A IA na Fisioterapia: O que é realidade e o que é exagero em 2026

The state of AI in physical therapy, category by category

"AI in physical therapy" describes three different technologies at three different stages of maturity, and lumping them together is how marketing outruns reality. Sorting them by category tells you what actually works today.

• Documentation assistance sits in live adoption. Ambient scribing tools draft notes from a session and clinics already run them daily.
• Movement analysis is maturing and validation-dependent. Computer vision and wearable sensors work technically, but peer-reviewed evidence for clinical use is still thin.
• Clinical decision support stays research-stage. Triage and outcome-prediction models live mostly in journals and academic health systems, not private practice.

One pattern runs through all three. Adoption tracks the strength of the published evidence, not how confident a vendor sounds. Where the research is solid, clinicians use the tool. Where it is not, they wait.

Assistência na elaboração de documentação: anotações em tempo real e geração de notas

Documentation is where AI has actually landed in physical therapy clinics, and it landed there because the underlying task suits what the technology does well. Ambient scribing tools listen to a visit through a phone or clinic microphone, transcribe the conversation, and generate a draft note structured into subjective, objective, assessment, and plan sections. The clinician then reviews, corrects, and signs. The machine drafts, and the human remains responsible for what enters the chart.

The pull toward these tools comes from documentation burden, which is well measured across healthcare. Physicians spend close to two hours on electronic records and desk work for every hour of direct patient care, according to a widely cited study in the Annals of Internal Medicine. Physical therapists face a similar drag, since note-writing often spills past scheduled hours and feeds documented burnout. Early health-system deployments of ambient scribing show modest, not dramatic, gains. A large-scale study covered by STAT found only a modest reduction in total electronic health record time and no significant change in after-hours charting, with primary care and female clinicians benefiting more than others.

Adoption in physical therapy specifically is still earlier than in physician primary care, but it is real and growing. The American Physical Therapy Association has begun addressing generative tools in its practice guidance, treating them as an emerging documentation aid that clinicians must supervise rather than trust outright. Most PT deployments today sit inside larger outpatient groups and hospital-affiliated clinics that already run enterprise records systems, because those settings have the compliance staff to vet a new note source before it touches billing.

The accuracy question is where honest description matters most. Ambient scribes produce fluent, plausible notes, and fluency is not the same as correctness. A generated note can misattribute a symptom, invent a detail the clinician never stated, or smooth over an ambiguous finding into false certainty. Because the draft reads clean, a rushed clinician can sign an error more easily than they would catch a blank field. The review step is not a formality, and clinics that treat it as one inherit the risk.

Liability follows the signature, not the software. When an AI-drafted note contains an error that affects care or an audit, the clinician who signed it owns that record, and current malpractice and licensing frameworks give no shelter for delegating the wording to a model. Vendors disclaim clinical responsibility in their terms, which means the accountability question has a clear answer that favors caution.

Payer and compliance exposure adds a third constraint. Documentation for physical therapy must support medical necessity, justify the plan of care, and match the codes billed, and generated notes can drift toward generic language that fails a payer review even when the clinical care was sound. A note that reads well but lacks the specific functional deficits and measurable goals a payer expects can trigger denials or clawbacks. Clinics adopting ambient scribing successfully treat it as a first draft that a clinician shapes into compliant documentation, not as a finished record they approve on volume.

Análise do movimento: visão computacional e sensores vestíveis para a análise da marcha e o movimento

Movement analysis is the category where the technology genuinely works and the clinical claims run far ahead of the evidence. Two distinct classes of tool get lumped together under the same marketing, and separating them is the first step to reading any vendor claim honestly. Lab-grade motion capture, the kind that uses marker arrays and force plates, has decades of biomechanics research behind it and produces measurements clinicians can trust. Consumer-grade video and pose-estimation tools, the ones a clinic can run from a tablet or a phone, use computer vision to infer joint positions from ordinary footage, and their accuracy varies widely with lighting, camera angle, clothing, and the movement being tracked.

The peer-reviewed base for the phone-and-tablet tools is thin and mostly small. Validation studies typically enroll a few dozen participants and report agreement with marker-based motion capture on specific tasks, such as a squat or a single-leg stance, under controlled conditions. Those studies tend to show acceptable agreement for large, planar movements and much weaker agreement for rotational motion, fine joint angles, or anything measured off-axis. A tool that estimates knee flexion within a few degrees during a filmed squat is doing real work. The same tool claiming to quantify subtle gait asymmetry in a cluttered clinic hallway is making a claim the published research does not support.

Where clinics actually deploy these tools today tells you more than the brochures. Sports performance and athletic screening lead adoption, because the movements are repeatable, the athletes are cooperative, and the stakes of a wrong measurement are lower than in medical rehab. Some orthopedic and neurologic rehab programs run pilots, often inside university clinics or hospital systems that can compare the output against their own instrumented labs. Mainstream private practice is barely touched. Most clinicians assessing gait or movement quality still rely on trained observation and validated functional tests, not because they reject the technology, but because no video tool has yet shown it improves an outcome the clinician couldn't already judge.

Wearable inertial sensors sit a step further along than pure video. Accelerometers and gyroscopes strapped to a limb or worn as an insole measure cadence, stride time, and step symmetry with better reliability than pose estimation, because they capture motion directly rather than inferring it from pixels. Gait research using wearables has produced more consistent published results, particularly for step counts and temporal parameters. The gap remains between measuring a parameter reliably and demonstrating that acting on it changes a patient's recovery, and that second study is the one most vendors have not run.

The honest read for 2026 is that movement analysis gives you a fast, repeatable way to capture certain measurements, not a validated substitute for clinical assessment. Treat any accuracy figure as conditional on the exact task and setup it was tested under, and ask whether the validation compared the tool against gold-standard motion capture or against nothing. A tool that speeds up an exercise-prescription workflow is a separate matter from a tool that claims diagnostic-grade movement measurement. Physitrack's program builder, for instance, uses smart search to help a clinician find exercises quickly, which is a search-and-workflow function rather than movement analysis, and worth distinguishing from the pose-estimation claims this category is really about.

Clinical decision support: triage, outcome prediction, and evidence retrieval

Clinical decision support is the least mature of the three categories, and it is easy to overstate. Most of what exists lives in research journals and health-system pilots, not in the software a private clinic buys. Vendors sometimes borrow the language of decision support to describe features that are really just search or filtering, so the gap between what gets published and what gets deployed matters more here than anywhere else.

The published work falls into three rough groups. Diagnosis-support algorithms attempt to flag likely conditions or triage patients toward the right care pathway from intake data. Outcome-prediction models estimate how a patient will respond to a given course of treatment, often using large datasets of prior cases to forecast recovery time or the odds of a good result. Evidence-retrieval and literature-summary tools help a clinician pull relevant research on a condition or intervention faster than a manual database search would. Each of these has produced peer-reviewed results, and each remains largely confined to academic medical centers and integrated health systems that have the data infrastructure and research staff to run them.

Mainstream commercial PT software rarely ships any of this. A model trained on one health system's patient population does not automatically transfer to a different clinic with different demographics and documentation habits. Outcome-prediction models in particular tend to degrade when applied outside the dataset they were built on, and validating them across new populations takes time and money that most vendors have not spent. The result is that decision support you can point to in a journal is not decision support you can buy and switch on next week.

Decision-support tools also face a higher bar than documentation tools, and the reason is straightforward. An ambient scribe drafts a note that a clinician reads, corrects, and signs, so the human stays in control of the final record. A triage algorithm or an outcome model influences clinical judgment directly, and it can nudge a clinician toward a decision before the note is ever written. When a tool shapes what you decide rather than how you document what you decided, the stakes for validation rise sharply.

That direct influence on judgment is also what makes the liability picture harder. If a triage suggestion steers a patient away from imaging that turned out to be necessary, the question of who answers for the miss becomes real in a way it never does for a mistyped note. Regulators and professional bodies have not settled how much a clinician can lean on an algorithmic recommendation before the accountability shifts, and that uncertainty keeps cautious clinics on the sidelines. The evidence for these tools may eventually catch up, but for now the honest reading is that clinical decision support in physical therapy is a research field, not a product category.

Por que a adoção continua ficando para trás em relação ao marketing

Three constraints explain why clinics adopt so much slower than vendors promise, and they apply across all three categories. The first is liability, and it stays unresolved. When an ambient scribe drafts a note that misstates a finding, or a triage tool suggests a wrong classification, the responsible party remains the clinician who signed off, not the software that generated the output. That accountability structure gives clinicians every reason to slow down and review, which erases much of the efficiency the tools promise. No regulator has drawn a clear line on where vendor responsibility ends and clinical responsibility begins.

Thin validation compounds the liability problem outside documentation. Ambient scribing borrows evidence from a large body of general healthcare research on clinician time and burnout, so it stands on firmer ground. Movement analysis and decision support do not. Most movement-analysis claims rest on small agreement studies against gold-standard motion capture, and most decision-support models live in journals rather than deployed practice. A clinic director asked to justify a purchase to a payer or a malpractice carrier has little peer-reviewed footing to stand on for those two categories.

Clinician skepticism in physical therapy is documented and reasonable, not reflexive resistance. Discussions in APTA forums and PT professional communities repeatedly surface the same objections, including notes that read plausibly but contain fabricated detail, tools that assume ideal recording conditions clinics rarely have, and vendors who present pilot results as settled proof. Clinicians who have watched a scribe invent a symptom or a pose estimator misread an obese patient's joint angles do not need a warning label. They have already priced the risk into their trust.

Those three forces together produce a structural pattern, not a temporary lag that a better product cycle will close. Adoption tracks evidence quality, and the evidence hierarchy across these categories is stable. Documentation has the strongest support and the widest use, movement analysis sits in sports and specialty rehab pilots where a controlled setting offsets thin validation, and decision support remains confined to academic medical centers with research staff to supervise it. A pilot succeeds precisely because a motivated team controls the conditions. Everyday practice removes that control, so results that held in the pilot degrade, and the tool stalls before it reaches the general caseload.

The pilot-to-practice gap therefore closes only where the evidence deepens and the liability question gets answered, category by category. Vendor confidence moves faster than either. That difference in pace, not any single missing feature, is why the marketing keeps running ahead of the clinic.

O que os fisioterapeutas devem realmente observar em 2026

When a vendor attaches "AI" to a rehab product, ask where the evidence comes from before you ask what the feature does. A demo and a case study are not the same as validation. Request the peer-reviewed study behind any accuracy or outcome claim, and check whether the authors are independent of the company selling the tool. Vendor-generated white papers and internal benchmarks tell you the product performs well under conditions the vendor chose.

Ask how the tool agrees with an established reference. For movement analysis, that means agreement with gold-standard motion capture in a published study, not a comparison to nothing. For documentation, that means measured error rates in generated notes, not a claimed time savings.

Pin down the human-in-the-loop requirement. Find out exactly which step requires your review and sign-off, and confirm who holds liability when the tool is wrong. A note you sign is your note, whatever software drafted it. A triage suggestion you act on is your clinical decision.

Real utility exists across all three categories, and ambient scribing has earned a place in daily workflows. The other two sit further back on the evidence curve than their marketing implies. Treat every "AI" label as a claim to be checked against published work, and you will separate the tools that help you from the ones that only market well.