Medtech 2025 Expert AI Predictions: ‘Biggest Leap Leveraging AI In Oncology,’ Says Verily’s Andrew Trister; Multimodal Integration, Use Of AI Agents

The area poised to make the biggest leap leveraging AI is in oncology, both for new target delineation and drug discovery, but also in patient-matching for improved clinical trial design and efficiency, and care delivery.  

As “bucket trials” mature, better use of molecular information collected outside of a clinical trial will be suited to provide access to the best first treatment in newly diagnosed patients. This concept of real-world data driving clinical decision support will be a simple application for generative AI tools in oncology. 

The other major trend in this space focuses on the agentic nature of models built on top of the large generative models.  

In 2025, we will see more infrastructure and personalized information across the spectrum in medicine to improve patient engagement and change how a person might gain access to care. 

For example, before the end of 2025, we will see prototypes of AI agents that remember a person, act on their behalf and to their individual goals using their own personal health records. This will lead to a renaissance in patient engagement for clinicians to focus on their care. In augmenting the care provider, these agents can enable a more personalized, effective, equitable, and cost-effective care system. 

Given the infrastructure and data available around Generative AI, it does seem that agentic models will make the biggest strides in chronic disease management. By incorporating the existing data available within the health system with data that is person-generated (through wearables, or other types of data) we will have an unprecedented approach to building proactive agents that can help a person better understand their own health and disease. Traditional AI systems (those that do not use pretrained data or encoder/decoder transformers) are more brittle to the individual needs but will continue to be important in most health applications for the foreseeable future. 

In 2025, responsible AI must be a priority for every organization and AI tools must be implemented in a way to not exacerbate already existing health inequities. Because large language models (LLMs) are only as good as the data they are trained on, we need to ensure we’re training and feeding AI on representative and accurate data. We also need to use LLMs purpose-built for health care. This means generating and leveraging health care-specific, unbiased data from a variety of sources to develop and validate these models.    

That said, in many cases, none of these AI systems can leverage the vast amounts of data generated in the health system today since those data are siloed and fragmented. This is something we’re focused on at Verily – stitching disparate data together and building the AI infrastructure to demonstrate the value of insights from these data to advance more personalized health care. 

Education. We need to help educate clinicians and patients on the opportunities for AI to help enhance workflows and deliver more personalized care. For example, how can we leverage AI to enhance the patient experience and the patient’s relationship with their physician beyond the four walls of the clinic? How can we leverage AI to increase a provider’s presence with their patient? Through thoughtful design and considerate implementation of machine learning models.   

We also must ensure we leverage AI in a way that maintains control for both the clinician and patient. This means ensuring both the clinician and the patient, not the technology, remain the mainstays of authority when it comes to what’s best for their care. 

Very few AI tools are reimbursed, and it doesn't feel like that's going to change any time soon. We don’t have the ability to fund really robust AI clinical trials. The other issue is that drug performance doesn't drift every few months or every few years. AI is much more variable. You can’t just approve an AI tool and expect it to work for the next 10 years, because the standard it is running on, the population, everything about how these tools are running on causes bias and drift.  

I think the current FDA approval mechanism just doesn't work for AI, not just in terms of funding. The reason why post-marketing surveillance is much more important than initial validation is because AI tools don't always reliably perform. The human body has been the same for centuries, so you can reasonably expect a drug that worked a year ago to work 10 years from now. You cannot expect the same of AI tools.  

We need to make sure that we're expecting validation of these AI tools on more diverse populations. That has been a stated goal and expectation by the FDA and others. I think that's important. Do you expect the AI tool to be valid on every single type of X-ray scanner that exists out there? That seems unfair to the developer, even though, honestly, like that has a huge impact on how these AI tools perform, right? You will never be able to break down all the variables that could impact AI performance. That is why we need post-marketing surveillance.

I think you’re going to see a lot of documentation automation that’s really been going on for the last decade, because of Generative AI. One thing that’s really exciting is multimodal Generative AI. Historically, we have had an AI tool for your X-ray, an AI tool for reporting, an AI tool for your billing. You’re seeing these things being integrated in multimodal ways where you have an AI tool for your X-ray that automatically generates a report and spits out the billing codes. A lot of these Generative AI solutions initially have just been documentation but now you can incorporate other types of findings. 

That presents a really interesting question for health systems, because, fundamentally, there's not going to be any one vendor. For [Ferrum Health], ultimately, there are going be some solutions that offer an end-to-end Generative AI workflow. We think of AI as a point solution, such as to help diagnose cancer or write a report, but Generative AI tools and agentive models (that are designed to act with a degree of autonomy), they're almost more powerful as a glue rather than the solution. I think the generalizability of the AI agents is going to allow them to leapfrog a generation of technology.  

With AI agents, you don't really need legacy players [like Epic and Cerner] to play ball, because you can just grab the data, whatever data they're putting out, and whatever format it is, and you can just get a screenshot of your Epic screen, give it into ChatGPT and it will tell me what the codes are.

You can pretty much get all the data you need to build the AI agent. That’s how you see all these tools entering the market. I feel like we are already at the point of deeper interoperability, being able to manipulate interfaces. There are things we can’t do, but at least in terms of access to data, we are really far along. The barriers of entry are lowered with more interoperability. They are further lowered with the advent of Generative AI and Generative AI agents. You can build a solution that automates huge portions of care coordination or sepsis diagnosis or you have solutions that can be easily built that generate a ton of value for providers across hundreds of use cases. 

The key for providers is going to be, “What is that glue that stitches it all together? How am going to build a coherent AI-powered system from all these different use cases of points solutions.“ The glue is unlikely to be electronic health records. EHR is a glue for patient data [and] provider communications and coding, but it’s not going to be the glue for AI application. That glue will be an AI vendor-neutral platform that is able to host those AI applications, able to monitor them, making sure the right data is going to the AI vendor at the right time and those outputs are being fed into the right subsequent AI agent or subsequent user. 

We’re moving from a world of patient data where the EHR has a chokehold on it to a world of AI-generated insights.

As a leader in breast and cervical cancer screenings, we’ve invested heavily in the use of AI within our Genius Digital Diagnostic and Genius 3D Mammography systems. These technologies help with diagnostic review to improve workflow and reduce the burden on health care providers and lab professionals.  

The AI technology used in Genius Digital Diagnostic identifies suspicious areas on slides, creating a gallery of clinically relevant cells for cytologists to review instead of manually searching for abnormalities. This saves time, reduces false negatives and shortens the wait for a patient’s diagnosis. Cytologists and pathologists can securely review cases remotely, so patients can benefit from the collective knowledge of geographically dispersed experts.  

Our recently introduced next-generation Genius AI Detection PRO solution, which features a new mammography AI assistant with features that can help improve specificity, accelerate image review and support radiologists.  

With AI, one additional cancer is being caught for every 10 noticed by humans, and we’re seeing a 70% reduction in false positives.  

AI has been a part of health care for decades. Today, the focus is on striking a balance between regulation, innovation and patient safety as new technologies emerge. Ensuring these devices are safe and effective for their intended use is critical, while also fostering advancements in the field.  

The FDA has approved nearly 700 medical devices incorporating AI, with most applications in radiology. As an industry, we continue to adapt and evolve as providers, lab professionals and patients grow more comfortable with integrating AI in the care continuum. Last year, Hologic put together a plan outlining 50 different use cases for AI across the company, and we’ll continue to build on that progress in 2025.