Hip surveillance in children with cerebral palsy in the UK

Listen to Andrew Duckworth, Katie Hughes and Daniel Perry discuss the paper 'Hip surveillance in children with cerebral palsy in the UK' published in the July 2025 issue of The Bone & Joint Journal.

Click here to read the paper.

Find out as soon as the next episode is live by following us on X (Twitter), Instagram, LinkedIn, Tik Tok or Facebook!

Transcript

[00:00:00] Welcome back everyone to our BJJ podcast series. I'm Andrew Duckworth and a warm welcome back from your team here at The Bone & Joint Journal. As always, we'd like to thank you all for your continued comments and support, as well as a big gratitude to our many authors and colleagues who take part in the series that highlights just some of the great work published by our authors each month.

So today for our monthly podcast, I have the pleasure of being joined by two authors from two annotations and one original paper that was published in the July edition of the BJJ. So firstly, I'm very pleased to be joined by one of our awesome trainees here in Edinburgh, Katie Hughes. Katie, thank you so much for taking the time to join us today.

Thanks for having us. Katie and I are delighted to welcome back our amazing Specialty Editor for paediatrics here at the BJJ, professor Dan Perry. Dan, always great to have you back with us. Always a pleasure, Andrew. So Katie, maybe I can kick off with yourself. I thought we'd start by discussing your really interesting annotation which was entitled 'Hip surveillance in children with cerebral palsy in the UK: history, challenges and future directions'.

Could you maybe start by giving us a brief background of why hip surveillance is so important in children with CP and some history of the [00:01:00] programmes that are there and how these have evolved? Yeah. Thank you so much. Thanks again for having us to talk about this amazing topic. So cerebral palsy is really common, the most common cause of childhood disability and about one in a thousand births worldwide.

And the severity of CP really ranges from children who have no significant physical limitation to those that need a lot of help with day-to-day activities. Maybe use a wheelchair to get around. And as orthopaedic surgeons, one of the main things we get involved in is hip health for these children because their muscles are often really tight, their femoral head can actually get pulled out of joint gradually over time.

And this can be really bad if left unchecked, the hip can completely dislocate and then they're left in chronic pain. They can't get comfortable in a wheelchair. It's really bad news. And there's no real clinical way of detecting this hip migration. Children might not be able to tell us if the hips becoming sore, and we can't necessarily tell if it's coming out from examining them.

So we rely on taking lots of radiographs of the hips and just visually inspecting them to see if they're starting to move out of joint. [00:02:00] And there's lots of angles that we can calculate, but the main one really uses this thing called Reimers migration percentage, where we quantify how lateral the femoral head is moving.

And this is kind of where hip surveillance has come in, because I think back in the day, we did this in a very reactive approach. We wouldn't know about a kid's hip dislocating until it was too late. But through a surveillance programmes, which allows a proactive planned programmes of looking at these hips, we'll know ahead of time if a hip is starting to move, and then we can intervene in a more kind of proactive way.

Great. And that's, that's really interesting. Katie is sort of in terms of. How, you know, the involvement we have with these patients, but obviously they have multiple aspects of medical care as part of their day-to-day lives. But in terms of like, because obviously you've talked about why we, why we need to surveil them, but can you just give us an idea about the difference between surveillance and screening though?

Because they're two different things, aren't they? Yeah, no, they definitely are, and it's important to make the distinction. Where we are currently with CP hip surgery, the, the operation that you're going to do for a hip that's half out and a hip [00:03:00] that's a hundred percent out isn't necessarily different.

It's still gonna be some kind of osteotomy surgery to put the femoral head back where it belongs. But the important thing is it's much often easier to do that operation when the hip is any halfway out than if it. All the way out and has been out for years. So the definition of a screening programmes is that it's able to detect and intervene in a disease process to change the trajectory of that disease.

This isn't necessarily what we're doing in a surveillance programmes. It's more about tracking changes over time and intervening in a more tailored, proactive way. Yeah. And in terms of what the, the benefits of that are, you know, there's, there's obviously the clear benefits to the patients, but there there potentially other benefits to that as well, aren't they?

For the health system as a whole? Yeah, definitely. I think you can't underestimate the value of, for these children and their families who are being enrolled in a programmes where they have a consistent point of contact with a health professional every year. And the amazing thing certainly about our UK-based surveillance programmess is that they're really multidisciplinary, so they're really led by specialist paediatric physiotherapists [00:04:00] who do yearly clinical examinations alongside the hip x-ray.

So it's this amazing kind of support network for families. Actually putting a number on the, the financial implications is challenging because of regional variations and challenges with coding. But in NHS England, they estimate that if programmes should save about six million pounds per annum, which is amazing.

That is an amazing number. In terms of, just before I come to Dan, in terms of the, the current surveilling programmes we have here, can you just give us a bit of background to that? Yeah, so we've probably done hip surveillance in ad hoc kind of way since the 1950s in the UK and then Mercer Rang in Toronto in the seventies.

Real really trailblazed it in Canada. And then the Swedish Group CPUP that started in the early nineties was probably the first to do it at a national level in a systematic way. We then started our Scottish surveillance programmes in 2013, followed by England and Wales in 2016, and then Ireland in 2019.

So we in Scotland have really good regional [00:05:00] capture of pretty much our whole Scottish paediatric CP population, which is, which is amazing. Yeah. That is really a really interesting and really impressive, and Dan, maybe if I could come to yourself now, you know. You know, Katie's given us a really brief over, really, really clear overview of the, the surveillance programmes and, you know, the potential benefits of it.

But what are some of the challenges these programmes have and maybe lead us into what, what, how can we use technology maybe to help with that? So I think Katie has already explained that, that, that these programmes are really, really good. But they rely very heavily on, on other people leading the programme.

I.e. primarily not, not doctors leading the programmes 'cause we're often the worst people to lead these sort of programmes. And so physiotherapists are, are very much the backbone of the programmes throughout the world and they, they really take the bull by the horns and, and look after these kids. I guess the biggest problem with hips though.

Is that whilst physiotherapists are great at looking at their, their overall kind of function and, and looking at the kids from a physical point of view, that they're not really trained to look at radiographs. [00:06:00] And it's often completely outside their comfort zone to, to look at the radiographs. And we're putting them in a, in a tricky position.

And so often we then have this kind of relationship where you've got a surgeon and a physiotherapist and the physiotherapist trying to get a surgeon or a radiologist or someone else to look at those x-rays. And that's kind of challenging. Some, some physios have been trained to, to do the extra interpretation, but at the same time, they're not always comfortable.

And of course they're not, they're not experts, at least at first when they start doing it, they're not experts. So we need a, we need a way to repeatedly look at these radiographs in a way that makes everyone comfortable. And I think that's the real opportunity that AI can bring. So, AI or other ways of doing this, it, it can

to use a fancy word, it can democratise how we, how we how we're able to do this so we can allow other people to do this in a repeatable way. That's, that's really easy. And that's simple to do and I, I think that's the real opportunity for AI to, to really, really get into this pathway.

I think that's a really interesting point, Dan. I think because I think, you know, as, as you and I know, and we all know you know, there's an explosion of the use of AI or [00:07:00] machine learning, whatever you wanna call it, in, in the literature and in the, and in the specialty and across medicine. And some of that is probably, probably misguided a little bit.

And some of it's very, very useful. And I think this is a perfect example of where actually you're freeing up the expertise of those people you've just described and using technology in a really positive way so that they can actually spend more time with their patients and do all the other things that we need to do day to day.

So Katie, I think that really brings us very nicely onto your, your work. Really interesting study. And this is looking at the use of fully automated measurements of paediatric CP pelvic radiographs with an ML tool, machine learning tool, I should say, called Bone Finder. So briefly, what were the aims of, of this study that that have been published?

Yeah, so exactly as Dan said, we were having this challenge in CPIP, so we basically had like a backlog of pelvic radiographs that needed to be analysed for our surveillance programme. This all very serendipitously worked out with this company called Bone Finder from the University of Manchester approaching us saying, well, we've developed this tool that can look at a CP pelvic radiograph and automatically do [00:08:00] these measurements, but

do you wanna give it a go? Do you wanna see if it actually works? And that really piqued our interest 'cause we thought, oh, well this sounds great. And also I was quite sceptical whether the tool would actually work and be beneficial. And so that's kind of the, the background of how it came about. We also gave us an opportunity to test another tool.

Something called HipScreen, which is a, a smartphone application that can measure hip RMP as well, which we were starting to explore maybe using them and CPIP. So we wanted to put that to the test as well. The Evelina group, group did a paper about that in the BJO couple of years ago. So the aims really were to put Bone Finder to the test against the clinical experts that currently do the measurements as part of the CPIP surveillance programme.

Perfect. And so this was sort of an, an external validation study and, and as you've said in your paper, it sort of followed the claim evaluation and the clinical practical integration of AI frameworks, which I think are really important to highlight. And so, could you tell us about the, you've sort of mentioned it already, but briefly, but so you tell us about the source of the external validation data that you [00:09:00] used.

Yeah. One of the amazing things about CPIP is it's got a really strong research background, already published a lot in this space, so they're always keen to, to take new projects on. So our essential aim was to take a random sample of pelvic radiographs from the historical CPIP database. So we took 509 radiographs totally at random over since its inception of the CPIPs database.

And we ended up with a really wide range of children of different severities of CP, different ages. Some had metal work in their hips, some didn't. Really kinda a veritable feast of children with CP that we thought was a representative sample to put in a machine learning tool like this to the test on.

And, and in terms of when, when the images were assessed, how, who, who did that and how was that done in the various sort of ways it was measured and automated. Sure. So each radiograph was essentially looked at or measured by six different people or things. The first measurement we had was the historical measurement from the CPIPs database, so that would've been made by, as Dan said, [00:10:00] probably by a paediatric orthopaedic surgeon, maybe your paediatric radiologist perhaps a physiotherapist that's basically been stored on the CPIPs database.

We then got Bone Finder, the machine learning tool to measure the radiograph. And then for the purpose of the study, we also got two clinicians to measure the same radiographs again. First of all, they did it manually using a a PAX tool that we most of us to be familiar with. And then they also used the HipScreen application as well.

Perfect. And I think I just for our listeners, I think it's really, really worth, obviously you going to read the paper anyway, but actually the methodology that you've used is so clear. And actually for when you were trying to integrate AI like we are, or using AI in the health, health setting. I think it's, it's a really, really something to be looked at and actually try and try and mirror.

And so in terms of what you found, the results, you know, what were the key sort of findings, I suppose as well, key findings, but any, in the context of any limitations you thought that were there? Definitely. So I think in short, the key finding is that the machine learning tool worked in an equivalent way to a clinician.

The statistics we did to look at the [00:11:00] degree of agreement between the raters was really, really positive. It showed them that the, the ICC value was nearly at 0.9%, so almost nearly nearing like perfect agreement between raters, but probably most importantly is looking at the absolute difference between raters in the RMP measurement itself, and that sat around five to 7% which we know is pretty much the amount of inherent natural variation that exists between raters anyway.

So really encouraging. The, the other thing that's I think really important for a tool like this is that it was really good at picking up a hip that was very significantly dislocated, especially if a machine learning tool like this was gonna be used as a kind of triage context, like pick out the hips that are really, really bad 'cause those are the ones we want to know about.

Bone Finder was really good at doing that. That's really interesting. And just before we move on, in terms of obviously, 'cause some of these children will have x-rays when there will be metal work there and other artefacts and was did, how did it sort of cope, cope with that? Yeah, we knew this was [00:12:00] gonna be a big challenge for Bone Finder 'cause it's actually, you know, to its credit has never been trained on looking at a hip radiograph with

pelvic metal work in or spinopelvic metal work or a baclofen pump or anything like that, we anticipated that was gonna throw it off the scent a little bit and it, and it did. There was about 14% of the hips in the cohort had metal work in situ. And when Bone Finder just looked at those radiographs, it often got a bit confused and disorientated.

So the amount of agreement between raters went down and the absolute difference in the measurements went up. But the, the quite amazing thing is that. The, the nature of machine learning tools is that you can then feed these radiographs back into the sort of training pool and you're able to refine the searcher tool.

So even in the process of doing this project, we've been able to feed it back to the team and they can make little tweaks and continue to make it better and better. Yeah, that's the, you say it's the great thing about it, isn't it? Continuously learning. So Dan, maybe if I come back to yourself now, know, bringing this all together I suppose first of all is, you know, the findings of this.

Is that what you sort of expected looking [00:13:00] at a tool like this? And what do you feel sort of the implications of our work are moving forward? So I guess I'm a little bit biased 'cause I, I was part of the, I'm part of the Bone Finder group that, that kind of, that created the initial models and we did that with some of the data from, from where, where I am.

But I mean, it was awesome to see it, see it work so well in a completely different setting. And, and I guess it gave us real confidence that we can do what we wanna do, democratise the care and you know, it, it was as good as experts, which is really, really cool. So for me it's a complete no brainer I think.

I think. Whoever should should, should fund it and should, should make it available to, to everyone to, to, to make their pathways work better. Absolutely. And do you think there's implications of this for, obviously this is looking at hips and CP done out with, from what your knowledge of Bone Finder.

 So, so I think there's huge potential so. So we, we actually, so CP is one of the uses of Bone Finder. So there's lots of different uses of Bone Finder. Bone Finder's been in this context has been specifically developed around kids hips to look at [00:14:00] it's been trained on Perthes disease, slipped epiphysis, CP and DDH.

Mm-hmm. And therefore I think there's a real opportunity to, one of my loads will be to diagnose SUFE early. We all know that the, the people often get an AP of a pelvis and then miss the fact there's a SUFE. If you've got an AI tool, which is looking for really subtle signs of, of SUFE you know, abnormalities around the, the, the physis and stuff that, then I think that'd be a really useful tool for a really useful use case for an AI tool.

So I think there's loads of potential and, you know, measuring different angles in, in DDH and stuff like that. There's loads of, loads of ways we could use it. So I think I think it's potentially great. I, I totally agree. Dan, just sort of a side question maybe just for that. Do you think, obviously these, as Katie has described and you described, they continuously learn.

Do you think there will ever be a point where actually it will pick up something that not even the most experienced clinician, isn't seeing, and will say this, this patient's at a risk. And then I think just wondering if that leaves you with [00:15:00] a little bit of a conundrum. I don't know. Because you don't, so I think that's hard, isn't it?

'cause everyone, so, so some people have said to me before, well, I'm only want to use it if it's better than us. Yeah. But I think, I think the thing is it can only be as good as who trains it. So it can only be as good as experts. And that's, that's the best comparison we can use. And at the same time. I think one of the risks in any kind of screening or whatever you wanna call it, tool, one of the risks is then it starts over detecting and is kind of over detecting stuff, which is actually just as harmful as, as under detecting in many contexts.

Yeah, yeah, yeah. So I, I don't think we wanna overtrain or overfit or whatever we do to the model. We just want it to be as good as experts 'cause. That's all we can hope for. Absolutely. Absolutely. No, I totally agree Dan. And Dan, that sort of leads us nicely just to briefly talk about your other paper in the, in the Journal for this month, you know, and that, and that was a scoping review that was look sort of more broader brush to it in terms of navigating the barriers and facilitators to the implementation, implementation of AI in healthcare.

And I think that's obviously links very well with what we've discussed here [00:16:00] already. So what, first of all, just quickly, what made you look at this in, in particular? I think I, I think it's partly, so it's partly I'm interested. Partly I've been working with Bone Finder and, you know, we want to get this stuff out there.

We want to see it used, but I think it's also partly frustration. So, so there are so many, so many AI tools and you read about them and we see them and then they never actually used, and there's, there's very few that are actually in routine clinical practice or, or making a difference. And you know, in innovation there's that whole 17 year lag between, between starting something and then get it using clinical practice and it's like, oh my gosh, how are we gonna overcome this?

Like, what, what, what are the barriers and how are we going to achieve, achieve some useful drive to go forward? Yeah, no, absolutely. And in terms of what that found, 'cause obviously this was a sort of scoping review of the available literature. What was sort of the key take homes from that? So I think it, I, I think the key takeaway for me for, so for someone that's trying to develop an AI tool I think it really made me so, so the, the so identified all [00:17:00] of the different areas where, where there may be, there may be barriers so that there may be pal, you know, that, I guess, I guess a lot of it is what, what we already know.

And I know that sounds really bad 'cause we're writing a paper about what we already know. But, but it really highlights and made me think about how we can overcome particular barriers. So, so for example, one of the barriers is that, that, that there's no real benchmarks for, for, you know, how we develop AI tools.

So, so people will, will develop an AI tool, say, oh, this, this tool detects 90% of fractures or 80% of fractures or whatever. But there's actually no standardised validating way of testing it. So all of these claims are just. A bit of phony claims. So, so I find that quite hard. So, so with that, we've gone out, we've, we've got in, in our hospital, we've, we've recently got an ethics permission where we can share all of our radiology data in an anonymised fashion with companies and with, with academic groups.

In order to try to try and be a source of validation for, for other, other tools. And I think that'd be really helpful if we had [00:18:00] some standardised data sets where, where we could actually put tools to the test and say, look, we, we, we, we validated our tool against this Alder Hey dataset. And then people had some sort of

some, some sort of confidence that, that, that dataset was a, was a, was a fair dataset to validate it against. 'Cause then we're not comparing apples to oranges, we're comparing oranges, oranges, which which makes a lot more sense. And then there's other things about, you know, organisational culture and how we overcome that and, you know, how do we overcome the ethical barriers and, you know, lots of different little barriers which.

Which, unless they're all written down in front of you, you don't necessarily move forward and address them all in a, in a kind of systematic way if you're trying to roll out an a AI AI tool. And I think that for me was the, the real opportunity in this paper, and I'm very grateful to, to, to Mohammed who, who wrote it

with me, but yeah, it was it was, I I think it's quite nice and it just lays out a framework. I totally agree with that and I think it's really helpful. And I think, you know, like, like you say with some scoping reviews, you think actually, are we just telling us what we know? But I think it really does lay out very nicely and what those barriers are we need to overcome.

And I think it also, to [00:19:00] me, it gave me a sense of actually we all have a duty here to try and get involved with this, 'cause actually the more people are involved, the more data these. No, in a, in a valid and, and and ethical way, the better it will get quicker. And actually, like you said, we won't be waiting that 17 years of you just described to actually use some really positive technology for, for our patients.

So I, I think it was really, really helpful from that, from that point of an aspect. Yeah, a lot of it's about workforce resistance and about the kind of the organisational culture. And if we're all gonna resist AI forevermore, then AI's not gonna happen. We do need to. To kind of, yeah, let, let the book out a little bit and, and start accepting things.

Do you think then, just a, maybe a slightly controversial question, but do you think is a, we need a broader brush in terms of, you know, rather than individual trust doing it? Do we need sort of a, a nationwide policy to say, look, we, we need to get involved with this, and if possible, you know, with, with appropriate consent and, and know, obviously patients being aware but do we, would that be a possibility thing or is that too big a barrier to overcome?

No, I think so. So I, I also sit with a kind of NIHR hat, so a National [00:20:00] Institute of Health Research hat and NIHR are very much focused about, about trying to get AI into, into healthcare and not just for the good of patients. Although of course it is very, very, very largely for the good of patients, but it's also about growing the economy.

Yeah, and so, so NIHR is all about the health and wealth of nation. So, so we see AI as a, a way of also growing the economy if we can, if we can have one of those massive AI healthcare companies in the UK. That's really cool. So I think, I think there's lots of flexibility within, within the NHS to say, look, how can we, how can we grow AI in, in our, in our

system and, and whilst there's many challenges with the NHS, the fact that all the data's so joined up, although you may not necessarily think so, but compared to all other healthcare systems, it's amazing and it could be a huge opportunity to have these massive training trading data sets, which could which could make the economy huge.

I totally agree Dan, Katie sort of, we've, we've, we've sort of touched on my last sort of point there with Dan about, you know, how do we bridge that gap between AI development of real world deployment. Anything you would sort of add to that to finish [00:21:00] up? Yeah, I think it's, it's really interesting to see the difference between tools that have come from a commercial background versus those that have come from an academic backgrounds.

I think Bone Finder is a tool that's been, had amazing, rigorous research put into it through years and years and years. But it's perhaps been challenging to get to the implementation phase just because of like financial backing and development. The science is amazing, but like the money isn't necessarily there.

Whereas on the flip side, you have lots of commercially driven tools where maybe the science isn't so great, but because they've had millions of pounds put in seed funding from the start, they've been able to get further along. So there's definitely something about getting funding from the right bodies to the right people.

The people who've been working hard, doing good science for years and years and years, and supporting those tools to get off the ground, I feel like. Absolutely, Katie. Totally agree. Well, both, I'm afraid that's all we have time for so, so thank you so much to you both for taking the time to join us and congratulations on some amazing work.

That's been published here in the journal and, and that you continue to do. I know, and I'm sure the [00:22:00] future's very bright in this area, and I think with people like yourselves guiding it, it's gonna be a really positive area. And I really look forward to more, more, more coming from you both in the future in, in this area.

So, and to our listeners, we do hope you've enjoyed joining us and we do encourage you all to share your thoughts and comments on the various platforms and the like. Feel free to post about anything we've discussed here today. And thanks again for joining us. Take care, everyone.