Gains:
- Understanding what video-based pose estimation tools can measure and where they are wrong in physiotherapy
- Ability to cross-validate AI-derived angle, symmetry and motion data with observational evaluation and actual measurement
- Ability to report video analysis while protecting patient privacy, knowing measurement errors and limits
The physical therapist's eye tells a lot by watching how a patient walks, squats, raises his arm. In recent years, artificial intelligence (AI) has introduced a new tool to support this eye: pose estimation; software that automatically marks the joints of the human body in a video or photo and calculates angle, distance and symmetry. It can estimate knee angle from a squatting video taken with a phone camera, and step symmetry from a walking video. This is a powerful aid, but it also contains serious measurement errors. In this unit you will learn what pose estimation can measure, where it is wrong, and how to safely verify its output.
What does pose estimation do and what it cannot do?
Pose estimation tools detect body points (such as shoulder, elbow, hip, knee, ankle) in an image and calculate values such as joint angle, body tilt, and symmetry difference from them. This is useful for comparing a repeated movement over time, giving visual feedback to the patient, and providing a rough scan. For example, visualizing a patient's squat depth increasing from week to week is motivating.
However, pose estimation does not replace clinical measurement. If the camera angle changes, the angle value changes; loose clothing hides the joint; bad light confuses the model; Extrapolating a three-dimensional (3D) motion from a two-dimensional (2D) video is inherently inaccurate (rotation and depth are lost). Measurement with a goniometer (clinical instrument that measures joint angle) is still the reference. Exposure estimation tells you "approximately"; It doesn't say "for sure".
Caution: An angle value resulting from pose estimation is not an official clinical measurement. Confirm with a goniometer or a validated method under standard conditions before writing in the report. The sentence "AI said 47°" is not clinical evidence.
Sources of measurement error
Knowing the factors that affect the accuracy of pose estimation is the key to reading the output correctly.
Error source
Effect
Reduction way
camera angle
Changes the angle value significantly
Keep the camera steady and perpendicular to the plane of motion
clothes
Hides/shifts the joint point
Tight or form-fitting clothing, landmarks
light
Model misses points
Good, uniform lighting; no shadow
2D-3D loss
Cannot see depth and rotation
Multi-angle shooting or 3D system in critical measurement
Distance/resolution
Low resolution increases error
Full body visible, sufficient resolution
Step by step: analyzing video safely
- Get consent. Explain to the patient that a video will be taken, how it will be stored and its purpose, and obtain explicit consent.
- Fix the standard. Make sure the camera position, distance, lighting and clothing are the same for each measurement (otherwise the comparison becomes meaningless).
- Pull and process. Run the tool, get output angle/symmetry values.
- Cross check with observation. Is your own clinical observation consistent with AI? If there is a discrepancy, rely on observation, not AI.
- Confirm the critical value with a goniometer. Every aspect that will be included in the report must be verified using the standard method.
- Report by specifying the limit. Like "Exposure estimate, with ±X margin of error." Protect privacy.
Three mini cases (in numbers)
Case 1 — Squat follow-up. 28-year-old patient, patellofemoral pain. The first week exposure estimation showed knee flexion at 62°, the goniometer measured 68° (6° difference, camera was slightly to the side). The physiotherapist corrected the camera, and in subsequent measurements the difference decreased to 2°. After 4 weeks, pain-free squatting depth increased significantly; The weekly comparison video to the patient was motivating.
Case 2 — Gait symmetry fallacy. 60-year-old patient after knee replacement. Pose estimation reported 18% step length asymmetry. But the patient was wearing baggy pants and the camera was in the corridor (the angle is not fixed). In the re-shoot with tight clothes and a fixed camera, the asymmetry was 7%. Lesson: bad conditions can double the error; It would be an error if the clinical decision was made based on the first incorrect value.
Case 3 — Shoulder elevation. Frozen shoulder at age 50. The 2D video overestimated the abduction because the arm was forward (the arm was not exactly in the lateral plane). The physiotherapist measured 95° with the goniometer, and said AI was 112°. The goniometer value was written in the report, the AI value was used only for trend tracking.
Weak prompt / Strong prompt
Here, "prompt" is the text that interprets both the tool and the result.
Weak: "From this video, tell the patient's knee angle and write it in the report."
Ignores measurement condition, margin of error and verification; can formalize an incorrect value.
Güçlü: "Comment the knee flexion angle in this pose estimation output. List possible sources of error due to camera angle, clothing, and 2D boundary. Label the value as 'estimated' and indicate that it should be confirmed with a goniometer. Recommend clinical decision."
Four copyable templates
Task: Interpret the exposure estimation output (not as a clinical measurement). Data: [angle/symmetry values + shooting condition: camera, distance, light, clothing] I want: (1) summary of values, (2) possible sources of error, (3) which values require goniometer confirmation, (4) label "estimated". Don't make a decision, just interpret and suggest verification.
Task: Produce standard acquisition protocol checklist.Context: [motion to be measured, e.g. squatting / shoulder abduction / walking] I want: Camera position, distance, angle, lighting, clothing, and an itemized checklist for repeatability. For comparable measurement.
Task: Compare exposure data from two different sessions.Data: Session 1 [values] / Session 2 [values] / are the conditions the same: [yes/no]I want: Summary of change + warning whether change may be due to measurement error or actual progress. If conditions are different, invalidate the comparison.
Task: Translate the video analysis finding into plain language for the patient. Finding: [technical angle/symmetry summary] Rule: Motivating but understated, state "estimated measurement", do not create fear, simply explain the next step.
Tip: The most powerful use of exposure estimation is the trend, not the absolute value. Watching change from week to week under the same standard condition is much more reliable and motivating than claiming a one-time "sure" angle.
Motivational value of visual feedback
Although the measurement accuracy of pose estimation is controversial, its value in providing visual feedback to the patient is great. When a patient sees side-by-side images of their squat depth or arm height increasing from week to week, they are motivated much more strongly than an abstract sentence of "you're progressing." This is based on the behavioral science principle that tangible and visible progress increases commitment. The same visual can also be used to demonstrate to the patient an incorrect movement pattern (e.g., knee slipping in when squatting); The patient learns to correct his own movements more easily when he sees them from the outside. But here too the rule remains: use the visual as a motivational and teaching tool, not as a claim to precise clinical measurement. Telling the patient "this is an estimated measurement, it shows your trend" maintains honesty and trust.
Tip: Base visual feedback on comparison with the patient's own history, not on the absolute number. The message “deeper and more aligned than last week” is both safer and more motivating than the message “bend knees 68°.”
Common mistakes
- Mistaking the AI angle for a clinical measurement. Pose estimation is guesswork; goniometer is the reference.
- Not standardizing conditions. Comparing videos shot with different cameras/lights/clothing is misleading.
- Forgetting the 2D limit. 2D video gives serious errors in movements involving rotation and depth.
- Not obtaining consent. Patient video is special data; Explicit consent and safe storage are essential.
- To invalidate the observation. If AI conflicts with clinical observation, trust the observation, then investigate the cause.
- Storing video insecurely. Keeping identifiable patient video unprotected in the cloud is a KVKK violation.
In summary
Video-based pose estimation is a valuable aid for tracking and comparing motion and giving visual feedback to the patient; However, it is prone to measurement error due to camera angle, clothing, lighting, and 2D-3D loss. Treat the output as an “estimate,” standardize conditions, verify critical values with a goniometer, and use most for trend tracking. Patient video is private data; Approval and safe storage are mandatory.
Application task
Choose a movement (squat, shoulder abduction, or walk). Produce standard shooting checklist with second template. Assume that the same movement was "shot" once in a standard way, once under deliberately bad conditions (side camera, loose clothing), and have the first template interpret the difference between the two outputs. Mark which value requires confirmation with the goniometer.
checklist
- [ ] I obtained explicit consent for the patient video and kept it safe.
- [ ] I standardized the shooting conditions (camera, distance, light, clothing).
- [ ] I labeled the AI angle as "estimated" and did not consider it an official measurement.
- [ ] I cross-checked with my clinical observation.
- [ ] I confirmed the critical values to be included in the report with a goniometer.
- [ ] I used exposure data mostly for trend tracking, not making a one-time definitive claim.