For a recent HTN interview, we caught up with Dr Sonam Vadera, specialty registrar in clinical radiology at University Hospitals of Leicester and fellow in clinical artificial intelligence at the AI Centre for Value Based Healthcare in London.
Sonam shared with us some insights and details around her current research, key learnings and takeaways from emerging AI applications in radiology, and the barriers in translating research on AI into practice.
To begin, Sonam offered a brief introduction to her role and background, including completing her medical education at UCL medical school, before returning to Leicester for her foundation years and radiology training.
“I’m currently an ST4 in radiology, specialising in MSK, and in July 2023 I started a fellowship in clinical AI. Currently I have a hybrid work pattern, so I work forty percent of the week doing my fellowship and sixty percent of the week in my usual radiology role. My fellowship project is based at University Hospitals of Birmingham.”
Research focus
Sonam told us that her focus at present is on AI safety monitoring frameworks, adding, “it’s really important that once an AI system is integrated into any particular clinical workflow, we continue making sure that it’s functioning as it should be, and that we’re picking up on any changes in its performance that need to be addressed”.
“It’s a method of real-world post-deployment safety monitoring, and we’re applying a framework called the medical algorithmic audit, which was developed at University Hospitals Birmingham. That provides a set of steps to follow to ensure that AI medical devices are monitored effectivelyand any deficiencies in performance are detected. We’ve applied the framework to real-world applications ofa couple of tools within the trust which we will then later publish, so that hopefully across other trusts we can start to integrate a uniform framework for safety monitoring that can be accessible to clinicians, and to multiple stakeholders working within the field.”
The benefits of this work, Sonam continues, are in helping to boost uniformity around AI safety monitoring, “because currently there’s a lot of a discord; people are working on different things, they’re doing safety monitoring in different ways, and there’s not really much uniformity across the system. It’s also really important to pick up any changes in performance over time due to data drift, and to detect any biases that may result in disparate subgroup performance.”
“My main role has been centred around AI safety monitoring of an autonomous chest X-ray reporting tool, which is essentially an AI tool that reports chest X-rays it feels are normal with high confidence, and removes them from the workflow, so they will not require a report from a clinician.”
Applying the tools to real-world data is another important step in the process, says Sonam, since “it’s one thing to have these tools and to know the sensitivity and specificity performance on the external test and training sets, but it’s another thing to then apply them to real-world data where they often perform differently, and that’s exactly what we’re doing with the audit framework”.
“There are a couple of other really interesting projects I’ve been involved with at the University Hospitals Birmingham, and one of those is the Standing Together project, which is focused on looking at the diversity of datasets and ensuring that they are diverse, generalisable and inclusive. Often there’s not transparency of the data that AI tools are trained on, and there are subgroups which might have been excluded from that particular dataset, which means that if we go and apply the tool to another population, it won’t perform the same and may not be representative. So the Standing Together team hase produced a set of standards to follow that would help ensure that there is transparency in reporting of the datasets.”
Another project Sonam is working on looks at the involvement of patients and the general public in decision-making around AI, setting up a “platform” where manufacturers can reach out to patients and members of the public with a background of literacy in AI in healthcare, to help them make decisions and get feedback.
Challenges around AI in the NHS
One of the challenges that Sonam noted specifically around AI in the NHS was the lack of education and training on the topic, since “most staff in the NHS receive very little education or specific training on digital healthcare”.
“At medical school I didn’t receive any training on digital healthcare.Although I know there are now more initiatives looking at how that can be incorporated, speaking from experience, most clinicians don’t have much insight into the applications of AI. There’s also a lot of anxiety from clinicians around adoption, especially among radiologists, in making sure that they are still valued and still kept in their role, and also knowing patient safety isn’t compromised. One phrase that always stands out to me is ‘learning to work with AI’, and I think that those clinicians who are open to working with AI might be more likely to succeed with this change in the future.”
In order to help the workforce become “more amenable” to adoption, Sonam states that it is important to make sure it’s not “overloading them with yet another thing that they need to learn, yet another thing they need to train in, in an already overstretched, overworked workforce”.
Another challenge Sonam mentioned was that “it’s really difficult for small and medium enterprises to actually break into the NHS, and I think that’s one of the biggest barriers, because they’re coming up with really novel, really interesting technologies, which could have a real impact”.
“When it comes to getting those technologies or solutions actually approved by the NHS, there are a lot of regulatory barriers, a lot of red tape, as well as massive financial constraints and IT issues. It’s one thing doing these things in a test environment, but when it comes to actually getting them into the real world, often SMEs don’t have the funds or the resources to necessarily be able to do that.”
“Things like the AI Diagnostic Fund and incentives to try to encourage these SMEs to enter the space I think are really important. There are really useful tools out there, and there is definitely a demand for them because the NHS needs help; but it’s a case of recognising those and giving them a way into hospitals, because that’s how they can thrive and gather the relevant data that they need and be applied in the right settings.”
There are also challenges when it comes to integrating AI tools into existing workflows, Sonam shared, “because we have these really defined, rigid workflows, and actually slotting in an AI system is difficult”.
“Furthermore, once you do change the workflow, and see the outcomes change, how do you go about monitoring what it was like with the AI system versus without, and are we being fair in saying that the device is having a certain impact? Not to mention the challenges aroundlarge language models and generative AI, which are theirown entity.There are enormous difficulties in safely bringing generative AI into healthcare, because how do you monitor something that has infinite outputs?”
Key learnings
We asked Sonam to share with us some of the key learnings she had already come away with from her fellowship. She discussed that one of these learnings was the importance of a collaborative approach in getting an AI tool “off the ground”, because “it’s not a one-person job, or even a one-team job; it requires the integration and involvement of multiple teams, as well as cross-site collaboration.”
“It’s important to learn from one another’s mistakes and experiences, and that’s particularly true of the NHS, where this sort of information isn’t always shared very well. The multidisciplinary team approach has always been valued in healthcare, but I think it’s never been more important, and that’s not just within the clinical teams, but also within the data science team, the administrative team, the IT teams. All of the meetings I’ve been involved with I’ve really enjoyed because we have so many different valuable voices to listen to.”
Also on teams, Sonam highlighted the importance of having “a solid digital transformation team”.
“At University Hospitals Birmingham, our digital transformation team’s role is to focus on bringing AI products to the hospital and looking at how to integrate them. But I know that a lot of hospitals don’t have a dedicated workforce for that, and I think that means that often the onus of bringing in these particular devices is left to the clinicians, or left to one particular person who’s interested in doing it, which can be somewhat impossible alongside clinical duties. So I think my experience has highlighted the importance of having a dedicated team who focuses on procuring and validating these tools.”
“Another thing I’ve learned is the importance of ensuring we recognise and addressbias within these AI tools,” Sonam continued, “as healthcare data is already biased, and having tools built using biased datasets will only amplify those biases”.
“One of the key things that I’ve taken away so far is how we can start to go about mitigating this bias, making sure datasets are diverse and transparent, ensuring that we do the safety framework monitoring to assess for disparate subgroup performance, and so on. Another thing I didn’t mention about the medical algorithmic audit is that a key component of that is a failure modes and effects analysis, which looks at all the different ways in which a tool could fail along the clinical pathway. We map that out, even prior to deployment, and the effects that could have downstream, to help us be prepared for those eventualities, and I think that is really important.”
AI for future healthcare
As a closing question, we asked Sonam what she was most excited about for the future of AI in healthcare, both within radiology and beyond.
“I’m excited about the impact that it could have on things like remote care, reaching rural populations, or improving diagnosis. I think it’s easy to focus on the UK and more developed or privileged countries, but actually I’m really excited about the impact on less developed or smaller countries that don’t have developed healthcare systems. Hopefully it can enable people to get faster diagnoses and get the healthcare that they need, because in some countries, you might have one radiologist serving thousands and thousands of people, and I’m sure that goes for other specialties as well.”
“From a local perspective, I’m excited to see how we can reallocate the time savings to streamlining the workflow. A lot of clinician time is spent doing administrative tasks, and I’m really excited to see how AI can smooth out those processes and take over some of these roles.I’m particularly excited to see the role of generative AI in that space. I think there is the potential to transform healthcare,releasing time for clinicians to give more time back to patients, and improve patient safety, not to mention give them time to focus on novel areas. I hope this will promote innovation in healthcare that we would never have time for without the AI assistance, and ultimately improve patient care.”
We’d like to thank Sonam for her time, and for sharing her insights with us.
