List of Papers Browse by Subject Areas Author List
Abstract
Video-based seizure detection is essential for the management of epilepsy patients, offering a non-invasive complement to electroencephalography. While several deep learning approaches have been developed for video-based seizure detection, none are inherently interpretable, limiting their adoption and translation into clinical practice. We present, to our knowledge, the first exploration of a neurosymbolic framework for video-based seizure detection that directly addresses this gap. Our approach (1) extracts patient-centric skeleton sequences from epilepsy monitoring units via a prompt-guided foundation model, (2) predicts seizure semiology clinically grounded spatio-temporal concepts, and (3) composes them via differentiable logic into interpretable Boolean rules with auditable contributions. Furthermore, to mitigate false positives arising from the traditional binary formulation (seizure vs.\ non-seizure), we sub-classify non-seizure segments into clinically relevant normal activities, providing the model with fine-grained discriminative supervision. Evaluated on two public seizure video benchmarks, our framework achieves 89.78% sensitivity with 0.06 false detections per hour on SAHZU and 85.27% / 0.09 on IEEE, while producing complete three-level interpretability: every prediction decomposes into which motor primitives were detected, how they were logically composed, and how much each rule contributed to the clinical decision. We publicly release all annotations, extracted pose sequences, our data pipeline and code. (https://gitfront.io/r/luffy/rELWxdnvy6q2/CDSD/)
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2068_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/2068_supp.zip
Link to the Code Repository
https://github.com/Mr-TalhaIlyas/CDSD
Link to the Dataset(s)
https://github.com/Mr-TalhaIlyas/CDSD
BibTex
@InProceedings{IlyTal_ANeurosymbolic_MICCAI2026,
author = { Ilyas, Talha AND Mehta, Deval AND Ge, Zongyuan},
title = { { A Neurosymbolic Framework for Interpretable Skeleton-Based Seizure Detection via Concept-Driven Logical Reasoning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of this paper is the introduction of an interpretability framework for the skeleton-based seizure detection task, along with its effective validation on two 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.
1.This paper yields a unified bank of 93 concepts, including 53 normal, 19 seizure spatial, and 21 shared temporal concepts. 2.Through a spatio-temporal encoder with a concept bottleneck, the proposed model learns differentiable Boolean rules, thereby forming an interpretable neurosymbolic framework.
- 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.The authors claim that leveraging the finer-grained annotation helped reduce the misclassification of movement-intensive non-seizure segments. However, although actions such as sitting, reading, and eating are movement-intensive, they do not exhibit the characteristic features of seizure episodes. The current fine-grained dataset is insufficient to support this hypothesis. 2.A more detailed description of the method is required, including which GCN is adopted in the spatio-temporal encoder, as well as the specific definitions of L_task and L_concept. 3.Ablation studies could be added to the experimental section to assess the individual contribution of each component. 4.Given that the method relies on accurate pose estimation, the authors should include an analysis of its performance in cases where the pose is severely occluded.
- 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?
Skeleton-based seizure detection methods have gained recognition due to non-contact nature. The authors innovatively introduce concept-driven logical reasoning, which significantly enhances the interpretability of skeleton-based seizure detection and and expands the opportunities for its clinical application.
- 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 proposes a novel integration of skeleton-based video modeling with neurosymbolic reasoning, where spatio-temporal motion concepts (grounded in ILAE semiology) are combined via differentiable Boolean logic layers to produce interpretable predictions. The approach achieves strong improvements over prior work on two benchmarks, particularly in reducing false detection rates while maintaining high sensitivity.
- 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 work is well-grounded in clinical practice (ILAE semiology), and directly addresses the key limitation of black-box models in medical AI. 2.The method achieves notable gains in sensitivity and especially false detection rate compared to prior work.
- 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.The evaluation is conducted on relatively small datasets (e.g., 14 patients in SAHZU), raising concerns about generalization to broader clinical settings. 2.The pipeline depends on several large models (SAM, Sapiens, VLMs, LLMs), which increases system complexity and may limit reproducibility and clinical deployment. 3.The proposed method appears to build heavily upon [7], but the paper does not sufficiently clarify the novel contributions. I recommend the authors to more explicitly clarify their contributions with respect to [7] both in the introduction section.
- 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 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?
See the weeknesses.
- 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 #3
- Please describe the contribution of the paper
This paper proposes a neurosymbolic framework for interpretable, video-based seizure detection. The framework leverages a prompt-guided foundation model to extract patient skeleton sequences, which are then processed through concept-driven logical reasoning to enable enhanced clinical interpretability for seizure detection. The work is evaluated on two public seizure video benchmarks. Additionally, it introduces a fine-grained, skeleton-based seizure dataset, contributing both a methodological advance and a new resource to the community.
- 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.Clinically Motivated Neurosymbolic Design: The integration of symbolic reasoning with deep learning is a principled and well-motivated choice for medical AI, where interpretability is critical. Unlike purely data-driven approaches, the neurosymbolic framework provides human-understandable decision pathways, which are particularly valuable for clinical adoption in epilepsy care. 2.Seizure Semiology and Localization Focus: Rather than treating the problem as a binary detection or coarse classification task, the paper engages with concepts of seizure semiology and localization that align directly with neurological clinical practice. This focus substantially enhances the clinical relevance and translational potential of the work. 3.False Positive Reduction via Discriminative Supervision: The paper addresses the critical challenge of false positives, a major practical barrier in seizure detection systems, through the use of normal activity subclasses that provide fine-grained discriminative supervision. This design choice reflects a thoughtful understanding of real-world deployment constraints. 4.Clear Conceptual Illustration: Figure 1 effectively communicates the overall framework and its contributions. It serves as an excellent entry point for readers and helps them understand the architecture at a glance. 5.Comprehensive Comparative Evaluation: The method is benchmarked against both unimodal and multimodal baselines, providing a thorough and fair assessment of its relative performance across different input settings. 6.Strong Ablation Study: Table 2 is particularly informative, offering a granular view of each component’s contribution to overall performance. This level of ablation clarity is commendable and supports the validity of each design choice. 7.Reproducibility: Code and data are provided via a well-organized repository, and the paper includes sufficient implementation details to support reproducibility. This is strongly appreciated and aligns with community best practices.
- 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.
Mentioned in additional comments.
- 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 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
1.Undefined Notation in Section 2.2 (Page 4): The binary membership variable M∈{0,1} of dimension ∣A∣×∣C∣ is introduced without defining ∣A∣. Since ∣C∣ refers to the number of concepts, |A| should be explicitly defined (e.g., number of anatomical joints or action primitives) at first use. This omission undermines notational clarity in a technically dense section. 2.Unexplained Variable in Section 2.3 (Page 6): The variable r is introduced without prior definition. Overall, Section 2.3 progressively becomes harder to follow. 3.Underperformance of Multimodal Variants (Table 1, Page 7): Table 1 shows that the multimodal approach underperforms relative to the skeleton-only modality model, which is counterintuitive given that additional modalities should, in principle, provide complementary information. The authors should provide at least 1-2 sentences in the results section discussing potential reasons. 4.Parameter Count Inconsistency (Table 1): HyperGCN [31] is reported with 2.4M parameters, yet the proposed variant (HyperGCN + Logic Layer + Multi-label) shows only 2.3M parameters. Adding components to a base model should not reduce the total parameter count. The authors should verify and clarify this discrepancy. 5.Incomplete Ablation Discussion (Table 2b): While the ablation study in Table 2b includes a discussion of [28], no comment is provided for [10], despite it appearing in the same comparison. For completeness and fairness, the limitation of [10] relative to the proposed method should be briefly addressed. 6.Absence of Failure Case Analysis: The paper does not discuss or illustrate cases in which the proposed method fails. An analysis of failure modes, e.g., ambiguous postures, occlusion, and atypical seizure presentations, is important for understanding the method’s limitations and guiding future work. Even a brief qualitative discussion or a dedicated figure would strengthen the paper.
- 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 makes a meaningful and clinically motivated contribution by introducing a neurosymbolic framework for interpretable seizure detection, with key strengths including its semiotic focus, false-positive reduction strategy, and reproducibility. However, several issues temper the score: unexplained notation reduces readability in key technical sections, a parameter count inconsistency in Table 1 raises reporting concerns, the under performance of the multimodal variant is left unexplained, and the absence of failure case analysis limits the depth of evaluation. These are addressable concerns, and if resolved, the paper would be a solid accept.
- 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
Author Feedback
We thank the AC and reviewers for the careful, constructive feedback. AC concerns are addressed inline via the reviewer responses below. Meta-Reviewer (AC).
- Small datasets / failure-mode analysis: Please see R2.W1 and R3.W6 / R1.W4.+ Reliance on large foundation models: Please see R2.W2.Reviewer 1. W1: Table 2(a) supports the claim binary to 9-class multi-label gives +6.47 sensitivity and 0.21→0.06 FDR/h (3.5×) on the same encoder; Fig. 3(c) t-SNE shows the resulting cluster separation. Ablation prose now states this attribution. W2: §2.3 names the adaptive Hyper-GCN [zhou2025adaptive]; L_concept is given in closed form (BCE between ĉ and c*=M[a,:]) after Eq. (1); L_task is the multi-class CE between ŷ and one-hot a ∈ A. W3: Table 2 already isolates four design choices (NTU pretraining; binary vs. multi-label; tracking; concept composition; latency); the surrounding prose now attributes each gain to its component. Further depth deferred to a journal extension. W4: SAM 3’s text-prompted segmentation with streaming memory (§2.1) maintains patient identity under transient occlusion. Severe bedding occlusion (≳60% body coverage), where Sapiens-2B keypoints destabilise, remains the dominant failure mode; due to the strict page limit, a dedicated failure-mode discussion is deferred to a journal extension. Reviewer 2. W1: Cohort size is a field-wide constraint, video-EEG seizure recordings are highly sensitive patient data and SAHZU/IEEE are among the only public sets. To partly mitigate cohort overfitting, IEEE results are reported under 5-fold patient-level CV (no patient appears in both training and test). Multi-centre prospective validation, contingent on dataset availability, is listed as future work. W2: The runtime is a 2.3M-parameter Hyper-GCN with logic layers; SAM 3, Sapiens-2B, and Qwen3-VL are used offline (one-time pose extraction and label refinement), not in the inference path. We have made this explicit at the end of §2.1.Reproducibility is supported by the public code/data release. W3: [7] (Gao et al., MICCAI 2025) itself builds on Wang et al. [26]; our work also builds on [26], extending it from static medical-image classification to spatio-temporal skeleton-based seizure understanding via (i) ILAE-grounded spatial concepts (vs. radiological findings); (ii) a shared spatio-temporal vocabulary capturing motion dynamics, absent in static-image CBMs; (iii) logic layers coupled with a fine-grained multi-label EMU formulation that our ablations show drives the 3.5× FDR reduction. Now stated explicitly in the Introduction. Reviewer 3. W1: A is defined at first use in §2.2 as the set of action classes (paired with the |A|×|C| association matrix M). W2: r is now introduced as the rule-activation vector r ∈ {0,1}^R, R=N_L, with each r_k indicating whether the k-th learned rule fires, plus a worked example. W3: The under-performance arises from EMU-specific appearance noise (bedding deformation, illumination, ambient staff motion) acting as label noise during fusion under limited cross-modal training data; the skeleton modality isolates patient kinematics and offers higher SNR. Due to the strict page limit, this discussion was kept brief in the Results section. W4: Hyper-GCN’s 2.4M includes its 121-class softmax head; we remove this head and add a compact D→|C| concept projection plus heavily ℓ₁-regularised binary logic layers totalling only 233K parameters, so the head savings outweigh the logic-layer overhead The fine-tuning paragraph in §3 now states this; both numbers re-verified from checkpoints. W5: A sentence in the ablation addresses [10]: Ilyas et al. reach 85.84% sensitivity but with 10× higher FDR/h (0.65 vs. 0.06), as their OF + lightweight pose front-end is sensitive to bedding motion under partial occlusion. W6: A sentence in the qualitative paragraph identifies severe bedding occlusion (≳60% body coverage) and atypical absence seizures as the dominant failure modes. Please see R1W4 (char limit).
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.
Paper summary: This paper studies the skeleton-based seizure detection task and proposes a novel and clinically relevant application of neurosymbolic AI to seizure detection, a topic of significant interest to the MICCAI community working at the intersection of interpretability and clinical neurology. The method is evaluated on two public datasets and compared with both unimodal and multimodal baselines, providing a thorough and fair assessment of its relative performance across different input settings.
Strengths The reviewers identified several strengths, including: Strength 1: The work is well-grounded in clinical practice (ILAE semiology) and directly addresses the key limitation of black-box models in medical AI. Strength 2: Strong empirical validation and benchmark results. The proposed method achieves notable gains in sensitivity and especially in false detection rate compared to prior work. and Strength 3: Relevance to MICCAI, as well as reproducibility and theoretical grounding. I broadly agree that the paper’s main contribution is significant.
Weaknesses: The main concerns raised by reviewers include: The evaluation is conducted on relatively small datasets, and there is no analysis of failure modes. The proposed approach also heavily depends on large models, raising concerns about increasing system complexity, which may limit reproducibility and clinical deployment.
Review Reconciliation: Reviewer opinions are largely consistent.
Decision: The paper is technically sound, addresses an important problem of significant interest to the MICCAI community working at the intersection of interpretability and clinical neurology in the context of skeleton-based seizure detection, and presents a meaningful contribution beyond existing work. The empirical validation is convincing, and no major technical weaknesses were identified that would materially affect acceptance. I therefore recommend Early Accept.