There's an AI physio seeing patients in the UK. Can it fix my back?

Back pain is one of the leading causes of disability in the world, but an AI-powered app is being used to help patients in the UK to manage their symptoms. But how does it measure up?

When I hurt my back in October 2024, the cause wasn't immediately clear. Perhaps, something went awry while hoisting one of my young nephews in the air. Or, more likely, I messed my back up while lifting a bed so my wife could slide a rug underneath.

Whatever the trigger, my overconfidence in my own strength left me with a lower back injury. Days later, the pain began radiating down the sciatic nerve in my left leg. Standing was fine, sitting was torture. In order to sleep, I had to slide a pillow underneath my legs to keep the downward-pulsing shockwaves at bay.

By mid-December, I saw a pain medicine specialist in the area of Washington, DC where I live. He ran some movement tests, sent me for X-rays, and wrote me a referral for physical therapy with a diagnosis of "lumbar radiculopathy". In the United States, this is a relatively quick process – I went from pain to starting treatment in about three months. But that's still a very long time for something so common – and debilitating – as lower back pain.

Since early January, I have seen human physical therapists who work wonders. Once a week, I go in, discuss my pain, get a soft-tissue massage, and do well-supervised exercises and stretches. My pain isn't gone, but it's finally under control.

I specified that these physical therapists are human because, well, I've also been seeing a new physio in the UK. But my new British physio is only partly human. Rather, she's a series of pre-recorded videos made by a real physiotherapist, powered by artificial intelligence that I can access via an app on my phone.

The service, provided by a company called Flok Health, is the first artificial intelligence clinic – of any kind – to be trialled by the UK's National Health Service (NHS). It's been approved for use by the country’s Care Quality Commission as a registered healthcare provider, and began seeing patients – just like me – at the end of 2024.

The reason: a lot of people suffer from lower back pain and struggle to get access to the care they need. As of September 2024, nearly 350,000 people in England alone were on waiting lists for treatment for musculoskeletal problems – the longest of all waitlists for any set of conditions. The UK government reported that 23.4 million work days were lost in 2022 alone due to these untreated problems, implying an enormous economic as well as human cost. Globally, lower back pain is a leading cause of disability and affects an estimated 223 million people worldwide. 

Flok's promise is to start seeing patients immediately, relieve the burden on the NHS, and head off patients' back pain before they go untreated and deteriorate. It's a noble mission, but I had one question: Was AI really the answer?

In addition to being a trained doctor, Finn Stevenson spent time as a professional rower in Great Britain's Olympic development programme. Anyone who has rowed before – or even spent time on a rowing machine at a gym – knows that the sport can be an incredible workout but also brutal on the back if your form isn't perfect. While a professional athlete, Stevenson had access to top doctors and physiotherapists who could treat his ailments immediately. But when he left rowing and his back pain flared up, he realised what it's like for everyone else trying to receive care.

"It was quite a big shock in terms of what access to care looked like for 99.5% of people," says Stevenson, now Flok's chief executive. "In theory, I should be pretty well placed to manage this. I had an academic background, I'd had three years of professional physio. If I was struggling, other people probably were too."

Stevenson and Ric da Silva, now his co-founder and Flok's chief technology officer, met when they were early employees at a company called CMR Surgical, a British startup where the pair worked on building robotics for soft-tissue surgery. 

But with Flok, they are aiming to provide treatment to people who don't need surgery. They are looking to help people with the easiest and most manageable cases. 

"There are quite a few conditions where that's true, where you don't need a scan, you don't need hands-on manipulation, you don't need medication," says da Silva. "What you need is 10 minutes of stretches a few times a week, and that will solve the problem."

The idea is that this should then free up clinicians to focus on the patients with more complex problems who really need their help.

When I first open Flok, I'm greeted by Kirsty, my physiotherapist. She appears in a minimalist setting with all of the accoutrements befitting a yoga instructor: black workout outfit, bamboo-coloured exercise mat, contemporary office chair, and a Monstera plant. She asks me questions about myself and my pain, and I respond by clicking multiple-choice options. Each response triggers a follow-up question or instruction as the underlying AI adapts to my answers.

But rather than being a live video call, this interaction is a series of pre-recorded video that are stitched together by AI based on my answers. 

While the artificial intelligence boom of the past few years has been spurred by advances in generative AI from large language models, Flok has a different use for AI. "We essentially developed a domain specific language for describing clinical reasoning," says Stevenson. Because it's not a ChatGPT-like chatbot predicting the next word in a sequence, there isn't a risk of the so-called hallucination problem in AI where an algorithm makes things up – a major concern about AI in the medical field.

Instead, this AI is more like a choose-your-own-adventure book where there are more than one billion "intervention combinations", says Stevenson. The power of AI makes it possible to deliver this in a seamless way.

It's a "perfect software problem", says Stevenson. "Software is an extremely good delivery model if you know what you're trying to deliver, and if you just need to do that in a more scalable way."

Flok is currently expanding its operations. Its service launched first in Scotland at the end of last year and has recently signed contracts to add locations in England, though Stevenson isn't yet able to reveal where. In the next 12 months, Stevenson said he hopes to cover at least half of the UK. And they're also expanding what pain areas they're treating hip and knee osteoarthritis and women's pelvic health. 

But it is also not alone in trying to use AI to tackle musculoskeletal pain. Another AI-based app that helps patients manage pain in the lower back and neck called selfBACK has been undergoing clinical trials in the US as an add-on to more typical care services.

Not entirely unlike Flok's approach, selfBACK works by analysing what treatments worked well for previous patients with similar conditions, then recommending personalised exercise plans. Studies have indicated that it can help to reduce pain among those who use it. But a recent trial found slow adoption rates – nearly one-third of trial participants never used the app, and another third only used it a few times. Patients and practitioners concluded that the app should be used to supplement traditional care rather than replace it. Other studies have also cautioned that self-management apps need to be provided in addition to care not as a replacement.

Even so, research seems to suggest that mobile health apps of this nature are a potentially promising approach to helping patients manage their back pain. But there are also no shortage of questions about AI's integration into healthcare more broadly. Elizabeth A. Stuart, a professor and chair of the biostatistics department at the Johns Hopkins University Bloomberg School of Public Health, says that AI healthcare tools should undergo the same rigorous evaluations as traditional medical interventions, such as regulatory approvals and peer review.

"In some ways, these AI tools are no different from other healthcare interventions that we want to learn about the effects of," says Stuart. "We need to assess how well they work, do they work, who do they work for, just as carefully as we have done for other healthcare interventions." 

Stuart says that while we don't need to reinvent the wheel when it comes to AI, we do need to recognise when these systems keep evolving even after they're deployed. "With some AI tools, they are sort of self-adapting," she says.

She contrasts AI tools with the annual flu vaccines, which undergo deliberate assessment processes each year to ensure they're targeted to the latest viral strain. But sometimes AI tools are a black box, "changing under the hood" in real-time and it's not always clear how, she says. 

Pranav Rajpurkar, an assistant professor of biomedical informatics at Harvard Medical School, says he sees promise in AI applications that help triage care for patients. "What's counterintuitive is that forced collaboration between AI and clinicians often underperforms compared to a clear division of labor," he says. "When doctors and AI review the same cases side-by-side, their combined accuracy is sometimes barely better than doctors alone. Clean separation of responsibilities side steps these issues."

In Flok's case, the company employs professional physios who appear in the videos that patients see, but are available to answer patient questions after their AI-enabled video sessions.

Rajpurkar predicts that eventually "we'll move from narrow AI tools to generalist medical AI systems capable of handling multiple tasks across different domains". It's an approach he's working on with his own company, a2z Radiology AI, which is developing AI-powered analytics for radiologists.

"The systems that succeed won't be those claiming to 'replace doctors', but those that thoughtfully redistribute clinical work to make healthcare more accessible, efficient, and human-centered," says Rajpurkar.

In our first 20-minute session, Kirsty asks me a long list of questions about myself, about my pain, and asks me to stretch in different ways – following her careful lead on the screen – in order to decide the best course of treatment. She gives me stretches and exercises to try for the week ahead, with careful instruction not to push too hard or do anything that feels uncomfortable. 

This isn't the experience I have in-person with my team of human physical therapists in the US. They frequently correct my posture and tell me when I'm doing something wrong – which is often. The big difference here is that Kirsty can't see me. Her pre-recorded videos don't watch my movements and stretches. They rely on me following her instructions correctly and reporting if something is amiss. There's also an option at the end of the session to leave a voice note for a real physiotherapist who will respond to my personal needs or questions – a real voice behind the curtain.

My back felt better after my sessions with Flok, but the app likely isn't for me. I'm clumsy and uncoordinated and need someone watching my form at all times – if not, I'm likely to hurt myself further. That ability to watch patients and offer feedback on posture could be something AI will enable in the future – but until then, I'll be sticking with the humans.

--

If you liked this story, sign up for The Essential List newsletter – a handpicked selection of features, videos and can't-miss news, delivered to your inbox twice a week. 

For more science, technology, environment and health stories from the BBC, follow us on Facebook, X and Instagram.