No sensor measures what you actually care about. An IMU measures orientation, not a joint angle; a sweat patch measures concentration, not fatigue; a camera measures keypoints, not force. Three 2026 papers each sit several inferential steps from their target — and the credibility of the final number depends on every step in between.
Table of Contents
Several steps from the raw signal
Each of these systems outputs a proxy, then reconstructs the quantity you want with a model — and every step adds error. That is not a flaw to be shocked by; it is the nature of inferred measurement. The useful habit is to ask, for any wearable number, how many modelled steps separate the raw signal from the claim on the label.
Paper 01 — Many IMUs, one clock
The first tackles the unglamorous foundation: time. A multi-IMU system with microsecond inter-sensor timestamping and Kalman filtering against drift was checked against a vision reference on knee motion, holding near-zero drift over a two-hour session. Before an IMU can become a joint angle, every sensor has to agree on when “now” is — get the clock wrong and the model downstream inherits the error.
Paper 02 — Reading fatigue in sweat
The second reads sweat — sodium, lactate, glucose, cortisol, pH — and maps those analytes to hydration, fatigue and stress status, with AI-assisted training and recovery advice on top. The chemistry is real; the inference is the hard part. Sweat composition varies between people and sessions, and calibration is the limiting factor between a concentration and a confident “you are fatigued.”
Paper 03 — From video keypoints to a health label
The third goes furthest down the inference chain: neural nets estimate ground reaction force from markerless-video kinematics, then classify health status without a force plate. When it works, it is remarkable — a force and a clinical label from ordinary video. But it is validated so far only in healthy adults, and each modelled step (keypoints → kinematics → force → label) is a place the final claim can drift.
Mind the gap between signal and state
The three papers rhyme: an IMU reports orientation, and a joint angle needs sync, filtering and a model; a sweat sensor reports concentration, and a fatigue state needs calibration; video reports keypoints, and a force or a health label needs a model. Mind the gap between signal and state — the sensor is not the measurement. For worked examples of that gap, see from kinematics to a modelled injury number and two pose pipelines, validated or not.
Frequently asked questions
Why isn’t a sensor reading the same as a measurement?
Because sensors output proxies. An IMU gives orientation, a sweat patch gives chemical concentration, a camera gives keypoints — and the quantity you actually want (a joint angle, a fatigue state, a force) is reconstructed from that proxy by a model. Every step in that chain adds error, so the final number is only as trustworthy as the weakest link.
Can wearable sweat sensors measure fatigue?
They measure analytes — sodium, lactate, glucose, cortisol, pH — that correlate with hydration, fatigue and stress. Turning those concentrations into a reliable fatigue state is the hard part, limited by how much sweat composition varies between people and sessions and by calibration. It is promising, not yet a settled measurement.
Can you estimate ground reaction force from video?
A 2026 system used neural nets to estimate ground reaction force from markerless-video kinematics and even classified health status without a force plate. It is a striking result, but validated so far only in healthy adults — and it stacks several modelled steps (keypoints → kinematics → force → label), each a potential source of drift.
References
[1] Samarasekera, T., Rathnayaka, P., Adhikari, S., et al. (2026). A Synchronized Multi-IMU Wearable System for Tracking of Joint-Angles in Sports Motion Analysis With Reference-Based Validation and Dynamic Task Characterization. arXiv:2607.26027.
[2] Yan, J., Shi, J., Zhang, Z., Yang, M., Zhang, M., & Wang, M. (2026). Applications of Wearable Sweat Biosensors in Sports Activities with Real-World Cases. Bioengineering, 13(8), 906.
[3] Rodriguez-Granados, E. J., Urriolagoitia-Sosa, G., Romero-Angeles, B., et al. (2026). Dual-Model Artificial Intelligence Framework Integrating AI-Based Markerless Motion Capture for Dynamic Gait Prediction and Health-Status Classification. Diagnostics, 16(16), 2588.
Takashi Fukushima — Sports Science & Pose Estimation.
▶ Subscribe on YouTube · Website · ORCID · Contact