[Abstract neural network image from Firefly]
2025 was predicted to be the “Agent Year,” referring to AI technology that allows users to perform autonomous tasks. And the trends are lined up. “I think the agents are already here,” said Raja Shankar of Machine Learning, vice president of IQVIA. AI agents also help streamline clinical trials as they have gained adoption in retail, financial services, customer service, software development and more. “We've already built agents and deployed them or offered them on a small scale to our customers,” Shankar said.
Earlier in the year, IQVIA announced that it had established a contract with the AI giant Nvidia to develop agent automation for “complex and time-consuming workflows throughout the treatment lifecycle.”
When asked about the backstory of their collaboration with Nvidia, Shankar said, “We are the largest provider of almost all kinds of services to life science companies, not just clinical trials, so the Tech giant contacted IQVIA.
Currently, IQVIA is building a variety of agent approaches with partners. “One of them will be produced soon, and at some point at the end of the year, we will be deploying some of these agents into our workflow,” Shankar said. Areas of promising Agent AI in clinical trials include early detection of potential compliance issues, analyzing registration rates, identifying registration delays tied to specific sites, and adapting as needed.
Faster and faster economics using AI/ML tools
Raja Shankar
Whether it's agent AI or more traditional ML, the applications of artificial intelligence are so multifaceted in life science that they can't be said to be almost out of reach in theory, Shankar said. “Everything we do can be used to use AI to make it better, faster, better quality, etc.”
With so many potential uses, it is of course helpful to focus on the economics of AI projects, including ROI. “The most important thing to focus doesn't come from the fact that 'Is this something that can be improved with AI?' It's where we get the most out of the back. ”
Shankar sees content generation as one of the first areas in which Genai can help. “All kinds of content generation – Informed consent form generation, CSR [clinical study report] Generation, Proposal Generation, You can do a literature review – all of these things are content generation that has great value from AI,” he said.
However, because content is being created in a highly regulated industry, it's not simply about giving bots complex allocations to finish at once. “You can't write the whole content at once, especially with Agent AI,” Shankar said. “Every content requires source material, extract information, create different sections, and have another agent check the work of the previous agent.”
While AI can accelerate many individual tasks in clinical trials, Shankar points to the fundamental constraint that technology alone cannot solve: patient recruitment. “If you have 10 companies that are testing with the same signs, they're all competing for the same patients,” Shankar said. “It's a finite pool of patients…you can't create a new patient.”
Get the genai system back on track
The central technical challenge of generator AI was the risk of “hastisation” or misinformation generation. Accuracy cannot be negotiated in high-stake environments such as clinical trials. Shankar explained how this could be managed. “The hallucination part is largely resolved by simply saying, “I'm just looking at the source and providing information, or I'm going to do an analysis based on this source,” he pointed out. Reliability increases dramatically by grounding AI to specific data, adding verification steps, and adding scaffolding using methods such as search and augmentation generation. “So you can actually make these work really well.”
Shankar points out that IQVIA will strictly assess the quality of the output of such systems before deploying such systems. “If you write 100 CSRs, you need to have a benchmark of 99% of the time, or whatever the benchmark is. “And you also need a human in the loop, because it might give us a very good draft. And even the agents, it's 70% of the way, even if they're there. And you have a human who does the final test to make sure it's right.
Towards the human/machine “Dream Team”
While keeping humans in a loop is important for AI workflows, Shankar also mentioned the promise to include people with sophisticated networks of AI and ML models. Therefore, human-machine collaboration is more multidimensional. For example, an agent framework can call other AI/ML models to create a multi-agent framework. This is a type of “dream team” that combines diverse AI models with human expertise to comprehensively address long-standing challenges of R&D.
Such a “dream team” is complicated to orchestrate. “Using a generated AI or foundation model allows you to leverage these models for multiple tasks instead of using one model per task,” Shankar explained. “We are investigating the use of generated AI models to make some of these predictions with a small number of shot learning that may be as good as using traditional single-task machine learning models.”
Ultimate Goal: Multimodal Foundation Model
Beyond the acceleration of individual tasks, the convergence of diverse data types via multimodal approaches promises to change the way disease biology is understood and predicted treatment outcomes. Because biology is inherently multimodal, the multimodal approach is essentially suitable for life sciences. For example, genes affect proteins, which affect cell function and affect tissues and organs. Additionally, at the macro level, many factors affect clinical trial results, such as demographics, environmental conditions, treatment history, and real-world clinical practice. In addition to that, there are various imaging modalities, electronic health records, real-world evidence, and patient-reported results. “If all of this can be combined to create a multimodal foundation model across multiple types of biological data at scale, this could potentially change the game,” Shankar said.
IQVIA has been building machine learning models for over a decade using approaches such as Xgboost and Random Forest in many applications. “We use these models to build a model for country and site selection, registration forecasting,” Shankar said. IQVIA uses traditional ML to inform clinical trial designs. “If you want to find a responder for different action mechanisms, you can build a machine learning model to find a responder,” he added. Such machine learning models can also help find people who are likely to progress from one treatment to another, predicting the success of clinical trials, likelihood of regulatory approval, and commercial potential.
Currently, IQVIA integrates individual machine learning models and generative AI approaches, from random forest tree-based approaches to neural networks. “What has changed in the Generated AI or Foundation Model is that instead of having one model for each task, we can utilize these models for multiple tasks,” he said.
IQVIA also explores the possibility that generative AI not only complements, but potentially replaces traditional machine learning approaches for several regression or classification tasks. “It's still in the early stages,” Shankar added. “However, there are some promises that using these generative AI models can make predictions as good as using traditional single-task machine learning models.”
Even with buy-in, a journey from a working model to a scalable enterprise-grade tool is a major commitment. A successful demonstration is one thing. A robust and production-ready system is another system. “What happens in a lot of companies is that there are really good data scientists who are ok laughing about something and building something that works really well. But if you need to deploy it and give it to a company or an outsider, you need the whole software engineering around it.”
Submitted below: Data Science, Machine Learning, AI
