XAI’s crucial role in GxP Manufacturing
XAI will unlock the power of AI for heavily regulated manufacturing industries
Artificial Intelligence in a manufacturing setting has the potential to unlock the massive value in enabling organizations to take rapid, accurate and autonomous action in managing their processing lines. In a joint report, Accenture and Frontier Economics estimated that Artificial Intelligence (AI) has the potential to increase manufacturing productivity by as much as 40 percent and it is projected that by 2035 manufacturing industry profits will increase of 39 percent.
However, in heavily regulated industries such as Life Sciences, the potential for patient harm means the bar for AI involvement in manufacturing processes is set quite high. It is imperative for these industries to adopt AI to unlock its potential, but to do so will require a high degree of trust and transparency in the integration of AI in to manufacturing processes. This is where Explainable Artificial Intelligence (XAI) can play a crucial role.
“Applying AI to biopharma manufacturing facilities and processes enables life sciences companies to stream factory and sensor data to analytics engines that generate novel insights. These insights can then help companies predict process bottlenecks, identify quality control issues, and proactively suggest corrective actions.”
What is XAI – and how does it differ to AI?
AI methods such as Artificial Neural Networks (ANN), while having the potential for unlocking massive value, function as black boxes through which no human can muster the ability to explain how they arrive at the outcomes they do.
XAI (Explainable Artificial Intelligence) solves this by providing transparency, interpretability and explainability to AI models. XAI has gained popularity in social settings where tackling social biases has become a hot topic. But in GxP manufacturing processes, it can also play a role in enabling teams to take advantage of AI’s supercharging of predictive modeling capabilities while allowing the operators to remain informed and in control.
XAI works by providing human-understandable explanations for the decisions made by AI models. This is achieved through various techniques such as local interpretable model-agnostic explanations (LIME), Shapley values and counterfactual explanations. By providing transparency and interpretability to AI models, XAI enables users to understand how the model arrived at its decisions and to identify any potential biases or errors in the model’s reasoning.
How can XAI play a crucial role in GxP manufacturing?
Human error remains a leading cause in manufacturing errors, with one estimate stating it is responsible for 80 percent of process deviations in pharmaceutical manufacturing. The more complex the process, with more moving parts to keep track of the higher the potential for human error to take hold.
Utilizing AI can help us lessen the burden of tracking the many interconnected parts of our process – while XAI ensures that the operators are always in control and in the know as to why the AI model is doing what it is.
AI models enable us to make accurate predictions and forecasts that take into account multivariate and non-linear relationships within our process. We can also monitor our individual controls at scale, never missing a beat. AI can be our all seeing eyes keeping track a virtually limitless set of concerns.
However when we use vanilla AI, and an anomaly is detected – while we are aware of this anomaly we are left with the task of diagnosing the deviation. Using the explainability of XAI, if an anomaly is detected we don’t just get alerted that the anomaly occurred but we get an insight into why – saving us valuable time in diagnosis.
Similarly, if we are using AI models to control parts of our process or recommend next best actions – a vanilla AI modeling approach would be challenging from a regulatory perspective. In a GxP setting we want to be able to adhere strictly to our manufacturing process and any deviation from that process must be justified and explainable.
This often means keeping a human in the loop. We can use an XAI approach to make intelligent recommendations complete with a justification – allowing us to act faster and with more confidence.
What are the limitations of XAI in GxP manufacturing setting?
There are many applications within the manufacturing setting where explainability may not be of importance and this is something that has to be weighted. XAI does come with additional overhead to apply in comparison to traditional AI approaches.
Where in the manufacturing process that you wish to apply Artificial Intelligence? Which of these use-cases may you expect to receive scrutiny on model outputs? Concentrating efforts here will see the largest impact. If model explainability is not important, it may not be worth the extra effort.
To understand if XAI will have a benefit within your manufacturing process it is important to have a thorough and holistic understanding of the manufacturing process. What stages of the process are the most difficult or sensitive? What are the pain points and how can AI address those? What would the AI model need to be able to do? Do we need to be able to explain the model in its entirety – or do we only want visibility on the data lineage?
To get a full understanding of this, stakeholders close to the manufacturing process are crucial to involve. Data and AI can be used in almost anything for any reason. Addressing the actual pain points of those controlling the manufacturing process, or carrying our investigations on incidents should remain the primary concern.
Conclusion
As with any Data Science project the focus should be to drive impact. In a GxP manufacturing setting that should mean improving the quality, efficiency or safety of the manufacturing process. Anything outside of that and it should be critically assessed if there is a need to apply XAI, AI or Data Science at large.