List of Papers Browse by Subject Areas Author List
Abstract
Accurate prediction of clinical symptoms and cognitive deficits from functional connectivity is essential for advancing precision psychiatry in schizophrenia (SZ).
However, predicting these clinical and cognitive measures from brain functional connectivity remains challenging. A key barrier is that age-related brain changes substantially overlap with SZ pathology, predisposing models to exploit age-correlated variance rather than genuine pathological signals. At the same time, hard removal of age signals degrades performance on cognitive tasks where aging and pathology interact. To this end, we propose a Latent Pathology-Age Decoupling (LPAD) framework that enforces explicit latent constraints to decouple age-related variations from pathological features. The framework introduces a dual-stream architecture with pathology-age orthogonal decoupling and supervised contrastive regularization to suppress age-driven shortcuts, followed by a clinically-informed gated modulation mechanism that reintroduces age as a subordinate contextual reference. Experiments demonstrate that LPAD achieves performance improvements across all eight clinical symptom and cognitive deficit tasks, with external validation suggesting cross-site generalizability, and linear probe analysis and spatially complementary region-of-interest profiles between the two streams corroborating effective pathology-age decoupling. Code is available at https://github.com/SCSE-Biomedical-Computing-Group/LPAD.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1376_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/SCSE-Biomedical-Computing-Group/LPAD
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanYig_Latent_MICCAI2026,
author = { Yan, Yige AND Chan, Yi Hao AND Cheng, Jun AND Chew, Qian Hui AND Sim, Kang AND Rajapakse, Jagath C.},
title = { { Latent Pathology–Age Decoupling in Functional Connectivity Graphs for Multi-task Schizophrenia Assessment } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a novel Latent Pathology–Age Decoupling (LPAD) framework to improve prediction of clinical symptoms and cognitive deficits in schizophrenia using functional connectivity (FC) data. LPAD explicitly disentangles age-related brain changes from pathological features through a dual-stream graph encoding with orthogonal latent constraints and supervised contrastive regularization. It then selectively reintroduces age as a gated contextual modulator for aging-sensitive tasks. The approach is validated on two datasets, showing consistent performance improvements across eight clinical and cognitive prediction tasks, with external validation demonstrating cross-site generalizability. Interpretability analyses further support effective pathology-age separation.
- 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.
- Introduces a novel and principled framework (LPAD) addressing a key challenge in schizophrenia FC analysis: confounding between age and pathology.
- Clinically-informed gated modulation (CIGM) thoughtfully reintroduces age as a contextual factor, preserving aging-sensitive cognitive signals.
- Extensive experiments across multiple symptom and cognitive domains demonstrate statistically significant improvements over strong baselines.
- External validation on an independent clinical dataset supports cross-site generalizability, which is critical for clinical translation.
- Comprehensive interpretability analyses, including linear probes and spatial ROI mapping, provide biological plausibility and insight into the pathology-age decoupling.
- 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 impact of preprocessing choices (e. g. , denoising, parcellation scheme) on model performance is not clearly discussed, limiting reproducibility and understanding of robustness.
- Sample size is relatively small.
- The study design and dataset usage, particularly which dataset serves as external validation, could be described more clearly to improve transparency.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
Overall, this work presents a well-motivated and technically rigorous approach to disentangling age and pathology effects in schizophrenia functional connectivity analysis. The LPAD framework’s novel latent decoupling strategy combined with clinically-informed gating demonstrates strong empirical performance and cross-site validation, supported by meaningful interpretability analyses.
- 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 #2
- Please describe the contribution of the paper
The paper proposes a Latent-Pathology-Age Decoupling (LPAD) framework that aims to disentangle age-related features from pathology-related features via an orthogonality constraint. The disentangled age representation is then reintroduced in a controlled manner through a Clinically-Informed Gated Modulation module to predict schizophrenia clinical symptoms.
- 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.
Main strength of the paper is the explicit reuse of age-related features after disentanglement, rather than discarding them. Additionally, the overall orchestration of the modules appears coherent, suggesting a well-designed pipeline.
- 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.
There are several concerns, mainly related to the clarity and completeness of methodological descriptions:
1 ) The description of the GPSConv-GINE layer is unclear and requires substantial clarification. Unlike widely established models such as GCN or GAT, this combination is not standard and cannot be assumed to be familiar to the reader. The GPS model is essentially a Graph Transformer framework that integrates local message passing with global attention, along with various encoding strategies while GINE model extends GIN by incorporating edge features into the aggregation process. It is not clear from the paper whether these two models are combined, and if so, how the integration is implemented, which components of the GPS framework are actually used and the motivation for choosing a Graph Transformer-based approach.
2 ) The baseline models lack sufficient explanation. For example, the abbreviation “MT” is introduced without definition. Some relevant prior works on disentangling age as a covariate (e.g., cited in [33]) are mentioned in the text but not included in the main comparison table. This makes it difficult to assess the fairness and completeness of the experimental comparison.
3 ) The paper states that two datasets are used, but it is not clear whether the datasets are merged and if merged, what are differences acquisition protocols etc..
- Please rate the clarity and organization of this paper
Poor
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
Overall, while the idea of disentangling and reusing age-related features is promising and well-motivated, the current presentation lacks sufficient clarity in key methodological components and experimental setup. Addressing these issues would significantly improve the paper’s reproducibility and impact.
- 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
This paper addresses the problem that age-related brain changes overlap with SZ pathology, and proposes a framework for clinical and cognitive prediction that enforces explicit latent constraints to decouple age-related variations from pathological features, and reintroduces age as a subordinate contextual reference through a clinically-informed gated modulation.
- 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 paper tackles the age confounding issue in schizophrenia, which is very relevant. The idea of removing age first and then reintroducing it as context is intuitive and aligns well with clinical reasoning. 2.The ablation study is thorough and shows each part actually helps. 3.The Interpretability analysis is thorough.
- 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 cosine orthogonality penalty only removes linear dependencies, whereas the aging process is inherently nonlinear. It is therefore possible that z^p still encodes age-related information or z^a encodes pathological information in a nonlinear form that this constraint cannot capture. 2.The node-level scalar gate g_i is computed solely from h^p. However, after being constrained by L_ortho, h^p is supposed to be free of age-related information, so how can it reliably decide how much contextual information from v_i should be injected? 3.The linear probe shows that z^p no longer contains age-related information, but there is no corresponding evidence that z^a is free of SZ-related pathological features.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
well-motivated, solid experiments.
- 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 reviewers and AC. We accept the meta-reviewer’s two limitations—no theoretical guarantee of nonlinear age removal, and limited sample size—and will state both in the camera-ready.
Nonlinear decoupling and bidirectionality (R3-1, R3-3). L_ortho enforces only linear orthogonality and cannot guarantee removal of nonlinear age–pathology coupling. The linear probe shows z^p does not predict age (R²=-0.176) but does not symmetrically test whether z^a predicts pathology. Both will be listed as limitations, with bidirectional probing and nonlinear extensions as future work. Existing analyses still jointly support bidirectional separation: (1) Fig. 3(b) shows zero overlap (Jaccard=0.000) between top-10% ROIs; age-stream ROIs concentrate in subcortical/temporal regions consistent with normal aging, while pathology-stream ROIs anchor in SZ-related DMN—if z^a retained pathological information, its ROIs would more likely overlap with SZ-related DMN. (2) z^a is driven by explicit age regression supervision, which with L_ortho exerts directional pressure toward age representation. (3) Fig. 3(a) shows clear t-SNE cluster separation, indicating separation persists in nonlinear spaces.
Gating on h^p (R3-2). The asymmetric gate g_i=σ(w_g^T h_i^p) is intentional. Under L_ortho the pathology branch suppresses age-related shortcuts while retaining pathology-severity information. The gate does not recover age from h^p; it determines, based on pathological representation, whether a node needs age-context modulation. Clinically, age–SZ pathology interactions are more pronounced in pathology-affected regions [20,8], so stronger-pathology nodes need more age modulation. Table 2 confirms this: removing CIGM yields the largest cognitive drop on WM (Δcorr=0.023), the task most sensitive to aging–pathology interactions.
GPSConv-GINE (R2-1). We will clarify in Sec. 2.1.GPSConv [25] is a hybrid graph transformer layer combining local message-passing with global attention. We instantiate the local module with GINE [12,32], which naturally exploits continuous FC edge weights w. GPSConv-GINE is thus a GPSConv layer using GINE as its local MPNN, motivated by GINE preserving continuous FC weights (often weakened in GCN/GAT) and GPSConv’s global attention modeling cross-network long-range dependencies disrupted in SZ.
Dataset usage (R2-3). Datasets are not merged. COBRE (n=135) is used solely for five-fold cross-validation in Table 1; the local dataset (n=35) serves only as independent external validation in Fig. 2.They differ in protocols and cognitive batteries (MCCB T-scores vs. BACS raw scores), so no merged training was performed. We will explicitly label the local set as held-out external validation in Sec. 3.1 and Fig. 2 caption.
Baseline coverage (R2-2). MT denotes Multi-Task; the full term will appear at first occurrence. [33] is built on sMRI-derived structural graphs with explicit covariate disentanglement; its modality, graph construction, and supervision differ from our FC-based subject-wise prediction, making fair reproduction under identical FC inputs infeasible. For comparability, Table 1 includes two directly corresponding FC-based age-deconfounding baselines: MT-SZ_ar (regression residualization [4]) and MT-SZ_adv (adversarial deconfounding [6,1]), reimplemented under identical FC inputs and task setup.
Preprocessing and sample size (R1). We agree preprocessing and parcellation affect FC modeling and that the paper underdiscusses this. Experiments use fMRIPrep [7] with FC on the Power-264 atlas [24]; the camera-ready will discuss denoising, parcellation, and FC construction. We acknowledge sample size as an important limitation; COBRE (n=135) and the external set (n=35) are comparable to recent multi-site SZ studies.
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.
The reviewers are all positive about the novelty of the proposal and well designed experiments. The authors should explicitly acknowledge limitations raised by the reviewers, e.g., inability to disentangle non-linear aging effects and small sample sizes, and improve clarity on the points raised by reviewer 2.
