The research, led by Peng Yang, Haifeng Zhang, Chenglong Feng and colleagues at Shanghai University of Engineering Science and Shanghai Yangzhi Rehabilitation Hospital affiliated with Tongji University, tackled one of the thorniest problems in wearable health monitoring: a model trained on one person’s muscle signals usually fails when applied to another. This cross-participant generalization problem has limited the clinical usefulness of electromyography-based fatigue detection for years, because every individual’s signal amplitude, electrode placement, and muscle anatomy differ enough to fool conventional machine learning classifiers.

To build their dataset, the team recorded twelve-channel surface electromyography from 42 participants, 28 healthy adults and 14 people with chronic low back pain, during rehabilitation-relevant tasks. Surface electrodes capture the summed electrical activity of underlying muscles through the skin, and as muscles fatigue, the frequency content of those signals shifts measurably downward. The researchers labeled fatigue using a dual-criterion approach based on the median frequency of the signal, a standard marker of muscle fatigue onset, and then evaluated their models with leave-one-subject-out cross-validation, meaning the model was repeatedly tested on participants it had never seen during training.

The core innovation lies in what the network was forced to learn. Rather than letting a deep neural network freely mine patterns from the data, the team embedded three biomechanical soft constraints drawn from established physiology: the Hill force-velocity relationship describing how muscle force output changes with contraction speed, a fatigue-dynamics steady-state model capturing how fatigue accumulates and recovers over time, and an EMG-force residual linking electrical activity to mechanical output. A fourth regularization term enforced smoothness across the ordering of the twelve recording channels. These constraints act like guardrails, penalizing predictions that fit the data but violate the physics of how muscles actually behave.

This physics-informed neural network, or PINN, approach represents a growing trend in biomedical machine learning. Pure data-driven models can achieve impressive benchmark scores yet behave implausibly when pushed outside their training distribution, a serious concern in clinical settings. By encoding known physical laws directly into the loss function, the Shanghai team’s model had to produce outputs consistent with muscle mechanics, which the authors argue improves both generalization to unseen participants and the interpretability of the learned representations.

The performance numbers are notable. The PINN achieved an area under the precision-recall curve of 0.923 and an F1 score of 0.919, with perfect recall, meaning it never missed a true fatigue event in the tested cohort. Its PR-AUC exceeded the strongest baseline, a support vector machine, by 0.029, a difference that was statistically significant with a 95 percent bias-corrected and accelerated confidence interval of 0.015 to 0.044 and a p-value of 0.004. Convolutional neural network and transformer baselines were also outperformed, suggesting that the biomechanical priors provided information those architectures could not extract from the signals alone.

Just as important as raw accuracy is calibration, the alignment between a model’s stated confidence and its actual correctness. A classifier that says it is 90 percent confident should be right about 90 percent of the time, especially when its output will guide clinical decisions. The researchers applied temperature scaling, a post-hoc calibration technique, which reduced the expected calibration error from 0.0508 to 0.0216 and improved the Brier score from 0.1329 to 0.1305. They then converted calibrated channel probabilities into discrete fatigue-risk states using a dual-threshold hysteresis scheme with consecutive-window confirmation, a design that prevents the system from flickering erratically between fatigue states when signals hover near a decision boundary.

The constraints paid off in physical consistency as well. Compared with an otherwise identical model lacking the prior terms, the exceedance rates for the three biomechanical residuals dropped by 24.1, 23.5, and 25.7 percentage points respectively, while the channel-order residual fell by 20.0 percentage points. In plain terms, the informed model’s predictions violated the encoded laws of muscle mechanics far less often, which the authors interpret as evidence that the priors genuinely shaped the network’s internal representations rather than merely adding a regularizing penalty.

Perhaps the most clinically forward-looking element is the sequential trigger-state generation. In a chronological replay evaluation, each fatigue-risk state was updated using only the current and preceding time windows, mimicking real-time deployment on a wearable device or rehabilitation platform. This streaming design, combined with the hysteresis logic, demonstrates a feasible pipeline for continuous, closed-loop fatigue monitoring in which a system could alert a therapist or adjust exercise intensity the moment a patient’s muscles cross a calibrated risk threshold.

The authors are careful to frame the scope of their claims. The gains were demonstrated within the studied cohort, task set, and comparator models, and the study was conducted at a single rehabilitation hospital under ethics approval from the Shanghai Yangzhi Rehabilitation Hospital Medical Ethics Committee, with all participants providing written informed consent. They emphasize that prospective clinician-in-the-loop studies are still required to establish whether the technology delivers measurable clinical benefit in closed-loop rehabilitation practice. The work was supported by the National Natural Science Foundation of China and a national clinical key specialty construction project. Even with those caveats, the study offers a compelling template for the next generation of rehabilitation wearables: neural networks that not only fit the data but also obey the physics of the body they are watching, and that know when to admit uncertainty. For the millions navigating chronic low back pain, a machine that reliably recognizes the moment muscles begin to fail could turn rehabilitation from an exercise in guesswork into a precisely dosed therapy.

Subject of Research: Physics-informed neural networks for surface electromyography-based muscle fatigue monitoring in chronic low back pain rehabilitation

Article Title: Biomechanics-informed PINN with calibrated trigger-state generation for sEMG fatigue monitoring in chronic low back pain rehabilitation

Article References: Yang, P., Wang, Z., Niu, W., Wang, Y., Zhang, H., & Feng, C. (2026). Biomechanics-informed PINN with calibrated trigger-state generation for sEMG fatigue monitoring in chronic low back pain rehabilitation. BioMedical Engineering OnLine. https://doi.org/10.1186/s12938-026-01637-z

Image Credits: AI Generated

DOI: 10.1186/s12938-026-01637-z

Keywords: surface electromyography, muscle fatigue, physics-informed neural network, chronic low back pain, rehabilitation, biomechanics, probability calibration, machine learning, Hill force-velocity model, wearable health monitoring, trigger-state generation, cross-participant generalization