{"id":"d5dafa05-f288-4187-92a5-2f9edf95c95d","arxiv_id":"2608.02408","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A CNN-BiLSTM model estimates vertical ground reaction forces from IMU sensors in Parkinsonian gait with reported inter-subject R2 around 0.91, and a two-IMU setup is proposed as a practical minimum.","lead":"This study tests a deep learning model that estimates foot forces during walking in Parkinson's disease from small body-worn motion sensors. It finds that a two-sensor setup can work, but the sensor choice was optimized on the test data, so the headline accuracy may be optimistic.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set selection of sensor configurations likely inflates the optimized two-/four-IMU R²; no nested hold-out validates them.","rationale":"The paper's central contribution beyond feasibility is an optimized minimal IMU set. The reader's critique targets the exact step where that contribution is established: all 1,023 configurations are scored on the same inter-subject test set, and the reported accuracies are the maxima of those scores. This is a textbook selection-bias problem. It is not an attack on the authors' integrity; it is a methodological gap that is directly testable. I considered whether the pressure-derived labels are a more fundamental concern, but the cross-dataset force-plate validation and the fact that the full-13-IMU claim is independent of sensor selection make the selection-on-test issue the more load-bearing one. The inconsistent four-IMU listing supports the reader's concern that the ablation results are not fully reliable. The reader's CONDITIONAL verdict is appropriate: the feasibility result stands, but the optimized-configuration accuracy claims require nested validation before they can be accepted as unbiased. Verdict should remain CONDITIONAL, so no adjustment is needed.","tokens_in":17402,"tokens_out":4419,"duration_ms":40883,"concrete_test":"Implement a nested evaluation: split PD participants into an outer test set (20%, subjects unseen during any ablation). On the remaining 80%, perform the 1,023-combination search using an inner validation split (or repeated CV) to choose the best 2-IMU and 4-IMU configurations. Then retrain the chosen models on the 80% training/validation data and evaluate once on the outer test set. Repeat with several random splits. If the selected FH+LD R² falls below 0.91 (or its CI overlaps the full 13-IMU result), the reported minimal-configuration accuracy is inflated. The same protocol should also resolve which four-IMU set (RA vs RT) is actually optimal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing element of the optimized-minimal-configuration claim is Section III.E, where the best 1- to 10-IMU configurations are chosen as the combination with the highest R² among 1,023 retrained models, all evaluated on the same inter-subject PD test set. The reported FH+LD result (R² = 0.91 ± 0.09) and the best four-IMU result (R² = 0.93) are therefore the maximum of a large set of correlated R² values, not an unbiased estimate of expected performance on new subjects. No nested validation is described: no outer hold-out is reserved to freeze the selected sensor set before final evaluation. The R² inflation may be modest because many configurations are correlated, but 1,023-way selection on the test set cannot be dismissed without a hold-out estimate. The internal inconsistency in the best four-IMU configuration (Section III.E: LD, RA, RS, FH; Section IV.B: LD, RS, RT, FH, R² = 0.93) further undermines confidence that the ablation results are being reported from a single protocol. The full-13-IMU feasibility claim and baseline comparisons do not depend on the selection step, but the central 'two IMUs are comparable' claim does.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports a deep-learning pipeline for estimating bilateral vertical ground reaction forces (vGRFs) in Parkinsonian gait from wearable IMU signals. Using a public multi-site dataset (61 PD patients and 65 age-matched healthy controls), the authors train a CNN-BiLSTM model on 13 IMUs and report intra-subject R2 = 0.98 for both cohorts and inter-subject R2 = 0.93 for HC and 0.91 for PD. They compare the proposed model against MLP, TCN, Transformer, CNN-only, and BiLSTM-only baselines under the full 13-IMU inter-subject setting, and they perform a cross-dataset external evaluation using OpenSim-derived virtual IMU signals from a separate PD dataset. The central novel contribution is a sensor-configuration ablation in the PD cohort: after screening the 13 single-IMU placements, the authors retrain the model for all 1,023 combinations of the remaining 10 IMUs and identify the best one- through ten-IMU configurations by test-set R2. They conclude that a two-IMU configuration (left dorsum and forehead) attains R2 = 0.91, comparable to the full set, and that a four-IMU configuration (reported with different sensor lists in different sections) attains R2 = 0.93.","tokens_in":17639,"tokens_out":7694,"duration_ms":68155,"significance":"If the inter-subject results survive scrutiny, the paper would be a valuable contribution: it is among the first to show that IMU-to-vGRF regression can generalize across unseen PD patients with cohort-level R2 above 0.9, and the systematic sweep of sensor configurations is a useful design resource. The study has clear strengths: subject-wise and trial-wise splitting before windowing, z-score standardization computed on training data only, a public data source, baseline comparisons under matched settings, and an explicit acknowledgement of the virtual-IMU limitation in the external validation. However, the headline minimal-sensor claim is currently an optimization on the test set rather than a predictive evaluation, and the label signal is pressure-derived rather than force-plate measured. These issues bear directly on the magnitude of the claimed two-IMU accuracy and on its comparison with prior force-plate-based studies.","major_comments":[{"comment":"The optimized sensor configurations are selected by taking, for each sensor count, the combination with the highest R2 among 1,023 retrained models, all evaluated on the same inter-subject PD test set used to report the final accuracy. The reported two-IMU (LD/FH) R2 = 0.91 ± 0.09 and four-IMU R2 = 0.93 are therefore maxima over a large set of correlated R2 statistics, not unbiased predictions for new subjects, and the 13-IMU baseline is the only result with an honest train/test split. No outer hold-out or nested validation is described to freeze the sensor configuration before final evaluation. Because the abstract's central claim is that a two-IMU setup provides estimation comparable to the full set, this selection bias must be addressed, for example by reserving an outer test set for the chosen configuration or by reporting selection-corrected estimates and reframing the sweep as exploratory.","section":"III.E; II.F"},{"comment":"The best four-IMU configuration is reported inconsistently: Section III.E states it is LD, RA, RS, and FH, while Section IV.B states it is LD, RS, RT, FH, both with R2 = 0.93. The Fig. 4/Fig. 5 summary should be checked against these lists. This discrepancy means the reader cannot tell which sensor set produced the headline four-IMU accuracy and suggests the optimal-configuration results are not being reported from a single protocol. Please reconcile the text, figure, and table, and ensure the accuracy value corresponds to the listed configuration.","section":"III.E vs. IV.B"},{"comment":"The initial screening step excludes XP, LW, and RW because their single-IMU R2 values fall below 0.8 in the PD inter-subject test set, and the 0.80 threshold is applied after inspecting the test-set results in Fig. 2. This is a further test-set-based selection, and it restricts the subsequent 1,023-combination search to the 10 sensors that already performed best on the test data. The screening decision should be made on training/validation data only, or the entire sweep should be treated as exploratory and validated on an untouched cohort.","section":"II.F; III.C"},{"comment":"The vGRF labels in the primary dataset are not force-plate measurements but estimates derived from pressure-walkway contact-area and pressure data via a static-standing calibration. The reported coefficient of variation of less than 0.03 is a repeatability measure and does not establish agreement with a gold-standard force measurement. Since every R2 and MAE in the paper is computed against these derived labels, the claims should be qualified as estimation of pressure-derived vGRFs, or a subset should be validated against simultaneous force-plate data, before comparing absolute error levels with force-plate-based studies.","section":"II.B"}],"minor_comments":[{"comment":"The text contains the typo \"one way ANOV A\" and should read \"one-way ANOVA\"; the same typo appears in the following sentence.","section":"II.F"},{"comment":"The caption uses \"Left Mid-Lateral Thigh (LM)\" and \"Right Mid-Lateral Thigh (RM)\", which are inconsistent with Table II's \"Left lateral thigh (LT)\" and \"Right lateral thigh (RT)\"; please use one abbreviation set throughout.","section":"Fig. 2 caption"},{"comment":"The caption phrase \"alone with the worst-performing\" should read \"along with the worst-performing\", and the conclusion's \"This study firstly proposes\" should be rephrased to standard English.","section":"Fig. 5 caption; Conclusion"},{"comment":"The claim that \"no study has investigated\" deep-learning-based GRF estimation in PD should be softened to \"to our knowledge\", and the closest prior works cited later (e.g., Eltoukhy et al. and Liu et al.) should be explicitly distinguished in the introduction.","section":"Abstract; I"},{"comment":"The statement that a non-significant difference in R2 between the primary and external datasets (p = 0.51) supports cross-dataset generalizability should be backed by an equivalence test or confidence intervals, since a non-significant test is not evidence of equivalence, particularly because the external test uses virtual IMU signals (a caveat acknowledged in Section IV.C).","section":"III.B; Table VI"}],"recommendation":"major_revision","confidential_remarks":"The optimized-configuration claim is the paper's main selling point, and the current test-set selection makes the central two-IMU accuracy an upper bound rather than an unbiased estimate. The inconsistency between Section III.E and Section IV.B regarding the four-IMU configuration must be resolved before any revision. The paper otherwise fits the journal's scope, and the public-data-based comparison is a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid feasibility study with an honest core and an inflated edge. The new thing is the application — first deep-learning estimate of bilateral vertical GRF in PD from wearable IMUs, on a decent public dataset (61 PD, 65 HC, 13 IMUs), with inter-subject evaluation and baseline comparisons. The 13-IMU inter-subject result (R2=0.91 PD, 0.93 HC) comes from a proper subject-level split and is worth taking seriously. The baseline comparisons are competent, the exhaustive 1,023-combination sensor search is a real contribution, and the citation pattern is fair — they explicitly note CNN-BiLSTM is not a new architecture.\n\nThe soft spot is exactly where the reader puts it: Section III.E selects the best one- to ten-IMU configurations as the maximum R2 over 1,023 retrained models, all evaluated on the same inter-subject PD test set. No outer hold-out is reserved to freeze the sensor set before final evaluation. So the headline two-IMU (LD+FH, R2=0.91) and four-IMU (R2=0.93) numbers are maxima of many correlated test-set estimates, not unbiased predictions for new subjects. The inflation might be modest, but the paper cannot claim 'comparable accuracy' from those numbers alone. There is also an internal inconsistency: the best four-IMU configuration is LD, RA, RS, FH in Section III.E and Fig. 4, but LD, RS, RT, FH in Section IV.B. That needs fixing. The external validation is weaker than it first looks: virtual IMUs generated from motion capture, and vGRF labels from a pressure walkway in the primary dataset rather than force plates. The limitations section admits the virtual-IMU issue, which is good, but it does not address the selection-on-test-set problem. No code is released, so verification is limited.\n\nNone of this breaks the main feasibility claim. The 13-IMU inter-subject result stands, and the sensor-ranking analysis is a useful start. What does not stand is the specific claim that two IMUs are 'comparable' — that conclusion is load-bearing for the practical message and it depends on the circular selection step.\n\nWho gets value: clinical biomechanics and wearable-sensing researchers, especially anyone building IMU-based GRF estimators for pathological gait. It deserves a serious referee: send it out, but with a request for nested validation (or at least an honest reframing of the optimized-set results as exploratory), resolution of the four-IMU inconsistency, and preferably code/data release.","headline":"Useful feasibility study of CNN-BiLSTM for PD vGRF from IMUs, with an honest 13-IMU core but an optimized-sensor claim that is selected on the test set and needs external validation.","tokens_in":18186,"tokens_out":3318,"would_cite":true,"duration_ms":28588,"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":"The paper claims that bilateral vertical ground reaction forces in Parkinsonian gait can be estimated from as few as two wearable IMUs with accuracy comparable to a 13-sensor setup, using a CNN-BiLSTM deep-learning model.","keywords":["ground reaction force estimation","Parkinson's disease","wearable IMU","CNN-BiLSTM","sensor placement optimization","vertical ground reaction force","gait analysis","deep learning"],"falsifier":"Run the sensor-selection procedure again with a strict nested design: use only training and validation folds to pick the best two- and four-IMU configurations, freeze those configurations, and evaluate them on a test cohort of Parkinson's patients that was never touched during either model training or sensor selection. If the frozen two-IMU model's $R^2$ falls below roughly 0.85 or is materially worse than the 13-IMU model's $R^2$, the paper's central 'two IMUs are comparable' claim is contradicted. A direct field test of the LD + FH configuration on new patients with real wearable IMUs, without retraining, would also settle it.","tokens_in":17183,"feed_emoji":"🦶","tokens_out":9048,"duration_ms":75701,"temperature":0.7,"pith_summary":"This paper claims that vertical ground reaction forces during walking in Parkinson's disease can be estimated accurately from wearable inertial sensors using a deep-learning model, without force plates. A hybrid CNN-BiLSTM network, trained separately on 61 people with Parkinson's disease and 65 healthy controls, reaches $R^2 = 0.98$ when the same person's data appear in training, and $R^2 = 0.91$ (PD) and $R^2 = 0.93$ (HC) on people never seen during training. The paper also argues that sensor placement matters more in Parkinsonian gait: the full 13-sensor configuration is not needed, four carefully chosen IMUs give the top accuracy, and a two-IMU setup on the left foot dorsum and forehead keeps accuracy at $R^2 = 0.91$ on unseen PD patients. If true, this makes continuous, home-based gait monitoring for Parkinson's disease practical without laboratory equipment.","feed_headline":"Two IMUs match 13-sensor accuracy for Parkinsonian gait forces","feed_subtitle":"CNN-BiLSTM keeps R² = 0.91 on unseen Parkinson's patients with just two body-worn sensors, opening home monitoring.","key_machinery":"The central object is the CNN-BiLSTM sequence-to-sequence regressor: one-dimensional convolutional layers extract local spatial features from the IMU channels, bidirectional LSTM layers model temporal dependencies across each 64-step sliding window, and fully connected layers output left and right vertical ground reaction forces normalized to body weight. The optimized-sensor analysis is carried by an exhaustive ablation over 1,023 IMU combinations (all subsets of 1 to 10 of the 13 placements after three low-accuracy single sensors are excluded), with every configuration retrained under the same inter-subject protocol. The two-IMU configuration at the left foot dorsum and forehead is the mechanism that turns the accuracy result into a practical wearable-monitoring claim.","core_discovery":"On the paper's own terms, the discovery is that bilateral vertical ground reaction forces in Parkinsonian gait are learnable from body-worn inertial data under a realistic inter-subject protocol. With 13 IMUs the CNN-BiLSTM model reaches $R^2 = 0.98$ under within-subject training and, on people never seen in training, $R^2 = 0.91$ for PD patients and $R^2 = 0.93$ for healthy controls. The paper further claims that optimal sensor placement is disease-specific: for PD, a four-IMU configuration gives the best accuracy ($R^2 = 0.93$), and a two-IMU configuration at the left foot dorsum and forehead retains $R^2 = 0.91$, comparable to the full 13-sensor model. In a matched 12-IMU comparison, the model transfers to an independent PD dataset with $R^2 = 0.86$ versus $R^2 = 0.88$ on the primary data, a difference reported as not statistically significant.","pith_inferences":["Because the same inter-subject test set was used both to select the best configurations and to report their $R^2$, the headline two-IMU and four-IMU accuracies are probably somewhat optimistic; a nested or fully separated selection hold-out would give the unbiased numbers.","The left-dorsum sensor's dominance in Parkinson's disease may reflect the cohort's left-dominant symptom laterality, which suggests that a mirrored LD + FH setup could be more appropriate for patients with right-dominant symptoms.","The same exhaustive ablation protocol could be transferred to other asymmetric or pathological gaits, such as stroke or osteoarthritis, to test whether disease-specific optimal sensor sets are a general rule rather than a Parkinson's peculiarity.","Real-world deployment would need to confirm that the two-IMU model tolerates sensor repositioning, day-to-day variability, and magnetic or mechanical disturbances, since the external validation used motion-capture-derived virtual IMUs rather than physical wearable recordings."],"forward_implications":["A two-IMU wearable (left foot dorsum plus forehead) can estimate bilateral vertical ground reaction forces on new Parkinson's patients at $R^2 = 0.91$, which is enough for gait monitoring without force plates.","Because the optimal sensor set differs between Parkinson's disease and healthy controls, wearable gait systems should be validated with disease-specific sensor placements rather than a universal configuration.","The cross-dataset transfer result suggests that a model pretrained on one Parkinson's cohort can be adapted to another cohort through transfer learning, retaining most of its accuracy.","Adding sensors beyond four yields only marginal accuracy gains, so a small, strategically placed set is sufficient for practical deployment in this population.","The intra-subject $R^2 = 0.98$ result implies that a short personalized calibration recording could support high-fidelity individual gait monitoring in a clinic."],"supporting_citations":[{"why":"Supplies the primary multi-site dataset of 61 PD patients and 65 healthy controls with 13 IMUs and pressure-walkway vGRF measurements.","marker":"[26]"},{"why":"Supplies the independent PD dataset used for cross-dataset external validation, from which virtual IMU signals were generated.","marker":"[27]"},{"why":"Provides an earlier deep-learning IMU-based GRF estimation method and the window-length guidance adopted for segmentation.","marker":"[14]"},{"why":"Shows two tibia-mounted IMUs plus machine learning estimate GRF in healthy adults, serving as the healthy-population baseline that the PD result extends.","marker":"[15]"},{"why":"Reports the only prior PD-inclusive vGRF estimation study using Newton-Euler equations; its two-shank MAE is the direct comparison point.","marker":"[11]"},{"why":"Provides the $R^2 > 0.80$ threshold used to screen single-IMU placements before the exhaustive sensor-combination ablation.","marker":"[37]"},{"why":"Supports the CNN-BiLSTM architecture for spatial-temporal IMU-to-lower-limb-joint prediction.","marker":"[24]"},{"why":"Supports the CNN-BiLSTM architecture for vertical ground reaction force prediction from wearable IMUs during running.","marker":"[25]"}],"fun_headline_variants":["Two IMUs replicate 13-sensor gait force estimates in PD","Parkinson's gait forces from just two wearables, R²=0.91","Deep learning trims gait force sensors from 13 to 2 for PD","Minimal IMU set predicts Parkinson's ground reaction forces","CNN-BiLSTM needs only two IMUs for Parkinsonian gait forces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical conclusion that two IMUs are enough rests on accuracy numbers ($R^2 = 0.91$) that come from the same test set used to pick the best sensor combination out of 1,023 possibilities, with no separate unseen data to confirm that specific choice.","fun_headline_variants_meta":{"raw":{"variants":["Two IMUs replicate 13-sensor gait force estimates in PD","Parkinson's gait forces from just two wearables, R²=0.91","Deep learning trims gait force sensors from 13 to 2 for PD","Minimal IMU set predicts Parkinson's ground reaction forces","CNN-BiLSTM needs only two IMUs for Parkinsonian gait forces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000824,"raw_usage":{"total_tokens":3675,"prompt_tokens":1091,"completion_tokens":2584,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":2487}},"tokens_in":707,"tokens_out":2584,"duration_ms":16159,"temperature":1.0,"reasoning_tokens":2487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:01:13.990991+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the sensor-selection procedure again with a strict nested design: use only training and validation folds to pick the best two- and four-IMU configurations, freeze those configurations, and evaluate them on a test cohort of Parkinson's patients that was never touched during either model training or sensor selection. If the frozen two-IMU model's $R^2$ falls below roughly 0.85 or is materially worse than the 13-IMU model's $R^2$, the paper's central 'two IMUs are comparable' claim is contradicted. A direct field test of the LD + FH configuration on new patients with real wearable IMUs, without retraining, would also settle it.","supporting_citations":[{"cited_title":"Weargait-pd: An open-access wearables dataset for gait in parkinson’s disease and age-matched controls,","cited_arxiv_id":null,"evidence_quote":"Supplies the primary multi-site dataset of 61 PD patients and 65 healthy controls with 13 IMUs and pressure-walkway vGRF measurements."},{"cited_title":"A public data set of walking full-body kinematics and kinetics in individuals with parkinson’s disease,","cited_arxiv_id":null,"evidence_quote":"Supplies the independent PD dataset used for cross-dataset external validation, from which virtual IMU signals were generated."},{"cited_title":"Predicting vertical ground reaction forces from 3d accelerometry using reservoir computers leads to accurate gait event detection,","cited_arxiv_id":null,"evidence_quote":"Shows two tibia-mounted IMUs plus machine learning estimate GRF in healthy adults, serving as the healthy-population baseline that the PD result extends."},{"cited_title":"An imu- based ground reaction force estimation method and its application in walking balance assessment,","cited_arxiv_id":null,"evidence_quote":"Reports the only prior PD-inclusive vGRF estimation study using Newton-Euler equations; its two-shank MAE is the direct comparison point."},{"cited_title":"Combining inertial sensors and machine learning to predict vgrf and knee biomechanics during a double limb jump landing task,","cited_arxiv_id":null,"evidence_quote":"Provides the $R^2 > 0.80$ threshold used to screen single-IMU placements before the exhaustive sensor-combination ablation."},{"cited_title":"Deep- learning model for the prediction of lower-limb joint moments using single inertial measurement unit during different locomotive activities,","cited_arxiv_id":null,"evidence_quote":"Supports the CNN-BiLSTM architecture for spatial-temporal IMU-to-lower-limb-joint prediction."},{"cited_title":"Prediction of vertical ground reaction forces under different running speeds: Integration of wearable imu with cnn-xlstm,","cited_arxiv_id":null,"evidence_quote":"Supports the CNN-BiLSTM architecture for vertical ground reaction force prediction from wearable IMUs during running."}],"review_version":1}