Medical Writing and AI…..Oh My!

AI is the technologie du jour. Reflecting on this topic, I suggest the role of the medical writer is vital to successful implementation of emerging AI technologies.

My conversations with folks who attended AMWA annual meeting in Baltimore drives this point home. I suggest AI should not be seen as the flavor of this month or this year. Keep in mind, AI has been around for some time. I recall having conversations with clients back in 2018 on how AI will (or will not) impact medical writing and the development of drug application submission documents. As I think about the notion of AI in regulatory writing, this is just like the conversations I had with clients about big data and the prospects of predictive analytics (think FDA and their tool Janus). Here we are again enjoying a moment of overheated speculation.

The greatest challenge in using AI technologies to support analyses and writing remains the same—what do we want AI tools to do for us? AI output will only be as good as the quality of the inputs. I see the issue with application of AI in writing as clearly defining the business need, the tool capabilities, and then packaging the data and context to fuel the analysis engine.

Here I want to emphasize the importance of context. Content will always be king, but context is the queen. From the beginning of the concept, AI vendors claimed that all you had to do was “train the AI tool to get what you need,” but the parameters of successful training were never defined (no different than with us human learners). “We need more training,” and “We need the right learning content,” are the often-repeated statements from AI developers. Interestingly, these two comments are just like what I get from clients seeking my group’s help with regulatory writing with "real" humans.

Some of my clients are aware that AI as currently constructed is only a content delivery platform. The issue to be addressed is in how the content is structured and more importantly, rationalized (contextualized) for the target audience. Rationalization of what data means to go "beyond the pale" of statistical analysis. Such an approach requires massive inputs from humans to create a workable ontology for the AI algorithm. I emphasize the word massive. The process is not easy to get any AI software up to speed to contextualize and summarize an already written clinical study report. AI can be pretty good at creating summaries. However, to contextualize data sets, well that is a wholly different task.

The reality is most Pharma houses have effectively checked the box when it comes to content of what we have to do to get a drug on the market. But they remain doing a poor job with the context of the content generated in their development studies. As I see it, Pharma houses do not carefully scrutinize the value of the content beyond the simple warehouse requirements. Not enough intellectual energy is placed in representing the "so what" in regulatory submission documents.

I know you’ll think I am weird with this next comment, but this is how I see the AI field right now. What is needed to make AI work is the CDC (Chief Director of Context). The CDC needs to help shepherd which silos contain what content and the “why” associated with each piece of content in the silo. Information alone won’t win the day with leveraging AI. Context as to why a specific piece of information is important, now that is the difference maker.

I do not hear enough in discussions with clients about: What’s the purpose of this data set? Why these exclusion criteria are appropriate for the field we are studying? Why this analysis approach is the most appropriate? Why should someone read this sub-section of our briefing book? I guarantee you that no AI engine is going to tell you why someone needs to read the content or why you generated a data set. All AI is going to do is cackle that the study was designed and conducted per ICH guidance.

Why do we create content in the first place (other than to comply with a regulation)? The answer should be, “To enable answers to questions.” The AI engine will create content because the engine is supposed to create content. That is the job as per software code. However, to fully contextualize all the content. Well, for AI we are not at this point, yet.

Why do regulatory review agency readers have problems? Usually because information is not designed correctly for them. I want to emphasize this point: FOR THEM. This is what counts. Not designed for you or what an AI software algorithm establishes as “correct”.

AI today is only a tool to solve a finite set of analyses and communication concerns. AI is not the magic ingredient sought since the time of the alchemists. Organizations are realizing this the hard way while they’re spending millions of dollars on AI technologies that will most likely fail. We’ll get there, eventually. But the "there" is not today.