Submit when the Science is Done
Learn how to make agentic AI trustworthy in clinical trials, with your experts in control, always.
One of the advantages of being semi-retired and independent, is that it affords time to reflect on the past, present and future of Statistical Programming. Not too long ago however, I recall a discussion in the Leadership team in which I sat as the Data Science VP where we looked at two scenarios regarding the future use of AI in Biostatistics and Data Science.
Scenario 1: Do nothing, continue as we are
Scenario 2: Consider how we would attempt to shape our future to take advantage of AI
We concluded that Scenario 1 amounted to extinction as sooner or later a 3rd party would harness AI in a way that works effectively enough to make our current way of working (and ourselves) largely redundant.
With Scenario 2, we would seek to understand and utilize AI well enough to actively promote best practice both within and outside the company. We all agreed that unconstrained AI was too risky but with the correct guardrails applied could be a seismic shift in terms of how we would work in the future.
This was over two years ago, and even though AI has yet to mature enough to bring about Scenario 1, solid advances have been made, and we are seeing a pathway for the changing role of both Statisticians and Statistical Programmers. So, two key questions arise, how do we get there and what does this look like?
There is no doubt as to the potential of AI to revolutionise our industry, remove the busy work from humans and suddenly you have the potential to do more with the same amount of resources, be this more trials, more indications to explore, faster analysis and submissions, faster approvals, and maybe even cheaper medicines and medical devices to the benefit of patients everywhere.
The risks of course can be summarised easily as “Who checks the AI?”. We are all patients at some point in our lives and blind acceptance of outcomes generated through AI is potentially dangerous to our individual health as the lessons of the past are still valid today hence why ICH and GCP exist.
The answer of course is that there must be some degree of human oversight and review and here is the opportunity here for both Statisticians and Statistical Programmers to embed themselves within this essential process to validate and help build confidence with the use and results from AI driven processes.
Another way of looking at this is to ask yourself - who will otherwise answer the inevitable regulatory questions post-submission and how easy will it be to understand what was originally submitted and led to those questions being asked? This of course includes any support for possible advisory committee meetings.
So here to me is the crux of the coming change, the need for Statisticians and Statistical Programmers to be hands-on with analysis and programming will reduce to almost nothing, but instead the need to verify, understand, explain and facilitate trust in the results will be stronger than ever. This includes correcting the AI when it inevitably gets things wrong.
Something TBD is that we might also see applied Statisticians and Statistical Programmers merging in a single Data Scientist role in the future as the need for separate roles may well diminish due to overlapping responsibilities.
Quality by design of processes, standards, validation of outcomes and understanding or explaining those outcomes is of course critical to success. This inevitably requires investment in training and upskilling of staff in the application of AI within Biometrics processes and importantly how LLMs are trained to be as good as they can be, recalling that garbage-in, garbage-out is every bit as true today. This would include at least some understanding of the following:
By doing so, and with, collectively, thousands of person-years of industry experience, Biometrics can finally operate less as a service function and more as a strategic partner to the industry whilst leveraging their current role as both guardians and explainers of the data.
So where could this go in the future?
Imagine that, instead of a dossier, your organization submits an integrated platform containing all of your trusted data, outcomes and interpretation. This platform would be supported by your own Data Scientists, contain embedded AI tools which in turn supports full traceability of data and results. The platform would allow reviewers to verify and interrogate trial outcomes including the ability to run their own independent confirmatory analyses.
In this scenario, reviewers can perhaps answer a lot of their own questions and even where needed have direct access to Data Scientists at the sponsor company to clarify anything as needed in almost real-time, potentially leading to faster approval times.
This is closer than it sounds, Verisian is building exactly this today. Its foundation is the Verisian Harness: code traceability that connects every result back through its derivation to the source data. AI agents work on top of it to plan, implement, validate and document the analysis, and experts validate their work at defined quality gates. So what the AI generates can be validated rather than assumed correct, and what a reviewer sees can be interrogated rather than merely read.
Disclaimer: The views expressed in this article are entirely my own, and should not be assumed to represent those of my client(s).
Andrew York has worked in statistical programming for nearly forty years, more than thirty of them in management, most recently as Vice President of Clinical Data Science at Novo Nordisk. Over that period he has contributed to the adoption of double programming, the implementation of CDISC standards, and the transitions to R and to the use of artificial intelligence in the discipline. He is a strategic advisor to Verisian.
See Verisian in action. Share a few details and we’ll be in touch to schedule your demo.
Your request has been received. A member of the Verisian team will contact you shortly to arrange a demo.
