“People should stop training radiologists now, it is just completely obvious that within five years deep learning will do better than radiologists,” said Geoffrey Hinton (often described as the ‘Godfather of AI’), in 2016. Ten years on, there is a lot of research showing that AI assistance can improve clinical outcomes when working alongside radiologists, but in most settings they are still a long way from replacing radiologists. In fact, the need for radiologists has increased substantially in the last 10 years, and there is currently a large, and growing shortage of radiologists in the US.
It’s in people’s nature to be optimistic, and to believe that the things they are working on will have a positive impact. To assess claims of AI in health, there needs to be high-quality evidence. Clinical trials are used to create high-quality data about how safe and efficacious different interventions are.
As I set out in a CGD paper published today, there are many reasons to believe that the current clinical trial system will not be able to adequately assess AI-based interventions. However, well-designed clinical trial networks and platform trials could help overcome this challenge.
Why existing trial infrastructure falls short
There are many challenges with regulating AI in healthcare. I will focus on the three largest problems here, but a long list of problems is outlined below in Table 1.
Frequent updates: AI models can change frequently, as they are updated with new data or improved technology. This creates a problem for clinical trials, which traditionally evaluate a specific version of a product. By the time a study is completed, the model being tested might already be out of date. Freezing models would overcome this problem but would mean losing one of AI’s main advantages—its ability to improve over time.
Comparative studies: Doctors and patients don’t just want to know whether a treatment works, but whether it is better than the alternatives. This is particularly difficult with AI, where the best available product might change within months, meaning that evidence comparing different products can quickly become outdated. The follow-on head-to-head academic studies that often look to compare different products will be too slow to be useful
Different contexts: AI products may perform very differently across countries and healthcare systems. A model trained using American patient data might not work as well in India or Kenya because of differences in language, clinical practices, or patient populations. This makes testing products across multiple settings particularly important.
Table 1. How AI impacts clinical evaluation
Challenge | Is this a new challenge or one exacerbated by AI? |
|---|
Interventions change during and after evaluation | Predominantly new |
Evidence becomes outdated quickly | Existing problem, substantially exacerbated |
Need for comparative effectiveness evidence | Existing problem, exacerbated |
Performance varies across settings and populations | Existing problem, likely to be exacerbated |
Need for continuous post-market evaluation | Existing problem, substantially exacerbated |
Determining when a modified product requires new evidence | Predominantly new |
Evaluating study results when the intervention changes during a trial | Predominantly new |
Three approaches could help overcome many of these challenges
- Continuously running clinical trials would allow researchers to monitor products as they change. These trials could operate with relatively few participants most of the time, but scale up recruitment when a model is updated, or when trial data suggest performance might have changed. Once researchers are satisfied that the product remains safe and effective, recruitment could be scaled back down.
- Head-to-head trials comparing competing products would help doctors and health systems identify which interventions work best.
- Multi-country trials would allow researchers to understand how well products perform across different populations and healthcare systems.
Together, these approaches would generate the robust evidence needed to make informed decisions about AI in healthcare. However, running continuous, comparative trials across multiple countries would be expensive and difficult to set up. There is a real risk that if the burden for evidence is too high, then companies will be discouraged from developing new AI tools, and the potential of this technology isn’t reached.
Platforms and networks are part of the solution
One way to reduce the cost and difficulty of running these studies would be to use platform trials and clinical trial networks. Both approaches involve creating research infrastructure that can be used to test multiple products, rather than setting up a new clinical trial from scratch every time a product needs to be evaluated. This could make it substantially quicker and cheaper to generate the evidence needed to assess AI health interventions.
Platform trials allow multiple products to be tested within the same study using a shared network of hospitals or clinics, a common trial protocol, and, where appropriate, a shared control group. New products can be added as they become available, whilst ineffective ones can be removed. This makes these trials particularly useful when many companies are developing products to address the same clinical problem, as competing interventions can be compared directly. They can also continue running over time, allowing new products or updated versions to be evaluated as they emerge.
There is good evidence that platform trials can reduce the cost and time needed to evaluate treatments. An analysis of the STAMPEDE prostate cancer trial estimated that testing 10 interventions through separate trials would have cost 57.5 percent more and taken over four times as long in cumulative duration as using a single platform trial.
Rather than running one large trial, clinical trial networks create a permanent network of hospitals and clinics that are able to participate in different trials. Each study has its own protocol and will recruit from sites they believe are most appropriate. This makes it easier for researchers to identify trial-ready sites. Networks can also standardise rules around price, data sharing, and ethical review. These networks offer greater flexibility, making them particularly useful for evaluating products that are too different to be tested within a single platform trial. The 2023 O’Shaughnessy Review found that much of the delay in commercial clinical research comes from repeated contracting and site set-up rather than from the scientific aspects of the trial itself. Clinical trial networks can greatly reduce this burden, although the savings are likely to be significantly smaller than platform trials.
An example using mental health
One area where platform trials could be particularly useful is AI-delivered mental health therapy. There are not enough trained mental health professionals to provide support to everyone who needs it, particularly in lower-income countries. AI could help overcome this shortage by providing therapy at a much lower cost. As the technology improves, we are likely to see multiple companies developing competing products.
A platform trial could compare these products directly, using common measures such as improvements in depression or anxiety symptoms. As new products become available, they could be added to the trial, whilst those that perform poorly could be removed. When an existing model is updated, the trial could also assess whether its effectiveness has changed. Running this across multiple countries would help establish whether these products work equally well in different languages, cultures, and healthcare systems.
What needs to happen next?
Further research is needed before these ideas can be implemented. Setting up these networks will require substantial funding and these costs need to be examined alongside any possible funding models. Companies have historically been reluctant to participate in head-to-head trials, and so might be reluctant to take part in a platform trial, though regulation or government procurement could be used to overcome this barrier. There are technical and ethical challenges around model updates, data security, and ensuring access to successful products in the regions they are tested in. Further work is therefore needed to understand the costs and benefits of these approaches, identify the setting where they would be most useful, build a design that overcomes technical and ethical concerns, and assess support for the project more broadly.
AI has the potential to transform healthcare, but to have these tools fulfill their potential, we need a system that can identify which interventions work, and in what settings. If we demand too little evidence, then ineffective or unsafe products risk being widely used. If generating evidence is too expensive or takes too long, promising technologies may never reach patients.
Platform trials and clinical trial networks offer a way to overcome this trade-off, making it easier and cheaper to generate robust evidence without lowering standards. We should start exploring the feasibility of these systems now, so that we have suitable trial systems in place before an influx of AI products in healthcare.
