pith. sign in

arxiv: 2604.13279 · v2 · submitted 2026-04-14 · 💻 cs.CV · cs.AI

Explainable Fall Detection for Elderly Monitoring via Temporally Stable SHAP in Skeleton-Based Human Activity Recognition

Mohammad Saleh , Azadeh Tabatabaei This is my paper

Pith reviewed 2026-05-10 15:50 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords fall detectionSHAPexplainable AIskeleton-basedLSTMtemporal stabilityelderly monitoringNTU RGB+D
0
0 comments X p. Extension

The pith

Linear smoothing of frame-wise SHAP attributions produces temporally stable explanations for LSTM-based skeleton fall detection.

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

The paper presents a lightweight LSTM model for detecting falls from skeleton pose sequences that also generates explanations. It introduces T-SHAP by treating standard SHAP values across video frames as a time series and applying a linear smoothing step to suppress erratic changes between frames. This step is presented as analogous to low-pass filtering that keeps the underlying Shapley values intact. On the NTU RGB+D dataset the model reaches 94.3 percent accuracy with under 25 milliseconds latency while the smoothed attributions score higher on perturbation faithfulness and show lower frame-to-frame variance than plain SHAP or Grad-CAM. The resulting explanations consistently point to lower-limb instability and trunk posture shifts that match known fall mechanics.

Core claim

Treating frame-wise SHAP attributions as a temporal signal and applying a linear smoothing operator reduces high-frequency variance while preserving the theoretical properties of Shapley-based attributions, producing more reliable and biomechanically interpretable explanations for real-time fall detection.

What carries the argument

T-SHAP: linear smoothing operator applied to frame-wise SHAP attributions viewed as a temporal signal.

If this is right

  • The combined LSTM and T-SHAP pipeline reaches 94.3 percent classification accuracy on the NTU RGB+D dataset.
  • End-to-end processing stays below 25 ms latency, enabling real-time deployment.
  • T-SHAP raises area-under-perturbation from 0.89 (SHAP) and 0.82 (Grad-CAM) to 0.91.
  • Temporal variance in the attribution signals decreases while biomechanical relevance is retained.
  • No extra model training is required to obtain the stabilized explanations.

Where Pith is reading between the lines

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

  • Stable attributions could support consistent alerts in continuous elderly monitoring without frequent false triggers.
  • The smoothing approach may transfer to other sequential skeleton tasks such as gait assessment.
  • Lightweight post-processing allows deployment on existing camera hardware without retraining.
  • Testing on additional datasets with varied fall types would check whether the highlighted motion patterns remain consistent.

Load-bearing premise

The linear smoothing operator preserves the theoretical properties of SHAP attributions and reduces variance without introducing bias or artifacts.

What would settle it

A faithfulness test in which T-SHAP attributions produce lower area-under-perturbation scores than unsmoothed SHAP or fail to highlight lower-limb instability in verified fall sequences.

Figures

Figures reproduced from arXiv: 2604.13279 by Azadeh Tabatabaei, Mohammad Saleh.

Figure 1
Figure 1. Figure 1: Overview of the proposed framework. A skeleton [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Fig.2. Architecture of the proposed lightweight temporal model for skeleton [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparative spatiotemporal analysis of feature attributions using SHAP, T-SHAP, and Grad-CAM. The top row shows heatmaps of joint-level importance over time, with the horizontal axis representing frames and the vertical axis indicating joint indices. Color intensity indicates contribution to the predicted fall class (Red indicates higher contribution). SHAP exhibits high temporal variance with fragmented a… view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of per-joint attribution magnitude across the skeleton. Higher attribution scores are observed in central joints (e.g., spine base (1), neck (3)) and distal joints (e.g., hands (24, 25)), suggesting that both biomechanical stability and full-body motion patterns contribute to fall detection. 4.7 Discussion We emphasize that T-SHAP is a structured post-hoc transformation tailored to sequential … view at source ↗
read the original abstract

Reliable fall detection in elderly care requires monitoring systems that are not only accurate but also capable of producing stable, interpretable explanations of motion dynamics, a requirement that existing post hoc explainability methods rarely satisfy when applied to sequential biosignals. This study introduces a lightweight framework for skeleton-based fall detection that combines a Long Short-Term Memory (LSTM) model with a temporally stabilized attribution mechanism. We propose Temporal SHAP (T-SHAP), which treats frame-wise SHAP attributions as a temporal signal and applies a linear smoothing operator to reduce high-frequency variance. From a signal processing perspective, this operation is analogous to low-pass filtering, enabling the extraction of consistent temporal patterns while preserving the theoretical properties of Shapley-based attributions. Experiments conducted on the NTU RGB+D dataset demonstrate that the proposed approach achieves 94.3% classification accuracy with an end-to-end latency below 25 ms, supporting real-time applicability. Quantitative evaluation using perturbation-based faithfulness metrics shows that T-SHAP improves attribution reliability compared to standard SHAP (AUP: 0.91 vs. 0.89) and Grad-CAM (0.82), while also reducing temporal variance in the attribution signals. The resulting explanations highlight biomechanically relevant motion patterns, such as lower-limb instability and changes in trunk posture, which are consistent with known characteristics of fall events. The resulting framework is computationally lightweight, requires no additional model training, and produces explanations that are both temporally stable and biomechanically meaningful, properties directly relevant to the reliability demands of AI-assisted clinical monitoring.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a lightweight LSTM-based framework for skeleton-based fall detection on the NTU RGB+D dataset, augmented by Temporal SHAP (T-SHAP): a post-hoc linear smoothing operator applied to per-frame SHAP attribution vectors to reduce high-frequency temporal variance. It claims this preserves Shapley properties, yields 94.3% classification accuracy with <25 ms end-to-end latency, improves perturbation-based faithfulness (AUP 0.91 vs. 0.89 for SHAP and 0.82 for Grad-CAM), and produces biomechanically relevant explanations such as lower-limb instability.

Significance. If the central claims hold, the work would offer a practical advance in temporally stable, post-hoc explainability for real-time sequential biosignal models without retraining. The reported accuracy, latency, and direct comparison to baselines on a standard dataset are concrete strengths; the emphasis on variance reduction addresses a genuine limitation of frame-wise attributions in activity recognition. Reproducible code or parameter-free derivations are not mentioned, but the perturbation evaluation provides a falsifiable metric.

major comments (2)
  1. [Abstract] Abstract (T-SHAP description): The assertion that the linear smoothing operator 'preserves the theoretical properties of Shapley-based attributions' lacks any derivation, proof, or post-smoothing verification that local accuracy continues to hold (i.e., that the sum of smoothed attributions still equals f(x) - E[f] for each frame or the full sequence). Because the operator mixes values across time steps, it generally does not commute with the summation axiom unless explicitly constructed to do so; no such construction or empirical check (e.g., summation error before/after smoothing) is provided. This property is load-bearing for both the AUP improvement claim and the assertion of biomechanical relevance.
  2. [Abstract] Abstract (experimental claims): The reported AUP gain (0.91 vs. 0.89) and temporal-variance reduction are presented without the perturbation protocol details, number of samples, choice of smoothing kernel width, or statistical test for the difference; these omissions prevent assessment of whether the 0.02 AUP margin is robust or an artifact of the specific smoothing operator.
minor comments (2)
  1. [Abstract] Abstract: Missing details on the LSTM architecture, exact smoothing operator (kernel, window size), baseline choice for SHAP, and full quantitative results (e.g., per-class metrics or variance values before/after T-SHAP).
  2. [Methods] The manuscript would benefit from a dedicated methods subsection with explicit notation for the linear operator and a figure comparing raw vs. smoothed attribution time series on example fall sequences.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and will revise the manuscript to strengthen the theoretical and experimental presentation.

read point-by-point responses

  1. Referee: [Abstract] Abstract (T-SHAP description): The assertion that the linear smoothing operator 'preserves the theoretical properties of Shapley-based attributions' lacks any derivation, proof, or post-smoothing verification that local accuracy continues to hold (i.e., that the sum of smoothed attributions still equals f(x) - E[f] for each frame or the full sequence). Because the operator mixes values across time steps, it generally does not commute with the summation axiom unless explicitly constructed to do so; no such construction or empirical check (e.g., summation error before/after smoothing) is provided. This property is load-bearing for both the AUP improvement claim and the assertion of biomechanical relevance.

    Authors: We agree that the abstract statement would be strengthened by explicit justification. The manuscript presents T-SHAP as a post-hoc linear smoothing operator applied to per-frame SHAP vectors. While we describe it as preserving properties from a signal-processing viewpoint, no derivation or verification of local accuracy is included. In the revision we will add an empirical check of the summation error (mean absolute deviation between summed attributions and f(x) - E[f]) before and after smoothing on the NTU RGB+D test set, and we will qualify the claim in the abstract and Methods section to reflect whether the error remains negligible in practice. This directly addresses the concern for the AUP and biomechanical claims. revision: yes

  2. Referee: [Abstract] Abstract (experimental claims): The reported AUP gain (0.91 vs. 0.89) and temporal-variance reduction are presented without the perturbation protocol details, number of samples, choice of smoothing kernel width, or statistical test for the difference; these omissions prevent assessment of whether the 0.02 AUP margin is robust or an artifact of the specific smoothing operator.

    Authors: We acknowledge that the abstract omits key protocol details. The full manuscript describes the perturbation-based faithfulness evaluation in the Experiments section, but these specifics are not summarized in the abstract. We will revise the abstract to include the perturbation protocol outline, number of evaluated samples, smoothing kernel width, and confirmation that a statistical test was applied to the AUP difference. This will make the 0.02 margin claim fully assessable and reproducible. revision: yes

Circularity Check

0 steps flagged

No circularity; T-SHAP is post-hoc linear smoothing on standard SHAP values

full rationale

The paper's core contribution is an LSTM-based fall detector plus T-SHAP, defined as applying an unspecified linear smoothing operator to per-frame SHAP attributions (described as analogous to low-pass filtering). This step is presented as a signal-processing post-processing technique that 'preserves the theoretical properties of Shapley-based attributions,' but the text contains no equations, self-referential definitions, fitted parameters renamed as predictions, or load-bearing self-citations that would make the reported AUP improvement (0.91) or 94.3% accuracy tautological by construction. The accuracy and faithfulness metrics are obtained from independent experiments on the NTU RGB+D dataset; the smoothing operator is not derived from the target results and does not reduce the claims to the inputs. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework depends on the unverified assumption that the smoothing operation enhances reliability without compromising the faithfulness of attributions, as no detailed derivation or sensitivity analysis is provided in the abstract.

axioms (1)
  • domain assumption Frame-wise SHAP attributions can be treated as a temporal signal amenable to linear smoothing while preserving Shapley properties.
    Invoked in the description of T-SHAP as analogous to low-pass filtering.
invented entities (1)
  • T-SHAP no independent evidence
    purpose: To provide temporally stable explanations for LSTM-based fall detection.
    Introduced in this work as a new mechanism without independent validation outside the reported experiments.

pith-pipeline@v0.9.0 · 5587 in / 1325 out tokens · 58115 ms · 2026-05-10T15:50:35.828052+00:00 · methodology

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

54 extracted references · 54 canonical work pages

  1. [1]

    Interpretable Passive Multi -Modal Sensor Fusion for Human Identification and Activity Recogniti on,

    L. Yuan et al., “Interpretable Passive Multi -Modal Sensor Fusion for Human Identification and Activity Recogniti on,” Sensors, vol. 22, no. 15, Art. no. 15, Jan. 2022, doi: 10.3390/s22155787

    work page doi:10.3390/s22155787 2022

  2. [2]

    Explainable Artificial Intelligence and Wearable Sensor -Based Gait Analysis to Identify Patients with Osteopenia an d Sarcopenia in Daily Life,

    J.-K. Kim, M.-N. Bae, K. Lee, J.-C. Kim, and S. G. Hong, “Explainable Artificial Intelligence and Wearable Sensor -Based Gait Analysis to Identify Patients with Osteopenia an d Sarcopenia in Daily Life,” Biosensors, vol. 12, no. 3, Art. no. 3, Mar. 2022, doi: 10.3390/bios12030167

    work page doi:10.3390/bios12030167 2022

  3. [3]

    Vision-based human activity recognition: a survey,

    D. R. Beddiar, B. Nini, M. Sabokrou, and A. Hadid, “Vision-based human activity recognition: a survey,” Multimed. Tools Appl., vol. 79, no. 41, pp. 30509–30555, Nov. 2020, doi: 10.1007/s11042-020-09004-3

    work page doi:10.1007/s11042-020-09004-3 2020

  4. [4]

    Artificial Intelligence for Elderly Fall Detection: State -of-the-art Methods, Applications and Challenges,

    M. J. A. Nahian et al., “Artificial Intelligence for Elderly Fall Detection: State -of-the-art Methods, Applications and Challenges,” Cogn. Comput., vol. 18, no. 1, p. 12, Feb. 2026, doi: 10.1007/s12559-026-10550-5

    work page doi:10.1007/s12559-026-10550-5 2026

  5. [5]

    Spatio-temporal neural network with handcrafted features for skeleton-based action recognition,

    M. Nan, M. Trăscău, and A.-M. Florea, “Spatio-temporal neural network with handcrafted features for skeleton-based action recognition,” Neural Comput. Appl., vol. 36, no. 16, pp. 9221–9243, Jun. 2024, doi: 10.1007/s00521-024-09559-4

    work page doi:10.1007/s00521-024-09559-4 2024

  6. [6]

    Modulated Periodic Activations for Generalizable Local Functional Representations , rights =

    Y. Chen, Z. Zhang, C. Yuan, B. Li, Y. Deng, and W. Hu, “Channel -wise Topology Refinement Graph Convolution for Skeleton -Based Action Recognition,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , Oct. 2021, pp. 13339 –13348. doi : 10.1109/ICCV48922.2021.01311

    work page doi:10.1109/iccv48922.2021.01311 2021

  7. [7]

    Efficient fall detection using Kalman filter -enhanced triaxial accelerometer signals and machine learning,

    K.-C. Liu and Y. -D. Lin, “Efficient fall detection using Kalman filter -enhanced triaxial accelerometer signals and machine learning,” Biomed. Signal Process. Control, vol. 114, p. 109304, Apr. 2026, doi: 10.1016/j.bspc.2025.109304

    work page doi:10.1016/j.bspc.2025.109304 2026

  8. [8]

    CABMNet: An adaptive two-stage deep learning network for optimized spatial and temporal analysis in fall detection,

    V. Soni, H. Yadav, S. Bijrothiya, and V. B. Semwal, “CABMNet: An adaptive two-stage deep learning network for optimized spatial and temporal analysis in fall detection,” Biomed. Signal Process. Control, vol. 96, p. 106506, Oct. 2024, doi: 10.1016/j.bspc.2024.106506

    work page doi:10.1016/j.bspc.2024.106506 2024

  9. [9]

    Rodriguez, Katherine R

    B. Crook, M. Schlüter, and T. Speith, “Revisiting the Performance-Explainability Trade-Off in Explainable Artificial Intelligence (XAI),” in 2023 IEEE 31st International Requirements Engineering Conference Workshops (REW) , Se p. 2023, pp. 316 –324. doi: 10.1109/REW57809.2023.00060

    work page doi:10.1109/rew57809.2023.00060 2023

  10. [10]

    Concept Embedding Models: Beyond the Accuracy -Explainability Trade-Off,

    M. Espinosa Zarlenga et al., “Concept Embedding Models: Beyond the Accuracy -Explainability Trade-Off,” Adv. Neural Inf. Process. Syst., vol. 35, pp. 21400–21413, Dec. 2022

    work page 2022

  11. [11]

    Risk prediction models for falls in hospitalized older patients: a systematic review and meta-analysis,

    A. Mao, J. Su, M. Ren, S. Chen, and H. Zhang, “Risk prediction models for falls in hospitalized older patients: a systematic review and meta-analysis,” BMC Geriatr., vol. 25, p. 29, Jan. 2025, doi: 10.1186/s12877-025-05688-0

    work page doi:10.1186/s12877-025-05688-0 2025

  12. [12]

    Building Trustworthy Multimodal AI: A Review of Fairness, Transparency, and Ethics in Vision-Language Tasks,

    M. Saleh and A. Tabatabaei, “Building Trustworthy Multimodal AI: A Review of Fairness, Transparency, and Ethics in Vision-Language Tasks,” May 08, 2025, arXiv: arXiv:2504.13199. doi: 10.48550/arXiv.2504.13199

    work page doi:10.48550/arxiv.2504.13199 2025

  13. [13]

    Trustworthy A lzheimer’s diagnosis: Integrating robustness, fairness, and explainability in neuroimaging based deep ensemble framework,

    M. Bashir, N. Rahim, S. El -Sappagh, O. A. El -Serafy, and T. Abuhmed, “Trustworthy A lzheimer’s diagnosis: Integrating robustness, fairness, and explainability in neuroimaging based deep ensemble framework,” Eng. Appl. Artif. Intell. , vol. 172, p. 114291, May 2026, doi: 10.1016/j.engappai.2026.114291

    work page doi:10.1016/j.engappai.2026.114291 2026

  14. [14]

    Ethics of trustworthy AI in healthcare: Challenges, principles, and practical pathways,

    P. Ahadian, W. Xu, D. Liu, and Q. Guan, “Ethics of trustworthy AI in healthcare: Challenges, principles, and practical pathways,” Neurocomputing, vol. 661, p. 131942, Jan. 2026, doi: 10.1016/j.neucom.2025.131942

    work page doi:10.1016/j.neucom.2025.131942 2026

  15. [15]

    World guidelines for falls prevention and management for older adults: a global initiative,

    M. Montero-Odasso et al., “World guidelines for falls prevention and management for older adults: a global initiative,” Age Ageing, vol. 51, no. 9, Sep. 2022, doi: 10.1093/ageing/afac205

    work page doi:10.1093/ageing/afac205 2022

  16. [16]

    Summary of the Updated American Geriatrics Society/British Geriatrics Society Clinical Practice Guideline for Prevention of Falls in Older Persons

    “Summary of the Updated American Geriatrics Society/British Geriatrics Society Clinical Practice Guideline for Prevention of Falls in Older Persons”, Accessed: Apr. 11, 2026. [Online]. Available: https://agsjournals.onlinelibrary.wiley.com/doi/10.1111/j.1532 - 5415.2010.03234.x

    work page doi:10.1111/j.1532 2026

  17. [17]

    A review of video-based human activity recognition: theory, methods and applications,

    T. F. N. Bukht, H. Rahman, M. Shaheen, A. Algarni, N. A. Almujally, and A. Jalal, “A review of video-based human activity recognition: theory, methods and applications,” Multimed. Tools Appl., vol. 84, no. 17, pp. 18499–18545, May 2025, doi: 10.1007/s11042-024-19711- w

    work page doi:10.1007/s11042-024-19711- 2025

  18. [18]

    Spatial tem poral graph convolutional networks for skeleton -based action recognition,

    S. Yan, Y. Xiong, and D. Lin, “Spatial tem poral graph convolutional networks for skeleton -based action recognition,” in Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, in AA...

    work page doi:10.5555/3504035.3504947 2018

  19. [19]

    Two -Stream Adaptive Graph Convolutional Networks for Skeleton -Based Action Recognition,

    L. Shi, Y. Zhang, J. Cheng, and H. Lu, “Two -Stream Adaptive Graph Convolutional Networks for Skeleton -Based Action Recognition,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , Jun. 2019, pp. 12018 –12027. doi: 10.1109/CVPR.2019.01230

    work page doi:10.1109/cvpr.2019.01230 2019

  20. [20]

    High-resolution Image Synthesis with Latent Diffusion Models,

    H. Duan, Y. Zhao, K. Chen, D. Lin, and B. Dai, “Revisiting Skeleton -based Action Recognition,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2022, pp. 2959–2968. doi: 10.1109/CVPR52688.2022.00298

    work page doi:10.1109/cvpr52688.2022.00298 2022

  21. [21]

    Advancing skeleton-based human behavior recognition: multi-stream fusion spatiotemporal graph convolutional networks,

    F. Liu, C. Wang, Z. Tian, S. Du, and W. Zeng, “Advancing skeleton-based human behavior recognition: multi-stream fusion spatiotemporal graph convolutional networks,” Complex Intell. Syst., vol. 11, no. 1, p. 94, Dec. 2024, doi: 10.1007/s40747-024-01743-2

    work page doi:10.1007/s40747-024-01743-2 2024

  22. [22]

    Is space- time attention all you need for video understanding?

    G. Bertasius, H. Wang, and L. Torresani, “Is Space -Time Attention All You Need for Video Understanding?,” Jun. 09, 2021, arXiv: arXiv:2102.05095. doi: 10.48550/arXiv.2102.05095

    work page doi:10.48550/arxiv.2102.05095 2021

  23. [23]

    Explainable AI: A Brie f Survey on History, Research Areas, Approaches and Challenges,

    F. Xu, H. Uszkoreit, Y. Du, W. Fan, D. Zhao, and J. Zhu, “Explainable AI: A Brie f Survey on History, Research Areas, Approaches and Challenges,” in Natural Language Processing and Chinese Computing , J. Tang, M. -Y. Kan, D. Zhao, S. Li, and H. Zan, Eds., Cham: Springer International Publishing, 2019, pp. 563–574. doi: 10.1007/978-3-030-32236-6_51

    work page doi:10.1007/978-3-030-32236-6_51 2019

  24. [24]

    Explainable artifi- cial intelligence (XAI) in deep learning-based med- ical image analysis.Medical Image Analysis, 79: 102470, 2022

    B. H. M. van der Velden, H. J. Kuijf, K. G. A. Gilhuijs, and M. A. Viergever, “Explainable artificial intelligence (XAI) in d eep learning- based medical image analysis,” Med. Image Anal., vol. 79, p. 102470, Jul. 2022, doi: 10.1016/j.media.2022.102470

    work page doi:10.1016/j.media.2022.102470 2022

  25. [25]

    Improving deep learning performance by using Explainable Artifi cial Intelligence (XAI) approaches,

    V. Bento, M. Kohler, P. Diaz, L. Mendoza, and M. A. Pacheco, “Improving deep learning performance by using Explainable Artifi cial Intelligence (XAI) approaches,” Discov. Artif. Intell., vol. 1, no. 1, p. 9, Oct. 2021, doi: 10.1007/s44163-021-00008-y

    work page doi:10.1007/s44163-021-00008-y 2021

  26. [26]

    Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization,” presented at the Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 618–626. Accessed: Jan. 10, 2025. [Online]. Available: https://openaccess.thecvf.com/content_iccv_201...

    work page 2017

  27. [27]

    A Unified Approach to Interpreting Model Predictions,

    S. M. Lundberg and S. -I. Lee, “A Unified Approach to Interpreting Model Predictions,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2017. Accessed: Mar. 25, 2026. [Online]. Available: https://proceedings.neurips.cc/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html

    work page 2017

  28. [28]

    M.et al.Explainable ai for trees: From local explanations to global understanding.Nat

    S. M. Lundberg et al., “From local explanations to global understanding with explainable AI for trees,” Nat. Mach. Intell., vol. 2, no. 1, pp. 56–67, Jan. 2020, doi: 10.1038/s42256-019-0138-9

    work page doi:10.1038/s42256-019-0138-9 2020

  29. [29]

    A novel human action recognition using Grad-CAM visualization with gated recurrent units,

    M. Jayamohan and S. Yuvaraj, “A novel human action recognition using Grad-CAM visualization with gated recurrent units,” Neural Comput. Appl., vol. 37, no. 17, pp. 10835–10850, Jun. 2025, doi: 10.1007/s00521-025-10978-0

    work page doi:10.1007/s00521-025-10978-0 2025

  30. [30]

    Axiomatic Attribution for Deep Networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic Attribution for Deep Networks,” in Proceedings of the 34th International Conference on Machine Learning , PMLR, Jul. 2017, pp. 3319 –3328. Accessed: Mar. 25, 2026. [Online]. Available: https://proceedings.mlr.press/v70/sundararajan17a.html

    work page 2017

  31. [31]

    Ethical Dimensions of Using Artificial Intelligence in H ealth Care | Journal of Ethics | American Medical Association

    “Ethical Dimensions of Using Artificial Intelligence in H ealth Care | Journal of Ethics | American Medical Association.” Accessed: May 23, 2025. [Online]. Available: https://journalofethics.ama -assn.org/article/ethical-dimensions-using-artificial-intelligence-health- care/2019-02

    work page 2025

  32. [32]

    ‘Help Me Help the AI’: Understanding How Explainability Can Support Human-AI Interaction | Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems

    “‘Help Me Help the AI’: Understanding How Explainability Can Support Human-AI Interaction | Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems.” Accessed: Jul. 12, 2025. [Online]. Available: https://dl.acm.org/doi/full/10.1145/3544548.3581001

    work page doi:10.1145/3544548.3581001 2023

  33. [33]

    Toward Real -Time, Robust Wearable Sensor Fall Detection Using Deep Learning Methods: A Feasibility Study,

    H. Yhdego, C. Paolini, and M. Audette, “Toward Real -Time, Robust Wearable Sensor Fall Detection Using Deep Learning Methods: A Feasibility Study,” Appl. Sci., vol. 13, no. 8, p. 4988, Jan. 2023, doi: 10.3390/app13084988

    work page doi:10.3390/app13084988 2023

  34. [34]

    Multiscale Deep Feature Learning for Human Activity Recognition Using Wearable Sensors,

    Y. Tang, L. Zhang, F. Min, and J. He, “Multiscale Deep Feature Learning for Human Activity Recognition Using Wearable Sensors,” IEEE Trans. Ind. Electron., vol. 70, no. 2, pp. 2106–2116, Feb. 2023, doi: 10.1109/TIE.2022.3161812

    work page doi:10.1109/tie.2022.3161812 2023

  35. [35]

    Deep Learning Based Systems Developed for Fall Detection: A Review | IEEE Journals & Magazine | IEEE Xplore

    “Deep Learning Based Systems Developed for Fall Detection: A Review | IEEE Journals & Magazine | IEEE Xplore.” Accessed: Jun. 29,

    work page

  36. [36]

    Available: https://ieeexplore.ieee.org/document/9186685?denied=

    [Online]. Available: https://ieeexplore.ieee.org/document/9186685?denied=

    work page arXiv

  37. [37]

    Using Deep Neural Networks for Human Fall Detection Based on Pose Estimation,

    M. Salimi, J. J. M. Machado, and J. M. R. S. Tavares, “Using Deep Neural Networks for Human Fall Detection Based on Pose Estimation,” Sensors, vol. 22, no. 12, Art. no. 12, Jan. 2022, doi: 10.3390/s22124544

    work page doi:10.3390/s22124544 2022

  38. [38]

    Explainability in medical image captioning - ScienceDirect

    “Explainability in medical image captioning - ScienceDirect.” Accessed: Aug. 05, 2024. [Online]. Available: https://www.sciencedirect.com/science/article/abs/pii/B9780323960984000181

    work page 2024

  39. [39]

    Combining Convolutional Neural Network (CNN) and Grad -CAM for Parkinson’s Disease Prediction and Visual Explanation,

    R. Dehghan, M. Naderan, and S. E. Alavi, “Combining Convolutional Neural Network (CNN) and Grad -CAM for Parkinson’s Disease Prediction and Visual Explanation,” Eng. Manag. Soft Comput., vol. 10, no. 1, pp. 1–13, Sep. 2024, doi: 10.22091/jemsc.2024.10828.1180

    work page doi:10.22091/jemsc.2024.10828.1180 2024

  40. [40]

    NTU RGB+D: A Large Scale Dataset for 3D Human Activity Analysis

    A. Shahroudy, J. Liu, T. -T. Ng, and G. Wang, “NTU RGB+D: A Large Scale Dataset for 3D Human Activity Analysis,” Apr. 11, 2016, arXiv: arXiv:1604.02808. doi: 10.48550/arXiv.1604.02808

    work page Pith review doi:10.48550/arxiv.1604.02808 2016

  41. [41]

    Video capture of the circumstances of falls in elderly people residing in long-term care: an observational study,

    S. N. Robinovitch et al., “Video capture of the circumstances of falls in elderly people residing in long-term care: an observational study,” Lancet Lond. Engl., vol. 381, no. 9860, pp. 47–54, Jan. 2013, doi: 10.1016/S0140-6736(12)61263-X

    work page doi:10.1016/s0140-6736(12)61263-x 2013

  42. [42]

    Digital Filtering in the T ime Domain

    “Digital Filtering in the T ime Domain.” Accessed: Apr. 11, 2026. [Online]. Available: https://www.sciencedirect.com/science/chapter/monograph/abs/pii/B9781904275268500178

    work page 2026

  43. [43]

    Filtering Biomechanical Signals in Movement Analysis,

    F. Crenna, G. B. Rossi, and M. Berardengo, “Filtering Biomechanical Signals in Movement Analysis,” Sensors, vol. 21, no. 13, p. 4580, Jul. 2021, doi: 10.3390/s21134580

    work page doi:10.3390/s21134580 2021

  44. [44]

    1€ Filter: A Simple Speed -based Low-pass Filter for Noisy Input in Interactive Systems,

    G. Casiez, N. Roussel, and D. Vogel, “1€ Filter: A Simple Speed -based Low-pass Filter for Noisy Input in Interactive Systems,” Conf. Hum. Factors Comput. Syst. - Proc., May 2012, doi: 10.1145/2207676.2208639

    work page doi:10.1145/2207676.2208639 2012

  45. [45]

    A human activity recognition method based on Vision Transformer | Scientific Reports

    “A human activity recognition method based on Vision Transformer | Scientific Reports.” Accessed: Mar. 29, 2026. [Online]. Av ailable: https://www.nature.com/articles/s41598-024-65850-3

    work page 2026

  46. [46]

    An Interpretable Deep Learning -Based Feature Reduction in Video -Based Human Activity Recognition,

    M. Dutt, M. Goodwin, and C. W. Omlin , “An Interpretable Deep Learning -Based Feature Reduction in Video -Based Human Activity Recognition,” IEEE Access, vol. 12, pp. 187947–187963, 2024, doi: 10.1109/ACCESS.2024.3432776

    work page doi:10.1109/access.2024.3432776 2024

  47. [47]

    TCN-Inception: Temporal Convolutional Network and Inception modules for sensor-based Human Activity Recognition,

    M. A. A. Al-qaness, A. Dahou, N. T. Trouba, M. Abd Elaziz, and A. M. Helmi, “TCN-Inception: Temporal Convolutional Network and Inception modules for sensor-based Human Activity Recognition,” Future Gener. Comput. Syst., vol. 160, pp. 375 –388, Nov. 2024, doi: 10.1016/j.future.2024.06.016

    work page doi:10.1016/j.future.2024.06.016 2024

  48. [48]

    Predicting the probability for falls in community-dwelling older adults using the Timed Up & Go Test,

    A. Shumway-Cook, S. Brauer, and M. Woollacott, “Predicting the probability for falls in community-dwelling older adults using the Timed Up & Go Test,” Phys. Ther., vol. 80, no. 9, pp. 896–903, Sep. 2000

    work page 2000

  49. [49]

    (PDF) Quantitative Falls Risk Assessment Using the Timed Up and Go Test,

    “(PDF) Quantitative Falls Risk Assessment Using the Timed Up and Go Test,” ResearchGate, doi: 10.1109/TBME.2010.2083659

    work page doi:10.1109/tbme.2010.2083659 2010

  50. [50]

    Biomechanical analysis of hip and knee joints during gait in elderly subjects,

    R. N. Kirkwood, H. de A. Gomes, R. F. Sampaio, E. Culham, and P. Costigan, “Biomechanical analysis of hip and knee joints during gait in elderly subjects,” Acta Ortopédica Bras., vol. 15, pp. 267–271, 2007, doi: https://doi.org/10.1590/S1413-78522007000500007

    work page doi:10.1590/s1413-78522007000500007 2007

  51. [51]

    Muscle Mechanics,

    “Muscle Mechanics,” in Biomechanics and Motor Control of Human Movement , John Wiley & Sons, Ltd, 2009, pp. 224 –249. doi: 10.1002/9780470549148.ch9

    work page doi:10.1002/9780470549148.ch9 2009

  52. [52]

    Gait Analysis: Normal and Pathological Function,

    “Gait Analysis: Normal and Pathological Function,” J. Sports Sci. Med., vol. 9, no. 2, p. 353, Jun. 2010

    work page 2010

  53. [53]

    Human fall detection on embedded platform using depth maps and wireless accelerometer,

    B. Kwolek and M. Kepski, “Human fall detection on embedded platform using depth maps and wireless accelerometer,” Comput. Methods Programs Biomed., vol. 117, no. 3, pp. 489–501, Dec. 2014, doi: 10.1016/j.cmpb.2014.09.005

    work page doi:10.1016/j.cmpb.2014.09.005 2014

  54. [54]

    FallAllD: An Open Dataset of Human Falls and Activities of Daily Living for Classi cal and Deep Learning Applications,

    M. Saleh, M. Abbas, and R. B. Le Jeannès, “FallAllD: An Open Dataset of Human Falls and Activities of Daily Living for Classi cal and Deep Learning Applications,” IEEE Sens. J., vol. 21, no. 2, pp. 1849–1858, Jan. 2021, doi: 10.1109/JSEN.2020.3018335

    work page doi:10.1109/jsen.2020.3018335 2021