Pith. sign in

REVIEW 5 major objections 6 minor 45 references

A transformer network, trained with a constant segment-length loss, predicts the next 76 frames of whole-body posture during load-reaching with a 41.4 mm RMSE—about 58% lower than a BLSTM.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 20:11 UTC pith:UPVIUHLB

load-bearing objection The paper shows a real architecture effect and a useful constraint, but the headline accuracy number is not the one to trust; it needs revision before the quantitative claims can be believed. the 5 major comments →

arxiv 2511.20615 v2 pith:UPVIUHLB submitted 2025-11-25 cs.CV cs.AI

Evaluating the Performance of Deep Learning Models in Whole-body Dynamic 3D Posture Prediction During Load-reaching Activities

Seyede Niloofar Hosseini , Ali Mojibi , Mahdi Mohseni , Navid Arjmand , Alireza Taheri This is my paper
classification cs.CV cs.AI
keywords deep learningtransformerBLSTMposture predictiontime series forecastingmanual material handlingkinematic constraintsbiomechanics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to establish that time-series deep networks can forecast the remaining three-quarters of a load-reaching movement from the first quarter and a few easy-to-measure task descriptors, using motion-capture data of 20 men performing 204 tasks each. It finds that a transformer architecture predicts the 76 future frames with a whole-body RMSE of 41.4 mm versus 100.0 mm for a bidirectional LSTM—roughly 58% better—even though the BLSTM is more accurate one step ahead. The paper also introduces a loss term that enforces constant forearm, upper-arm, and shank lengths, cutting arm error by about 9% and leg error by about 27% while making predicted skeletons more anatomically consistent. The practical payoff, if correct, is that ergonomic spine-load assessments could start from a short motion sample instead of full motion capture.

Core claim

The central claim, stated on the authors' terms, is that a transformer with an encoder-decoder structure and multi-head self-attention, trained per body segment, can recursively generate the full 101-frame trajectory of a load-reaching task once the first 25 frames and seven task/subject features (load position, lifting technique, handling technique, body height and mass) are given. In the test split the transformer reaches a whole-body long-term RMSE of 41.4 mm, versus 100.0 mm for the BLSTM; with the kinematic-constraint loss, arm RMSE falls from 49.4 to 45.4 mm and leg RMSE from 28.7 to 22.7 mm. Under leave-one-subject-out cross-validation the whole-body error is 54.4 mm, which the author

What carries the argument

The load-bearing machinery is a transformer encoder-decoder for one-step-ahead frame prediction, used recursively at test time to forecast the entire remaining sequence, with separate models for four body segments (head, arms, trunk-pelvis, legs) to keep parameter counts manageable. The novel ingredient is a modified loss function that adds to mean squared marker error a penalty proportional to the squared deviation of each segment's length from its constant per-subject value, applied to forearms, upper arms, and shanks. The attention mechanism is what the paper credits for the transformer's stability over long rollouts; the segment-length term is what keeps the predicted skeleton from drift

Load-bearing premise

The reported accuracy assumes that the first 25 frames of each task are measured (or perfectly predicted); in the proposed deployment the seeds must come from a separate static-frame model whose error is unquantified and could compound through recursive prediction.

What would settle it

Run the full deployment pipeline: generate the first 25 frames with the authors' static-frame model for a held-out subject, then recursively predict the remaining 76 frames and compare to motion capture; if the whole-body RMSE exceeds the 41.4 mm obtained with measured seeds by a substantial margin, the system-level claim does not hold for field use.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • For load-reaching and similar manual-material-handling tasks, the transformer architecture is the better choice when multi-second recursive whole-body forecasts are needed; the BLSTM remains competitive only for single-frame lookahead.
  • Adding the constant segment-length penalty improves marker-position RMSE by roughly 9% for arms and 27% for legs while also aligning the predicted distribution of segment lengths more closely with measured motion capture data.
  • A complete movement can be generated from the first 25% of its frames plus seven easy inputs (load position, technique codes, height, mass), so an ergonomist would not need full motion capture for the predicted portion.
  • Both models operate on standard plug-in-gait marker coordinates, so their outputs can be fed directly into musculoskeletal models that estimate spinal loads and moments for injury-risk assessment.
  • The leave-one-subject-out result of 54.4 mm is comparable to earlier frame-by-frame static models, but the new approach adds temporal consistency and smoother, more plausible limb motion.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The headline 41.4 mm figure is reported for an evaluation that starts from available measured seed frames; the paper's deployment path instead generates those first 25 frames from a separate static-frame model. Since that seed model's own error is not quantified and would propagate through the recursive rollout, end-to-end field accuracy could be materially worse than reported.
  • The segment-length penalty is one instance of a broader principle—enforcing known kinematic invariants during training can suppress drift in recursive predictors. The same idea could be extended to joint-angle limits, bilateral symmetry, or foot-floor contact, and would likely help other time-series pose models beyond this architecture.
  • The transformer's long-term advantage over the BLSTM may be specific to the training procedure: because the recurrent model is trained with teacher forcing (true inputs) but tested on its own predictions, its hidden states accumulate error. Training the BLSTM with scheduled sampling or a long-term loss is a direct test of whether the gap is inherent or an artifact of exposure bias.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper compares BLSTM and transformer networks for recursive whole-body 3D posture prediction during load-reaching tasks, using a dataset of 20 subjects performing 204 tasks each. Inputs are the first 25 frames (25% of the normalized task) plus task/subject metadata; models predict the remaining 76 frames. A kinematic loss term enforcing constant arm/leg segment lengths is proposed. The central claims are that the transformer achieves long-term whole-body RMSE of 41.4 mm, approximately 58% better than the BLSTM (100.0 mm), and that the kinematic constraint reduces arm/leg RMSE by about 8% and 21%. The paper also reports leave-one-subject-out (LOSO) results and an external validation with three subjects.

Significance. If the reported results are robust, the paper would be a useful step toward practical deep-learning-based posture prediction for ergonomic analysis, combining temporal modeling with a simple biomechanical plausibility constraint. The public code release, LOSO cross-validation, and external-subject attempt are strengths. However, the headline quantitative claims are compromised by internal inconsistencies, a favorable single-split evaluation, and missing external results. The directional finding that the transformer generalizes better than BLSTM over long horizons is plausible and supported by Table V, but the magnitude of the advantage and the absolute accuracy numbers are not yet established.

major comments (5)
  1. [Abstract; Table V; Table VI; Section IV.A] The headline accuracy is not a cross-validated estimate and is internally inconsistent. The abstract (and the full-text abstract in different places) cite 41.4 mm or 47.0 mm, while Table V reports 41.414 mm from a single 14/4/2 split with only 2 test subjects. The LOSO cross-validation in Table VI gives 54.39 ± 17.59 mm whole-body RMSE for the same transformer. The Discussion itself acknowledges 54.4 mm when comparing with prior work. With no per-subject or per-fold error bars for the BLSTM comparison, the 'approximately 58% more accurate' claim rests on a favorable split and is not statistically supported. Moreover, the transformer was selected as the final model based on this same test set (Section III.B), so the reported comparison is optimistic. Please present per-subject results and LOSO for both models, with confidence intervals, and state which number is intended as the headline.
  2. [Table V; Section II.F] The architecture comparison is confounded with the kinematic constraint. Table V gives BLSTM results without a kinematic-constraint variant, while the transformer is reported both without and with constraints. The 41.414 mm whole-body number appears to be from the unconstrained transformer; no whole-body RMSE is provided for the constrained version that is the final proposed model. Thus the claimed 58% margin over BLSTM could be partly due to the constraint, not the architecture. Please train BLSTM with the same constraint or report the constrained transformer's whole-body RMSE, and compare under identical conditions.
  3. [Section II.C; Abstract; Section IV.B] There is an evaluation/deployment mismatch regarding the first 25 input frames. The abstract states that inputs are '3D coordinate data of the body posture from the first 25% of the task duration' (implying measured data), but Section II.C says these first 25 frames 'were provided using our previously developed load-reaching whole-body posture prediction model based on static frames [16].' This means the reported long-term RMSEs include errors from that seed model, yet the paper never quantifies the seed model's contribution or compares measured versus generated seeds. If the intended deployment uses [16], the accuracy is end-to-end but the seed error is unknown; if the numbers assume measured seeds, real-world performance is unstated. Clarify and ideally report both settings.
  4. [Section III.E; Table VII] The external validation claim is unsubstantiated. Section III.E states that performance on three external subjects is summarized in Table VII, but Table VII as printed contains the same BLSTM/Transformer values as Table V rather than external-subject results. No external RMSE/nRMSE/R² numbers appear anywhere. Figure 6's KL-divergence analysis mixes 2 local test subjects and 3 external subjects, but no external posture-error metrics are given. Please provide the actual external performance table, or remove the claim that external validation was performed.
  5. [Section IV.A; Table V; Abstract] The claimed improvement from the kinematic constraint is misreported. The abstract says 'approximately 8% and 21%' for arms and legs; from Table V, arms 49.370→45.437 is a 7.96% reduction and legs 28.690→22.669 is a 20.99% reduction. The Discussion states '8.9% improvement' and '27% improvement', neither of which matches Table V. Please correct the percentages and ensure all derivations are consistent.
minor comments (6)
  1. [Abstract/Full text] The two versions of the abstract in the manuscript give different RMSE values: 41.4 mm in one place and 47.0 mm in another. Table V supports 41.414. Fix the discrepancy.
  2. [Section II.C.2; Table II] The transformer description says '16 self-attention heads', but Table II lists the selected number as 32. Align the text and table.
  3. [Table VI] The header reads 'RSME' (typo for RMSE). Also, reporting only mean ± std over subjects is helpful, but the large standard deviations (e.g., whole-body 17.59 mm) should be discussed as evidence of high inter-subject variability rather than 'robustness.'
  4. [Section III.E / Table VII] The caption of Table VII refers to external subjects, but the table content is a duplicate of Table V. This must be corrected.
  5. [Section III.C / Fig. 6] The statistical test used for the KL-divergence p-value (p=0.0226) is not specified. State the test (e.g., paired t-test, Wilcoxon) and justify it.
  6. [Section II.D] The definition of whole-body RMSE is not explicit beyond Eq. (2). Clarify how the 'whole-body' rows in Table V are computed from the four segment models.

Circularity Check

1 steps flagged

A secondary validation of the segment-length constraint is self-referential (KL compares the very quantity the loss enforces), while the main transformer-vs-BLSTM accuracy claim is benchmarked externally against measured motion capture.

specific steps
  1. self definitional [Section II.F, Eq. (3)-(4); Section III.C, Fig. 6]
    "To quantitatively evaluate the effect of this constraint, we computed the Kullback-Leibler (KL) divergence between the empirical distribution of segment lengths from the motion capture data distribution (P) and the predicted distributions from the models with and without the kinematic loss term ... The results showed that the model with the kinematic loss Qkin exhibited significantly lower KL divergence compared to the MSE-only model, indicating a closer match to the real segment length distributions. [Loss Eq. 3:] ℒ = ... + a × Σ (ℓ_m − ℓ̂_m)^2"

    The added loss term directly minimizes (ℓ_m − ℓ̂_m)^2, the squared error between predicted and measured segment lengths. The KL analysis then measures how close the predicted segment-length distribution Q_kin is to the measured distribution P. Because the loss is defined as pulling ℓ̂_m toward the same ℓ_m values that define P, Q_kin is forced toward P by construction; reporting a lower KL for Q_kin than for the MSE-only model is equivalent to reporting that the constraint term did what it was defined to do. This makes the KL claim self-definitional rather than an independent test of biomechanical plausibility. The RMSE improvements on marker coordinates remain externally benchmarked and are not affected by this circularity.

full rationale

The paper's central comparisons are not circular: long-term recursive RMSE (Table V) is computed against measured Vicon marker coordinates, and the claimed transformer advantage over BLSTM and the arm/leg RMSE reductions from the kinematic constraint are evaluated on those external coordinates. The kinematic loss term (Eq. 3) is a legitimate regularizer; only the accompanying KL-divergence validation (Eq. 4, Fig. 6) is self-referential, since it evaluates the model on the same segment-length invariant that the loss explicitly enforces. Other same-group dependencies—the dataset from [18], the seed frames from [16], and the prior MLP baseline—are contextual inputs and acknowledged limitations rather than steps that reduce the present result to its inputs by construction. The abstract/test-set 41.4 mm versus LOSO 54.39 ± 17.59 mm discrepancy is an evaluation-protocol and generalization-estimation issue, not circularity. Overall score reflects one secondary self-definitional validation while the main claims retain independent empirical content.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claims rest on the measured mocap dataset (external ground truth), a standard biomechanical invariant (constant segment lengths), and standard ML hyperparameter fitting. No invented entities. The main fitted knob is the constraint weight a (validation-tuned), plus TPE-tuned architecture choices; the headline improvements are reported without subject-level error bars.

free parameters (4)
  • Kinematic-constraint weight a = a = 10 (arm model), a = 1 (leg model)
    Selected by trial-and-error among {1, 10, 100} on validation loss; the reported 8%/21% improvements depend on this choice (Sec. II.F, Sec. III).
  • Input window length l_in = 25 frames
    Sliding-window length set to 25 following prior gait-prediction work [26]; defines the temporal conditioning of every prediction.
  • Transformer hyperparameters (heads, layers, dropout, expander width) = 32 heads (Table II) or 16 (Sec. II.C.2); 1 enc/3 dec; dropout 0.25
    Tuned via TPE on validation loss; the paper is internally inconsistent about the final head count, which affects reproducibility.
  • Body segmentation into four networks = head, arms, body-pelvic, legs
    Design choice to cut parameters; segments are trained as separate models and errors are combined, so the 'whole-body' RMSE depends on this split.
axioms (4)
  • domain assumption Body segment lengths (upper arm, forearm, shank) are constant within a task.
    Standard biomechanical invariant invoked to build the constrained loss (Sec. II.F).
  • domain assumption The first 25 frames contain enough information to recursively predict the remaining 76 frames of a load-reaching task.
    Structural to the whole forecasting approach (Sec. II.C); no information-theoretic or empirical support beyond the test results.
  • domain assumption The 20-subject dataset (young, normal-weight, right-handed men) is representative enough for training and for subject-level generalization claims.
    Authors themselves flag narrow demographics (Sec. IV.B); LOSO across 20 same-profile males does not test demographic generalization.
  • domain assumption Marker-based plug-in-gait mocap, filtered at 10 Hz and resampled to 101 frames, is an adequate ground truth for whole-body posture.
    Ground truth for all reported errors (Sec. II.A-B); marker placement error and resampling artifacts are not quantified.

pith-pipeline@v1.3.0-alltime-deepseek · 13547 in / 16791 out tokens · 166972 ms · 2026-08-03T20:11:29.762764+00:00 · methodology

0 comments
read the original abstract

This study aimed to explore the application of deep neural networks for whole-body human posture prediction during dynamic load-reaching activities. Two time-series models were trained using bidirectional long short-term memory (BLSTM) and transformer architectures. The dataset consisted of 3D full-body plug-in gait dynamic coordinates from 20 normal-weight healthy male individuals each performing 204 load-reaching tasks from different load positions while adapting various lifting and handling techniques. The model inputs consisted of the 3D position of the hand-load position, lifting (stoop, full-squat and semi-squat) and handling (one- and two-handed) techniques, body weight and height, and the 3D coordinate data of the body posture from the first 25% of the task duration. These inputs were used by the models to predict body coordinates during the remaining 75% of the task period. Moreover, a novel method was proposed to improve the accuracy of the previous and present posture prediction networks by enforcing constant body segment lengths through the optimization of a new cost function. The results indicated that the new cost function decreased the prediction error of the models by approximately 8% and 21% for the arm and leg models, respectively. We indicated that utilizing the transformer architecture, with a root-mean-square-error of 41.4 mm, exhibited approximately 58% more accurate long-term performance than the BLSTM-based model. This study merits the use of neural networks that capture time series dependencies in 3D motion frames, providing a unique approach for understanding and predict motion dynamics during manual material handling activities.

Figures

Figures reproduced from arXiv: 2511.20615 by Ali Mojibi, Alireza Taheri, Mahdi Mohseni, Navid Arjmand, Seyede Niloofar Hosseini.

Figure 1
Figure 1. Figure 1: The process of generating inputs to the neural networks from time series data: for each segment, the 3D coordinates of the markers in lin = 25 input frames and the task characteristics are combined to generate one sample for input to the networks. This means that they did not consider any correlation or relationship between the movement frames as time series. Recent advancements in deep learning, however, … view at source ↗
Figure 2
Figure 2. Figure 2: Skeleton for the human body with attached markers: (a) four different segments used for posture prediction and (b) important markers. for each dataset were generated using the conventional sliding window method with a length of 25 frames (i.e., 3D coordinates of 25 frame motions) and a stride of 1 frame [26]. A sample X ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Transformer-based model for posture prediction. Input sequences are embedded through linear layers, then processed by encoder-decoder blocks. The output is passed to a fully connected layer to generate future posture predictions using self-attention mechanisms. D. Performance metrics The short- and long-term performances of the model were evaluated in terms of RMSE, normalized RMSE (to the range of measure… view at source ↗
Figure 3
Figure 3. Figure 3: BLSTM architecture for posture prediction. The model processes time-series data bidirectionally, capturing both past and future context [37]. Outputs from forward and backward LSTM layers are concatenated and passed to a fully connected layer to predict posture at each time step. C.2. Transformer The transformer architecture, with encoder-decoder blocks and multi-head self-attention, has outperformed RNNs … view at source ↗
Figure 5
Figure 5. Figure 5: Predicted vs. actual dynamic postures across six key time frames using the Transformer model. Each subplot shows a comparison between the predicted posture (green) and the ground truth (red) during a load-reaching task at different time frames (1, 25, 50, 75, 90, and 101). The associated RMSE values (in mm) indicate the prediction accuracy at each frame. The task involved stoop lifting with both hands, tar… view at source ↗
Figure 6
Figure 6. Figure 6: KL divergence values for arm and leg segments under two training conditions: models trained with MSE loss only (gray bars) and models trained with MSE loss plus kinematic constraints (white bars). Error bars indicate standard deviation. A significant reduction in KL divergence for the leg segment when using kinematic constraints is denoted by the red asterisk (p = 0.0226). This analysis was performed over … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 12 canonical work pages

  1. [1]

    Risk factors for first-ever low back pain among workers in their first employment,

    A. Van Nieuwenhuyse et al., “Risk factors for first-ever low back pain among workers in their first employment,” Occupational Medicine, vol. 54, no. 8, pp. 513–519, Dec. 2004, doi: 10.1093/occmed/kqh091

    work page doi:10.1093/occmed/kqh091 2004

  2. [2]

    Intradiscal pressure together with anthropometric data – a data set for the validation of models,

    H.-J. Wilke, P. Neef, B. Hinz, H. Seidel, and L. Claes, “Intradiscal pressure together with anthropometric data – a data set for the validation of models,” Clinical Biomechanics, vol. 16, pp. S111– S126, Jan. 2001, doi: 10.1016/S0268-0033(00)00103-0

    work page doi:10.1016/s0268-0033(00)00103-0 2001

  3. [3]

    Estimation of loads on human lumbar spine: A review of in vivo and computational model studies,

    M. Dreischarf, A. Shirazi-Adl, N. Arjmand, A. Rohlmann, and H. Schmidt, “Estimation of loads on human lumbar spine: A review of in vivo and computational model studies,” Journal of Biomechanics, vol. 49, no. 6, pp. 833–845, Apr. 2016, doi: 10.1016/j.jbiomech.2015.12.038

    work page doi:10.1016/j.jbiomech.2015.12.038 2016

  4. [4]

    Musculoskeletal injury risk assessment in a car dashboard assembly line using various quantitative and qualitative tools,

    M. Bahramian, M. A. Shayestehpour, M. Yavari, H. Mehrabi, and N. Arjmand, “Musculoskeletal injury risk assessment in a car dashboard assembly line using various quantitative and qualitative tools,” in 2021 28th National and 6th International Iranian Conference on Biomedical Engineering (ICBME), Nov. 2021, pp. 310–316. doi: 10.1109/ICBME54433.2021.9750385

    arXiv 2021

  5. [5]

    Comparison and evaluation of ten widely-used ergonomic risk assessment tools in various manual material handling activities,

    S. Zargarzadeh, M. R. Bahramian, M. Mohseni, and N. Arjmand, “Comparison and evaluation of ten widely-used ergonomic risk assessment tools in various manual material handling activities,” Scientia Iranica, 2024, doi: doi:%2010.24200/sci.2024.63530.8451

    arXiv 2024

  6. [6]

    Accuracy of Computer Vision-Based Pose Estimation Algorithms in Predicting Joint Kinematics During Gait,

    A. A. Hulleck et al., “Accuracy of Computer Vision-Based Pose Estimation Algorithms in Predicting Joint Kinematics During Gait,” Aug. 31, 2023. doi: 10.21203/rs.3.rs-3239200/v1

    work page doi:10.21203/rs.3.rs-3239200/v1 2023

  7. [7]

    Marker-less versus marker-based driven musculoskeletal models of the spine during static load-handling activities,

    F. Asadi and N. Arjmand, “Marker-less versus marker-based driven musculoskeletal models of the spine during static load-handling activities,” Journal of Biomechanics, vol. 112, p. 110043, Nov. 2020, doi: 10.1016/j.jbiomech.2020.110043

    arXiv 2020

  8. [8]

    Spinal segment ranges of motion, movement coordination, and three-dimensional kinematics during occupational activities in normal-weight and obese individuals,

    M. Ghasemi and N. Arjmand, “Spinal segment ranges of motion, movement coordination, and three-dimensional kinematics during occupational activities in normal-weight and obese individuals,” Journal of Biomechanics, vol. 123, p. 110539, June 2021, doi: 10.1016/j.jbiomech.2021.110539

    arXiv 2021

  9. [9]

    Daroudi, N

    S. Daroudi, N. Arjmand, M. Mohseni, M. El-Rich, and M. Parnianpour, “Evaluation of ground reaction forces and centers of pressure predicted by AnyBody Modeling System during load reaching/handling activities and effects of the prediction errors on model-estimated spinal loads,” Journal of Biomechanics, vol. 164, p. 111974, Feb. 2024, doi: 10.1016/j.jbiome...

    arXiv 2024

  10. [10]

    BlazePose-Seq2Seq: Leveraging Regular RGB Cameras for Robust Gait Assessment,

    A. A. Hulleck et al., “BlazePose-Seq2Seq: Leveraging Regular RGB Cameras for Robust Gait Assessment,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 32, pp. 1715–1724, 2024, doi: 10.1109/TNSRE.2024.3391908

    arXiv 2024

  11. [11]

    In- field assessment of change-of-direction ability with a single wearable sensor,

    S. Apte, H. Karami, C. Vallat, V. Gremeaux, and K. Aminian, “In- field assessment of change-of-direction ability with a single wearable sensor,” Scientific Reports, vol. 13, Mar. 2023, doi: 10.1038/s41598- 023-30773-y

    doi:10.1038/s41598- 2023

  12. [12]

    Artificial neural networks to predict 3D spinal posture in reaching and lifting activities; Applications in biomechanical models,

    A. Gholipour and N. Arjmand, “Artificial neural networks to predict 3D spinal posture in reaching and lifting activities; Applications in biomechanical models,” Journal of Biomechanics, vol. 49, no. 13, pp. 2946–2952, Sept. 2016, doi: 10.1016/j.jbiomech.2016.07.008

    work page doi:10.1016/j.jbiomech.2016.07.008 2016

  13. [13]

    A new scheme for the development of IMU-based activity recognition systems for telerehabilitation,

    A. M. Nasrabadi, A. R. Eslaminia, P. R. Bakhshayesh, M. Ejtehadi, L. Alibiglou, and S. Behzadipour, “A new scheme for the development of IMU-based activity recognition systems for telerehabilitation,” Medical Engineering & Physics, vol. 108, p. 103876, Oct. 2022, doi: 10.1016/j.medengphy.2022.103876

    arXiv 2022

  14. [14]

    A Systematic Review: Advancing Ergonomic Posture Risk Assessment Through the Integration of Computer Vision and Machine Learning Techniques,

    Z. Yang, D. Song, J. Ning, and Z. Wu, “A Systematic Review: Advancing Ergonomic Posture Risk Assessment Through the Integration of Computer Vision and Machine Learning Techniques,” IEEE Access, vol. 12, pp. 180481–180519, 2024, doi: 10.1109/ACCESS.2024.3509447

    arXiv 2024

  15. [15]

    Coupled artificial neural networks to estimate 3D whole-body posture, lumbosacral moments, and spinal loads during load-handling activities,

    F. Aghazadeh, N. Arjmand, and A. M. Nasrabadi, “Coupled artificial neural networks to estimate 3D whole-body posture, lumbosacral moments, and spinal loads during load-handling activities,” Journal of Biomechanics, vol. 102, p. 109332, Mar. 2020, doi: 10.1016/j.jbiomech.2019.109332

    arXiv 2020

  16. [16]

    AI-based human whole- body posture-prediction in continuous load reaching/leaving activities,

    R. Ahmadi, M. Mohseni, and N. Arjmand, “AI-based human whole- body posture-prediction in continuous load reaching/leaving activities,” Journal of Biomechanics, vol. 185, p. 112681, 2025, doi: 10.1016/j.jbiomech.2025.112681

    arXiv 2025

  17. [17]

    An artificial neural network for full- body posture prediction in dynamic lifting activities and effects of its prediction errors on model-estimated spinal loads,

    N. Hosseini and N. Arjmand, “An artificial neural network for full- body posture prediction in dynamic lifting activities and effects of its prediction errors on model-estimated spinal loads,” Journal of Biomechanics, vol. 162, p. 111896, Jan. 2024, doi: 10.1016/j.jbiomech.2023.111896

    arXiv 2024

  18. [18]

    Multi-task artificial neural networks and their extrapolation capabilities to predict full- body 3D human posture during one- and two-handed load-handling activities,

    M. Mohseni, S. Zargarzadeh, and N. Arjmand, “Multi-task artificial neural networks and their extrapolation capabilities to predict full- body 3D human posture during one- and two-handed load-handling activities,” Journal of Biomechanics, vol. 162, p. 111884, Jan. 2024, doi: 10.1016/j.jbiomech.2023.111884

    arXiv 2024

  19. [19]

    Improved artificial neural networks for 3D body posture and lumbosacral moment predictions during manual material handling activities,

    M. Mohseni, F. Aghazadeh, and N. Arjmand, “Improved artificial neural networks for 3D body posture and lumbosacral moment predictions during manual material handling activities,” Journal of Biomechanics, vol. 131, p. 110921, Jan. 2022, doi: 10.1016/j.jbiomech.2021.110921

    arXiv 2022

  20. [20]

    Prediction of the thorax/pelvis orientations and L5–S1 disc loads during various static activities using neuro-fuzzy,

    S. H. Mousavi, H. Sayyaadi, and N. Arjmand, “Prediction of the thorax/pelvis orientations and L5–S1 disc loads during various static activities using neuro-fuzzy,” J Mech Sci Technol, vol. 34, no. 8, pp. 3481–3485, Aug. 2020, doi: 10.1007/s12206-020-0740-0

    work page doi:10.1007/s12206-020-0740-0 2020

  21. [21]

    Predicting Standing Reach Postures using Deep Neural Networks,

    M. P. Reed, S. M. Ebert, C. D’Souza, and M. L. H. Jones, “Predicting Standing Reach Postures using Deep Neural Networks,” Proceedings of the Human Factors and Ergonomics Society Annual Meeting, vol. 66, no. 1, pp. 662–666, Sept. 2022, doi: 10.1177/1071181322661150

    work page doi:10.1177/1071181322661150 2022

  22. [22]

    Lifting Posture Prediction With Generative Models for Improving Occupational Safety,

    L. Li, S. Prabhu, Z. Xie, H. Wang, L. Lu, and X. Xu, “Lifting Posture Prediction With Generative Models for Improving Occupational Safety,” IEEE Transactions on Human-Machine Systems, vol. 51, no. 5, pp. 494–503, Oct. 2021, doi: 10.1109/THMS.2021.3102511

    arXiv 2021

  23. [23]

    A neural network model for predicting postures during non-repetitive manual materials handling tasks,

    M. A. Perez and M. A. Nussbaum, “A neural network model for predicting postures during non-repetitive manual materials handling tasks,” Ergonomics, vol. 51, no. 10, pp. 1549–1564, Oct. 2008, doi: 10.1080/00140130802220570

    work page doi:10.1080/00140130802220570 2008

  24. [24]

    Predicting Human Postures for Manual Material Handling Tasks Using a Conditional Diffusion Model,

    L. Qing, B. Su, S. Jung, L. Lu, H. Wang, and X. Xu, “Predicting Human Postures for Manual Material Handling Tasks Using a Conditional Diffusion Model,” IEEE Transactions on Human- Machine Systems, vol. 54, no. 6, pp. 723–732, Dec. 2024, doi: 10.1109/THMS.2024.3472548

    arXiv 2024

  25. [25]

    Lower Limb Kinematics Trajectory Prediction Using Long Short-Term Memory Neural Networks,

    A. Zaroug, D. T. H. Lai, K. Mudie, and R. Begg, “Lower Limb Kinematics Trajectory Prediction Using Long Short-Term Memory Neural Networks,” Frontiers in Bioengineering and Biotechnology, vol. 8, 2020, Accessed: July 15, 2023. [Online]. Available: https://www.frontiersin.org/articles/10.3389/fbioe.2020.00362

    arXiv 2020

  26. [26]

    Prediction of gait trajectories based on the Long Short Term Memory neural networks,

    A. Zaroug, A. Garofolini, D. Lai, K. Mudie, and R. Begg, “Prediction of gait trajectories based on the Long Short Term Memory neural networks,” PLOS ONE, vol. 16, p. e0255597, Aug. 2021, doi: 10.1371/journal.pone.0255597

    work page doi:10.1371/journal.pone.0255597 2021

  27. [27]

    Deep Learning Models for Stable Gait Prediction Applied to Exoskeleton Reference Trajectories for Children With Cerebral Palsy,

    R. Kolaghassi, G. Marcelli, and K. Sirlantzis, “Deep Learning Models for Stable Gait Prediction Applied to Exoskeleton Reference Trajectories for Children With Cerebral Palsy,” IEEE Access, vol. 11, pp. 31962–31976, 2023, doi: 10.1109/ACCESS.2023.3252916

    arXiv 2023

  28. [28]

    An Attention-Based CNN-LSTM Model with Limb Synergy for Joint Angles Prediction,

    C. Zhu, Q. Liu, W. Meng, Q. Ai, and S. Q. Xie, “An Attention-Based CNN-LSTM Model with Limb Synergy for Joint Angles Prediction,” in 2021 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), July 2021, pp. 747–752. doi: 10.1109/AIM46487.2021.9517544

    arXiv 2021

  29. [29]

    Continuous Angle Prediction of Lower Limb Knee Joint Based on sEMG,

    C. Li, H. He, S. Yin, H. Deng, and Y. Zhu, “Continuous Angle Prediction of Lower Limb Knee Joint Based on sEMG,” in 2021 IEEE International Conference on Recent Advances in Systems Science and Engineering (RASSE), Dec. 2021, pp. 1–6. doi: 10.1109/RASSE53195.2021.9686816

    arXiv 2021

  30. [30]

    Deep learning with an attention mechanism for continuous biomechanical motion estimation across varied activities,

    G. Ding, A. Plummer, and I. Georgilas, “Deep learning with an attention mechanism for continuous biomechanical motion estimation across varied activities,” Front. Bioeng. Biotechnol., vol. 10, Oct. 2022, doi: 10.3389/fbioe.2022.1021505

    arXiv 2022

  31. [31]

    Deep Transformer Models for Time Series Forecasting: The Influenza Prevalence Case,

    N. Wu, B. Green, X. Ben, and S. O’Banion, “Deep Transformer Models for Time Series Forecasting: The Influenza Prevalence Case,” Jan. 22, 2020, arXiv: arXiv:2001.08317. doi: 10.48550/arXiv.2001.08317

    Pith/arXiv arXiv doi:10.48550/arxiv.2001.08317 2020

  32. [32]

    Effect of Gait Speed on Trajectory Prediction Using Deep Learning Models for Exoskeleton Applications,

    R. Kolaghassi, G. Marcelli, and K. Sirlantzis, “Effect of Gait Speed on Trajectory Prediction Using Deep Learning Models for Exoskeleton Applications,” Sensors, vol. 23, p. 5687, June 2023, doi: 10.3390/s23125687

    work page doi:10.3390/s23125687 2023

  33. [33]

    Effect of obesity on spinal loads during load-reaching activities: A subject- and kinematics-specific musculoskeletal modeling approach,

    M. Bahramian, N. Arjmand, M. El-Rich, and M. Parnianpour, “Effect of obesity on spinal loads during load-reaching activities: A subject- and kinematics-specific musculoskeletal modeling approach,” Journal of Biomechanics, vol. 161, p. 111770, Dec. 2023, doi: 10.1016/j.jbiomech.2023.111770

    arXiv 2023

  34. [34]

    Dehghan and N

    P. Dehghan and N. Arjmand, “The National Institute for Occupational Safety and Health (NIOSH) Recommended Weight Generates Different Spine Loads in Load-Handling Activity Performed Using Stoop, Semi-squat and Full-Squat Techniques; a Full-Body Musculoskeletal Model Study,” Hum Factors, vol. 66, no. 5, pp. 1387–1398, May 2024, doi: 10.1177/00187208221141652

    work page doi:10.1177/00187208221141652 2024

  35. [35]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press, 2016. [Online]. Available: http://www.deeplearningbook.org

    2016

  36. [36]

    Stacked bidirectional and unidirectional LSTM recurrent neural network for forecasting network-wide traffic state with missing values,

    Z. Cui, R. Ke, Z. Pu, and Y. Wang, “Stacked bidirectional and unidirectional LSTM recurrent neural network for forecasting network-wide traffic state with missing values,” Transportation Research Part C: Emerging Technologies, vol. 118, p. 102674, Sept. 2020, doi: 10.1016/j.trc.2020.102674

    arXiv 2020

  37. [37]

    Bidirectional recurrent neural networks,

    M. Schuster and K. K. Paliwal, “Bidirectional recurrent neural networks,” IEEE Transactions on Signal Processing, vol. 45, no. 11, pp. 2673–2681, Nov. 1997, doi: 10.1109/78.650093

    doi:10.1109/78.650093 1997

  38. [38]

    Attention is All you Need,

    A. Vaswani et al., “Attention is All you Need,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2017. Accessed: July 26, 2023. [Online]. Available: https://papers.nips.cc/paper_files/paper/2017/hash/3f5ee243547dee91f bd053c1c4a845aa-Abstract.html

    2017

  39. [39]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    A. Dosovitskiy et al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” June 03, 2021, arXiv: arXiv:2010.11929. doi: 10.48550/arXiv.2010.11929

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.11929 2021

  40. [40]

    Adam: A Method for Stochastic Optimization,

    D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” Jan. 29, 2017, arXiv: arXiv:1412.6980. doi: 10.48550/arXiv.1412.6980

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2017

  41. [41]

    Optuna: A Next-generation Hyperparameter Optimization Framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A Next-generation Hyperparameter Optimization Framework,” July 25, 2019, arXiv: arXiv:1907.10902. doi: 10.48550/arXiv.1907.10902

    Pith/arXiv arXiv doi:10.48550/arxiv.1907.10902 2019

  42. [42]

    A Learning Algorithm for Continually Running Fully Recurrent Neural Networks,

    R. J. Williams and D. Zipser, “A Learning Algorithm for Continually Running Fully Recurrent Neural Networks,” Neural Computation, vol. 1, no. 2, pp. 270–280, June 1989, doi: 10.1162/neco.1989.1.2.270

    doi:10.1162/neco.1989.1.2.270 1989

  43. [43]

    Professor Forcing: A New Algorithm for Training Recurrent Networks,

    A. Lamb, A. Goyal, Y. Zhang, S. Zhang, A. Courville, and Y. Bengio, “Professor Forcing: A New Algorithm for Training Recurrent Networks,” Oct. 27, 2016, arXiv: arXiv:1610.09038. doi: 10.48550/arXiv.1610.09038

    Pith/arXiv arXiv doi:10.48550/arxiv.1610.09038 2016

  44. [44]

    Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks,

    S. Bengio, O. Vinyals, N. Jaitly, and N. Shazeer, “Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks,” Sept. 23, 2015, arXiv: arXiv:1506.03099. doi: 10.48550/arXiv.1506.03099

    Pith/arXiv arXiv doi:10.48550/arxiv.1506.03099 2015

  45. [45]

    Performance of Deep Learning Models in Forecasting Gait Trajectories of Children with Neurological Disorders,

    R. Kolaghassi, M. K. Al-Hares, G. Marcelli, and K. Sirlantzis, “Performance of Deep Learning Models in Forecasting Gait Trajectories of Children with Neurological Disorders,” Sensors (Basel), vol. 22, no. 8, p. 2969, Apr. 2022, doi: 10.3390/s22082969

    work page doi:10.3390/s22082969 2022