In developed economies, it is expected, and assumed, that any authorized drug or medicinal therapy designed for human use is unequivocally safe. Pharmacovigilance processes – the discipline of continuously monitoring the effects of drugs once on the market – are designed to uphold that position over time, once a product has been authorized for patient consumption.
Approaches to post-market Safety monitoring have changed little in decades, however, despite soaring volumes of available information – submitted in an increasing array of formats, via a proliferating range of channels.
Many companies still take the approach of “booking” cases or determining whether the mandatory elements are present, to quickly assign an identifier. This approach doesn’t take into account the actual contents of a case, forcing Pharmacovigilance teams to apply the same treatment to all information. This means all cases are assigned the same priority in the early stages; there is no discretion to allocate teams’ bandwidth according to a potential case’s complexity or risk. Keeping track of all of the potential signals, assessing their validity, and responding swiftly to relevant cues, is both an absolute mandate and a very costly and labor-intensive administrative burden.
In particular, adverse event (AE) case intake represents one of the most overlooked and broken workflows in biopharma in its current form – costing companies dearly to deliver, without being sufficiently timely or accurate to reliably contain emerging safety issues or stave off costly recalls.
At last: robotic automation gives way to human-like deductions
Technology-enabled process automation has long promised to transform the speed, efficiency, and accuracy of AE case intake and triage, by capturing and assessing relevant Safety signals arising via a wide range of channels (including self- or clinician-reported AEs submitted by email, post, phone call, or web portal, as well as mentions via online forums).
Up to now, machine intelligence has not come close to mimicking human powers of data extraction, filtering, inference, or deduction. The early excitement about this potential – while not misplaced – was premature. Early automation systems had to be highly structured and painstakingly trained to recognize every possible format and variant of how important data might show up – from basic information such as the patient’s date of birth to richer detail such as the combination of possible contributors to the adverse event (from the individual’s age and general state of health to the other drugs they may be taking). This compromised the systems’ ability even just to recognize and extract the right data. This in turn limited the scope for real step changes in process efficiency and Safety delivery.
But now Generative AI (GenAI) and Large Language Models (LLMs) are beginning to fundamentally transform biopharma R&D adverse event data collection and their associated workflows, with powerful results. In early pilots, data extraction accuracy and quality have exceeded 90 percent, and overall efficiency gains related to the intake process have topped 65 percent. And that’s from a standing start; results will only improve with human oversight and AI adjustments.
The urgent appetite for advanced automation
New advanced automation solutions, which transform the data collection part of the AE case intake process and associated workflows, are resonating hard in an industry that has been crying out for a modern, more efficient way to execute case intake/safety data collection, as volumes of case data soar and pressure mounts to accelerate analysis times.
A recent industry survey (1) shone a light on the biopharma industry’s growing appetite for AI-powered automation, revealing that over 75% of biopharma R&D organizations already use some form of advanced automation within daily processes today, and more than 70% plan to expand business process automation over the next 18 months.
This hunger for viable solutions has intensified in line with a maturation of AI-powered process automation technology, from early robotic process automation aligned to regimented processes (guided by strict structure and rules, and specified workflow around exceptions management), to a less inhibited approach where the technology understands much more about what it is looking for (irrespective of format), and what to do with it.
GenAI technology, using LLMs, can quickly identify and infer what’s relevant and important and reliably summarize key findings for the user – and even extrapolate from them to make predictions about future scenarios. It can do this without the need for painstaking ‘training’ (from scratch) by overstretched teams, and without protracted system validation. Indeed, specialized applications can now be developed that can apply GenAI-type techniques, contextually, to data they haven’t seen before – learning from and processing the contents on the fly.
This is a huge leap forward that has seen biopharma companies start to put GenAI AE intake solutions to the test in their operations, under the watchful eye of their Safety professionals. The ability to simply instruct a system to “Scan X document for Y contents” paves the way to faster, higher-quality extraction of more relevant data, no matter how much greater in volume this is, or how much more diverse or complex the sources – reducing the risk of something significant being missed, and improving downstream efficiency.
Removing bias from deductions
A strong aspect of the business case for harnessing GenAI in AE case intake management comes from the scope for handling first-line capture and processing of very high volumes of data – relieving Safety professionals from that labor intensity and allowing them to delve deeper into the findings and what they might mean.
However letting GenAI take the strain of case intake also removes human limitations such as fatigue, mental overload, distraction, data blindness, and unconscious bias. An AI-powered tool can more efficiently detect patterns and determine trends, with reliable consistency using approaches that are based on precedence. It can draw on the findings of millions of prior cases and assessments, to make credible predictions and unbiased assessments regarding causality (the likelihood of a direct link between a product and a reported adverse event), that are based on probability rather than a gut feel.
Next-generation cognitive computing in the form of GenAI and LLMs – might be considered to be in an adolescent state of maturity currently (largely ready for the world, with some guidance and controls still needed), but the early output is proving very encouraging. Teams are now seeing that the level of oversight, quality review, and sampling that is required to satisfy regulators, develop a track record, and build trust in the technology (e.g. its process of learning and decision-making) – is a relatively low hurdle to clear. It helps that the links back to the sources are readily traceable for checking.
Maximizing the transformation potential
Going forward, as biopharma companies look to capitalize on GenAI and LLMs to advance their process automation goals, they mustn’t focus solely on the potential bottom-line benefits. After all, this is a much-needed chance to re-allocate resources; to elevate Safety professionals’ roles from data management to adding new, strategic insight-driven value to R&D decision-making.
This then requires provision for change management and transformation ‘readiness’, not just a choice of the right technology for the job.
In the short term, it is AE case intake that has captured companies’ imagination – where unprecedented new insights as well as greater process efficiency promise to revolutionize the function, and its role and value, starting right now. But over time there will be other powerful use cases too, so it’s a good idea to allow scope for additional applications in due course (e.g. by deploying an enabling ‘platform’ rather than a single-use application). Strong next contenders for GenAI/LLM treatment include real-time pharmacovigilance assessments and associated decision-making (e.g. the earlier identification of unexpected benefits/discovery of new indications); harnessing international Regulatory intelligence to transform marketing authorization applications and maintenance; and clinical trial modeling, reducing the reliance on traditional clinical studies.
The key to whether GenAI/LLM treatment is appropriate will be the high volumes of data involved in the target processes. Certainly, the more opportunities there are for the advanced automation system to be exposed to information, the faster it will learn to identify, categorize, assess, and deduce what to do, driving ever greater trust in – and reliance – on the technology to do the heavy lifting.
Where an organization has not already started on a path toward increased utilization of automation techniques, a recommended first step would be to break down how processes are currently managed, the core requirements driving those processes, and where the pain points are. The next priority should be to review and rewrite standard operation procedures so that they can evolve with and be improved by advanced technology, both today and as AI capabilities continue to evolve.
References and notes
- ArisGlobal’s 2024 Industry Survey Report, Life Sciences R&D Transformation: Ambitions for Intelligent Automation & Today’s Reality.
