{"id":"b116cf12-af6a-4b72-8745-054b6c70965f","arxiv_id":"2506.14815","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Supervised models, especially Support Vector Regression, predict lean mass, bone density, and body fat from 3D optical scans with roughly 8-11% error, while the p-Laplacian model stays around 10-20% when only 10% of patients are labeled.","lead":"Body composition metrics such as muscle mass, body fat, and bone density are usually measured by expensive DXA scans. This study tests whether machine learning on cheap 3D optical body measurements can predict them, and finds support vector regression is most accurate while a newer semi-supervised model still works with very few labeled patients.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing 10%-label supervised baseline makes the p-Laplacian data-constrained advantage untestable; the semi-supervised claim needs a same-label-rate comparison.","rationale":"The reader identifies graph smoothness as the weakest assumption. I agree that smoothness is the core modeling assumption and that BFP's low correlations make it doubtful; the paper's own discussion acknowledges BFP is harder. However, the decisive threat to the central claim is the missing same-label-rate supervised baseline. The p-Laplacian is proposed as a solution to data scarcity, so its usefulness can only be evaluated by comparing it with supervised methods at equal label counts. Without this, even a perfectly smooth graph would not establish the paper's conclusion. I would keep the reader's CONDITIONAL verdict: the paper should be conditionally accepted only after adding low-label supervised baselines and moving hyperparameter selection into an inner validation loop. The absolute supervised results at 80% remain credible, but the semi-supervised contribution is currently unverified.","tokens_in":19350,"tokens_out":6130,"duration_ms":65962,"concrete_test":"Train ridge regression, linear SVR, and LSSVR on exactly the same 10% training folds used for p-Laplacian in Tables 7-12, with hyperparameters chosen by inner cross-validation on that 10% only. Compute RMSE% with Equation 1 on the held-out 90%. Repeat for the 5% condition if feasible. If any supervised baseline attains an RMSE no worse than the p-Laplacian value for the same dataset and target (e.g., Combined ALM at 10%: 13.21 vs baseline), the data-constrained advantage is not established. Additionally, rerun p-Laplacian hyperparameter selection (p, k) inside the training folds and report the resulting RMSE to check how much of the reported gain comes from test-set selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's stated motivation for the p-Laplacian model is that supervised algorithms are data-intensive (Abstract, Section 1), and its headline evidence is the low-training-percentage performance in Section 8.5 (Tables 7-12): about 13% RMSE for ALM, 10% for BMD, and 20% for BFP at 10% training data. But Section 6 fixes all supervised models at an 80-20 split, and Tables 13-15 contain no supervised results for 10% or 5% training. Consequently, the p-Laplacian numbers are never compared against a supervised learner with the same label budget. This is not merely a missing ablation; it is the comparison that would justify the central 'promising in a data-constrained environment' conclusion. Given that the ALM biomarkers have correlations above 0.9 (Table 4), a ridge or linear SVR trained on the same ~52 labeled subjects might match or beat the p-Laplacian's 13-14% ALM RMSE. If that happens, the apparent value of the semi-supervised machinery evaporates. A secondary but related issue is that Tables 7-12 report the minimum RMSE over p and k for each training percentage, suggesting hyperparameters were selected on the test folds; a fixed inner-validation protocol is needed to ensure the low-label numbers are not optimistically selected.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper compares supervised machine-learning algorithms and a semi-supervised p-Laplacian regression model for predicting ALM, BFP, and BMD from 44 anthropometric biomarkers extracted from 3D optical images. After cleaning, 515 records are split into combined, male, and female subsets. All supervised models (linear/polynomial regularized regressions, SVR, LSSVR, random forest, XGBoost, MLP) are evaluated with five-fold cross-validation at an 80/20 training/test split. The p-Laplacian model is evaluated at training proportions of 80%, 50%, 33%, 25%, 20%, 10%, and 5%, using either all 44 biomarkers or the top-10 correlated biomarkers. The paper reports best RMSEs around 6-8% for ALM and BMD and about 11% for BFP with SVR/LSSVR at 80% training, and about 10-13% for ALM, 9-10% for BMD, and 14-25% for BFP with the p-Laplacian at 10% training, depending on the dataset. It concludes that the p-Laplacian model is promising in data-constrained healthcare settings.","tokens_in":19728,"tokens_out":10544,"duration_ms":99437,"significance":"If the reported accuracies hold, the paper would provide a useful benchmark for replacing DXA-based body-composition measurement with 3D imaging plus machine learning, which is practically relevant because DXA is costly and time-consuming. The empirical coverage is broad: seven model families, three targets, three population partitions, and seven training proportions for the semi-supervised method. The paper also makes the graph-smoothness hypothesis explicit and reports errors separately for male, female, and combined data, which is helpful. However, the central data-constrained claim is not yet supported: the p-Laplacian results at 5-50% training are never compared with any supervised model at the same label budget, and the selection of both biomarkers and hyperparameters appears to use test-fold information. The paper does not ship code or an external validation set, so reproducibility rests on the textual description and the GraphLearning package. The supervised benchmark results themselves are plausible and internally consistent.","major_comments":[{"comment":"The headline claim that the p-Laplacian model is 'promising in a data-constrained environment' is not tested against any supervised method at the same label rate. All supervised models are fixed at an 80-20 split (Section 6, Tables 13-15), while the p-Laplacian is evaluated at 80%, 50%, 33%, 25%, 20%, 10%, and 5% training (Tables 7-12). The p-Laplacian's 10%-training RMSEs (e.g., 13.21% for ALM and 9.65% for BMD on the combined data) are therefore never compared with, say, a linear SVR or ridge model trained on the same roughly 52 labeled subjects. Because the ALM biomarkers have correlations above 0.9 (Table 4), a simple linear model might match or beat those numbers at the same label budget. Please add supervised baselines at the same training percentages, using identical fold splits and hyperparameter-selection protocols, and report the mean and standard deviation over the 10 runs.","section":"Section 6; Section 8.5; Tables 7-15"},{"comment":"The top-10 biomarker subsets for p-Laplacian-2 are selected using correlations computed on the full dataset (Section 4.3), before the cross-validation splits are made. This means test-fold information enters the graph construction for the results in Tables 10-12, making the p-Laplacian-2 RMSEs optimistically biased. The biomarker selection must be performed inside each training fold, or the authors should demonstrate that the results are insensitive to this choice by recomputing Tables 10-12 with training-fold-only selection.","section":"Section 4.3; Section 8.5; Tables 4-6, 10-12"},{"comment":"The hyperparameter protocol for the p-Laplacian results is unclear, and the presentation suggests test-set selection. The text states that the models were run 'for different combinations of p and k for a fixed Training %' and Figure 6 plots the 'lowest RMSE' for each training percentage; Tables 7-12 report the 'optimal' p and k with the corresponding RMSE. If p and k are chosen by minimizing RMSE on the test folds, all reported semi-supervised errors are optimistic. Please specify a nested validation scheme (e.g., an inner hold-out or a separate validation split) and report the RMSE at fixed, pre-specified parameters as well as the selected-parameter RMSE.","section":"Section 8.5; Tables 7-12"},{"comment":"It is not stated whether the StandardScaler/z-score normalization is fitted on the training fold only or on the entire dataset before splitting. If the scaler is fitted on all data, label information from the test patients is used in preprocessing for every model, including the p-Laplacian graph. Please state explicitly that all preprocessing statistics are computed from training folds only, and if that is not currently the case, rerun the experiments with a proper per-fold scaling pipeline.","section":"Section 5; Section 7.1"}],"minor_comments":[{"comment":"The description of the modified K-fold scheme is ambiguous because K=5 is used both for the standard 80/20 split and for the modified one-fold-training scheme, which would give 20% training rather than 80%; please clarify which protocol produced each column in Tables 7-12.","section":"Section 8.5"},{"comment":"Immediately after Eq. (2), the text calls eta a 'nondecreasing function' after defining it as a nonnegative decreasing function and choosing eta(t)=exp(-t^2); correct this contradiction.","section":"Section 7.1"},{"comment":"The cross-references to the figures are inconsistent; for example, 'Figure 12d' is cited, but the figure is a nine-subplot grid with no 'd' label described. Please renumber or describe the subplots precisely.","section":"Section 8.1; Figures 11-13"},{"comment":"The 'Regression' rows in Tables 13 and 15 list parameter names without specifying which regression variant (traditional, Ridge, or Bayesian) was used for each dataset; please complete these rows.","section":"Tables 13 and 15"},{"comment":"The caption reads 'Optimal Parameters and corresponding parameters'; it should read 'Optimal Parameters and corresponding RMSEs'.","section":"Table 15 caption"},{"comment":"Please use 'DXA' consistently instead of alternating with 'DEXA'.","section":"Abstract; Section 9"},{"comment":"The MLP description is incomplete; please report the number of hidden layers and units, the learning rate, and any regularization so that the neural-network results are reproducible.","section":"Section 8.4"},{"comment":"The claim that 'This paper is the first to demonstrate the application of a p-Laplacian model for regression' is too strong, since reference [10] already develops p-Laplacian regression; rephrase as the first application to body-composition prediction from 3D optical biomarkers.","section":"Abstract"},{"comment":"The statement that DXA 'could be potentially replaced in the future to estimate ALM and BMD' goes beyond the evidence, because the reported RMSEs are not compared with DXA measurement error or with clinical decision thresholds; please soften the conclusion and add these caveats.","section":"Section 9"},{"comment":"There are typos in the text, including 'kernrel' in 'linear kernrel' (Section 8.2) and 'p-Lapacian' in the Section 8.5 heading; these should be corrected.","section":"Sections 8.2 and 8.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an empirical benchmark rather than a methodological contribution; the theoretical novelty is limited to applying an existing model from the authors' own prior work (refs. 9, 10, 12). The referee report focuses on the missing equal-label-rate supervised baseline because that is the only evidence for the data-constrained claim. If the authors can add that comparison and fix the leakage issues, the manuscript would be suitable as an applied machine-learning / health-informatics paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short take on arXiv:2506.14815. The supervised benchmark is credible; the semi-supervised headline is not yet supported. The paper applies p-Laplacian regression, from Calder and Drenska, to predict ALM, BFP, and BMD from 3D optical biomarkers, and it's the first time that model is used for regression in this domain. That's a real application-level novelty, and the authors are honest that the method comes from prior work and is implemented in the public GraphLearning package.\n\nWhat works: the supervised results are plausible and consistent with prior body-composition studies—SVR and LSSVR land around 8% RMSE for ALM and BMD, 11% for BFP at an 80/20 split. The per-sex analysis and correlation tables are useful. The paper is also candid about BFP's weak biomarker correlations and about the need for larger datasets before clinical conclusions.\n\nThe main soft spot, and it's load-bearing: the p-Laplacian numbers at 10% training are never compared against a supervised model trained on the same 10%. Since the paper's motivation is that supervised methods are data-hungry, you need that same-label-rate comparison. With ALM biomarkers correlating above 0.9, a ridge or linear SVR on the same ~52 labeled subjects might match the p-Laplacian's ~13%—or beat it. The stress-test note is right about this; it's not a minor missing ablation, it's the comparison that would justify the central claim.\n\nTwo smaller problems: the top-10 biomarker selection for p-Laplacian-2 uses the full dataset, which leaks test information; and Tables 7-12 report the minimum RMSE over p and k for each training percentage, suggesting hyperparameters were selected on the test folds. Both need fixing with inner validation and within-fold feature selection.\n\nThe \"DEXA could be potentially replaced\" sentence in Section 9 is overreach at 8% RMSE; the Discussion's own caveat about larger datasets is the right level of caution.\n\nBottom line: this deserves peer review, but it needs a revision that adds low-label supervised baselines and cleans up the validation protocol. A serious referee could turn this into a solid benchmark paper. If I were in this area I'd want to see the fixed version before citing it.","headline":"A credible supervised benchmark for 3D-optical body composition, but the semi-supervised p-Laplacian claim is untestable as reported—no same-label-rate supervised baseline.","tokens_in":20257,"tokens_out":2893,"would_cite":false,"duration_ms":26089,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","68R10","35J92"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that 3D optical body scans plus machine learning can predict appendicular lean mass and bone mineral density with about 8% RMSE, and that a semi-supervised p-Laplacian model holds BMD error near 10% using only 10% of…","keywords":["semi-supervised learning","p-Laplacian regression","3D optical imaging","appendicular lean mass","body fat percentage","bone mineral density","support vector regression","graph-based label propagation"],"falsifier":"Permute the training labels among patients on the same k-NN graph and re-run the 10% training p-Laplacian; if RMSE stays near 10% for BMD, the graph similarity is not carrying the prediction. A second check is to hold out an entire scanner type or age group and see whether the 10% training errors grow beyond the reported values.","tokens_in":19187,"feed_emoji":"🩻","tokens_out":6024,"duration_ms":61572,"temperature":0.7,"pith_summary":"This paper claims that three body-composition measurements normally obtained from DXA scans, namely appendicular lean mass (ALM), body fat percentage (BFP), and bone mineral density (BMD), can be estimated from 44 biomarkers extracted from 3D optical body images. It reports that supervised support vector regression predicts ALM and BMD with about 8% root-mean-square error, and least-squares SVR predicts BFP with about 11% error, using 80% of the data for training. The central advance is the first regression application of a semi-supervised p-Laplacian model, which reaches about 13% error for ALM, 10% for BMD, and 20% for BFP when only 10% of patients are labeled. If these numbers hold, clinicians and researchers could reduce dependence on costly, time-consuming DXA scans, especially in data-constrained settings.","feed_headline":"3D scans plus ML hit ~8% error on lean mass and bone density","feed_subtitle":"A graph-based semi-supervised model keeps errors near 10% even when only 10% of patients are labeled.","key_machinery":"The central object is the game-theoretic p-Laplacian regression model, solved by the tug-of-war dynamic programming principle. Patients are vertices of a k-nearest-neighbor graph, edge weights decay with the Euclidean distance between z-score-normalized biomarker vectors, and the predicted value at each unlabeled vertex solves $L_p u = 0$ subject to labeled values on the training vertices; the parameter $p$ interpolates between a weighted random-walk average ($p=2$) and a min-max tug-of-war as $p \\to \\infty$. This graph-based label propagation carries the argument: when biomarker similarity implies target similarity, the labeled patients' values flow through the graph to unlabeled patients, which is what allows the model to work with a 10% training set.","core_discovery":"The central claim is that p-Laplacian regression on a patient similarity graph can propagate body-composition labels from a small labeled set to the rest of the population, and that this propagation is accurate enough to be clinically useful. In the paper's own terms, the authors show that a graph whose nodes are patients and whose edges encode similarity of 3D-derived biomarkers supports the hypothesis that similar patients have similar ALM, BFP, and BMD values. With 10% of the data labeled, the model predicts ALM within about 13% RMSE, BMD within about 10% RMSE, and BFP within about 20% RMSE; with 80% labeled, supervised SVR and LSSVR do better on all three targets. The paper concludes that DXA could potentially be replaced when estimating ALM and BMD, pending further validation on larger datasets.","pith_inferences":["A straightforward external test would be to run the same pipeline on a cohort scanned with a different 3D device, since the paper used three internal scanner types, to see whether the similarity graph transfers across hardware.","Because the graph weights all 44 biomarkers equally, weighting edges by target-specific correlations, which the paper itself floats as future work, would likely lift BFP accuracy given its top-ten correlations of only 0.4 to 0.6.","The relative RMSE metric means that a 10% error on BMD is a small absolute error on a variable whose mean is near 1.03 g/cm²; clinicians would need absolute comparisons against DXA repeatability before judging replacement.","If the smoothness hypothesis is the true driver, deliberately violating it, for example requiring the model to extrapolate to an age or body-mass range absent from the labeled set, should produce a measurable accuracy cliff."],"forward_implications":["If the results replicate on larger cohorts, 3D optical imaging plus supervised SVR/LSSVR could replace DXA for routine ALM and BMD estimation, cutting cost and scan time.","The p-Laplacian model's roughly 10% BMD error at 10% labeled data suggests that body-composition prediction can proceed with only a small curated labeled set, easing data-privacy and annotation bottlenecks.","BFP remains the hardest target across every model, indicating that the current 44-biomarker set carries weaker signal for adipose distribution than for muscle and bone.","The reported convergence of p-Laplacian RMSE as $p$ grows justifies treating large-$p$ behavior as a stable regime for this regression method.","Training sex-stratified models generally improves accuracy for ALM and BFP, so population stratification is a built-in component of the method."],"supporting_citations":[{"why":"Supplies the p-Laplacian regression model and the tug-of-war dynamic programming equation used as the solver.","marker":"[10]"},{"why":"Provides the computational implementation of the p-Laplacian regression that the experiments run.","marker":"[9]"},{"why":"Supplies the k-NN graph construction method that defines patient similarity.","marker":"[11]"},{"why":"Proves that low label rates make the $p=2$ case degenerate, used to explain why optimal $p$ rises at low training percentages.","marker":"[12]"},{"why":"Describes the 3D optical imaging system and biomarker extraction procedure that produced the dataset.","marker":"[25]"},{"why":"Provides the closest supervised baseline, a neural network predicting ALM from 3D digital anthropometry.","marker":"[21]"},{"why":"Provides the prior machine-learning prediction of BFP that shows gender effects and LSSVR strength.","marker":"[2]"},{"why":"Supplies prior semi-supervised BMD estimation, the context for the claim that SSL is new for these targets.","marker":"[33]"}],"fun_headline_variants":["3D scans + ML predict lean mass and bone density","AI from 3D scans estimates body composition accurately","Semi-supervised ML on 3D scans predicts body metrics","3D imaging ML: body composition from few labels","Data-efficient ML predicts body composition from 3D scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that patients with similar 3D body measurements have nearly the same ALM, BFP, and BMD values, so that labels can flow across the similarity graph; if that smoothness fails, the 10% training results have no basis.","fun_headline_variants_meta":{"raw":{"variants":["3D scans + ML predict lean mass and bone density","AI from 3D scans estimates body composition accurately","Semi-supervised ML on 3D scans predicts body metrics","3D imaging ML: body composition from few labels","Data-efficient ML predicts body composition from 3D scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001051,"raw_usage":{"total_tokens":4470,"prompt_tokens":1056,"completion_tokens":3414,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":3333}},"tokens_in":672,"tokens_out":3414,"duration_ms":25220,"temperature":1.0,"reasoning_tokens":3333,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:44:54.918021+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Permute the training labels among patients on the same k-NN graph and re-run the 10% training p-Laplacian; if RMSE stays near 10% for BMD, the graph similarity is not carrying the prediction. A second check is to hold out an entire scanner type or age group and see whether the 10% training errors grow beyond the reported values.","supporting_citations":[{"cited_title":"Consis- tency of semi-supervised learning, stochas- tic tug-of-war games, and the p-laplacian","cited_arxiv_id":null,"evidence_quote":"Supplies the p-Laplacian regression model and the tug-of-war dynamic programming equation used as the solver."},{"cited_title":"Improved spectral convergence rates for graph laplacians on ϵ-graphs and k-nn graphs","cited_arxiv_id":null,"evidence_quote":"Supplies the k-NN graph construction method that defines patient similarity."},{"cited_title":"Graphlearning python package, January 2022","cited_arxiv_id":null,"evidence_quote":"Proves that low label rates make the $p=2$ case degenerate, used to explain why optimal $p$ rises at low training percentages."},{"cited_title":"Dechenaud, Jerome A","cited_arxiv_id":null,"evidence_quote":"Describes the 3D optical imaging system and biomarker extraction procedure that produced the dataset."},{"cited_title":"Advances in digital anthropometric body composition assess- ment: neural network algorithm prediction of appendicular lean mass","cited_arxiv_id":null,"evidence_quote":"Provides the closest supervised baseline, a neural network predicting ALM from 3D digital anthropometry."},{"cited_title":"Recent develop- ments in data science: Comparing linear, ridge and lasso regressions techniques using wine data","cited_arxiv_id":null,"evidence_quote":"Supplies prior semi-supervised BMD estimation, the context for the claim that SSL is new for these targets."}],"review_version":1}