. Also, in the course of the production lifecycle, bioprocesses tend to change piece by piece. As a result, the process might deviate beyond the original guidelines and exceed a threshold, at some point.
So, how do you ensure that your bioprocesses, which were thoroughly vetted during the design and development phases, do not deviate from their original validated state even years after the launch? The answer is Continued Process Verification.
What’s Continued Process Verification?
FDA’s Stage 3 CPV guidelines aim to move the industry from traditional process verification techniques — such as plant audits and product sampling, which are merely snapshots in the history — to continuous monitoring of vast amounts of data to assure stable, high product quality.
Continued Process Verification is the collection and analysis of end-to-end production components and processes data – during the ongoing process – to ensure product outputs are within predetermined quality limits.
In essence, its central purpose is to ensure that processes are in a constant state of control, thus ensuring final product quality. Central to effective CPV is a method with which to identify unwanted process inconsistencies to execute corrective or preventive measures. Once quality standards are set in place, they must be monitored with regular frequency to confirm those parameters are being met.
Traditional statistical applications in CPV include:
- Monitoring of Critical Process Parameters (CPP) and Critical Quality Attributes (CQA)
- Customer complaint analysis for ongoing monitoring
- Setting up and ongoing use of control charts
- Ongoing monitoring of process capability
- Implementation of measurements and controls
- Integration of Annual Product Review (APR) statistics and CPV
Benefits of continued process verification
The primary reason for creating a Continued Process Verification is not only to achieve compliance with regulations. It also helps to prevent batch discards, mitigate process vulnerabilities, and enables systematic improvement possibilities.
Continued process verification not only improves patient and customer safety, but it also brings other benefits to pharmaceutical companies. It reduces process-related failures and quality costs, improves control, and reduces failure rates.
Some might consider Continued Process Verification simply as a cost. However, continuous process verification delivers concrete value from the very beginning – improved quality, patient safety, systematic process improvement possibilities, and operational cost savings. So, CPV is a safe investment and the payback time can be very short – avoiding just one incident can cover the entire investment.
What are the issues with traditional CPV methods?
The traditional statistical approach has relied on the Design of Experiments (DoE) approach, where you gather samples to do analysis. This inherently means that you will gain a batch-based analysis of the process, instead of continuous analysis. As opposed to a continuous, real-time data analysis, a batch-based analysis provides you with historical information, limited sampling interval, and lower accuracy.
So far, the batch-based approach has been the only option in pharmaceuticals because of limited computational power. Now, the computing power is no longer an issue, and more advanced methodologies based on Machine Learning and Predictive Analytics can be used for CPV.
3-step approach to Machine Learning-based CPV
According to a century-old wisdom, you shouldn’t repeat the same mistake twice. The same principle holds true with Machine Learning. The aim is to learn from the incidents in your bioprocesses systematically, and rectify the processes on the fly to prevent the same incidents from happening again.
The key here is to combine Machine Learning and expert knowledge in your organization. This enables operators to learn from the process, reduce the time needed for root cause analysis, and to choose the right corrective actions in a systematic manner.
Here is a rundown of the 3-step approach for implementing CPV with Machine Learning.
1. Learn from your process history
You have probably collected a rich data set of your finished product runs. Also, your personnel has refined their expertise by solving the problems and incidents that occurred during these product runs. The corrective actions taken at that time have affected the process. What if you extract this information contained in the data and combine it with an expert’s knowledge?
By analyzing your process history for anomalies, you can leverage Machine Learning to systematically find similar anomalies, label them through data, and add descriptions and suggested corrective actions based on the expert’s knowledge.
Each anomaly has a “fingerprint.” With Machine Learning, you can find similar anomalies from the dataset. By labeling and describing these anomaly classes, you can create a process diary that can be used to learn from your entire process history and previous actions systematically.
2. Detect anomalies in real-time
By monitoring your process, you can encounter a continuous flow of signals. You must be able to filter out the relevant information from the signals, often in real-time. With the process diary approach, you can continuously run the same anomaly detection that you used for scanning the process history. This way, whenever an anomaly is encountered, you can compare it to history to find similar anomalies.
The process history works as a library of suggested root causes and actions. This allows you to quickly determine if you have encountered a real deviation in the process or a technical problem, e.g., a malfunctioning sensor. In either case, the history data tells you what the correct actions are.
3. Incorporate predictive modeling
Predictive modeling is another crucial tool for continuous process monitoring. For example, you can start with yield predictions – they can serve as a tool for setting expectations for the current process run, and provide you with an assurance that the process is on the right track. The predictive models can also give you important insights – including which variables affect the process outcome, and how. This helps you to adjust and optimize the process in the correct way.
The model performance itself also serves as an anomaly indicator. Decreasing model performance gives you an indication that there is an unobserved variable in the process, and you should investigate which changes in the process might have caused that extra variability.
Predictive modeling is not obligatory in the beginning. However, by combining predictions with the process diary, you can understand how parameter changes affect your quality and yield, and can you generate scenarios for alternative actions.
Benefits of using Machine Learning for CPV
Implementing the CPV system using modern Machine Learning and leveraging high computing power can benefit you in several ways.
- Machine Learning enables you to improve your processes systematically over a long time
- Machine Learning scales better when the amount of data grows. The conventional statistical methods do not scale well.
- It enables automation – you can use the data in the learning process in the real-time
- Machine Learning is a systematic and transparent method for detecting anomalies faster – and, deriving the root causes quickly
- It decreases the effect of the ‘human factor’ coming from the training and experience of operators. It evens out variations in the process quality due to differences in individual skills in the organization
- Machine Learning enables data-driven decision making and makes data-based evidence visible for reporting purposes
How can you take the first step towards better CPV?
Elisa IndustrIQ team helps global pharmaceutical companies to increase efficiency, reduce costs, and improve product quality – Contact us and let’s take the first step towards smarter CPV!
How To Improve Resilience And Grow Your Business Sustainably
Unexpected disruptive events in manufacturing always interrupt normal production conditions and cause production loss. How resilient are your planning and manufacturing processes? Are you prepared to deal with the unexpected?
Take the quiz below to find out how your organisation stacks up in the face of crisis and get expert insights on how to make your manufacturing more resilient.