List of Papers Browse by Subject Areas Author List
Abstract
Precise decoding of intended joint movements from brain
activity is essential for next-generation neurorehabilitation interfaces, yet most motor imagery studies focus on single joints or coarse movement categories rather than systematically evaluating all upper-limb joints. We present a joint-aware decoding framework that systematically quantifies joint-level separability across eight upper-limb motor imagery tasks through binary pairwise classification in a unified machine learning framework. All preprocessing, feature selection, dimensionality reduction, and hyperparameter optimization were performed
within subject-wise nested cross-validation to eliminate any risk of data leakage. Evaluation on a public dataset of eighteen participants yielded consistent performance across all joint pairs with mean accuracies ranging from 83% to 88% (SD = 0.04). Logistic regression (LR) and linear discriminant analysis (LDA) achieved the strongest and most stable performance. Statistical testing using the Wilcoxon signed rank test confirmed that linear discriminant analysis significantly outperformed the baseline model with p < 0.001. These findings establish a systematic and
reproducible framework for assessing fine-grained upper-limb joint separability in fNIRS-based motor imagery decoding, enabling interpretable and clinically meaningful evaluation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5943_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
FineMI dataset (fNIRS and EEG motor imagery): https://figshare.com/articles/dataset/Data/24123303
BibTex
@InProceedings{NosTay_JointLevel_MICCAI2026,
author = { Nosheen, Tayyaba AND Jurkojć, Jacek AND Wodarski, Piotr AND Jawed, Soyiba AND Malik, Aamir Saeed},
title = { { Joint-Level Motor Imagery Decoding in fNIRS via a Unified Pairwise Machine Learning Framework } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The primary contribution of this work is the proposal of TSCIMLMI, a rigorously validated, unified machine learning framework designed to systematically decode joint-level motor imagery (MI) from fNIRS signals. Addressing the limitations of previous studies that often focus on coarse paradigms like MI versus rest, this paper introduces a structured pairwise binary classification approach that transforms eight upper-limb joint tasks into 28 distinct binary decoding problems. The framework integrates a rich set of time, spectral, and HbO/HbR coupling features within a subject-wise nested cross-validation pipeline to eliminate data leakage and ensure robust evaluation across all joint pairs. Through a comprehensive benchmark of six classifiers under identical conditions, the authors demonstrate that linear discriminant analysis (LDA) consistently achieves accuracies exceeding 83%, significantly surpassing a simple baseline. Statistical tests confirm the robustness of these gains across subjects, while SHAP analysis reveals the physiological basis of the discrimination, indicating that frequency-specific deoxygenation dynamics and coordinated oxygenation–deoxygenation interactions drive joint separability. Ultimately, this work paves the way for more granular, clinically meaningful BCI systems for neurorehabilitation by establishing a systematic and reproducible foundation for assessing fine-grained upper-limb joint separability.
- 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.
One major strength of this work is the systematic and exhaustive evaluation of joint-level separability. By decomposing the problem into all 28 possible pairwise combinations of eight upper-limb joints, the authors provide a granular view of decoding difficulty that is often overlooked in broader MI studies. This approach allows for a more nuanced understanding of which specific joint movements are intrinsically harder to distinguish via fNIRS, particularly highlighting anatomical overlaps such as shoulder movements. Additionally, the methodological rigor is commendable; the use of subject-wise nested cross-validation ensures that preprocessing, feature selection, and hyperparameter optimization do not lead to data leakage, a common pitfall in BCI research. The inclusion of SHAP value analysis further strengthens the paper by offering interpretability, revealing that spectral components and HbO–HbR coupling features are more discriminative than static amplitude features alone. Another significant strength is the substantial performance improvement over the baseline. The proposed framework achieves an accuracy improvement of approximately 22 to 34 percentage points compared to a conventional SVM baseline using standard features. This demonstrates that the multi-domain feature representation effectively captures complementary signatures of motor imagery, enhancing the signal-to-noise ratio for fine-grained decoding. The statistical validation using Wilcoxon signed-rank tests confirms that the performance gains are consistent across subjects.
- 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.
A notable weakness is the lack of clinical validation despite the stated motivation of neurorehabilitation for stroke patients. The introduction emphasizes the burden of stroke and the potential for neurorehabilitation, yet the experiments are conducted exclusively on healthy participants. This limits the immediate clinical feasibility and generalizability of the findings, as hemodynamic responses in stroke patients can differ significantly due to neurovascular coupling deficits. While the public dataset used is suited for methodological benchmarking, the absence of patient data leaves a gap regarding clinical generalizability.
The application of the framework is systematic; however, the core algorithmic components are not novel. The feature extraction methods (temporal statistics, FFT, Pearson correlation) and classifiers (LDA, SVM, Random Forest, etc. ) are standard tools in the BCI literature, as seen in prior works such as recent MI-fNIRS-BCI reviews [1] and other existing motor imagery decoding studies [2]. The novelty lies primarily in the unified evaluation protocol rather than the development of new mathematical models or deep learning architectures.
Although this reliance on established methods is a minor weakness in terms of algorithmic novelty, the comprehensive benchmarking effort is highly valuable for establishing reproducible performance baselines in this area.
References: [1] R. Finnis, A. Mehmood, H. Holle, und J. Iqbal, „Exploring Imagined Movement for Brain–Computer Interface Control: An fNIRS and EEG Review“, Brain Sci. , Bd. 15, Nr. 9, S. 1013, Sep. 2025, doi: 10.3390/brainsci15091013. [2] A. Vavoulis, P. Figueiredo, und A. Vourvopoulos, „A Review of Online Classification Performance in Motor Imagery-Based Brain–Computer Interfaces for Stroke Neurorehabilitation“, Signals, Bd. 4, Nr. 1, S. 73–86, März 2023, doi: 10.3390/signals4010004.
- Please rate the clarity and organization of this paper
Good
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The recommendation to Accept is based on the paper’s solid methodological contribution to the field of fNIRS-based BCI, specifically in the area of fine-grained motor imagery decoding. The systematic evaluation of 28 joint-pair tasks provides valuable insights into the intrinsic separability of upper-limb movements that are not available in coarse-grained studies. The rigorous use of nested cross-validation and the demonstration of significant performance improvements over a conventional baseline establish the validity of the proposed multi-domain feature framework. While the lack of patient data and the incremental nature of the machine learning components prevent a higher score, the work addresses a clear gap in the literature regarding structured evaluation of joint-level MI. The clarity of the presentation and the use of a public dataset further support the decision, making it a valuable addition to the conference proceedings for researchers interested in neurorehabilitation interfaces.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
This paper presents a rigorously validated framework for fine-grained motor imagery decoding from fNIRS signals, addressing a clear gap in the literature regarding systematic evaluation of joint-level upper-limb separability. The strength of this work lies in its exhaustive pairwise benchmarking across 28 joint combinations and SHAP-based interpretability revealing physiologically meaningful spectral and coupling features as key discriminators. The authors appropriately acknowledge that clinical validation with stroke patients and multimodal integration (e.g., EEG–fNIRS fusion) represent important future directions that would expand beyond the current scope of this fine-grained motor imagery evaluation framework, which I agree is a reasonable characterization of the work’s boundaries. In addition, the authors have addressed concerns about variance reporting and hard-to-classify pairs, confirming that performance variability and pair-specific difficulties are well-characterized in the manuscript. The paper’s methodological rigor, clear presentation, and reproducible pipeline make it a valuable addition to MICCAI 2026 for researchers developing neurorehabilitation BCI systems, and I therefore maintain my recommendation for acceptance.
Review #2
- Please describe the contribution of the paper
This paper points out that existing motor imagery studies are limited to single joints or coarse movement classifications, lacking a systematic evaluation of the entire upper limb joints. To address this, the authors propose a framework for fine-grained and joint-level decoding. The study introduces an evaluation approach that reconstructs 8 upper limb joint motor imagery tasks into 28 binary classification pairs and builds a robust pipeline for fNIRS signals using multi-domain feature extraction. The proposed method evaluates the performance of various machine learning models on a single public dataset. Furthermore, SHAP analysis demonstrates that spectral and neurovascular coupling dynamics features play a key role in fine-grained movement classification.
- 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.
1.The reliability of the classification results is secured by conducting comparative performance analyses across multiple machine learning models. 2.The paper reformulates the conventional problem of single-joint and coarse movement classification, proposing a new fine-grained decoding evaluation method for the upper limb joints. 3.The SHAP analysis identifies the key features contributing to the fine-grained movement classification, providing interpretability.
- 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.
1.While fixed feature extraction benefits interpretability, an experimental analysis of which feature combinations yield optimal performance, or the specific rationale for selecting these features, is missing. 2.Although various conventional machine learning models are compared, the academic implications of the results are unclear. If the LDA model achieves the highest performance, a detailed analysis based on its classification results should be provided. 3.Although the classification performance metrics for each of the 28 binary pairs are presented, an in-depth analysis of these results is insufficient. For a systematic evaluation of the entire upper limb joints, specific discussions on the implications of individual outcomes such as the reasons for performance variations across specific joint pairs are required. 4.The evaluation relies on a single dataset, which limits the confirmation of the generalization capability and universal applicability of the proposed pipeline. 5.Performance evaluation is conducted using within-subject 5-fold cross-validation, but there is no analysis on whether there are differences in important features across subjects. 6.A deeper discussion is required on what the identified important features specifically signify in terms of fine-grained movement classification and physiological mechanisms. 7.For fine-grained feature analysis, it is necessary to focus on the classification performance between highly correlated pairs rather than simply presenting the performance of all pairs. The current focus on average performance weakens the connection to the paper’s main objective of fine-grained decoding.
- Please rate the clarity and organization of this paper
Good
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the originality of the proposed performance evaluation framework is acknowledged, its specific role and contribution are unclear. For fine-grained feature analysis, it is necessary to focus on binary classification and important feature analysis specifically for hard-to-classify joint pairs. Due to remaining concerns regarding the validity of the current experimental design and analysis methodology, I have assigned the current evaluation score.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
After reviewing the authors’ rebuttal, I maintain my original recommendation of Weak Reject. The authors explicitly position their main contribution as a unified, leakage-controlled evaluation framework for fine-grained fNIRS motor imagery decoding. Since the framework itself is the core contribution, its validity must be demonstrated, and internal methodological consistency alone is insufficient. Establishing this validity requires substantive analysis such as whether the performance differences across the 28 joint pairs reflect genuine neurophysiological representations, and whether the identified features align with known fNIRS biomarkers. However, the authors attribute the hardest pairs to anatomical overlap and proximal to distal interactions, but this claim lacks neurophysiological grounding given that wrist and shoulder joints occupy clearly distinct regions of the motor homunculus. The narrow accuracy range of 83 to 88 percent across all 28 pairs may in fact suggest that the framework is capturing a general motor imagery signal rather than fine-grained joint-level representations, yet this alternative interpretation is not addressed. Most importantly, no comparison is provided between the identified features and known fNIRS findings, leaving it unclear whether the framework genuinely captures joint-level neural representations or merely fits classifiers to variance in the data. As the analyses required to support the framework’s validity are not provided in the rebuttal, I maintain my recommendation of Weak Reject.
Review #3
- Please describe the contribution of the paper
The authors introduce a framework for using channel x channel correlation of fNIRS time series acquired for eight upper-limb joints to do pairwise binary decoding of 28 joint-pairs.
- 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.
-
As far as I can tell, there isn’t a comparable joint-pair decoding framework; this novelty adds a different perspective to the solution space. The framework’s pairwise discriminability is neurophysiologically meaningful.
-
Good use of signal processing for time series-based classification while being limited to one dataset. I don’t think there’s an exactly comparable one for the task the authors are interested in solving.
-
- 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.
-
I am confused as to why the authors didn’t use the EEG signals in the Yi et. al (2025) dataset for a doing a comparative analysis; it seems like an accessible and logical next step for this paper instead of leaving that for future work.
-
Related to the above, the methods compared for the framework are insufficient as they’re not providing validation or evaluation of the framework itself. It would make the results more meaningful had the authors explored EEG signals as well, or EEG+fNIRS together. The combined spatio-temporal resolution from EEG+fNIRS would make up for the relative delay between neural activity and hemodynamic responses as measured by fNIRS.
-
The authors mention 22-34% improvement over baseline, but I am not convinced that the best performing models are not overfit; the mean classification shown in Table 1 doesn’t mention variance, and fNIRS is known to be variable across subjects. The Wilcoxon signed-rank test and rank-biserial correlation of -1 imply that every subject improved over the baseline (which is chance), but by how much? It’s unclear what the efficacy of joint-pair decoding is if there’s no indication of generalization to held-out subject-wise data.
-
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
I think there is sufficient novelty in the approach, but I am not convinced it has been validated thoroughly enough.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
This paper is still a weak accept for me because while I appreciate the novelty of the framework, I would like to see proof that these results generalize beyond the sample, and that the models are not overfit, e.g., the paper could report per-subject performance and the distribution across subjects.
Author Feedback
We thank the Meta-Reviewer and Reviewers for their constructive feedback and for recognizing the methodological rigor and relevance of our work. We address the key concerns regarding scope, feature rationale, pair-specific interpretation, validation, and generalization. Contribution and scope (AC, R1, R2, R5). Our contribution is not a new classifier, but a unified, leakage-controlled evaluation framework for fine-grained fNIRS motor imagery. By reformulating eight upper-limb joint imagery classes into all 28 binary joint-pair tasks, we provide a systematic benchmark for intrinsic joint-level separability under identical preprocessing, feature extraction, selection, and classifier evaluation. This directly addresses the gap in coarse MI paradigms and clarifies the role of the framework. Feature rationale and LDA interpretation (R2, R5). The multi-domain feature bank was selected a priori to ensure physiological interpretability: temporal/statistical features capture hemodynamic magnitude and shape, spectral features isolate frequency-specific dynamics within the slow fNIRS band, and HbO–HbR coupling features capture oxygenation–deoxygenation interactions. Correlation pruning and ANOVA ranking were executed strictly within training folds, ensuring no feature selection utilized held-out test data. The strong performance of linear models (LDA/LR) indicates that the proposed representation yields largely linearly separable joint-pair structure, consistent with the small per-subject sample regime. Hard-to-classify pairs and physiological insight (R2, R5). As shown in Table 1 and Fig. 4, performance varies across joint pairs, with the most challenging cases consistently involving proximal shoulder-related movements. In particular, pairs such as WFE–SAA and SPS–SAA exhibit the lowest LDA accuracies (83%) and higher off-diagonal confusion (16–17%), with additional difficulty observed in WAA–SPS, EPS–SAA, and EPS–SFE. These pairs reflect anatomically overlapping or functionally coupled movements and mixed proximal–distal interactions, which plausibly produce less separable hemodynamic responses. In contrast, pairs involving more distinct motor representations (e.g., HOC–SFE, SAA–SFE) show higher separability. Importantly, SHAP analysis indicates that discriminability is driven primarily by spectral HbR dynamics and HbO–HbR coupling rather than static amplitude features, supporting a physiological interpretation based on dynamic neurovascular signatures. Validation, variance, and overfitting (R1, R2, R5). The pipeline enforces subject-wise nested cross-validation, with all feature selection, redundancy removal, and hyperparameter tuning confined to training folds. The dataset includes 18 subjects (320 trials each). Performance is consistent across all 28 pairs (83–88%, SD = 0.04), and Fig. 2 reports variability. The Wilcoxon signed-rank test (W = 0, z = −3.72, p < 0.001) confirms consistent improvement over baseline across subjects. Single dataset, patient data, and EEG (R1, R5). We agree that clinical validation with stroke patients and cross-dataset validation are vital next steps. Regarding the concurrent EEG data in the Yi et al. dataset, incorporating EEG or EEG–fNIRS fusion introduces a distinct multimodal problem and would fundamentally shift the central scope of the paper. We consider multimodal fusion, cross-subject transfer, and patient validation as future work. Reproducibility (R2). The manuscript specifies the dataset, preprocessing, MBLL conversion, feature construction, train-fold-only selection, classifiers, and statistical testing. We will make the complete source code publicly available upon acceptance in accordance with MICCAI policies.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
The paper presents a solution to classify single upper-limb joint (imagined) movement from motor imagery in fNIRS signal with potential applications to stroke rehabilitation by means of BCI.
The solution is built from 8 original classes corresponding to upper-limb joints encoded as 28 binary classification problems. The results could misleadingly look modest (83-88% acc.) in the era of AI-fuelled pipelines, but I’m still impress especially because this is built with good old fashion tools that we know inside out (e.g. LDA, rnadom forest, etc), so interpretable as hell -although just in case, the authors throw SHAP on top (nice touch!) and solved with simplistic elegance. And I may not have been the only one to be favourably impressed; there is a majority of reviewers leaning to acceptance.
In a nutshell, this paper makes excellent use of signal processing and time series classification (R3/AC), old-fashion performance comparative analysis (R2), as well as reasonable use of hypothesis testing and inferential stats (R1), resulting in a solid pipeline and experimentation -methodological rigour (R1/AC)-.
The lack of clinical validation (R1) and the missed opportunity of further analysis (R2) especially when EEG signals appear to be available as well (R3), together with the fact that all individual algorithmic pieces are old thus limited technical novelty (R1), have been listed among the weaknesses.
The decision to invite for a rebuttal is only because of the split decision among reviewers, but otherwise, I’m keen on accepting this one.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
I was somewhat enthusiastic about this paper in my initial meta-review, but unexpectedlyly, the reaction to the authors’ response was less so, with R2 hardening the score to Reject -and I must say with an insightful comment about how the narrow accuracy may reflect the capture of a general motor imagery signal rather than fine-grained joint-level representations. Anyhow, the split in the appreciation by the reviewers remain which tips the balance toward rejection at this stage.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers raised concerns regarding the limited algorithmic novelty, reliance on a single dataset, absence of clinical validation, and the depth of the physiological interpretation of the identified features. The rebuttal addressed these concerns appropriately by clarifying that the primary contribution is a unified, leakage-controlled evaluation framework for fine-grained joint-level motor imagery decoding rather than the introduction of a new classifier architecture.
The paper demonstrates strong methodological rigor through subject-wise nested cross-validation, train-fold-only feature selection, comprehensive evaluation across all 28 joint pairs, statistical validation, and SHAP-based interpretability analysis. The authors provided a clearer rationale for the selected feature representations and discussed the physiological interpretation of hard-to-classify joint pairs. The paper fills an important gap in the systematic evaluation of fine-grained motor imagery decoding using fNIRS and provides a reproducible and well-validated benchmark framework.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have provided detailed justifications to the reviewer concerns. It would be suggested to change the title in order to better focus on the actual paper contribution.