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

The aim is to develop and evaluate a physics-informed neural network (PINN) incorporating biomechanics-informed priors for cross-participant surface electromyography (sEMG) fatigue monitoring and sequential trigger-state generation during chronic low back pain (CLBP) rehabilitation. Twelve-channel sEMG was recorded from 42 participants, including 28 healthy adults and 14 participants with CLBP. After preprocessing and dual criterion median frequency labeling, time-series samples were evaluated by leave-one-subject-out cross-validation against support vector machine, convolutional neural network, and transformer baselines. Hill force–velocity, fatigue-dynamics steady-state, and EMG–force residuals were incorporated as biomechanical soft constraints, together with channel-order smoothness regularization. Temperature scaling, dual-threshold hysteresis, and consecutive-window confirmation converted channel probabilities into fatigue-risk states. The PINN achieved an area under the precision–recall curve (PR-AUC) of 0.923, an F1 score of 0.919, and recall of 1.000. PR-AUC exceeded that of the strongest baseline, the support vector machine, by 0.029 (95% bias-corrected and accelerated confidence interval, 0.015–0.044; p = 0.004; Cliff’s δ = 0.26). Temperature scaling reduced the expected calibration error from 0.0508 to 0.0216 and the Brier score from 0.1329 to 0.1305. Relative to the model without prior constraints, the exceedance rates for the three biomechanical residuals and the channel-order residual decreased by 24.1, 23.5, 25.7, and 20.0 percentage points, respectively. Within the studied cohort, task, and comparator set, the proposed method showed complementary gains in discrimination, probability calibration, and consistency with the encoded priors. Chronological replay, in which each risk state was updated using only the current and preceding windows, demonstrated the feasibility of sequential fatigue-state generation. Prospective clinician-in-the-loop studies are required to establish clinical closed-loop benefit.