{"id":"41321cba-7ad7-4145-a72a-07a52b64c36c","arxiv_id":"2507.02908","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hyperbolic kernel graph fusion framework, using tangent-space random features and cross-modality coupling, improves SMC vs CN and ANI vs CN classification over Euclidean and hyperbolic graph baselines.","lead":"The authors propose a framework that maps structural and functional brain connectivity graphs into hyperbolic space and fuses them through new kernel-based graph neural networks to detect early neurocognitive decline. It reports higher classification accuracy than standard graph and hyperbolic graph baselines on memory-concern and HIV-related cognitive impairment cohorts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The hyperbolic geometry in HKGF is not isolated: with c=0.001 and near-identity log maps, the gains may come from the cosine branch and coupling, not from hierarchy preservation.","rationale":"The paper's empirical core is substantial: two independent cohorts, over 4,000 subjects used for pretraining, and consistent gains over many Euclidean and hyperbolic baselines. The cross-modality coupling module appears to contribute meaningfully, as shown by the w/oC ablations. However, the central scientific claim is that hyperbolic geometry preserves and exploits hierarchy. That claim rests on the assumption that the log-map-plus-Euclidean-activation construction is a genuinely hyperbolic mechanism. The implementation details undercut this: with c=0.001 and typical fMRI feature norms, the log map is nearly the identity, and Fig. 5 shows curvature has little effect while the cosine branch weight λ matters more. The ablations do not isolate the geometric component because they change the backbone entirely. The reader's weakest assumption is exactly this point, and the proposed c=0 ablation would settle whether the hyperbolic label is doing real work or is decorative. The manuscript also contains unresolved inconsistencies in task descriptions and a code link that points to a different repository, but these are secondary to the mechanism question. Given that the empirical results may still be reproducible and practically useful, the conditional verdict is appropriate; no change is needed.","tokens_in":26070,"tokens_out":6017,"duration_ms":62418,"concrete_test":"Rerun the ADNI SMC-vs-CN and HAND ANI-vs-CN experiments with HKGF1 and HKGF2 after replacing log_c_0(·) in Eqs. (13)-(18) and (21) with the identity map (the c→0 limit), keeping all hyperparameters, including λ and the cosine branch, fixed. If the AUC/ACC change by less than the reported cross-validation standard deviation, the hyperbolic curvature is not the operative mechanism; the improvement comes from the cosine branch and coupling module rather than from hierarchy-preserving geometry. To fully settle the attribution, additionally compute the Gromov δ-hyperbolicity of the learned embeddings under c=0.001 and under c=0; if both are similarly non-hyperbolic, the claim that the embeddings are tree-like is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"HKGF's central claim is that HKGNNs preserve hierarchical brain-network structure by operating in hyperbolic space. The actual computation in Eq. (14) and Eq. (18) maps nodes with the Poincaré log map (Eq. (4)) and then applies ordinary Euclidean GCN/GAT operations with ReLU/ELU plus a λ-scaled cosine term. Because c=0.001, and because fMRI node features (Pearson correlation rows with norm <10) are not rescaled by the projection in Eq. (13), log_c_0(z) ≈ z + O(c) ≈ z; the model is effectively a Euclidean GNN with an extra cosine branch. Figure 5 confirms the curvature parameter has only a minor effect on AUC/ACC, while λ matters more. The ablations (HKGF-G vs HKGF1) change both the log map and the cosine branch simultaneously, so they cannot attribute the improvement to hyperbolic geometry. No δ-hyperbolicity, tree-likeness, or hierarchy-preservation metric is reported for the learned embeddings. The kernel derivation in Eqs. (7)-(11) is also not faithfully implemented: Eq. (8)/(11) use random features with fixed p(w), while Eq. (14) uses learned W and graph aggregation, so the link from kernel theory to the architecture is informal. Without isolating the geometric component, the central claim that hyperbolic structure drives the results is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HKGF, a multimodal fusion framework for neurocognitive decline classification that combines DTI- and fMRI-derived brain graphs (and, in a generalization experiment, ASL-derived graphs) using a family of hyperbolic kernel graph neural networks (HKGCN and HKGAT). The HKGNNs project node features into a Poincaré ball and apply logarithmic maps to a tangent space before performing graph aggregation with ReLU/ELU and cosine activations; a cross-modality coupling graph and a hyperbolic neural network are used for fusion and prediction. The authors report state-of-the-art results on SMC vs. CN classification on ADNI (fMRI+DTI) and ANI vs. CN classification on HAND, with a transfer-learning strategy pretraining models on 3,806 auxiliary fMRI scans. The paper includes ablations, hyperparameter sensitivity, generalization to ASL, computational complexity analysis, and visualization of discriminative brain regions.","tokens_in":26363,"tokens_out":5638,"duration_ms":58901,"significance":"If the reported results hold, HKGF would be a practically useful multimodal fusion framework for challenging neurocognitive decline tasks, with the notable strengths of released source code and pretrained models, extensive baselines, a large auxiliary pretraining corpus, and a computational-cost analysis showing parity with GCN/GAT. The central scientific claim, however, is that hyperbolic geometry is what drives the gains by preserving hierarchical brain-network structure; this claim is currently not established, because the hyperbolic component is not isolated and the paper's own sensitivity analysis weakens it. The contribution would be more credible as an empirical fusion method than as a demonstration of hyperbolic representation benefits.","major_comments":[{"comment":"The hyperbolic component is not isolated, and the paper's own evidence suggests it is not the driver of performance. With the implemented curvature c=0.001 and the projection in Eq. (13) leaving most features unchanged, the logarithmic map in Eq. (4) satisfies log_c^0(z) ≈ z + O(c), so Eqs. (14) and (18) are effectively Euclidean GCN/GAT layers with an extra λ-scaled cosine branch. Figure 5 confirms that the curvature parameter c has only a minor effect on AUC/ACC, while λ has a more noticeable impact. The ablations HKGF-G and HKGF-A replace HKGCN/HKGAT with plain GCN/GAT, changing both the log map and the cosine branch simultaneously, so they cannot attribute the improvement to hyperbolic geometry. The paper also reports no δ-hyperbolicity, tree-likeness, or hierarchy-preservation metric for the learned embeddings. To support the central claim, the authors should compare against an identical architecture with the log map replaced by the identity (or a large c range), and measure whether the embeddings actually reflect hierarchical structure.","section":"§5.2, Eqs. (14) and (18)"},{"comment":"The kernel derivation is not faithfully implemented. The HAC and HRBF kernels are defined via random feature expansions with fixed weights W drawn from p(w), as in Eqs. (8) and (11), but the HKGCN and HKGAT layers in Eqs. (14) and (18) use trainable weights and apply nonlinear activations after graph aggregation, and the cosine term is applied to the aggregated features rather than to pointwise random Fourier features. No analysis or theorem is provided to show that Eq. (14) approximates the proposed hyperbolic kernels on graphs. Since the method is named and motivated by hyperbolic kernel theory, this gap between the theoretical formulation and the implemented architecture should be addressed or the kernel framing should be softened.","section":"§3.2.2, Eqs. (7)–(12) vs. Eq. (14)"},{"comment":"There is a direct contradiction in the experimental setup. Section 4.2.2 defines Task 1 as SMC vs. CN classification on ADNI with fMRI and ASL data from 29 SMC and 15 CN subjects, but Table 2 reports SMC vs. CN results on ADNI with fMRI and DTI data from 46 SMC and 48 CN subjects, and the ASL experiment appears only later in Table 6. This inconsistency makes it unclear what data underpin the main ADNI claim and must be corrected for reproducibility.","section":"§4.2.2 vs. Tables 2 and 6"}],"minor_comments":[{"comment":"There is a typo in the first sentence: \"Twp prediction tasks\" should be \"Two prediction tasks.\"","section":"§4.2.2"},{"comment":"The notation b^T in f(Â(log_c^0(X̃)W + b^T)) is confusing: since b ∈ R^M is defined as a vector, adding a transposed vector to an N×M matrix is not well-defined. Please clarify whether this is meant to be a broadcast bias term, and use consistent notation (e.g., 1b^T).","section":"Eq. (14)"},{"comment":"The caption refers to \"different settings of c and α,\" but the hyperparameter discussed in the text and figures is λ. Please correct the symbol.","section":"Fig. 5 caption"},{"comment":"In the description of ablations, \"HKGFC-A\" appears to be a typo for \"HKGF-A.\"","section":"§5.1"},{"comment":"The word \"prepossessing\" in the ASL generalization section should be \"preprocessing.\"","section":"§5.4"},{"comment":"The text attributes a dynamic-FC GNN framework to \"Liu et al. [22],\" but reference [22] is Ereira et al., Nature Mental Health 2024, which appears to be a different work. Please verify the citation or replace it with the intended reference.","section":"§2.1, reference [22]"},{"comment":"The reported HKGF2 results differ slightly between Tables 2 and 4 (e.g., AUC 80.42 vs. 80.36, ACC 81.26 vs. 81.83). Please explain whether these are different random seeds or a reporting error.","section":"Tables 2 and 4"}],"recommendation":"major_revision","confidential_remarks":"The paper relies on the authors' own ACTION toolbox [56] and HAND dataset definition [23] for pretraining and target data. This is not circular in a technical sense, but the reviewer should be aware that the pretraining infrastructure is shared with the authors' prior work, which may make it easier to tune the pipeline to the target tasks. The deeper concern is novelty: with c=0.001, the proposed HKGNN is essentially a Euclidean GNN with a cosine branch, and the hyperbolic framing may overstate the contribution. If the authors can provide a convincing isolation of the geometric effect, the paper would be a solid empirical study; otherwise, it may be better positioned as an empirical fusion framework without strong hyperbolic claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:2507.02908. First, the empirical package is substantial: HKGF beats a long list of baselines on two brain-imaging cohorts, with ablations, hyperparameter sensitivity, a transfer-learning recipe, and an extra modality. Second, the central claim—that hyperbolic geometry is what does the work—is not supported by the evidence as written.\n\nWhat's genuinely new is the assembly: HAC and HRBF random feature kernels grafted onto GCN/GAT backbones, a two-stage SC-FC coupling graph, and a hyperbolic predictor. I don't see that exact combination in the cited literature. The experiments are thorough: roughly 4,000 pretraining subjects, two target cohorts, multiple backbones, statistical tests, and a generalization to ASL. The authors also report parameter counts and FLOPs. That is real effort and real evidence the framework works as a fusion method.\n\nThe soft spot is the geometry. With c=0.001, the Poincaré log map is very close to the identity (log_c_0(z) is approximately z). So HKGCN/HKGAT are effectively Euclidean GNNs with an extra cosine branch. The paper's own sensitivity analysis shows curvature c has a minor effect on AUC/ACC, while lambda matters. The ablations (HKGF-G vs HKGF1) change both the log map and the cosine branch at once, so they can't attribute the gain to hyperbolic structure. No delta-hyperbolicity or hierarchy-preservation metric is reported. The honest reading is that the cosine branch and the coupling graph drive the improvements, not the negative curvature.\n\nThere are also consistency issues: Section 4.2.2 describes Task 1 as fMRI+ASL with 29 SMC and 15 CN, but Table 2 uses fMRI+DTI with different subject counts; text and Table 1 disagree on the ADNI CN count. The code link points to a general toolbox, not a dedicated repository, so I couldn't verify the exact implementation. And the HAND cohort is all male, which limits generalization claims.\n\nNone of this kills the paper. The framework and the empirical comparisons are useful, and the authors are honest about some limitations. But the hyperbolic motivation needs a clean isolation: add a Euclidean random-feature baseline with the same cosine branch, subtract only the log map in an ablation, and measure tree-likeness or delta-hyperbolicity of the learned embeddings. Then the hierarchy claim can be tested.\n\nWho should read this: people working on multimodal brain network fusion, and anyone considering hyperbolic GNNs for medical imaging. It deserves a serious referee—the experiments are extensive and the method is presented clearly, even if the geometric claim needs major revision. I'd send it to review with a request for major changes.","headline":"Solid empirical fusion framework; the hyperbolic geometry claim is not isolated and likely not what drives the gains.","tokens_in":26888,"tokens_out":3389,"would_cite":true,"duration_ms":34890,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Hyperbolic kernel graph networks outperform Euclidean baselines on early neurocognitive decline classification.","keywords":["hyperbolic kernel graph neural networks","multimodal neuroimaging","brain connectivity","neurocognitive decline","DTI and fMRI fusion","Poincaré ball","graph classification","transfer learning"],"falsifier":"Train HKGF alongside a matched Euclidean control that replaces the logarithmic map with the identity (equivalently, sets curvature $c$ to 0) while keeping the same ReLU-plus-cosine activations and the same coupling graph; if ADNI SMC-vs-CN AUC stays at the reported 80.42%, then the hyperbolic geometry is not what produces the improvement.","tokens_in":25878,"feed_emoji":"🧠","tokens_out":10464,"duration_ms":100883,"temperature":0.7,"pith_summary":"The paper claims that brain connectivity graphs—structural networks from diffusion MRI (DTI) and functional networks from resting-state fMRI—are hierarchically organized, and that Euclidean graph neural networks miss this organization. It introduces HKGF, a framework that encodes these graphs with hyperbolic kernel graph neural networks (HKGNNs), fuses structural and functional embeddings through a learned SC-FC coupling graph, and classifies with a two-layer hyperbolic network. On two small clinical cohorts, the attention-based variant HKGF2 reaches AUC 80.42% on SMC vs CN classification in ADNI and 69.74% on ANI vs CN classification in HAND, outperforming Euclidean GCN/GAT variants and a prior hyperbolic GCN. The paper also shows the framework transfers to fMRI+ASL inputs and benefits from pretraining on 3,806 auxiliary fMRI scans.","feed_headline":"Hyperbolic kernels lift brain-scan decline detection to 80% AUC","feed_subtitle":"Fusing DTI and fMRI in hyperbolic space catches subtle memory and HIV-related impairment better than plain graph nets.","key_machinery":"The load-bearing object is the hyperbolic kernel graph layer, instantiated as HKGCN (Eq. (14)) and HKGAT (Eq. (18)). Both versions project node features into the Poincaré ball, apply the logarithmic map to reach the tangent space, perform adjacency- or attention-based aggregation there, and then apply a two-term activation: a nonlinearity $f(\\cdot)$ plus a scaled cosine $\\lambda\\cos(\\cdot)$. The first term is a topology-aware approximation of the hyperbolic arc-cos (HAC) kernel and is meant to capture global relationships; the cosine term approximates the hyperbolic RBF (HRBF) kernel and captures local similarity. The SC-FC coupling graph, built as the inner product of normalized structural and functional embeddings, is the mechanism that fuses modalities before a second HKGNN pass. This design avoids Möbius addition and multiplication, which is why the paper reports HKGCN matching GCN in FLOPs and HKGAT matching GAT.","core_discovery":"The paper's central discovery is that replacing Euclidean message passing with two curvature-aware kernels—a hyperbolic arc-cos kernel for global similarity and a hyperbolic RBF kernel for local similarity—lets a graph neural network represent brain networks in a Poincaré ball at roughly the computational cost of a standard GCN. Node features are projected into the ball, pulled back to the tangent space with the logarithmic map $\\log_0^c$, aggregated over the graph, and then activated by $f(\\cdot)+\\lambda\\cos(\\cdot)$, where the cosine term approximates an HRBF kernel. The same kernel layers are applied a second time to a data-driven coupling graph whose edges are the inner product of normalized structural and functional embeddings, and a two-layer hyperbolic network performs prediction. The paper reports consistent gains over Euclidean baselines and over HGCN in both tasks, with HKGF2 (the attention backbone) giving the best numbers.","pith_inferences":["The paper does not measure tree-likeness or hyperbolicity of the learned embeddings, so a direct test of the mechanism would compare HKGF against its Euclidean counterpart with identical two-term activations and coupling; if the gap collapses, the performance gain is not coming from negative curvature.","The sensitivity analysis shows curvature $c$ has only a minor effect while the cosine weight $\\lambda$ matters, which suggests the practical boost may come from the residual cosine branch and the coupling graph rather than from hyperbolic geometry per se.","The inner-product coupling graph could be reused for other heterogeneous graph pairs, such as gene co-expression with protein interaction or behavioral measures with connectivity, giving a general fusion primitive beyond neuroimaging.","Because HKGF requires complete modality pairs and the authors flag missing-modality handling as future work, a clinical deployment would need imputation or a partial-observation extension before the framework can be used on incomplete records."],"forward_implications":["If correct, DTI-fMRI fusion for early cognitive decline can run in hyperbolic space at essentially the same FLOPs as Euclidean GCN/GAT, removing the main practical barrier of prior hyperbolic GNNs.","The framework is modality-agnostic: the authors show fMRI+ASL classification reaches AUC 76.85%, so the same coupling and kernel machinery can absorb perfusion imaging alongside connectivity data.","Pretraining on large auxiliary fMRI cohorts adds more than 5% AUC on the HAND task, suggesting transfer learning can offset the tiny target cohorts typical of clinical neuroimaging studies.","The discriminative regions the method highlights—frontal-parietal ROIs for SMC, subcortical and cerebellar ROIs for ANI—are consistent with prior literature, so the model can point toward plausible biomarkers rather than only a scalar risk score."],"supporting_citations":[{"why":"It provides the seven functional subnetworks used to motivate and label the brain's hierarchical organization in the t-SNE visualizations.","marker":"[9]"},{"why":"It supplies the hyperbolic GCN baseline that motivates the work and that HKGF is compared against and outperforms.","marker":"[16]"},{"why":"It is the earlier hyperbolic multimodal fusion method for mild cognitive impairment that HKGF extends and competes with on the same imaging modalities.","marker":"[19]"},{"why":"It establishes the Poincaré ball as a representation space for hierarchical data, justifying the choice of hyperbolic geometry.","marker":"[34]"},{"why":"It provides the Poincaré logarithmic map and hyperbolic neural network machinery used by the HKGNN layers and the HNN predictor.","marker":"[35]"},{"why":"It supplies the Poincaré kernel constructions and the curve-length equivalence theorem that justify treating tangent-space distances as hyperbolic distances.","marker":"[36]"},{"why":"It is the arc-cosine kernel that the proposed hyperbolic arc-cos (HAC) kernel generalizes to the Poincaré ball.","marker":"[40]"},{"why":"It gives the random feature and random Fourier feature approximations used to turn the HAC and HRBF kernels into neural-layer activations.","marker":"[41]"},{"why":"It is the standard graph convolutional network that HKGCN extends with hyperbolic kernels and a primary Euclidean baseline.","marker":"[44]"},{"why":"It is the graph attention network that HKGAT extends with hyperbolic kernels and a primary Euclidean baseline.","marker":"[45]"}],"fun_headline_variants":["Hyperbolic kernels hit 80% AUC in brain decline detection","Multimodal brain nets in hyperbolic space aid decline diagnosis","Fusing DTI and fMRI with hyperbolic GNNs improves decline reads","Curved geometry captures hierarchical brain networks for decline","Hyperbolic graph neural nets yield 80% AUC on decline tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's load-bearing premise is that mapping brain-network nodes into a Poincaré ball and computing kernels in the tangent space captures the hierarchical organization that Euclidean graph networks miss; the paper does not directly measure whether the learned embeddings are actually more hierarchical or tree-like.","fun_headline_variants_meta":{"raw":{"variants":["Hyperbolic kernels hit 80% AUC in brain decline detection","Multimodal brain nets in hyperbolic space aid decline diagnosis","Fusing DTI and fMRI with hyperbolic GNNs improves decline reads","Curved geometry captures hierarchical brain networks for decline","Hyperbolic graph neural nets yield 80% AUC on decline tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000366,"raw_usage":{"total_tokens":1973,"prompt_tokens":958,"completion_tokens":1015,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":930}},"tokens_in":574,"tokens_out":1015,"duration_ms":10573,"temperature":1.0,"reasoning_tokens":930,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:06:54.515945+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train HKGF alongside a matched Euclidean control that replaces the logarithmic map with the identity (equivalently, sets curvature $c$ to 0) while keeping the same ReLU-plus-cosine activations and the same coupling graph; if ADNI SMC-vs-CN AUC stays at the reported 80.42%, then the hyperbolic geometry is not what produces the improvement.","supporting_citations":[{"cited_title":"The organization of the human cerebral cortex estimated by intrinsic functional connectivity,","cited_arxiv_id":null,"evidence_quote":"It provides the seven functional subnetworks used to motivate and label the brain's hierarchical organization in the t-SNE visualizations."},{"cited_title":"Hyperbolic graph convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"It supplies the hyperbolic GCN baseline that motivates the work and that HKGF is compared against and outperforms."},{"cited_title":"Multimodal deep fusion in hyperbolic space for mild cognitive impairment study,","cited_arxiv_id":null,"evidence_quote":"It is the earlier hyperbolic multimodal fusion method for mild cognitive impairment that HKGF extends and competes with on the same imaging modalities."},{"cited_title":"Poincar ´e embeddings for learning hier- archical representations,","cited_arxiv_id":null,"evidence_quote":"It establishes the Poincaré ball as a representation space for hierarchical data, justifying the choice of hyperbolic geometry."},{"cited_title":"Hyperbolic neural networks,","cited_arxiv_id":null,"evidence_quote":"It provides the Poincaré logarithmic map and hyperbolic neural network machinery used by the HKGNN layers and the HNN predictor."},{"cited_title":"Poincar ´e kernels for hyperbolic representations,","cited_arxiv_id":null,"evidence_quote":"It supplies the Poincaré kernel constructions and the curve-length equivalence theorem that justify treating tangent-space distances as hyperbolic distances."},{"cited_title":"Kernel methods for deep learning,","cited_arxiv_id":null,"evidence_quote":"It is the arc-cosine kernel that the proposed hyperbolic arc-cos (HAC) kernel generalizes to the Poincaré ball."},{"cited_title":"Random features for large-scale kernel machines,","cited_arxiv_id":null,"evidence_quote":"It gives the random feature and random Fourier feature approximations used to turn the HAC and HRBF kernels into neural-layer activations."},{"cited_title":"Graph attention networks,","cited_arxiv_id":null,"evidence_quote":"It is the graph attention network that HKGAT extends with hyperbolic kernels and a primary Euclidean baseline."}],"review_version":1}