{"id":"258f9084-24f3-48b4-88f5-8d042687986c","arxiv_id":"2506.05118","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Gaussian process trained on simulated skin wave experiments predicts stress and pre-stretch from two surface wave speeds with R2=0.957, but the full method lacks in vivo validation.","lead":"This paper tests whether a machine learning model can predict the tension and natural stretch of skin from the speed of surface waves measured without cutting the skin. The model is trained on thousands of simulated wave experiments and checked on synthetic skin, with partial success that stops short of validating the full method on real tissue.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The full stress/pre-stretch inversion depends on the supersonic shear wave speed, which the paper's own experiment could not detect; without it, pre-stretch predictions collapse (R²=0.1405), so the central in vivo claim is untested.","rationale":"The reader's weakest assumption correctly identifies the same load-bearing concern: the supersonic shear wave measurement is indispensable for the pre-stretch prediction, yet the experimental device could not detect it. The paper itself flags this in Section 2.5 ('it was only possible to extract information about the Rayleigh wave as the supersonic wave was not visible') and in the Discussion ('the supasonic shear wave was not visible... it was not possible to validate the GP model presented in Section 3.3'). This is not an external or manufactured objection; it is the manuscript's own admitted limitation. The FE cross-validation is honest and the emulator is nearly perfect (R²=0.9993), with good analytical agreement, so the numerical pipeline is credible as a proof of concept. However, the abstract and conclusions overstate the clinical readiness: no experiment has yet produced the input pair (v_R, v_s) on any skin-like material, let alone in vivo skin. The proposed concrete test—using a more sensitive sensor to measure v_s on a synthetic sample with known pre-stretch and then evaluating the Section 3.3 model—directly settles whether the central inversion can work with physically obtainable measurements. If it cannot, the method reduces to a stiffness predictor with known pre-stretch, not a skin tension estimator. Thus the reader's CONDITIONAL verdict is appropriate, and no change in verdict is needed.","tokens_in":20686,"tokens_out":3874,"duration_ms":46850,"concrete_test":"Repeat the Section 2.5 experiment using a sensor setup with sufficient bandwidth and sensitivity to resolve the supersonic shear wave (e.g., higher-frequency piezoelectric transducers or optical coherence elastography as in Li et al. [22]) on a synthetic skin sample with known uniaxial pre-stretch values (1.12, 1.19, 1.22, 1.27). Measure both v_R and v_s, feed them into the GP inverse model from Section 3.3, and compare predicted λ_1 and S_11 to the known applied stretch and the stress computed from the constitutive model. If v_s cannot be reliably detected, or if the median absolute error in predicted λ_1 exceeds 0.05, the central inversion claim is not supported by experiment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that two surface wave speed measurements (v_R and v_s) suffice to infer steady-state stress S_11 and natural pre-stretch λ_1 in vivo. The GP inverse model achieves R²=0.9570 in cross-validation, but this is entirely dependent on v_s: with v_R alone, λ_1 prediction drops to R²=0.1405 (Section 3.3). The experimental validation in Sections 2.5 and 3.4 explicitly states that the supersonic wave was not visible with the piezoelectric sensors, so the authors instead validated a different GP model that inputs known pre-stretch and v_R to predict Young's modulus. That validation does not exercise the central inverse model at all. The Discussion acknowledges this limitation, but the abstract and conclusions nevertheless state that the method provides 'real time non-invasive access to in vivo stretch and stress.' Since the only measurement that carries the pre-stretch information was not experimentally obtained, the central clinical claim remains unsupported. The FE-based cross-validation is internally sound and the emulator's agreement with the analytical Rayleigh solution (R²=0.9951) is independent support, but it only shows that the GP can invert the FE simulator, not that real in vivo skin produces a detectable v_s or that the simulated mapping survives experimental noise, anisotropy, layering, and viscoelasticity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a machine-learning pipeline for non-invasive estimation of in vivo skin stress and pre-stretch from surface wave speed measurements. The authors simulate 10,000 uniaxial pre-stretch and wave-propagation finite element experiments over a Latin hypercube of material parameters, train a Gaussian process emulator mapping (E, β, ρ, λ1) to Rayleigh and supersonic shear wave speeds, and then train a second Gaussian process to invert (v_R, v_s) to steady-state stress S11 and natural pre-stretch λ1. The inverse model achieves 10-fold cross-validation R²=0.9570. An experimental device with piezoelectric sensors is used on synthetic skin at four pre-stretch levels; Rayleigh wave speeds are measured and agree reasonably with the FE/analytical predictions, but the supersonic wave is not detected. A separate GP trained on FE data with pre-stretch and v_R as inputs predicts the synthetic skin's Young's modulus within one standard deviation of an independent measurement. The paper concludes that elastic wave measurements combined with machine learning provide a viable non-invasive method for determining in vivo skin tension.","tokens_in":20927,"tokens_out":7796,"duration_ms":88779,"significance":"If the two-input inverse mapping were validated on physical skin, the approach would be a meaningful advance: a cheap, fast, non-invasive measurement of skin pre-stretch and stress could inform surgical planning. The paper's strengths are its large simulator-based dataset, the GP emulator's agreement with the analytical Rayleigh solution (R²=0.9951, Section 3.1), stable 10-fold cross-validation with a small standard deviation, and a successful experimental prediction of Young's modulus for synthetic skin using a FE-trained GP. However, the headline inverse model (v_R, v_s) → (S11, λ1) is not tested on experimental data because v_s could not be measured; the experimental validation exercises a different model. The paper is therefore best read as a simulation-based proof of concept with a partial experimental feasibility demonstration.","major_comments":[{"comment":"The central inverse model (Section 2.4) has not been validated experimentally. Section 2.5 states that \"the supersonic wave was not visible\" with the piezoelectric sensors, and Section 3.4 consequently trains a new GP that uses pre-stretch and v_R to predict Young's modulus, explicitly noting that the Section 2.4 model could not be validated. Because Section 3.3 shows that λ1 prediction collapses to R²=0.1405 when only v_R is used, the usefulness of the method depends critically on v_s detectability. The abstract's statement that the method provides \"real time non-invasive access to in vivo stretch and stress\" and the corresponding concluding claim are therefore not supported by the experimental results; either demonstrate v_s detection with a more sensitive receiver or explicitly scope the claims to the simulated setting.","section":"Section 2.5, Section 3.4, Conclusions"},{"comment":"The 10-fold cross-validation R²=0.9570 is computed on noiseless, deterministic simulator outputs. The authors acknowledge in Section 4 that no noise is present in the training data. Real wave-speed measurements on synthetic skin show 1–3 m/s scatter for a fixed Young's modulus (Figure 12a), and the experimental device could not resolve the weaker v_s signal. Since the reported R² does not include measurement noise, it likely overstates the real-world performance of the inverse model. A noise-robustness study (e.g., adding realistic jitter to the simulated inputs during training, or reporting prediction intervals under measurement error) would substantially strengthen the generalization claim.","section":"Section 3.3 and Section 4"},{"comment":"The inverse problem is described as \"ill-posed\", but the identifiability of the target pair (S11, λ1) from the two wave speeds is not examined. Because the forward simulator maps four inputs (E, β, ρ, λ1) to two outputs, different parameter combinations can in principle produce identical (v_R, v_s); in that case the GP would regress to a conditional mean and the R² value would hide the non-uniqueness. A check of the Jacobian of the forward map over the input hypercube, or an analysis of how many training points with similar (v_R, v_s) have different targets, would clarify the actual degree of ill-posedness.","section":"Section 2.4 and Section 3.3"}],"minor_comments":[{"comment":"The third formula should read D1 = (9−18ν)/(E(1+ν)); as written, \"ν=\" is a typo that obscures the intended relationship.","section":"Section 2.2, Eq. (6)"},{"comment":"The distance between the two sensors (17.13 mm) appears only in the Figure 5 caption; please state it in the text, along with the distance from the impact site to the first sensor.","section":"Section 2.5"},{"comment":"\"transistion\" should be \"transition\" and \"feasability\" should be \"feasibility\".","section":"Section 3.2 and Section 4"},{"comment":"Specify the units of all fitted coefficients (e.g., E in kPa, wave speeds in m/s, S11 in Pa) and note that Equations 9 and 10 include a quadratic term in E, so they are not strictly linear regression models as described in the text.","section":"Equations 9–12"},{"comment":"Clarify that the GP used for Young's modulus prediction is the newly trained model from Section 3.4 (inputs: pre-stretch and v_R), not the two-input inverse model of Section 2.4.","section":"Figure 12b"},{"comment":"When comparing the two-input and one-input GP models, specify whether the same kernel and hyperparameter settings were used for both, so that the comparison between R²=0.9570 and R²=0.1405 is meaningful.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent proof-of-concept, but the gap between the simulated inverse model and the experimental demonstration is larger than the abstract and conclusions suggest. In its current form, the manuscript risks overclaiming clinical readiness. I would encourage the editor to require a revision that either detects the supersonic shear wave or carefully re-scopes the central claim to an in silico feasibility study with experimental support for the forward wave-speed–stiffness relationship."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a competent proof-of-concept for inverting surface wave speeds with Gaussian process surrogates trained on FE simulations. The genuinely new bit is the combination: a GP emulator for the forward FE model plus a GP inverse model that maps Rayleigh and supersonic shear wave speeds to steady-state stress and pre-stretch. The cross-validation (R²=0.957) is strong, and the emulator agrees with the analytical Rayleigh solution (R²=0.995), which is a nice independent check.\n\nI agree with the reader's conditional verdict, and the stress-test note holds up on reading. The central inverse model depends on the supersonic wave speed—drop it and pre-stretch prediction collapses to R²=0.14—yet the experiment could not detect the supersonic wave at all. The authors explicitly acknowledge this in Section 3.4 and the Discussion, and they validate a different GP that predicts Young's modulus from pre-stretch and Rayleigh speed. That validation works, matching the independent 146 kPa from Kho et al. within one standard deviation, but it does not exercise the stress/pre-stretch inversion. So the abstract's claim of 'real time non-invasive access to in vivo stretch and stress' is ahead of the evidence.\n\nOther soft spots are real but secondary: no code or data shared, the FE model is 2D and ignores anisotropy, viscoelasticity, layering and noise, and the simplified parametric equations are fits to emulator output, not physics. The authors list most of these limitations themselves, which is honest.\n\nWho is this for? Researchers working on elastography or surgical planning who want a cheap surrogate-based inverse method. It deserves a serious referee, though that referee should push for the experimental supersonic detection or a tempered claim. I'd recommend acceptance of a revised version that either reports the supersonic wave experimentally or reframes the central claim as a simulation-based proof of concept.","headline":"A solid simulation-based proof of concept for GP inversion of skin wave speeds, but the central in vivo claim is untested because the supersonic wave was never detected experimentally.","tokens_in":21506,"tokens_out":2980,"would_cite":true,"duration_ms":34668,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Gaussian process trained on 10,000 finite-element skin simulations maps Rayleigh and supersonic wave speeds to stress and pre-stretch, with cross-validated $R^2 = 0.9570$.","keywords":["skin tension","Rayleigh surface wave","supersonic shear wave","Gaussian process regression","finite element simulation","pre-stretch","non-invasive measurement","inverse problem"],"falsifier":"Run an in vivo study on human skin with a sensor that can resolve both the Rayleigh and supersonic arrivals, then compare the GP-predicted $S_{11}$ and $\\lambda_1$ against an independent measurement such as excision retraction; if the supersonic wave is undetectable or the predictions disagree beyond the training range, the central claim fails.","tokens_in":20467,"feed_emoji":"🩺","tokens_out":10600,"duration_ms":111525,"temperature":0.7,"pith_summary":"This paper aims to show that skin tension and natural pre-stretch can be inferred from surface wave speeds without cutting the skin. It builds a dataset of 10,000 simulated wave-propagation experiments from a simplified two-dimensional finite-element model of pre-stretched hyperelastic skin, then trains a Gaussian process regression on that dataset. The inverse model takes the Rayleigh wave speed $v_R$ and the faster supersonic shear wave speed $v_s$ as inputs and returns the steady-state principal stress $S_{11}$ and the natural pre-stretch $\\lambda_1$, reaching $R^2 = 0.9570$ in cross-validation. The supersonic wave is essential for this: with only $v_R$, pre-stretch predictions degrade to $R^2 = 0.1405$. Experimental Rayleigh wave speeds from stretched synthetic skin match the simulations, and a simulation-trained model estimates the sample's stiffness close to its independently measured value, supporting the feasibility of a cheap non-invasive clinical measurement.","feed_headline":"Two wave speeds can read skin tension without cutting","feed_subtitle":"A Gaussian process trained on 10,000 simulated skin experiments turns surface wave speeds into stress and pre-stretch estimates.","key_machinery":"The machinery is a two-dimensional finite-element model of a pre-stretched skin block that simulates a surface wave propagation experiment, generating 10,000 training cases across two hyperelastic material descriptions, coupled with a multi-output Gaussian process regression using a radial-basis-function kernel. A Gaussian process is a statistical model that predicts an output value along with an uncertainty from previously seen training examples. The GP serves two roles: as an emulator it reproduces the finite-element wave speeds with $R^2 = 0.9993$ at about four orders of magnitude lower computational cost; as an inverse solver it maps $(v_R, v_s)$ to $(S_{11}, \\lambda_1)$. The inversion works because increasing pre-stretch raises the Rayleigh speed while lowering the supersonic shear speed, so the speed pair carries far more information about stretch than either speed alone.","core_discovery":"The central claim is that the ill-posed inverse problem of determining in vivo skin stress and pre-stretch from surface wave speeds can be solved in real time with a statistical surrogate. Using the pair of measurable speeds, $v_R$ and $v_s$, a Gaussian process regression trained on finite-element outputs predicts the steady-state stress $S_{11}$ and natural pre-stretch $\\lambda_1$ with cross-validated $R^2 = 0.9570$. The supersonic wave is load-bearing: dropping it leaves pre-stretch almost unidentifiable ($R^2 = 0.1405$), while including it makes pre-stretch recoverable. Experimental Rayleigh wave speeds from uniaxially stretched synthetic skin agree with the FE predictions up to moderate stretch, and a model trained only on simulations predicts the synthetic skin's Young's modulus near the 146 kPa measured by independent destructive characterisation. The paper concludes that elastic wave measurements combined with machine learning are a viable non-invasive route to patient-specific skin tension.","pith_inferences":["Beyond the paper: if clinical sensors cannot detect the supersonic shear wave, the two-speed inversion presented here cannot be deployed as-is, and the Rayleigh-only fallback gives stiffness but not pre-stretch.","Beyond the paper: a decisive engineering test would move receivers closer to the impact site and increase excitation amplitude to see whether the supersonic wave survives attenuation in living human skin.","Beyond the paper: the same emulator-plus-inverse-GP recipe transfers to other pre-stressed soft tissues with expensive forward simulations and cheap wave measurements, such as tendon or arterial wall.","Beyond the paper: because the training data are deterministic and noise-free, adding realistic sensor noise or retraining on real waveforms is the natural next step toward clinical robustness, as the paper itself notes but does not test."],"forward_implications":["With two measured wave speeds, the inverse GP returns $S_{11}$ and $\\lambda_1$ in near real time, replacing iterative inverse finite-element fitting for this problem.","A Rayleigh-only device can still estimate stiffness, as demonstrated on synthetic skin, but cannot recover pre-stretch ($R^2 = 0.1405$), so device design should target reliable detection of the supersonic shear wave.","The simplified fitted equations for $S_{11}$ and $\\lambda_1$ provide closed-form approximations that portable clinical devices could evaluate without running the GP.","The training pipeline extends by adding new simulator runs, so additional material models or three-dimensional geometries can be incorporated as future training data."],"supporting_citations":[{"why":"Reports in vivo elastic wave measurements showing subject-specific skin tension direction and magnitude, motivating the non-invasive device concept.","marker":"[11]"},{"why":"Provides the analytical Rayleigh wave speed formula for a pre-stressed hyperelastic half-space against which the FE outputs are validated ($R^2 = 0.9951$).","marker":"[21]"},{"why":"Gives the two-wave description of Rayleigh and supersonic shear waves under prestress, supplying the physical basis for using $v_s$ as a second input.","marker":"[22]"},{"why":"Supplies the mean failure strain of excised human skin used to set the sampled pre-stretch range.","marker":"[29]"},{"why":"Supplies the standardised-regression-coefficient sensitivity analysis method used to rank the influence of each material input.","marker":"[42]"},{"why":"Defines Gaussian process regression, the model class used for both the emulator and the inverse solver.","marker":"[47]"},{"why":"Provides the independently measured Young's modulus of the synthetic skin sample used as the experimental validation target.","marker":"[49]"}],"fun_headline_variants":["Skin tension from wave speeds: ML model does it non-invasively","Gaussian process maps wave speeds to skin stress and pre-stretch","Non-invasive skin tension: two wave speeds, one ML model","Wave-speed ML predicts skin tension without cutting","Surface waves + GP model = non-invasive skin tension"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the supersonic shear wave speed can be measured reliably in living human skin; the paper's own experimental device did not detect it, and without it pre-stretch predictions degrade to $R^2 = 0.1405$.","fun_headline_variants_meta":{"raw":{"variants":["Skin tension from wave speeds: ML model does it non-invasively","Gaussian process maps wave speeds to skin stress and pre-stretch","Non-invasive skin tension: two wave speeds, one ML model","Wave-speed ML predicts skin tension without cutting","Surface waves + GP model = non-invasive skin tension"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000885,"raw_usage":{"total_tokens":3857,"prompt_tokens":1013,"completion_tokens":2844,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":2761}},"tokens_in":629,"tokens_out":2844,"duration_ms":23378,"temperature":1.0,"reasoning_tokens":2761,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:24:41.753843+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an in vivo study on human skin with a sensor that can resolve both the Rayleigh and supersonic arrivals, then compare the GP-predicted $S_{11}$ and $\\lambda_1$ against an independent measurement such as excision retraction; if the supersonic wave is undetectable or the predictions disagree beyond the training range, the central claim fails.","supporting_citations":[{"cited_title":"Nagle, S","cited_arxiv_id":null,"evidence_quote":"Reports in vivo elastic wave measurements showing subject-specific skin tension direction and magnitude, motivating the non-invasive device concept."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the analytical Rayleigh wave speed formula for a pre-stressed hyperelastic half-space against which the FE outputs are validated ($R^2 = 0.9951$)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the independently measured Young's modulus of the synthetic skin sample used as the experimental validation target."}],"review_version":1}