The diagnoses physiotherapy keeps missing.

A patient walks into a physiotherapy clinic with what looks like a simple knee complaint. Six weeks of pain, worse on stairs, no obvious incident. The physio examines the knee. Patellofemoral, maybe. A short treatment block is prescribed. Three weeks later the patient is back, no better. Two months after that, an orthopaedic referral finally surfaces a hip labral tear. The knee was never the problem.

Most physiotherapists have a story like this. The diagnosis nobody made, until somebody finally did. It rarely makes for an interesting story because there is no villain in it. The clinician was competent, the patient was honest, the examination was thorough enough. And yet something was missed.

For a long time the profession has treated these moments as bad luck. Hard cases, unusual presentations, a referral that arrived late. The quiet truth is that misdiagnosis in musculoskeletal practice is more common than the profession likes to say.

What the literature actually says

The honest picture from clinical research is uncomfortable. A systematic review of red flags for cauda equina syndrome found pooled sensitivity between 0.19 and 0.43, and specificity between 0.62 and 0.88, meaning the screening questions clinicians are taught to ask miss most cases of the very thing they are designed to catch.¹ A nationwide register-based cohort study of musculoskeletal physiotherapy patients found that the prevalence of serious pathology in this group exceeded guideline-endorsed estimates.² A mixed-method study of primary care physiotherapists found that correct management decisions in critical medical categories were associated with experience (odds ratio 2.73) and quality audit participation (odds ratio 2.90), suggesting that without those structural supports, accuracy drops.³

The misses tend to follow patterns. Hip pathology presenting as knee pain. Cervical artery dysfunction tucked inside what looks like a stiff neck. Cauda equina red flags that don't announce themselves with saddle anaesthesia, but with vague urinary hesitancy the patient didn't mention until the third visit. Referred pain patterns that match three different sources. Conditions a clinician sees once or twice a year and so does not hold easily in working memory.

It is rarely a knowledge problem

The instinct is to blame the clinician. Read more, study harder, do another course. Most physiotherapists already do this. The bottleneck isn't what they know. It is what they can pull up in the moment.

A 30-minute slot, sometimes 20, is not enough time to systematically consider every plausible diagnosis. So the clinician anchors on the first reasonable hypothesis and works to confirm or rule it out. This is anchoring bias, and it is well-documented in clinical reasoning research. A 2023 study of physician decision-making confirmed that even when clinicians are aware of the bias, the anchoring effect persists.⁴ Roughly three quarters of diagnostic errors have a cognitive component, and up to 80% of adverse events related to misdiagnosis are judged preventable.⁵ It isn't a personal failure. It is what happens to anyone making decisions under time pressure with incomplete information.

Layer on the patient who arrived already half-diagnosed by a Google search, the WhatsApp message from the front desk about a billing issue, the next appointment waiting outside the door, and the cognitive surface area available for a careful differential narrows.

Where AI actually helps

AI is good at three specific things in this context, and bad at most others. Knowing which is which matters.

The first is breadth. A well-built clinical model can hold every one of those 2,000 conditions in active memory at once and compare them against the signs and symptoms in front of it. It does not get tired in the afternoon. It does not anchor on the first plausible diagnosis. It can surface the rare condition, the one the clinician sees twice a year, as a ranked possibility for a human to consider.

The second is red flag detection. Cauda equina, vertebrobasilar insufficiency, malignancy referred to bone, infection. These conditions kill people if missed, and they hide inside ordinary-looking presentations. They are also exactly the kind of pattern matching an AI can do without forgetting. A clinician under pressure can.

The third is consistency. Two physiotherapists examining the same patient on the same day will often arrive at slightly different assessments. This is not incompetence, it is human variability. Structured AI-assisted screening forces a baseline that every patient gets, regardless of which clinician they happen to see and what kind of morning that clinician is having.

Where AI doesn't help, and shouldn't pretend to

AI cannot palpate a joint. It cannot watch a patient walk into a clinic and notice the antalgic gait before the patient has even spoken. It cannot read the small tells that tell an experienced physio that the real story isn't the one being narrated.

It also cannot be held accountable for a treatment decision. A physiotherapist can. That accountability is part of what makes the relationship work. The patient is putting their body in someone's hands. They are not handing it to a model.

Any product that frames AI as a replacement for the clinician is selling something that won't work in a real clinic. The clinician is still the centre of the encounter. Always.

What this should look like in a clinic

In practice, the framing changes the design. AI assistance shouldn't arrive as a verdict on the screen. It should arrive as a structured prompt during the encounter. Did you screen for these red flags. Have you considered these alternative diagnoses given what the patient reported. Here are the outcome measures most validated for this presentation.

The clinician accepts, rejects, or modifies each suggestion. The final diagnosis is theirs. The final treatment plan is theirs. The AI's job is to make sure no obvious thing was overlooked and to free up cognitive space for the things only a human can do.

This is what we have tried to build at Healui. An AI clinical layer that sits underneath every patient encounter, not on top of it. Screening that adapts to the patient. Differential diagnosis that ranks 2,000+ conditions with confidence scores. Outcome measures the AI auto-selects from 47 validated instruments. Treatment protocols generated in one pass, then reviewed and customised by the physiotherapist.

None of it tries to replace the clinician. All of it tries to give the clinician back the time and attention that diagnosis genuinely needs.

The honest closing

Some diagnoses will still be missed. AI does not change that. What it can change is how often, and how badly. Fewer late referrals. Fewer patients walking out of a clinic with a treatment plan for the wrong condition. Fewer red flags hidden in plain sight. Fewer cases where the physiotherapist goes home and replays the encounter for the third time wondering what they missed.

The case for AI in physiotherapy is not that it makes the clinician obsolete. It is that it lets the clinician do what they trained for, with the cognitive load that should never have been on one person in the first place.