- 10 Monocular knee-angle error runs 9.3 to 21.9 degrees (Nature Scientific Reports, 2025). So a pose model reporting a crisp joint angle from a half-occluded hip is guessing. We built an exercise-verification engine that refuses to guess. 🧵
- 10 Pose estimation is free. BlazePose and MoveNet will stream you a 33-keypoint skeleton and count reps all day. The hard, billable layer is what happens after the keypoints: turning a raw pose stream into proof a clinician can act on.
- 10 Take Maria, a synthetic post-ACL patient, 15 prescribed squats. On rep 12 the hip keypoint drops below the visibility gate. The engine returns "Rep not scored. Hip not visible." The skeleton greys out. No angle gets invented.
- 10 That abstention is the whole point. An occluded rep that gets scored anyway is, by definition, a fabricated angle. So we treat abstention as a measured invariant, not a nice-to-have you hope holds up in production.
- 10 The number, on a fixed labeled synthetic set: of 3 occluded reps, 3 abstained, 0 fabricated angles. It is derived from the data, not asserted. An occluded rep carrying a confident angle would itself be the violation.
- 10 Why it matters: a raw pose library will happily log 15 of 15 good squats on a knee caving medially on rep 9. Confident and wrong is exactly the failure mode you do not want baked into a patient's billing record.
- 10 Mechanically, the confidence gate sits after 1-Euro smoothing and before any biomechanical feature is computed. Below the visibility threshold, no feature runs. Plain, unit-tested NumPy. No LLM anywhere in the scoring path.
- 10 Agents advise, code decides. The optional longitudinal monitor and clinical scribe only interpret the deterministic numbers. Any figure an agent introduces that is not already in the evidence gets rejected before it ships.
- 10 Honest question for the pose and signal folks: a keypoint goes low-confidence mid-rep. Do you interpolate, drop the rep, or flag and abstain? We chose abstain. Where would you draw that line, and why?
#PoseEstimation #ComputerVision #DigitalHealth #RTM #HealthTech - 10 The whole thing runs keyless, no live camera, on synthetic patients plus one real captured squat clip. Go find where it abstains: veriprajna.com/demos/ai-biomechanics-exercise-verification
Published on X · August 7, 2026
On social media
See this post on its original platform
In our archive