Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Balance function assessment is vital for proactive health monitoring and fall prevention. However, conventional clinical methods are heavily influenced by subjective judgment, leading to potential inconsistencies in evaluation. While the rapid development of wearable technology has provided a significant research hotspot for objective monitoring, current methods rarely identify stability states through activity semantics. Furthermore, the lack of publicly available datasets containing true imbalance states severely hinders the development of robust assessment models. To address these challenges, we introduce MBFAD, a multi-modal benchmark encompassing a ternary paradigm of static, dynamic, and reactive balance activities. MBFAD collects synchronized whole-body kinematics and plantar pressure data from 29 subjects and induces true imbalance states using an environmental perturbation scheme. Based on this, we propose the M^2-Balance Framework, a synergistic multi-task architecture for joint activity recognition and stability analysis. Within M^2-Balance, the HSConv module captures multi-scale features through recursive convolution to provide semantic priors, while the CSConv module utilizes an expected channel damage matrix to mitigate high-dimensional redundancy and identify stability-sensitive channels. Experimental results on SisFall, PDS and MBFAD datasets demonstrate that M^2-Balance outperforms existing models, establishing a fundamental paradigm for modeling synergistic activities and stability in the Balance function assessment domain. The code and dataset will be released at https://github.com/huaingyingqin/MBFAD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5325_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/huaingyingqin/MBFAD

Link to the Dataset(s)

SisFall: https://github.com/BIng2325/SisFall/releases/download/dataset/SisFall.zip PDS: https://figshare.com/ndownloader/articles/4525082/versions/1 MBFAD: https://github.com/huaingyingqin/MBFAD

BibTex

@InProceedings{GeZha_MBFAD_MICCAI2026,
        author = { Ge, Zhaoyang AND Qin, Huaiying AND Lei, Bingchao AND He, Ziyang AND Ma, Jingzhe AND Xu, Mingliang},
        title = { { MBFAD: Benchmarking Balance Function Assessment with Environmental Perturbation and Synergistic Tasks } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    • This manuscript addresses balance function assessment using wearable multimodal sensing, motivated by limitations of subjective clinical evaluation and by the lack of public datasets with authentic instability events. The dataset and problem formulation are distinctive, while the model is a competent integration of existing ideas with moderate methodological novelty.

    • The strongest contribution is the benchmarking and data-collection design. The authors introduce MBFAD, a multimodal benchmark comprising synchronized whole-body kinematics and plantar pressure data from 29 subjects across 14 activities under both normal and perturbation conditions. This stands as a new dataset production, an original way to use data.

    • Also, the authors propose the M2-Balance, a framework that uses human activity recognition as a prior for balance assessment. Prior work treats activity recognition, fall detection, and balance analysis as separate pipelines, whereas the authors fuse this information.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    • MBFAD combines whole-body kinematics, plantar pressure, reactive tasks, and perturbation-based acquisition in a single dataset. Table 1 compares MBFAD with prior datasets and shows that MBFAD uniquely includes kinematics, pressure, reactive activities, and perturbation together. The acquisition setup is also clearly illustrated in Figure 1.

    • Using activity recognition as a prior for balance assessment is sensible because stability risk is context-dependent across activities. The paper explicitly formulates this distinction in Section 3.1.It operationalizes it through parallel HAR and BFA streams in Section 3.2, along with an architecture diagram, making the design easy to follow (although some details are missing).

    • The comparison against multiple baselines on SisFall and PDS includes modality and task ablations on MBFAD in Tables 2 to 4.This is good experimental practice and strengthens the argument that both multimodality and joint learning matter.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    • The most substantial contribution to the field is the MBFAD, which contains 29 subjects, all young adults around 20 to 22 years old. The motivation of the paper is about fall prevention and balance health. The absence of older adults or clinical populations is a serious limitation for translational claims. The data may still be useful as a technical benchmark, but the current cohort is not representative of the target populations most at risk.

    • The paper presents only in Figure 2, the necessary data preprocessing steps (coordinate alignment, timestamp, sliding window, etc). None of these steps is explicitly presented. The details needed to reproduce the results are not provided, so the explanation of data collection and processing is limited. Also, no description of the training and inference details (optimizer, epochs, learning rate, etc. ). There are no details in the data or the splits (are they the same as in [22]?)

    • It is not clearly explained how the balance function labels, especially the high/ low risk labels, were assigned. The authors mentioned that, with two rehabilitation experts and without a precise label-generation protocol, it is difficult to fully assess whether the labels are strong enough to support future clinical claims.

    • The conclusion discusses the broader implications of moving toward context-aware stability analysis and suggests future value in this direction. However, the paper does not address key limitations, including a young, small cohort, a lack of clinical validation, and insufficient evidence for generalization to real patient populations.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    • Explicitley clarify how the BFA labels were defined. What determines high risk versus low risk for each window or sample, especially under perturbation?

    • Clarify the split protocol. Was the evaluation subject-independent?

    • Please report the preprocessing settings explicitly. Figure 2 shows these stages conceptually, but not enough detail for replication.

    • Report the full training details, including the optimizer, batch size, learning rate, number of epochs, early stopping policy, and the values of alpha and beta in the joint loss.

    • Explain how p-values were computed in Tables 2 to 4.

    • Please The paper should evaluate older adults or clinical populations, or at least discuss why results from young adults should transfer.

    • H ∈ R S×L -> L is never explicitly defined (the interpretation is that each modality feature h(s) is an L-dimensional vector, where L is the feature length).

    • Similar comment to F ∈ RH×W ×C;

    • Minor typos/inconsistencies: Introduction: “quality of life [14] . the primary”; “enter of Pressure data , while Wu et al. [24]”; “3.2 Overview The M2-Balance Framework pipeline (Fig. 2)”

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The recommendation is justified by the enumerated points, both in the major strengths and limitations.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper’s main contribution is the introduction of the MBFAD benchmark, a multimodal dataset for balance and fall assessment structured around static, dynamic, and reactive activities, along with a novel sensory perturbation setup (air cushion) to simulate balance disturbances. In addition, the authors propose the M²-Balance Framework, a multitask model that jointly learns human activity recognition and balance assessment, and demonstrate its effectiveness on both the proposed dataset and existing public datasets.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper presents several notable strengths. First, it introduces a novel multimodal benchmark (MBFAD) for balance and fall assessment that explicitly models static, dynamic, and reactive activity paradigms, which is more comprehensive than many existing datasets. Second, the incorporation of sensory perturbation via an air cushion is an original idea that enables controlled simulation of balance disturbances, providing a more realistic experimental setup compared to standard recordings. Third, the proposed M²-Balance Framework leverages a shared encoder for jointly learning human activity recognition (HAR) and balance/fall assessment (BFA), which is a meaningful multitask formulation that can improve representation learning. Finally, the authors validate their approach on both the collected dataset and public benchmarks, demonstrating consistent improvements over baseline methods, which supports the generalizability of the approach.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The paper has several important weaknesses. First, the evaluation protocol is insufficiently described and potentially flawed, as it is unclear whether subject-independent validation (e.g., leave-one-subject-out or subject-wise splits) is used, raising concerns about possible data leakage. Second, the reported results appear to rely mainly on sliding window-level metrics, without providing sequence-, trial-, or subject-level evaluation, which limits practical relevance. Third, although the paper reports p-values, the statistical analysis is not properly described. The authors do not specify which statistical tests were used, whether assumptions were checked, or how multiple comparisons were handled, making it difficult to assess the validity of the claimed significance. Finally, the clinical applicability is limited, as the dataset appears to be collected primarily from younger adults, while fall risk assessment is most relevant for geriatric populations. The lack of validation on elderly or clinical cohorts reduces the translational impact of the work.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents an interesting idea with a novel benchmark and the use of sensory perturbation for balance assessment. However, the decision is mainly driven by concerns regarding evaluation rigor and clinical relevance. The evaluation protocol is unclear, particularly with respect to subject-independent validation, raising potential data leakage issues. Results are primarily reported at the sliding window level, without sequence- or subject-level analysis, which limits their real-world applicability. In addition, although p-values are reported, the statistical methodology is not sufficiently described, making it difficult to assess the validity of the findings. Finally, the lack of validation on geriatric or clinical populations weakens the clinical impact, which is critical for a MICCAI submission. Overall, while promising, the work requires stronger experimental design and validation to meet MICCAI standards.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    Authors collect a new dataset (29 subjects, 14 activities) for balance function assessment and for human activity recognition. Also, authors propose a new AI model, called M^2-balance for (presumeably) 14-class classification for activity recognition, and binary classification for balance function assessment.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The newly acquired dataset MBFAD that consists of data of 29 subjects for 14 activities, using 16 IMU sensors on the body and 32 in-sole sensors for each shoe, is a good contribution. In addition, authors also implement and evaluate a new AI model, called M^2-balance for (presumably) 14-class classification for activity recognition, and binary classification for balance function assessment.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    This reviewer has the following comments:

    • Consider additional age groups, of middle age people and possibly elderly
    • Authors may want to define/describe all 14 activities done by each of the 29 subjects, under berg or brunel scale methods. They should especially explain the sensory perturbation-based motions and reactive motion.
    • In section I, authors state that they have collected kinematics data from 59 body locations, while in section II, they state that they have used 16 9-axis IMU sensors for kinematics data collection. Is each IMU sensor at a given location measuring more than one body organ/element? Please reconcile both statements.
    • The variables L, H,W,C represent the dimensions of the feature maps, but are not defined in Eqs. 1,2.- In section 3.3, authors make mention of G groups? Is G=14, corresponding to 14 different activities that subjects do? Or, does it mean something else?
    • Authors may want to explain the rationale/reasoning behind Eq. 2.In order words, why did they decide to split the intermediate feature map into two maps and then to concatenate the first half O_1^1 with the next group F^{i+1}?
    • Is F^i from Eq. 2 and F_n from Eq. 3 the same? If so, then better to use a consistent notation. Or, are they different?
    • What is N in Eq. 3? What is the motivation for Eq. 3 that shows us the absolute difference of two convolutions (of the feature map with the kernel), that is then averaged over N?
    • Eq. 6: Authors may want to do an ablation on parameters alpha and beta, to learn more about their individual contributions. Also, authors may want to define the two loss functions.
    • Overall, the mathematical description of the proposed AI model (M^2-balance) needs improvement in terms of rigor.
    • Authors seem to have fully omitted the key data pre-processing details which is something very important to better understand the contribution of this work, and for the reproducibility of this work. They show a glimpse of their pre-processing pipeline in Fig. 2, but do not talk about it in the paper at all.
    • Similarly, authors make mention of the 2 feature extractors MotionNet and PressureNet in Fig. 2, but again, they don’t mention important details such as backbone architecture of these 2 encoders in the paper at all.
    • Authors present their results in results section, but without specifying the precise details of the problem being solved. That is, for activity recognition problem, they probably have solved a 14-class classification problem (as hinted in Fig. 2). If that is the case, they need to mention it in the results section (Table 2) with clarity.
    • Table 2: Authors do not describe their data splitting methodology, and of the related works. Thus, it is not possible to infer something meaningful out of these results unless and until we get to know the data splitting methodology for all the works (related works, and authors’ work). And as authors may already know, the leave some subjects out cross validation methodology is the preferred data splitting method, as otherwise there will be data leakage between train and test sets.
    • For the BFA task, authors report the results in table 3, 4.However, I would like to know how authors have computed the accuracy values for this task. I mean, Fig. 2 tells me that authors solve a binary classification problem here (low risk, high risk). How do authors compute the fall risk to be high or low, for a given subject? I would also like to know the data splitting methodology for this task as well.
    • What is the fate of the collected MBFAD dataset? Will it be made public for the benefit of fellow researchers? Overall, the paper presents an interesting idea, i.e., a new MBFAD dataset and M^2-balance AI model. However, there is an ambiguity regarding the results reported in the results section, because authors have not mentioned the data splitting methodology, as well as the key data pre-processing details. Also, the mathematical explanation of proposed M^2-balance framework requires polishing and improvement. Finally, it is not clear whether authors plan to release their collected dataset to the public.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Overall, the paper presents an interesting idea, i.e., a new MBFAD dataset and M^2-balance AI model. However, there is an ambiguity regarding the results reported in the results section, because authors have not mentioned the data splitting methodology, as well as the key data pre-processing details. Also, the mathematical explanation of proposed M^2-balance framework requires polishing and improvement. Finally, it is not clear whether authors plan to release their collected dataset to the public.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

Responses to Editors and Reviewers: We would like to sincerely thank all reviewers for their thoughtful and constructive feedback on our paper titled “MBFAD: Benchmarking Balance Function Assessment with Environmental Perturbation and Synergistic Tasks.” We greatly appreciate the time and effort these experts have devoted to evaluating our work. We have carefully read and considered all suggestions and comments. While the paper has been provisional accepted, we recognize that these valuable suggestions are of significant importance to the quality of our paper, the rigor of our work, and its clinical application. Specifically:

To Meta-Review: We sincerely thank you for your thorough and insightful evaluation of our work. We highly value the points you and the reviewers have raised, and we will carefully address each of them in the camera-ready version.

To Reviewer 1: Thank you for recognizing the innovativeness of the dataset and framework of this study. Regarding population representativeness, we recognize that our study is currently limited to younger individuals; future investigations will seek to incorporate older participants, provided that ethical standards and safety considerations are fully satisfied. We will briefly supplement the final version with hyperparameters, preprocessing, and the protocol for independent subject partitioning. Also, thank you for pointing out the errors in the label definition; we will clarify in the camera-ready version that BFA labels are defined as steady-state labels caused by supporting surface conditions. Finally, thank you for your review; we will supplement the p-value test explanation and mathematical symbol definitions, and correct formatting errors.

To Reviewer #2: Thank you for your valuable comments. Your comments on preventing data leakage are very insightful. We will further explain the use of independent participant segmentation in the manuscript. We also agree that subject level indicators are meaningful for clinical interpretation. In this study, the reported indicators are computed at the window level because the analysis was performed on short temporal segments of imbalance responses. We will make this evaluation setting clearer in the manuscript and acknowledge participant-level assessment as an important consideration for future clinical validation. We will revise the manuscript to clarify this rationale and discuss subject level assessment as an important direction for future clinical validation. Furthermore, thank you for your suggestions on statistical rigor and clinical applicability; we will revise the p-value annotations in the tables and add testing plans for the elderly population. We will also release relevant data and source code in the camera-ready version.

To Reviewer #3: Thank you for your detailed feedback and constructive suggestions. Regarding the discrepancies in the activity description and sensor quantity representation in the manuscript, due to space limitations, we will briefly mention the activity definition and clarify the calculation logic of the 59 kinematic nodes in the final version. Thank you for your requirement for technical rigor; for omissions in mathematical description and architectural details, we will supplement the relevant definitions and motivations in the text. Regarding the ambiguity in the presentation of results, we will revise the title of Table 2 and clarify the accuracy calculation protocol. Finally, thank you for your suggestions regarding the diversity of participants. We plan to expand to include middle-aged and elderly individuals in the future, and will open-source the data and code during the camera-ready version.

We once again express our sincerest gratitude for the valuable feedback from all reviewers, which will help us improve the quality of the paper and further refine and expand our work in the future.

With best regards

All Authors



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    I recommend an Accept based on the consensus of the reviewers, who all provided positive evaluations. The reviewers unanimously praised the introduction of the MBFAD benchmark, noting its novel integration of synchronized whole-body kinematics, plantar pressure, and reactive activities, along with the original use of sensory perturbation via an air cushion to simulate authentic balance disturbances. Furthermore, the proposed M²-Balance framework was commended for its synergistic, multi-task approach that jointly learns human activity recognition and balance assessment to improve representation learning.

    While the core contributions are strong, the reviewers raised several methodological and clinical points that must be addressed in the camera-ready version. The authors must clarify the evaluation protocol, specifically detailing whether subject-independent data splitting was employed to prevent data leakage between the training and testing sets. They also need to enhance reproducibility by including thorough descriptions of the data preprocessing steps, training hyperparameters, and the specific backbone architectures of the feature extractors. Additionally, the mathematical descriptions require refinement to explicitly define all variables, such as L, H, W, C, G, and N , and clearly explain the rationale behind the feature map splitting and the joint loss function parameters. From a clinical perspective, the authors must acknowledge the dataset’s limitation of comprising primarily young adults and discuss the generalizability and translational applicability of these results to older, at-risk geriatric cohorts. Finally, the manuscript must explicitly clarify the protocol used by the rehabilitation experts to assign the high-risk and low-risk balance function labels and detail the specific statistical tests utilized to compute the reported p-values. Addressing these clarifications will significantly strengthen the final manuscript.



back to top