Is AI Poised To Take Over Outpatient Physical Therapy?
AI has been making some big waves in the musculoskeletal space this past year. First, with Sword Health confirming its status as the “largest care provider through artificial intelligence”. And most recently with the National Health Service in the UK employing Flok Health as its first ever AI-powered physio clinic to help the overburdened universal health system reduce its 3+ month wait list for physical therapy visits. This raises a lot of questions about how AI will affect the future of conservative management of musculoskeletal care.
Will patients prefer to engage with artificial intelligence? Or will they miss the “human touch”?
Can AI provide equal (or better) care compared to the average Doctoral-trained physical therapist?
Will clinics be able to outsource and bill for services performed by AI?
What about the regulatory aspect? What will happen if there is a problem that requires legal action? Who is responsible?
Will AI put physical therapists out of business? How will this affect the market dynamics?
In general, is this good or bad news for clinicians? Managers? Patients?
Flok Health
Given the long wait lines (apparently more than 300,000 citizens on a waitlist) and the high demands for conservative musculoskeletal care, it’s easy to see why the NHS was incentivized to be among the first to roll out a completely autonomous AI solution. Instead of waiting months for a physical therapy appointment, Flok Health provides UK citizens with same-day video appointments with an “AI physiotherapist”. Flok is the first technology to be granted regulatory clearance for automating the triage, assessment, and treatment of back pain (the only condition that it will treat for the time being).
Patients can access the AI Physio via a smartphone app.The AI program uses a clinical decision engine to provide treatment recommendations based on the latest clinical evidence. The aim is to offer UK patients the option to choose treatment from an AI physio immediately instead of waiting 3 months for an appointment. Even if a modest percentage agree to it, it will help those seeking conservative care to get access immediately, while simultaneously shortening the wait time for those who need in-person care.
Did they achieve the goal?
In short, the answer is yes...at least according to Cambridge University Hospital trial data, which demonstrated not only that the technology helped reduce wait times for physical therapy, but when they ceased using the AI at the conclusion of the experiment, the waitlist for in-person appointments increased by 50%.
So How Does Flok Health Work?
One of the issues with physical therapy is that it is incredibly labor-intensive. This is even more true for PT compared to other healthcare services given that time spent with a physical therapist can range from 20-60 minutes and recurs multiple times per week in many cases. And the need for continuous follow-up with a highly educated professional makes it even more difficult to deliver this type of care at the scale needed in most health systems. Thus, Flok Health is able to create immediate access to a service that is typically gated by a human resource component.
Using a video assessment from the app, the AI physical therapist provides an evaluation of the patient's symptoms, determines if the patient is appropriate for AI treatment, and then prescribes exercises and pain management techniques over weekly, 30-minute structured “video calls”, adjusting the treatment based on the patient's progress, as well as subjective information provided to the AI. The prescribed techniques are unique to each patient profile and selected based on the patient’s individual symptoms and movement assessment.
What About The “Patient Experience”?
Flok’s AI uses patient reported outcome measures to test how well the system works in producing a clinical improvement based on its intervention, as well as measuring the patient’s confidence in self-managing their condition. And according to an NHS survey of those patients receiving care from an AI physical therapist, all respondents reported their experience with Flok had been at least equal to seeing a human physical therapist, and 57% of patients said they thought the AI experience was actually better. Is this hard to believe?
Flok Health makes it easy for the patient. Once the patient opts for the digital pathway instead of waiting to see an in-person therapist, they are registered and onboarded to Flok’s system. The first visit is a triage assessment. This is a structured video visit where the patient answers questions, and depending on their answers, the AI will respond accordingly, modifying its line of questioning.
Under the Hood
The technology itself consists of two components. The first component is an AI decision engine, which chooses what should happen next to the patient. The second involves a system which assembles and streams video to communicate the decision with the patient. This is the avatar of the AI therapist presented to the patient.
The AI physical therapist is actually a real person who was recorded in a professional studio. However, this second software component will pull specific sequences apart and reassemble a series of frames in a unique order based on the specific interaction it has with each patient. This unique blend is necessary because the decision-making core is just one piece of the puzzle, while presenting that information to the patient is also equally important.
There are other tools on the market that allow for the question-and-answer format via text-based interaction. However, Flok Health’s decision to use a video-based technology creates a much more engaging experience for the patient and allows for communicating nuance and clarity in a manner that is difficult to achieve with a simpler text-based platform.
For the system to work, it has to interact with the patient in a way that the patient can relate to, while simultaneously affording that clinical subtlety that occurs during a hands-on physical therapy assessment. This is one of the reasons Flok has chosen to use a video-based format. It’s not enough that the software chooses the correct course of clinical action, it also has to make the patient feel as though they are receiving that information and having an experience with a real human.
Of note, the system isn’t simply playing a recorded playlist of videos. Instead the system is stitching together video frames in unique sequences in real-time based on the patient’s response. This differentiates Flok’s AI from other software available which uses a static decision-tree for digital triage, which is useful but not quite sophisticated enough to cover every scenario, nuance, and range of clinical pathways that we see in an evidence-based physical therapy encounter.
Is It Safe And Effective?
Flok determines whether a patient is appropriate for their service, and if not, the technology will refer to in-person physical therapy or an appointment with a physician. Their triage process has undergone testing, and (so far), there have been no reports of missed red flags in their assessment. The AI will also flag to the clinical team if it is unsure if the patient is an appropriate candidate.
Whether a condition responds well to conservative care is difficult to measure. In part this is due to the heterogeneous nature of physical therapy as a discipline. What constitutes “physical therapy” is largely determined by how the therapist assesses and which interventions they choose. This will vary greatly between therapists. Patients tend to see physical therapy as a commodity, believing that each organization and each provider is offering the same service. But the service (and the outcome) will vary greatly between providers even in the same organization.
Even worse, most clinicians and organizations do not quantify their clinical outcomes or provide any level of quality assurance that is meaningful. So you have a recipe where each clinician is doing something different, and of the interventions they are doing, they are not typically collecting and analyzing data on the outcomes. And those that do collect data very rarely go back and review this data to make interventional changes in real-time or after the fact so as to modify their clinical practice over time.
But it gets even worse. Most human physical therapists aren’t even using evidence-based recommendations. A 2019 systematic review conducted by Zadro and Ferreira found that more physical therapists today are providing treatments of unknown value in this decade compared to the two previous decades.
Indeed, I’ve walked into many clinics where the so-called medically necessary physical therapy intervention rendered was nothing short of a series of exercises that a patient could find on YouTube if they wanted (combined with a 10 minute massage to “get those knots out”). In this sense, Healthtech companies could easily replicate and provide these cookie-cutter, glorified personal training sessions that are done in many clinical settings and outsource a large portion of physical therapy visits. But this isn’t because the software is necessarily better than the human. It is simply replicating the same non-skilled intervention that the therapist is doing.
A sophisticated AI system, however, could in theory follow the latest clinical practice guidelines and recommendations, and will do so consistently for every patient seen so long as it is programmed with that knowledge in its decision-making engine. So over time, I predict that we will eventually reach a point where an autonomous AI physical therapist is indeed more skilled and more consistent than a human counterpart.
Do Patients Really Want To Work With A Machine Instead Of A Human?
On face value, you might think the answer is an emphatic ‘no’. I would have thought the same. Until I saw this video of Sundar Pichai demonstrating Google’s AI assistant calling a restaurant and making a reservation, and doing so with a level of sophistication such that the person answering the phone was not aware that they were speaking with a machine. Have a look:
…and this was 6 years ago! As the technology develops, I have no doubt that users will eventually be unable to tell humans and AI apart, except for the fact that AI will undoubtedly be more reliable, readily available, and far more consistent. It’s not hard to imagine a time when someone will prefer to get on the phone with an AI program that will listen to their concerns and respond in an empathetic manner, while answering all of their questions and concerns.
This type of thought may seem far-fetched, until you step back and look at the cultural shift that is taking place in our society where more and more of us humans are interacting with autonomous software systems.
Think about this for a second. How often do you ask a virtual assistant like Amazon’s Alexa for an answer to a question? Or to set a timer? Companies like Tesla are working steadily towards developing fully autonomous vehicles on the road. Many organizations are using chatbots and customer service platforms to handle customer inquiries and provide support. In the financial world, AI systems monitor and analyze transactions in real-time to detect and prevent fraudulent activities. Streaming services like Netflix and Spotify use AI algorithms to suggest content based on user preferences and viewing/listening history.
In healthcare, surgeons can now use autonomous surgical robots such as the Da Vinci Surgical System to assist in performing operations. It’s not difficult to see that a time will come where AI will be part of the mainstream zeitgeist. The Overton Window is literally moving in real-time and this will affect musculoskeletal care in a big way!
The heterogeneity in MSK practice patterns is astounding. It is quite difficult and takes an extraordinary amount of time to adopt musculoskeletal clinical guidelines and evidence-based practices at scale and applied consistently in a traditional clinical environment because doing so would require behavior change across thousands of clinicians. Autonomous systems can take newly published clinical practice guidelines and implement them immediately and consistently across a population and analyze the effectiveness. And it can be implemented far quicker than would be the case in a traditional clinical care model.
What About The “Black Box” Phenomenon?
The so-called “Black Box Phenomenon” refers to the well-known issue of transparency, clarity, and interpretability when it comes to the internal workings of an AI system and its decision-making process. In other words, the model's internal processes are opaque, thus clinicians (and even developers) cannot easily see or understand how inputs are transformed into outputs. This is often due to how advanced these systems are, especially the deep neural networks, which involve numerous layers and parameters, making it difficult to trace the specific pathways that lead to a particular decision or prediction.
This creates a trust issue. If clinicians do not understand how the AI system is choosing a decision, how can a health system trust the AI, particularly in high-stakes medical situations, where the consequences of a wrong decision could be harmful to the patient?
Flock Health avoids the black-box phenomenon through a language-based approach which makes the system predictable and observable with a system that can be audited with full access to every encounter, transcripts of conversations, records of clinical decisions and why those decisions were made.
Flok Health Isn’t The Only Player In The Game
Meanwhile, Sword Health has already delivered in excess of 1 million sessions of AI-based physical therapy care in the previous year. In Sword’s hybrid approach, the human therapist oversees the patient’s progress while the AI care specialist, Phoenix, is available 24/7 to guide patients through their individual sessions suggesting modifications and form correction through their motion sensor technology. In this way, Phoenix functions almost as a physical therapist assistant. Like other AI programs, Phoenix can interact directly with patients, asking them about their symptoms and how they feel while performing each exercise, and respond accordingly, thus providing for a more “human” experience. Moreover, Phoenix will gather data, provide insights, and suggestions that it gleans from these treatment sessions to the supervising human physical therapist.
Will AI Replace Human Physical Therapists?
AI is highly unlikely to replace traditional physical therapy as a whole, at least not at this point in time. However, it can be a great tool as an extension and adjunct to in-person physical therapy, especially for those diagnoses that are simple, straight-forward, and do not require intensive hands-on care. It will also temper the burden for many patients that are unnecessarily gated by availability of physical therapists and costs of attending frequent sessions. It is not possible (at least not yet) for AI to cover the full range of patient conditions and populations (neurological, post-operative, pediatric,...etc) that exist to deliver care to the population. But it can take common orthopedic issues with relatively simple and routine clinical encounters off the board for a large swath of the population.
I do not believe physical therapists will have to worry about being replaced, especially as the demand for their services greatly outpaces the number of therapists on the market. Perhaps this will decrease the number of those 4-6 patient-per-hour “PT mills” that still exist. However, one thing is certain- therapists will have to up their game as a whole.
We need fewer interventions of questionable evidence and/or curative value and more treatment that is based on clinical reasoning, high quality research, and subclassifications which lead to outcomes that can be standardized and quality-assured. This is especially true for very common conditions such as low back pain and knee osteoarthritis that affect large portions of the population, where clinical evidence is largely in favor of conservative care, and where these types of scalable clinical applications expedite care delivery care delivery to patients without the cost and human resource constraint associated with standard physical therapy.