Industries / Health / Healthcare / Physical therapy
Exercise programs
Exercise programs are exercise given as treatment in a clinic, hospital ward or gym under a therapist's eye: choosing the exercises, setting the dose, and raising it as the patient gets stronger.
7
Milestones achieved, from 1945 to 2022.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1945
Heavy weights for injured soldiers
Thomas DeLorme, a US Army doctor, had injured servicemen lift their 10-repetition maximum, the heaviest load they could lift ten times, at a time when slow rehabilitation was keeping wounded men in military hospitals. By 1948 he had settled on three progressively heavier sets of 10, which he called progressive resistance exercise, and military and civilian physical therapy adopted it as the standard.
Credit
Thomas L. DeLorme, US Army; history by J. S. Todd, J. P. Shurley and T. C. Todd (Journal of Strength and Conditioning Research, 2012)
2009
Therapy for ventilated patients
At two US university hospitals, 104 adults on breathing machines in intensive care were given physical and occupational therapy from the first days, during daily breaks in sedation, or therapy as their doctors ordered. At discharge 59% of the early group could walk and manage six daily tasks on their own, against 35% of the others, with one serious adverse event in 498 sessions.
Credit
W. D. Schweickert, J. P. Kress and colleagues (The Lancet, 2009)
2011
Treadmill harness no better
Body-weight-supported treadmill training holds a stroke patient in a harness over a moving belt while therapists move the legs. The LEAPS trial gave it to 408 people two or six months after a stroke and compared it with progressive exercise at home set by a physical therapist. At one year 52% of all participants walked better, with no difference between the groups, and the treadmill groups had more dizziness and, among the most impaired, more falls.
Lessona costly machine was no better than simpler exercise that kept increasing in difficulty.
Credit
P. W. Duncan and the LEAPS investigators (New England Journal of Medicine, 2011)
2013
Knee and hip classes nationwide
GLA:D, short for Good Life with osteoArthritis in Denmark, trains physiotherapists to give people with knee or hip osteoarthritis three education sessions and 12 sessions of supervised exercise, twice a week for six weeks, and records every patient in a national registry. From 2013 to 2015, 9,825 registry patients reported pain lower by 12.4 points on a 0 to 100 scale after 3 months; these are the program's own registry figures. By 2018 the program had been set up in Canada, Australia, China and Switzerland.
Credit
Søren Skou and Ewa Roos, University of Southern Denmark
2015
Too much, too soon
The AVERT trial, in 56 stroke units in five countries, got 2,104 patients out of bed within 24 hours of a stroke, more often and for longer than usual care. Fewer did well at three months: 46% against 50% in usual care, by the modified Rankin scale, a 0 to 6 score of disability after stroke.
Lessona practice in guidelines worldwide had never been tested for its dose, and the larger dose did harm.
Credit
the AVERT Trial Collaboration group (The Lancet, 2015)
2016
Exercise matches knee surgery
In Norway, 140 adults aged 35 to 60 with a worn, torn meniscus, the cartilage pad inside the knee, got either 12 weeks of supervised exercise or keyhole surgery to trim the tear. After two years their knee scores differed by 0.9 points, too little to matter, and the exercise group's thigh muscles were stronger at three months; 19% of the exercise group later had surgery anyway, with no extra gain.
Credit
Nina Jullum Kise, May Arna Risberg and colleagues (BMJ, 2016)
2022
More mobilizing, more harm
The TEAM trial gave 750 ventilated patients in intensive care units in six countries about 21 minutes a day of active mobilization, against about 9 minutes in usual care. Days alive and out of hospital over 180 days did not differ (143 against 145), and adverse events that could be due to the exercise, such as abnormal heart rhythms and low oxygen, were more common: 9.2% against 4.1%.
Lessonthe gain seen in a small early trial did not hold when the dose was pushed higher across many hospitals.
Credit
C. L. Hodgson and the TEAM investigators with the ANZICS Clinical Trials Group (New England Journal of Medicine, 2022)
Watch for
Finding the right dose
After AVERT, the AVERT DOSE trial randomly gave 1,000 people with mild or moderate stroke, in 50 hospitals in seven countries, one of four doses of mobility training starting within 48 hours, to find which dose is safe and works best. The trial was stopped early for lack of funding; results for the two groups are due after data lock at the end of February 2026.
Achieved whenpeer-reviewed results show one regimen clearly better than the reference regimen, and stroke guidelines adopt that dose.
Claimed by
Julie Bernhardt and the AVERT DOSE Trialist Collaboration, in a European Stroke Organisation Conference abstract, 2026
A hip exoskeleton for stroke
An exoskeleton is a wearable frame with motors that helps the joints move. A University of Utah team reports that a 2.5 kg (5.5 lb) battery-powered hip exoskeleton, tuned to each walker, cut the hip's workload by nearly 30% and the energy cost of walking by 18% in seven people with weakness on one side after a stroke, tested on a treadmill.
Achieved whena team outside Utah repeats the result in a larger group, and a trial shows that walking improves after weeks of training with it.
Claimed by
Kai Pruyn, Bo Foreman and colleagues, University of Utah, in Nature Communications, reported May 2026
Open
01
An open exercise dose table
A public, machine-readable table of the exercise dose in every trial of a large Cochrane review, starting with the 2021 review of exercise for chronic low back pain, which holds 249 trials. For each exercise group the table gives the exercises, sets, repetitions, load or effort, rest, sessions a week, number of weeks, who supervised and how the dose was raised, coded against CERT, the Consensus on Exercise Reporting Template, a checklist of what an exercise trial report should state. An AI model drafts each row with a page citation and two people check it.
The briefCredit · unclaimed02
Band force you can check
An open set of measured forces for the flat elastic bands sold for rehabilitation, covering at least eight brands, five colors each and three batches, new and after 1,000 stretches, at fixed amounts of stretch. Alongside it, a phone tool that reads how far a band is stretched from two printed markers on it and gives the force from the table.
The briefCredit · unclaimed
Open 01 · the brief
An open exercise dose table
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, machine-readable table of the exercise dose in every trial of a large Cochrane review, starting with the 2021 review of exercise for chronic low back pain, which holds 249 trials. For each exercise group the table gives the exercises, sets, repetitions, load or effort, rest, sessions a week, number of weeks, who supervised and how the dose was raised, coded against CERT, the Consensus on Exercise Reporting Template, a checklist of what an exercise trial report should state. An AI model drafts each row with a page citation and two people check it.
Why it is still open
A 2026 preprint by Alexandra Bodnaruc and David Moher, which checked 130 exercise descriptions with AI-assisted extraction, found that trial reports gave a median of 53.2% of CERT items and systematic reviews only 16.7%. It already paired one human extractor with OttoSR, an AI extraction tool built on OpenAI's o3-mini-high model, and settled every disagreement against the papers, but it sampled half the trials of nine reviews, scored what was missing rather than recording the doses, and released no dose table. Searches of Europe PMC, the web and GitHub found no public dose table for a whole review.
Built with what exists
Open-access trial PDFs from PubMed Central and Unpaywall; AI models that read long documents and cite the page, such as Claude, GPT or Gemini (for a document-review example, see the Note "Seven more flagged"); Label Studio or a spreadsheet for the human checks; Zenodo for publishing with a DOI.
First step
Have two people code 10 trials by hand, run one AI model on the same 10, and compare them field by field.
Done looks like
At least 200 of the 249 trials coded. On a random 50, the AI's entries match the agreed answer of two independent human extractors on at least 95% of fields, with every disagreement listed. The table, prompts and code are published with a DOI, with a count of how many exercise groups are described fully enough to repeat.
Open 02 · the brief
Band force you can check
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open set of measured forces for the flat elastic bands sold for rehabilitation, covering at least eight brands, five colors each and three batches, new and after 1,000 stretches, at fixed amounts of stretch. Alongside it, a phone tool that reads how far a band is stretched from two printed markers on it and gives the force from the table.
Why it is still open
A 2026 review by Guzmán-Muñoz and colleagues says band color alone is an unreliable guide to force, because resistance varies across manufacturers, distributors, band widths, production batches and wear. The published tables cover single brands, such as the 2025 AGUEDA equations for seven Thera-Band colors, or heavy loop bands from four sports distributors (Fuentes and colleagues, 2019). Consumer Reports stretched bands from Bala, Theraband, TRX and other makers in its lab for its 2026 ratings, but it publishes scores for how evenly a set's levels step up, not an open table of forces. Searches found research prototypes of bands with built-in sensors, such as the 2025 Intelligent-Resistance-Band project on GitHub, but no open table of therapy bands across brands, batches and wear.
Built with what exists
A 50 kg load cell with an HX711 amplifier and an Arduino or Raspberry Pi Pico; a wall-mounted rail and tape for set lengths; a small linear actuator for the 1,000-stretch wear test; OpenCV with printed ArUco markers for the phone tool.
First step
Build the load-cell rig and repeat the AGUEDA protocol on one Thera-Band color to check the rig against the published equation.
Done looks like
Force at 25%, 50%, 100%, 150% and 200% stretch for every band tested, measured twice on different days with agreement (intraclass correlation) of at least 0.9; the change after 1,000 stretches reported per brand; the phone tool within 10% of the load cell on 100 test stretches; data, rig design and code published under an open license.