{"id":"b3fd4a51-5d69-4d21-8aff-69a8847f282f","arxiv_id":"2412.04213","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A physics-informed neural network predicts muscle forces from unlabeled sEMG by embedding the Hill muscle model as a loss, matching supervised baselines on wrist and knee data.","lead":"This paper trains a neural network to predict muscle forces from surface EMG signals without needing muscle force labels, using a physics-based Hill muscle model as an extra training constraint. It reports wrist and knee predictions comparable to supervised methods, plus estimates of individualized muscle parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is not yet supported because the OpenSim CMC ground truth and the physics loss are generated from the same Hill-type muscle model, so the reported RMSE/R2 may measure self-consistency rather than physiological accuracy; the 'unlabeled' claim is also overstated since Eq.","rationale":"The reader's conditional verdict is appropriate: the core idea is plausible and the experiments show internal consistency, but the central claim as stated overreaches. My stress test agrees with the reader's weakest assumption and sharpens it: the ground-truth and loss share a model family, the 'physiological range' check is based on the same initial guesses used to seed the optimization, and the method uses joint-angle labels despite claiming to be label-free. These issues do not prove the method is wrong, but they mean the reported accuracy could reflect model self-consistency rather than real physiological identifiability. A cross-model synthetic validation would settle whether the physics-informed constraint recovers individual muscle forces or only reproduces the torque-level behavior of the embedded model, so I recommend no change to the conditional verdict.","tokens_in":15739,"tokens_out":10445,"duration_ms":117087,"concrete_test":"Generate a synthetic dataset with a deliberately mismatched musculoskeletal model: use OpenSim CMC with compliant-tendon muscles and muscle strengths perturbed ±30% from the initialized generic values, drive the model with the recorded sEMG-derived activations, and produce joint angles plus individual muscle forces. Then train the proposed method on only sEMG and joint angles and compare the recovered individual muscle forces with the mismatched ground truth. If the individual forces are still recovered at the reported RMSE (~3-7 N), the concern is resolved; if only the net joint torque matches or RMSE degrades markedly, the reported accuracy is a model-consistency artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A states that ground-truth muscle forces are produced by OpenSim's CMC tool, a Hill-type musculoskeletal simulation with parameters initialized from the same scaled generic model used to build the physics loss. The training loss embeds the same modeling assumptions: Eq. (3) enforces Hill-based forward dynamics, Eq. (4) computes torque from Hill muscle forces, and Eq. (8) defines those forces using the same force-length-velocity relation. The reported RMSE/R2 in Tables III-IV therefore measure consistency with the model family the network was trained against, not agreement with independently measured muscle forces. The parameter identification is likewise not independently validated: the 'physiological ranges' in Table II are constructed as ±50% (F_m0) and ±0.01 m (l_m0) around the OpenSim generic initial guesses (Table I), so lying inside them is a boundedness check; several identified values sit at these bounds (FCR F_m0=205.2 N vs lower bound 203.5 N; ECU=286.6 N vs upper bound 288 N), suggesting bound-constrained optimization rather than identifiable subject-specific parameters. Additionally, Eq. (2) supervises the network with measured joint angles, so the abstract's 'without any label information' is internally inconsistent; the accurate claim is 'without muscle-force labels.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-informed deep learning framework for predicting muscle forces from sEMG signals and simultaneously identifying muscle-tendon parameters. The network is trained by combining an MSE loss on joint angles with two physics-based losses derived from a Hill-type muscle model and forward-dynamics torque balance. Experiments are conducted on a wrist flexion/extension dataset from six subjects, with additional knee data, and the method is compared against LSTM, GRU, CNN, FNN, SVR, and ELM baselines that use labeled muscle forces. The authors claim that the method works with unlabeled sEMG data and achieves comparable or better RMSE/R2 than the baselines.","tokens_in":16030,"tokens_out":5329,"duration_ms":57734,"significance":"If the claims were fully supported, the paper would offer a practically valuable contribution: a physics-informed approach that avoids muscle-force labels, runs in real time, and yields interpretable subject-specific parameters. The paper has genuine strengths: a carefully designed composite loss, a multi-subject wrist experiment with six baselines, additional knee generalization, sensitivity analyses over architectures/hyperparameters, and convergence plots. However, the significance is conditional on resolving two central issues: the method is not label-free as claimed because joint angles are supervised through Eq. (2), and the evaluation is performed against OpenSim CMC ground truth generated from a Hill-type model of the same family as the physics loss, so the reported accuracy may measure self-consistency with a simulation model rather than physiological accuracy.","major_comments":[{"comment":"The paper claims training \"without any label information\" and \"unlabeled sEMG data,\" but Eq. (2) computes an MSE loss between the predicted and actual joint angles, so measured joint angles are used as supervised labels during training. The claim should be narrowed to \"without muscle-force labels\"; as written, the label-free claim is internally inconsistent.","section":"Abstract and Section II-C, Eq. (2)"},{"comment":"Ground-truth muscle forces are generated by OpenSim's Computed Muscle Control from a scaled generic Hill-type musculoskeletal model, while the physics losses embed a Hill-model forward dynamics of the same family: Eq. (8) is the same force-length-velocity relationship, and Eqs. (3)-(4) impose the same torque-balance structure. The reported RMSE/R2 values in Tables III-IV therefore quantify agreement with a model-generated target that shares the model class used for regularization, not agreement with independently measured muscle force. Independent validation (e.g., dynamometer joint torques, instrumented implant forces, or a CMC model built with deliberately different muscle parameters) is needed before \"effectively predict muscle forces\" and \"accurately identify\" parameters can be claimed.","section":"Section III-A and Section II-D, Eqs. (3)-(4), (8)"},{"comment":"The \"physiological ranges\" in Table II are defined relative to the OpenSim initial guesses of Table I (±50% for F_m0 and ±0.01 m for l_m0). Thus, the finding that all identified parameters fall within these ranges is only a boundedness check around the initialization, not evidence of accurate or identifiable subject-specific parameters. Several identified values sit at or effectively on the bounds (FCR F_m0=205.2 N vs. lower bound 203.5 N; ECU F_m0=286.6 N vs. upper bound 288 N), which suggests bound-constrained optimization rather than identifiable estimates. No identifiability analysis or uncertainty quantification is provided.","section":"Section IV-B, Table II"},{"comment":"The conclusion that the proposed method performs \"comparable and even better\" than the labeled baselines is based on single point estimates without repeated-seed variability or significance testing. Several individual comparisons favor the baselines (e.g., S5 FCR RMSE 5.41 vs. LSTM 2.75; S6 ECU RMSE 3.97 vs. LSTM 1.23; S2 FCR RMSE 7.01 vs. LSTM 6.15). Aggregate claims need statistical support, such as confidence intervals over random seeds/partitions or paired tests across subject-muscle cells.","section":"Section IV-C, Tables III and IV"}],"minor_comments":[{"comment":"The S1 ELM row under ECRB reports R2 = 10.90, which is impossible for a coefficient of determination; this appears to be a typo and should be corrected.","section":"Section IV-C, Table IV"},{"comment":"All comparison results are reported as point estimates from a single 70/30 random split; reporting means and standard deviations over multiple splits or seeds would make the comparisons more convincing.","section":"Tables III and IV"},{"comment":"The intrasession (cross-speed) scenario is evaluated only with representative prediction curves; quantitative RMSE/R2 values for this scenario would substantially strengthen the robustness claim.","section":"Section IV-D, Fig. 6"},{"comment":"The data description does not specify how many trials per subject are used in the main results or whether the reported metrics are averaged over trials or over the two speeds beyond the statement that \"the same flexion speed\" is used for training and testing; this should be clarified.","section":"Section III-A"},{"comment":"The three loss terms Lq, Lfd, and LF have different physical units (rad^2, N^2 m^2, and N^2) and are summed without explicit weighting; the authors should justify this choice, especially given the small magnitude of Lq noted in Section IV-A.","section":"Section II-C"}],"recommendation":"major_revision","confidential_remarks":"The core issue for the editor is the mismatch between the paper's \"unlabeled\" claim and the actual use of joint-angle supervision, combined with the shared model family between the OpenSim CMC target and the physics losses. A revision that reframes the claims to \"without muscle-force labels\" and adds an independent validation or a model-disagreement analysis could make the contribution publishable; without such changes, the central claim remains unsupported. The paper is otherwise within the scope of the journal and shows careful experimental work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new thing here is that they train a network to predict muscle forces without muscle-force labels, replacing them with a Hill-model forward-dynamics residual plus joint-angle supervision, and they identify muscle-tendon parameters at the same time. That is a real step beyond the cited physics-informed musculoskeletal papers, which either need force labels or require post-processing the network output with the dynamics model.\n\nThe paper does some things well. The loss design is clean, they report convergence, they compare against a sensible set of baselines, include an intrasession generalization check, and show the method transfers to the knee. They also state their simplifications openly in the discussion.\n\nThe soft spots are real and they align with the reader's report. First, 'unlabeled' is an overstatement: Eq. (2) supervises the network with measured joint angles, so the accurate phrasing is 'without muscle-force labels.' That is still a useful contribution, but the abstract as written is wrong. Second, and more important, the ground truth for muscle forces comes from OpenSim CMC, which is a Hill-type model of the same class as the one embedded in the physics losses. The network is being asked to match a target generated by a model it is also being penalized against. The reported RMSE/R2 therefore measure consistency with a model family, not agreement with independently measured forces. Third, the parameter identification is validated only by checking that values fall inside ranges that were defined as ±50% around the OpenSim initial guesses, and several identified values sit right at the bounds. That suggests bound-constrained fitting rather than identifiable subject-specific parameters. The paper does not claim independent validation, but the 'accurate identification' phrasing goes beyond what the evidence supports.\n\nThese issues are significant but not fatal. The method could still be valuable if reframed as a way to use kinematics-only supervision and avoid the expensive force-labeling step, and if the evaluation is presented as model-consistency rather than physiological accuracy. To strengthen it, they should validate against at least inverse-dynamics joint moments or a different musculoskeletal model, and report parameter estimates with uncertainty or sensitivity analysis.\n\nI would send this to peer review. It deserves serious referee time, but the authors need to fix the label claim and temper the accuracy claims.","headline":"Drops force labels but not all labels; the evaluation measures consistency with the same Hill model family used to generate the targets, so the accuracy claims need tempering.","tokens_in":16576,"tokens_out":2839,"would_cite":false,"duration_ms":29974,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Muscle forces can be predicted from sEMG without muscle-force labels by using the Hill muscle model as a physics-based training loss.","keywords":["physics-informed deep learning","muscle force prediction","unlabeled sEMG","Hill muscle model","muscle-tendon parameter identification","wrist joint","musculoskeletal forward dynamics","Computed Muscle Control ground truth"],"falsifier":"Measure actual muscle forces on the same wrist task with a direct method (for example an implanted tendon transducer or instrumented joint replacement) and compare the predicted forces in absolute units; alternatively, create a synthetic dataset with known Hill parameters and known true forces, and check whether the recovered parameters match the known values. If the predictions deviate systematically from direct measurements, the reported RMSE and $R^2$ comparisons do not transfer to real physiology.","tokens_in":15557,"feed_emoji":"💪","tokens_out":6957,"duration_ms":64316,"temperature":0.7,"pith_summary":"The paper proposes a physics-informed deep learning method that predicts muscle forces from sEMG signals without muscle-force labels during training. Instead of supervising the network with measured forces, it embeds the Hill muscle model's forward dynamics as additional loss terms that constrain the network's outputs. On a six-subject wrist flexion/extension dataset, the method reports RMSE and $R^2$ comparable to or better than LSTM, GRU, CNN, FNN, SVR, and ELM baselines that train on labeled muscle forces. The same optimization also identifies subject-specific Hill parameters inside physiological ranges. The authors argue this removes the expensive labeling bottleneck while keeping fast inference.","feed_headline":"Muscle force prediction works without force labels","feed_subtitle":"A physics loss based on the Hill model matches supervised baselines on wrist sEMG data.","key_machinery":"The load-bearing object is the Hill-based musculoskeletal forward dynamics used as a physics regularizer inside the loss. sEMG is mapped to activation through the exponential relation $a_{t,n} = (e^{A e_{t,n}}-1)/(e^{A}-1)$; contraction dynamics then give muscle-tendon force from activation and force-length-velocity factors, with pennation angle updated as $\\varphi_{t,n} = \\sin^{-1}(l^m_{o,n}\\sin\\varphi_{o,n}/l^m_{t,n})$. The forward-dynamics residual $L_{fd}$ and the implicit force residual $L_F$ together couple the network's predicted forces and the kinematic predictions to physiological parameters $\\kappa$, so the network learns force without ever seeing a force label. Identifying $\\kappa$ simultaneously makes the model subject-specific.","core_discovery":"The paper's central claim is that a fully connected network trained with the composite loss $L_{\\mathrm{total}} = L_q + L_{fd} + L_F$ can predict muscle forces from sEMG signals without any muscle-force labels, while also identifying subject-specific parameters of the Hill muscle model. The data-based term $L_q$ is the MSE between measured and predicted joint angles, $L_{fd}$ enforces the musculoskeletal forward-dynamics equation $M(q)\\ddot{q} + C(q,\\dot{q}) + G(q) = \\tau(\\kappa)$, and $L_F$ penalizes the difference between the network's force output and the force computed by the embedded Hill model. The authors report that this training scheme produces wrist flexion/extension force predictions with RMSE and $R^2$ comparable to or better than LSTM, GRU, CNN, FNN, SVR, and ELM baselines that train with labeled forces, and that the recovered parameters (maximum isometric force, optimal fiber length, activation shape factor) stay within physiological ranges. Knee examples are included to indicate generalization beyond the wrist.","pith_inferences":["The term 'unlabeled' is partial: the method still uses measured joint angles through $L_q$, so a fully label-free version would need a kinematic-free supervision or physics-only loss.","Because the ground-truth muscle forces are produced by a musculoskeletal model that contains the same Hill model family used in the physics loss, the reported accuracy partly measures self-consistency with that model rather than absolute physiological truth.","A natural extension is to weight the physics residuals relative to $L_q$ and test whether the identified parameters remain physiological when sEMG is noisy or training data are scarce.","The framework could be combined with transfer learning across subjects to shorten the long per-subject training time, a direction the paper names as future work."],"forward_implications":["Muscle-force estimators could be trained from sEMG plus joint kinematics alone, removing the need for force/torque labels that are hard to acquire in practice.","A single training run yields both force predictions and personalized muscle-tendon parameters, so subject-specific modeling does not require a separate parameter-estimation stage.","Because inference is a one-pass network forward, the trained model is fast enough for real-time biofeedback even though training is slower than standard baselines.","The physics losses act as a regularizer, which the paper shows improves stability when tested on data of a different wrist speed (intrasession generalization).","If the approach transfers to other joints as the knee examples suggest, the same recipe could reduce labeling costs across the body."],"supporting_citations":[{"why":"Provides the rigid musculotendon contraction dynamics used in the physics loss.","marker":"[39]"},{"why":"Supplies the muscle-tendon force calculation with activation and force-length-velocity terms.","marker":"[40]"},{"why":"Gives the musculoskeletal simulation tool that provides moment-arm polynomials and the Computed Muscle Control ground-truth forces.","marker":"[41]"},{"why":"Defines the EMG-to-activation nonlinearity used in the loss.","marker":"[38]"},{"why":"Provides the polynomial moment arm equations for the wrist model.","marker":"[37]"},{"why":"Established the physics-informed deep learning framework for muscle forces that this work extends to unlabeled data.","marker":"[32]"},{"why":"Supplies physiological ranges used to assess whether identified parameters are plausible.","marker":"[43]"}],"fun_headline_variants":["Physics-informed AI predicts muscle forces without labels","Unlabeled sEMG muscle force prediction via physics loss","Hill-model physics loss trains muscle force predictor label-free","No force labels needed: physics-informed net matches baselines","Deep learning plus Hill model predicts muscle force from sEMG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that muscle forces computed by the Computed Muscle Control tool of a scaled generic musculoskeletal model are valid ground truth, even though the same Hill-model machinery used to create those targets is embedded in the training loss.","fun_headline_variants_meta":{"raw":{"variants":["Physics-informed AI predicts muscle forces without labels","Unlabeled sEMG muscle force prediction via physics loss","Hill-model physics loss trains muscle force predictor label-free","No force labels needed: physics-informed net matches baselines","Deep learning plus Hill model predicts muscle force from sEMG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1414,"prompt_tokens":999,"completion_tokens":415,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":337}},"tokens_in":615,"tokens_out":415,"duration_ms":4365,"temperature":1.0,"reasoning_tokens":337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:37:42.106934+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure actual muscle forces on the same wrist task with a direct method (for example an implanted tendon transducer or instrumented joint replacement) and compare the predicted forces in absolute units; alternatively, create a synthetic dataset with known Hill parameters and known true forces, and check whether the recovered parameters match the known values. If the predictions deviate systematically from direct measurements, the reported RMSE and $R^2$ comparisons do not transfer to real physiology.","supporting_citations":[{"cited_title":"Flexing computational muscle: Modeling and simulation of musculotendon dynamics,","cited_arxiv_id":null,"evidence_quote":"Provides the rigid musculotendon contraction dynamics used in the physics loss."},{"cited_title":"Neuromusculoskeletal modeling: Estimation of muscle forces and joint moments and movements from measurements of neural command,","cited_arxiv_id":null,"evidence_quote":"Supplies the muscle-tendon force calculation with activation and force-length-velocity terms."},{"cited_title":"OpenSim: Simulating musculoskeletal dynamics and neuromuscular control to study human and animal movement,","cited_arxiv_id":null,"evidence_quote":"Gives the musculoskeletal simulation tool that provides moment-arm polynomials and the Computed Muscle Control ground-truth forces."},{"cited_title":"An EMG-driven musculoskeletal model for estimating continuous wrist motion,","cited_arxiv_id":null,"evidence_quote":"Defines the EMG-to-activation nonlinearity used in the loss."},{"cited_title":"Muscle moment arm and normalized moment contributions as reference data for mus- culoskeletal elbow and wrist joint models,","cited_arxiv_id":null,"evidence_quote":"Provides the polynomial moment arm equations for the wrist model."},{"cited_title":"Physics-informed deep learning for musculoskeletal modeling: Predicting muscle forces and joint kinematics from surface EMG,","cited_arxiv_id":null,"evidence_quote":"Established the physics-informed deep learning framework for muscle forces that this work extends to unlabeled data."},{"cited_title":"Benchmarking of dynamic simulation predictions in two software platforms using an upper limb musculoskeletal model,","cited_arxiv_id":null,"evidence_quote":"Supplies physiological ranges used to assess whether identified parameters are plausible."}],"review_version":1}