List of Papers Browse by Subject Areas Author List
Abstract
Understanding the role of the brain-body axis in the causal relationship between psychiatric disorders and endocrine disorders is crucial for advancing multi-organ disease modeling and precision medicine. Despite the importance of understanding these interactions, the development of a mature framework has been hindered by traditional models’ inability to move beyond mere statistical associations to uncover true causal directionality in high-dimensional, population-scale data. In this study, we proposed a graph neural network (GNN)-based framework to construct causal graphs capturing interactions among brain imaging features, organ-level physiological indicators, and disease outcomes. Leveraging the directed acyclic graph (DAG)-GNN model, we perform causal structure learning on the comprehensive UK Biobank cohort, enabling the discovery of DAGs that reflect potential causal dependencies. Experimental results demonstrate that our method effectively uncovers interpretable brain-body-disease pathways, revealing shared white-matter vulnerabilities across distinct disorders and highlighting how intermediate physiological systems mediate these interactions, providing insights into psychiatric and endocrine conditions. Our code is available at https://github.com/NIMGroup/UKBCausal.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0603_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/NIMGroup/UKBCausal
Link to the Dataset(s)
UK Biobank Resource (controlled-access biomedical database; used under Application No. 419012): https://www.ukbiobank.ac.uk/
BibTex
@InProceedings{WanZho_Unveiling_MICCAI2026,
author = { Wan, Zhonghua AND Zhang, Chengzhe AND Xie, Yu AND He, Yifei AND You, Qingwei AND Wang, Peng AND Qin, Jiaolong AND Zhang, Huifeng AND Wu, Ye},
title = { { Unveiling Brain-Body Axis Interactions in Psychiatric-Endocrine Disorder with Deep Graph Causal Neural Network } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper addresses an important problem with an interesting technical framework. The integration of MAE for missing data, PCA for dimensionality alignment, and DAG-GNN for structure learning is well-motivated. However, the causal claims are substantially overstated given the cross-sectional observational design, and the lack of any quantitative validation of the learned causal graph is a significant gap. With more careful framing of what the method can and cannot establish, this could become a strong contribution.
- 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.
This paper proposes a GNN-based causal discovery framework to model directed interactions among 27 white matter tracts, 5 organ systems, and 11 psychiatric/endocrine disorders using UK Biobank data. The authors adopted a masked autoencoder handles missing organ-level features, PCA compresses brain dMRI features, and DAG-GNN learns a directed acyclic graph representing putative causal relationships. They report interpretable brain-body-disease pathways, including convergent white matter vulnerabilities and mediated systemic cascades.
The brain-body axis is becoming important more and more for both psychiatric and endocrine diseases. The paper addresses a genuine gap: most prior work examines brain-body links in isolated pairs or uses purely correlational methods. Modeling the full multi-organ, multi-disease interaction network simultaneously is a valuable direction for precision psychiatry and computational medicine. The organ-specific masked autoencoder is a brilient solution to use with UK Biobank analyses. Training on complete cases with synthetic masking, reconstructing only masked entries, and using the learned latent space (rather than imputed values) for downstream analysis is methodologically cleaner than standard mean imputation or deletion approaches. The comparison against Mean+PCA and SoftImpute baselines provides empirical support. The paper carefully ensures that brain (PCA), organ (MAE latent), and disease (embedding) nodes share a consistent feature dimensionality before graph construction. This is a necessary and non-trivial design choice.
- 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.
Since the entire framework learns from cross-sectional data, argument of causal structure is severely limited. The authors should clearly mention this limitation, and future plan or suggestion for followup study to overcome this. Importantly, due to same reason, the “causal directions” should be carefully re-addressed, otherwise it is well-validated (and known) direction. The discovered DAG is not validated against any external ground truth, alternative causal discovery methods (PC algorithm, GES, FCI, NOTEARS), or known causal relationships from randomized trials or Mendelian randomization studies. Without any form of validation. Although the UKB provides relatively comprehensive data, the sample is not still representative for our species. So, if external validation is not feasible, clearly mention limitation to prevent the over-interpretation of the finding.
Minor issues
- The sample size reduction from 501,000 to 11,628 (due to requiring both imaging and complete organ data) may introduce selection bias that is not discussed.
- The threshold values for edge filtering (0.5 and 0.2) appear arbitrary; sensitivity analysis would be informative.
- 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.
(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 approach is novel, the data and analysis is straightforward. But need some ‘tone down’ to prevent potential ‘causal’ misunderstanding.
- 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 unified causal-discovery pipeline to jointly model brain microstructure, organ-level physiology, and psychiatric/endocrine disease outcomes from the UK Biobank.
The pipeline has three stages: 1- Masked autoencoder (MAE) that learns a compact latent representation robust to missing physiological entries. 2- PCA-based alignment step that projects each brain ROI’s high-dimensional dMRI features to the same latent dimensionality (d = 3) as the organ embeddings, so brain, organ, and disease nodes can share a unified node-feature tensor. 3- DAG-GNN variational graph model with a NOTEARS-style differentiable acyclicity constraint and an augmented-Lagrangian objective that learns a weighted adjacency matrix over 27 JHU white-matter tracts, 5 organ systems, and 11 diseases.
The application contribution is a post-hoc inspection of the learned DAG at two edge thresholds (w > 0.5, w > 0.2), describing: 1- Convergent transdiagnostic vulnerability of the cingulum of the cingulate gyrus and the posterior limb of the internal capsule. 2- Mediated cascades such as overweight → metabolic → cingulum and hypothyroidism → depression → superior cerebellar peduncle.
- 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 author proposed that a directed representation of joint brain–microstructure, organ, and disease relations is an underexplored and clinically meaningful direction, rather than relying on pairwise correlational brain–body studies.
2- The study leverages a large-scale, multimodal dataset such as the UK Biobank, combining diffusion metrics across 27 JHU white matter bundles with organ-level phenotypes spanning musculoskeletal, immune, metabolic, renal, and hepatic systems, along with 11 disease labels derived from ICD-10 and self-report records.
3- Rather than using mean or low-rank imputation, the author used a masked autoencoder for organ-level representation learning to handle missing or zero values.
4- The author shared an anonymized code link (https://anonymous.4open.science/r/UKBCausal) for reproducibility.
- 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 paper does not propose a new loss, architecture, or theoretical result. The contribution is largely a pipeline assembly, indicating limited methodological novelty.
2- The author used random thresholds for edges (w > 0.5 for the most essential and w > 0.2 for the second most important) with no sensitivity analysis or justification for these cutoffs.
3- The author did not take into account basic confounders such as age, sex, and head motion. Could you please justify this?
4- Temporal misalignment between brain and organ phenotypes (critical): Section 2.3 explicitly states that brain dMRI features are taken from UK Biobank imaging instance 2.In contrast, Section 2.1 does not specify which instance the organ-level phenotypes (HbA1c, C-reactive protein levels, creatinine, albumin, etc.) come from. Most UK Biobank data for these variables are available only in instances 0 and 1, not the imaging instance. e.g: https://biobank.ndph.ox.ac.uk/showcase/field.cgi?id=30600
In the UK Biobank dataset, there are four instances:
- Instance 0: Initial assessment visit
- Instance 1: First repeat assessment visit
- Instance 2: Imaging visit
- Instance 3: First repeat imaging visit
5- The author reports a sample size of 11,628 subjects, but does not provide detailed sample sizes for individual phenotypes after filtering and intersection, making it difficult to assess the effective sample size.
6- The graph is trained and tested on the same dataset (UK Biobank). It would be beneficial to evaluate it on an external cohort as well.
7- All three figures have only one-line titles instead of proper captions. Additionally, in Figure 2 (right), the y-axis is labeled “Number of ROIs (CSV files)”; what does “CSV files” mean in this context?
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an interesting integrative question, and the UK Biobank application is ambitious; however:
1- The methodological contribution is a straightforward pipeline of published components (MAE, PCA, and DAG-GNN). 2- The causal graph is evaluated only qualitatively, without structural metrics, stability analysis, baselines, or confounder control. 3- Brain and organ phenotypes are selected from different UK Biobank visit instances, which structurally undermines the interpretation of causal directionality.
With quantitative DAG evaluation, a clarified disease-node specification, and instance-aligned phenotypes, this could become a clear 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
Review #3
- Please describe the contribution of the paper
The paper presents a causal graph learning framework to uncover interactions between brain white-matter microstructure, body organ systems, and psychiatric diseases using large-scale observational data.
The motivation is significant since most of analyses mainly identify statistical associations, while this paper focuses on a casual relationship which is crucial for mechanistic insights into psychiatric disorders.
- 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.
It is an interesting interdisciplinary problem where it integrates brain imaging, organ physiology and disease. It uses large-scale dataset like UK Biobank to provide strong statistical power. The idea of integrating brain + body + disease nodes is novel and it can handle missing organ features y using masked autoencoder. It identifies shared white-matter vulnerability hubs for potential biological insights.
- 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.Causal interpretation is weak. The paper claims causal discovery, but data is purely observational cross-sectional. No temporal ordering was mentioned. No mention of potential confouding variables like age, sex. No interventin or instrumental variables were used. DAG-GNN may infer statistical correlation, but interpreting edges as causal mechanism is risky. The authors should clarify assumptions (causual sufficiency, no hidden confouders)and discuss limitations more carefully.
2.Evaluation is weak. The paper mainly evaluates the reconstruction error of MAE, but no quantiative validation of causal structure. It worths doing some prediction evalution or even do some statistical enrichment anlaysis or replication in another dataset. It also lacks comparison with other causal models.
- 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?
The causal validity and evaluation needs to be further improved.
- 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 #4
- Please describe the contribution of the paper
This paper proposes a DAG-GNN-based framework that integrates a masked autoencoder (MAE) for missing value imputation and PCA for dimensionality alignment to uncover interpretable causal brain-body-disease interactions from incomplete observational data.
- 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 tackles a clinically important problem: causal inference in brain body axis interactions. Unlike traditional correlation based studies, this work explicitly models directed dependencies. The combination of MAE for missing data, PCA for feature alignment, and DAG GNN for graph learning is sensible and technically sound.
- 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.Experimental details are insufficient for reproducibility. “Repeated 5 times” is not specified (e.g., 5-fold cross-validation or 5 random splits). The MAE training mask ratio is not reported, while only the test-time 30% masking is described. Key sample statistics, including total participant count, disease-positive numbers per condition, and final organ-specific sample sizes after 70% filtering, are missing from the method section. 2.Figure 2 has poor readability: font sizes are too small, only a subset of the 27 ROIs are shown without selection criteria, and the y-axis label of the third subplot contains a “CSV” artifact, suggesting the figure was generated without proper post-processing. 3.Sensitivity analyses are lacking. There is no investigation of different MAE training mask ratios, no justification for choosing the 70% observed feature threshold for subject filtering, and no explanation for setting the edge weight thresholds (w > 0.5 for “most essential” and w > 0.2 for “second essential”) in the causal graph. 4.The baseline methods “Mean+PCA” and “SoftImpute” are mentioned without any description of their implementation details or the rationale for their inclusion, making the comparison less transparent.
- 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?
The core idea and framework are novel and clinically promising. However, the paper requires major clarifications in experimental protocols, sensitivity analyses, and figure quality to ensure reproducibility and scientific rigor.
- Reviewer confidence
Somewhat confident (2)
- [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 reviewers for their constructive feedback, which has improved the rigor and clarity of our work.
Response to R#1: We agree that the causal interpretation from observational UKB data should be treated with substantial caution. We will tone down strong causal wording and describe the learned DAG as candidate-directed dependencies/hypothesis-generating brain–body–disease pathways, rather than as confirmed biological causality. Since ground-truth brain–body causal graphs are unavailable, longitudinal cohorts, MR, interventional studies, and independent datasets are needed for confirmation. We will also discuss potential UKB selection bias arising from cohort overlap.
Response to R#2: We agree that methodological contribution is more accurately framed as an integrated, missingness-aware brain–body–disease framework rather than a new loss, architecture, or theory. Importantly, the integration is not a sequential assembly: MAE, ROI-wise PCA, and DAG-GNN are coupled to build modality-aligned node features for one heterogeneous graph, enabling joint brain–organ–disease structure learning [1,8,23]. Regarding confounders and UKB instances, age, sex, head motion, and temporal differences are important limitations. Organ biomarkers were taken from instance 0, while dMRI features were from imaging instance 2; thus, although organ measures generally precede imaging, edges should not be interpreted as definitive temporal or interventional causal effects. Instead, they represent candidate-directed dependencies requiring further validation. Phenotype-specific sample sizes, filtering steps, and effective sample counts should be reported more transparently, and external-cohort validation remains an important area for future work.
Response to R#3: We agree that the causal interpretation from cross-sectional observational data should be handled cautiously. DAG-based structure learning relies on assumptions such as causal sufficiency, the Markov condition, and faithfulness [23], while hidden confounding from age, sex, head motion, medication, lifestyle, and selection effects may remain. Thus, the learned edges should be interpreted as candidate-directed dependencies for hypothesis generation rather than validated causal mechanisms. The MAE reconstruction primarily evaluates the missing-data representation module rather than the causal graph itself; this limitation should be stated more clearly to avoid overclaiming graph validity. Additional prediction tasks, enrichment analysis, external replication, and comparisons with other causal discovery methods would be valuable avenues for future work to validate the inferred structure quantitatively.
Response to R#4: We agree with R#4 that several experimental and reproducibility details require clarification. Repeated 5 times means five independent random masking trials for held-out reconstruction evaluation. MAE training used random masking ratios from 10% to 40%, while 30% was used for evaluation as the most common organ-missingness level. The 70% observed-feature threshold was used to balance organ-feature reliability and retained sample size [12]. The mask protocol, disease counts, and organ-specific sample sizes should be stated more explicitly. Our initial bootstrap analysis showed stable recovery of the main pathways, suggesting robustness beyond a single split or threshold. These results were omitted for space because they are supporting rather than central evidence, but can be included in the camera-ready version. Mean+PCA and SoftImpute were included as common linear and low-rank missing-data baselines; their implementation and rationale will be clarified.
Minor issues: Figure/caption/ROI issues and the “CSV files” artifact will be corrected in the camera-ready. The thresholds w>0.5 and w>0.2 were used solely to visualize the learned continuous adjacency matrix, and threshold-dependent findings will be interpreted with greater caution in the discussion.
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 study proposes a method that probes brain-body association with a wide array of non-imaging phenotypes, validates on a large-scale dataset, and aims to provide mechanistic insights beyond simple association. These are major strengths, timely, and will be of great interests to medical AI researchers. However, reviewers also raise concerns on missing details on methodological and experimental design. A major concern is that whether it is possible to better validate the derived causal structure.
- 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’ concerns were mainly about clarity of methodological and experimental design, which were large resolved in the rebuttal and will likely be resolved in the final version. The causal claim indeed needs to use caution, but findings are still causally insightful as long as the authors can clarify assumptions and when the causal claims can become invalid. I believe MICCAI should put more emphasis on multi-modal medical data investigation beyond imaging and single-organ analysis, so I believe the study will potentially become a valuable asset to the community.
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.
After considering the reviews and rebuttal, I recommend acceptance. The paper addresses an interesting brain–body–disease interaction problem using UK Biobank data, and proposes a useful integrated framework combining missingness-aware organ representation learning, imaging feature alignment, and DAG-based graph modeling.
The rebuttal appropriately acknowledges that the causal interpretation should be toned down, and that the learned graph should be framed as hypothesis-generating candidate directed dependencies rather than validated causal mechanisms. My acceptance recommendation assumes that the final version will clearly reflect this more cautious framing.
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.
Three of four reviewers originally assigned Weak Accept, while one reviewer (R2) rejected the submission during the review phase based on concerns about methodological depth, comparison to existing causal discovery methods, and insufficient ablation studies. No reviewer submitted a post-rebuttal opinion, and specifically, the absence of a post-rebuttal opinion from R2 means I cannot assess whether the authors’ clarifications were satisfying. From what I can see, the rebuttal was substantive and specifically addressed several of R2’s points, including clarifying the causal modelling assumptions and the distinction between their approach and standard GNNs.
