The Why, How, and Hope of Generative AI and Data Technology With Bryn Roberts
By Irene Yeh
September 26, 2023 | As artificial intelligence continues to rapidly develop, it is tempting for organizations to rush into using AI—or at least an AI buzzword. In the latest episode of the Trends from the Trenches podcast, host Stan Gloss channels Simon Sinek and challenges Bryn Roberts to start with why.
Roberts, Global Head of Data & Analytics, Roche Information Solutions, highlights the huge unmet needs in healthcare. “Ultimately for us working in the healthcare and life science space, we’re trying to improve the wellbeing of humankind,” he says, including drug development, healthcare, and preventative well care.
Editor’s Note: Bryn Roberts and Stan Gloss will speak again—live!—at the Bio-IT World Conference and Expo Europe in late November. There, they will focus on Data Citizenship and Changing Data Culture. To join the conversation live, register by October 20 for advanced savings and use code TFT10 for an additional 10% off your registration fee. You are also welcome to submit questions for the two to address at the event.
With decades of experience in pharma R&D, Roberts returns to Target Product Profiles—TPPs—tools to, “really understand who a drug is for, what benefit it will bring, how will it be delivered and experienced, and what kind of side effects to avoid for particular patient population.” TPPs should apply to digital products and approaches as well. “As we, perhaps, talk a bit more about AI, we have to start with the Target Product Profile,” he says. “What are we trying to achieve with this approach? Many of the ways we might develop, train, tune, monitor, and manage those AIs or machine learning models will depend on that Target Product Profile definition.” Is the model signal-seeking for a discovery setting, or is it more stringent for use in a clinical setting? What sort of biases or confounding factors are in the data? Are the data high-quality and well-curated? Building with the customer in mind—whether the customer is a physician or patient or even an internal scientist—leads to a better product.
It’s a “brilliant time” for AI, Roberts said. “Over the past six to twelve months there’s been this huge explosion, particularly in the generative AI space with large language models creating such a storm!” While Roberts described his group’s work as formerly, “doing cool things with data and analytics, but happily left a bit in the background,” now, thanks to ChatGPT, AI models and data analytics are “main stage.” Now executives across the various Roche business groups are asking, “What can we do with generative AI or AI generally” Roberts reports. “It’s actually giving us new opportunities, and also, new challenges,” he says. “There’s a lot of hype, a lot of, perhaps, false expectations—or expectations that are very difficult and expensive to realize.”
Artificial intelligence and machine learning began decades ago, he reminds listeners, with algorithms tuned for specific, specialized tasks. For example, digital pathology began by looking at oncology images; convolutional neural networks were well-suited for image analysis. Later, digital biomarkers in the form of time-series data from Parkinson’s Disease patient clinical trials were assessed by recurrent neural networks to track motor changes.
Today’s push into generative AI is broadening those use cases, gathering data from both the public domain and internal data sources and building or fine-tuning foundational models of which we can ask “a whole range of very general questions.” Generative AI also excels at creating synthetic data, which Roberts describes as a form of “privacy preserving technology”. He predicts synthetic twins of datasets that can be explored without compromising anonymity.
These use cases—and many others—promise to change the nature of work, Roberts says, but he’s not foreboding. For software developers, for instance, leveraging large language models can ease both coding initially—he specifically mentions GitHub’s 2023 Bio-IT World Best of Show winner CoPilot—and testing code. “I don’t think that does away with the need for great software developers or data scientists, but it has the potential to make them ridiculously more productive,” he says. “The nature of the work will change slightly and what we value may, over time, change.”
We have always had more product ideas than we can realize, Roberts says. Generative AI will help us explore more ideas, test them, and scale the winners more quickly, “bringing significantly more products and better products at a dramatically reduced cost.”
At Roche, this is the next logical step in the company's existing analytics. The four classical types of analytics are descriptive questions (seeing and searching data), diagnostic questions (Why is this happening in a system?), predictive questions (What will happen if a parameter is changed?), and prescriptive (recommendations or clinical guidelines). “For a long time, we’ve been using things like natural language processing and different forms of advanced analytics to help scientists to grapple with very large datasets and information sources like published literature, big historical banks of experimental data, clinical trial data, and so on,” Roberts says. Now AI can serve as augmented intelligence to empower a physician to bring together or grapple with the data on an individual patient or group of patients, answering the same descriptive, diagnostic, predictive, and prescriptive questions.
The volume of healthcare data is daunting, far more than a person can manage. “An AI model that really is ingesting these regularly and keeping up to date can make a great companion for that physician in augmenting decision making,” Roberts says, “making sure they don’t miss things, maybe testing, confirming, or challenging their decisions... more like a virtual healthcare assistant or decision support tool.”
“We’re at the beginning of a very exciting journey,” says Roberts. “We’ve only done a fraction of what we hope to do in the coming years.”