Story follows beneath the quote from the extended video below
“There’s very little need and this sounds cynical and I’m not really, I hope not, a very cynical person, but there really isn’t a need for a lot of manual interface when you can have it done by AI, except in the loop. For now, at least, humans in the loop is required. I wonder why that always is a requirement because we don’t have explainability of the human brain and we assume we always do things better, which is not the case.”
Thomas Senderovitz, senior vice president of data science, Novo Nordisk
The introduction of any new technology can be disruptive. Think about the printing revolution and how scribes all but became obsolete soon after. Or typists before the personal laptop gained foothold. Or these eight jobs that have disappeared over the past 50 years. Humans were central to the loop in each of these until they weren’t.
One life science executive posed a provocative question about that human centrality at an event organized by consulting firm BCG during JPM last week.
Artificial intelligence is coming for many more jobs big and small, but its threat/potential is all the more scary/exciting because it aims to replace not simply a physical skill that humans possess but rather the one capacity that has propelled us to the top of the food chain: our ability to think and make decisions.
In the field of medicine, the advent of AI therefore comes with soothing words from its developers – “it augments not replaces,” “there’s always a physician in the loop” “this will boost your efficiency”, and permutations and combinations of the above sentiment.
But machine learning is advancing at a head-spinning pace, and the new word floating around at every event at JPM was “Agentic AI.” Think chatbots but on steroids having more agency and able to act by itself without human intervention. A kind of AI that has the ability to mimic and thereby replace human judgment.
At the BCG event that sought to explore how digital health and AI are changing the healthcare industry, a Novo Nordisk executive —Thomas Senderovitz, senior vice president of data science — talked about agentic AI in the context of the Danish company’s efforts in building and automating a clinical trials infrastructure. Called FounDATA, it is a repository where all data from completed clinical trials are
pooled and prepared for insights-generation by applying a variety of AI algorithms.
“We have now 20 billion data points and we’re going to get around 1500 RCT or randomized control trial data onto the platform,” Senderovitz said. “We’re adding images, multi omics [data], we’re going to add real world data all the way up to the the claims and outcomes data and then upstream to research data. So we have .. one place for real time analytics, all agentic AI set up and and that we have done ourselves.”
The system is set up on Microsoft’s Azure Cloud and Novo is partnering — whether it be academic institutions or other companies — to bring analytical applications to gain insights from that pool of data. The system is designed to be interoperable and Senderovitz explained the goal is to make the entire value chain “automated, AI-powered”. And then he said something very interesting and thought provoking.
“There’s very little need and this sounds cynical and I’m not really a very cynical person, but there really isn’t a need for a lot of manual interface when you can have it done by AI, except in the loop,” As for now, at least, humans in the loop is required. I wonder why that always is a requirement because we don’t have explainability of the human brain and we assume we always do things better, which is not the not the case.” [bolded for emphasis]
So, where is the automation happening in Novo’s clinical trial infrastructure repository?
“So, I think we are going to see [automation] all the way from the scientific design of the protocol, the center of the protocol; the electronic data capture will disappear, [we] will pull data directly out for electronic health records. It will go directly into a flow,” Senderovitz said. “The statistical analysis plan will be automated, the analytical code will be generated, the results will go automatically and they already do into starter report.”
He noted that Novo doesn’t write starter reports manually anymore.
“Ultimately that could also be, ‘don’t submit reports, submit your data and all your code’ and then they can replicate,” he speculated about the future. “So that process we are building and it will come sooner than we believe, including scientific manuscript writing.”
He added that Novo has done GenAI manuscript writing, which he couldn’t distinguish from humans though Novo hasn’t submitted them yet.
“It’s only the New England Journal of Medicine’s AI Journal that would accept, as far as I know, Gen AI [articles], but it will come,” he said. “It’s just our resistance.”
He added that to be able to do this all this AI automation and insights-generation properly, Novo Nordisk has created an a data ethics council internally, so that these issues aren’t just “an ad-hoc discussion.” Novo also has an data governance layer to oversee information transfer.
“So every single AI which is deployed in the regulated area and/or versus patients in the real life, will have to go through that governance before [in order] to go out,” he said before noting that there are a whole host of issues that are technical, ethical and related to legal compliance that need to be addressed in such a system.
The task is even greater — from a trust perspective — as there happen to be fewer and fewer humans in the loop in the future.
“There’s a new area by which I would call explainability science or decision science, because not all models would be will be able to explain. But we have to be able to completely track how we make decisions and how decisions are made. And, the, the less we have humans in the loop, the more decisions aren’t made by humans, the more we need to at least track and be able to have that transparency.”
But Senderovitz also acknowledged a challenge given how AI technology is rapidly changing.
“You know, a year ago, we didn’t think about agentic AI or infrastructure. In half a year, agentic AI will already be a little bit outdated. It’ll be something else, right? When you’re in the regulated space that that I sit in, at a certain point of time, we have to lock something and say, this is now what we do and validate that [in such a way that] regulators and and authorities can accept. But the technology keeps evolving. So how do we balance, and I don’t have the answer yet, how do we balance that? On one hand, the technology evolves so fast. On the other hand, we need to make sure that it’s trustworthy and that we feel safe enough to deploy.”
