{"id":"6acace41-5cc2-4fb7-9f0b-33610662e6d4","arxiv_id":"1908.03264","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Causal search algorithms FASK and Two-Step identify reproducibly consistent voxel subregions inside medial temporal lobe ROIs that appear to drive ROI-to-ROI connectivity.","lead":"The paper tests whether two causal search algorithms, FASK and Two-Step, can pick out the small groups of brain voxels inside larger brain regions that actually send signals between the regions. Using repeated scans of one person, the methods found fairly consistent subgroups across sessions, suggesting that useful detail is lost when brain regions are averaged together.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Concatenated-session iid assumption is load-bearing; without a surrogate null, stable subregion maps could be driven by session-level artifacts rather than true signaling.","rationale":"The reader's CONDITIONAL verdict is appropriate. The paper's central proof strategy is empirical consistency, not formal correctness, and the consistency metrics are computed on datasets whose construction violates the algorithms' exchangeability assumptions. The most direct way the claimed subregions could be spurious is if session-specific offsets, drifts, and autocorrelation create shared variance across voxels that FASK and Two-Step misread as direct adjacencies. Since both methods share the same input and similar first steps, their agreement does not control for this. The proposed surrogate test would distinguish true signaling from artifact while leaving the entire pipeline untouched, so it directly addresses the weakest assumption. I agree with the reader's identification; no verdict change is needed, though the paper should be revised to include such a control or explicitly qualify the claim.","tokens_in":16482,"tokens_out":5065,"duration_ms":57899,"concrete_test":"Using the same preprocessed sessions and the same concatenation scheme, generate phase-randomized surrogate time series per voxel and session (Fourier surrogates preserving the power spectrum, mean, and marginal distribution of each voxel while destroying cross-voxel dependence), then run the exact FASK and Two-Step pipelines with the paper's parameters. Compare (a) graph and subgraph density, (b) Jaccard consistency across the eight surrogate datasets, and (c) spatial overlap between surrogate subregions and the reported subregions (e.g., Dice coefficient for S_CA32DG and S_CA1). If surrogate results exhibit comparable Jaccard values or substantial overlap with the real maps, the reported consistency is not evidence for true subregions; if surrogate graphs are sparse and inconsistent, the concatenation concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The empirical case for connectivity subregions rests on treating 4610 concatenated time points from ten sessions as iid samples for FAS-stable's BIC conditional-independence tests and for Two-Step's ICA step. BOLD signals are autocorrelated, and each session contributes its own baseline, drift, and motion profile; appending sessions creates artificial mean shifts and boundary discontinuities. Any such session-level factor is a common cause of many voxels, so it can create the exact adjacency structure FASK/Two-Step are designed to avoid, but now as artifact. Because both algorithms consume the same concatenated data and share step-one machinery (FAS-stable or adaptive lasso), their agreement cannot independently validate the subregions. Tables 1-4 report consistency only; the paper's own Jaccard discussion concedes that high consistency can reflect uninformative dense graphs, and no null baseline is provided. The Discussion also concedes that single-scan stability 'seems advisable,' undercutting the concatenation design. If the iid/exchangeability assumption fails, the inferred S_P sets and their across-session/hemisphere stability may be properties of acquisition artifact, not of effective connectivity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a procedure to identify 'connectivity subregions' within anatomically defined ROIs by applying two causal search algorithms, FASK and Two-Step, to voxel-level BOLD time series from a single individual's medial temporal lobe. For each pair of ROIs, subregions are defined as the sets of voxels that are adjacent to at least one voxel in the other ROI in the inferred directed graph. Using eight datasets formed by concatenating ten resting-state sessions each, the authors report Jaccard consistency of the inferred subgraphs and subregions across datasets, with high consistency for CA32DG-CA1 and CA1-SUB and lower consistency for entorhinal-related pairs. They conclude that the methods recover stable, spatially localized subregions, including across hemispheres, and suggest that most MTL inter-regional signaling is reciprocal.","tokens_in":16713,"tokens_out":6725,"duration_ms":70155,"significance":"If the assumptions underlying FASK and Two-Step hold, this work offers a principled alternative to seed-based correlation or partial-correlation parcellation, potentially recovering intra-ROI functional differentiation from effective-connectivity information. The paper is clearly written, uses publicly available longitudinal data, and builds on previously published, code-released algorithms. However, the empirical demonstration rests on a concatenation scheme that violates the algorithms' iid assumptions, and the consistency metrics lack null baselines and direct cross-method/cross-hemisphere comparisons. As a proof-of-concept, the idea is promising but the current evidence is insufficient to support the strong claims in the abstract.","major_comments":[{"comment":"The concatenation of ten temporally ordered sessions into a single 4610-point time series violates the iid/exchangeability assumptions required by FAS-stable's BIC conditional independence tests and Two-Step's ICA step. BOLD time series are autocorrelated, and each session has its own baseline, drift, and motion profile; appending sessions creates artificial mean shifts and boundary discontinuities. Any such session-level factor is a common cause of many voxels, so it can create the exact adjacency structure FASK and Two-Step are designed to avoid, but as artifact rather than as true signaling. The Discussion's concession that 'investigation of stability over multiple scans of the same subject seems advisable' (p. 15) conflicts with the concatenation design. Please provide evidence that the inferred subregions are stable when analyses are performed on single sessions or when session effects are explicitly modeled (e.g., nuisance regressors, session-specific means), or temper the causal claims accordingly.","section":"Data and Voxel level graph results (pp. 6-8)"},{"comment":"All consistency claims are based solely on Jaccard indices without any null baseline. The paper correctly notes that an overfitting complete graph yields Jaccard = 1, so high values such as 0.89 for S_CA32DG and 0.90 for S_CA1 in Table 4 do not by themselves indicate meaningful stability. Please report expected Jaccard values under a null model matched for graph density (or permutation-based nulls) and, if possible, confidence intervals, so the reader can assess whether the observed consistency exceeds chance.","section":"Tables 1-4 and Jaccard discussion (pp. 7-11)"},{"comment":"The abstract and text claim that FASK and Two-Step 'recovered similar subsets of voxels' and that subregions show 'spatial consistency across different scanning sessions and across hemispheres,' but no quantitative comparison is provided between the two methods for the same data, nor between the left and right hemispheres. Tables 3-4 only report within-method consistency across the eight concatenated datasets. Please add quantitative cross-method (FASK vs Two-Step) and cross-hemisphere similarity measures (e.g., Jaccard or Dice coefficients) for the identified subregions and report them in the main text or supplementary material.","section":"Abstract and Connectivity subregions section (pp. 11-13)"},{"comment":"Both algorithms' correctness rests on the absence of unmeasured confounders: FAS-stable is a PC variant that assumes causal sufficiency, and Two-Step explicitly ignores the M/C terms ('In our application, we ignore the possible unobserved confounders C,' p. 4). Resting-state fMRI is subject to well-known physiological and scanner confounds. If latent common causes are present, the inferred adjacencies—and hence the subregions S_P—may reflect shared nuisance signals rather than effective connectivity. Please justify this assumption for the present data or provide robustness analyses (e.g., with global signal regression or inclusion of physiological regressors).","section":"FASK and Two-Step Methods (pp. 3-5)"}],"minor_comments":[{"comment":"There is a typographical error in step 4: the notation 'v_P ∈ V_P' is repeated; the second instance should be 'v_Q ∈ V_Q'.","section":"Connectivity Subregions Algorithm (p. 7)"},{"comment":"The in-text citation 'Pearl, Glymour and Jewel' should be 'Pearl, Glymour and Jewell', and the reference list entry should be checked for consistency.","section":"References (p. 16)"},{"comment":"The citation 'Jaccard et al., 1912' should be 'Jaccard, 1912', since the work is single-authored.","section":"References (p. 17)"},{"comment":"Please clarify whether the 57 volumes removed 'due to drifting' are removed from each session individually, and how the concatenation handles the session boundaries (e.g., any overlap, gap, or detrending).","section":"Data (p. 6)"},{"comment":"The parameters for FASK (c = 1, alpha = 1e-7, Delta = 0.3) and Two-Step (sigma = 75 log(n), lambda = 20, t = 0.05) are set 'to match the sparsity obtained with FASK'; no sensitivity analysis is provided. Since the consistency results may depend on these choices, a brief robustness check or a reference to prior simulation calibration would strengthen the report.","section":"Voxel level graph results (p. 7)"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and the paper is clearly written, but the empirical support is currently weaker than the abstract suggests. The concatenation issue, the missing null baseline, and the absence of direct cross-method and cross-hemisphere comparisons are all fixable with the existing data, so I encourage a major revision rather than rejection. The single-subject design also limits generalizability, which should be stated more prominently."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper gives a simple definition of a 'connectivity subregion' — voxels in ROI P that share an estimated adjacency with voxels in ROI Q — and applies it to high-resolution MTL data from one person across 80 sessions. The new contribution is the adjacency-based subregion concept, not the search algorithms, which are the authors' own FASK and Two-Step. That distinction matters: the paper is essentially an application note with a clean formalization.\n\nThe good parts: The definition is transparent and easy to reuse. The authors explicitly say Jaccard is a consistency measure, not an informativeness measure, which is the right caveat. They also concede that single-scan stability 'seems advisable,' which is an honest acknowledgment of a weakness. Both algorithms are open source, so the work is reproducible. For two of the five ROI pairs (CA32DG–CA1 and CA1–SUB), the subregion Jaccard indices are high (around 0.85–0.90) and consistent across the two methods, so the central idea has some empirical support.\n\nThe soft spots are real and load-bearing. The eight datasets are made by concatenating ten temporally ordered sessions into one 4610-point time series. FASK's FAS-stable step and Two-Step's ICA step assume exchangeable, stationary samples. BOLD data are autocorrelated, and each session has its own baseline, drift, and motion. Appending sessions creates artificial discontinuities and mean shifts that can act as common causes, producing exactly the kinds of adjacencies the algorithms are designed to find — except these are artifacts. Because those artifacts would be present in every concatenated dataset, the across-dataset Jaccard consistency could be driven by repetition of the same artifacts, not by true signaling. The paper provides no null baseline (e.g., phase-randomized data or session-shuffled voxels) to show that the observed Jaccard values exceed what an artifact-only process would produce. The right-hemisphere results are described as 'similar' but no quantitative cross-hemisphere overlap is given, so the cross-hemisphere claim is asserted rather than tested. The low consistency for ENT-related subregions is acknowledged in the text, which is honest, but it also weakens the claim that the method works generally.\n\nThe authors' own Discussion suggests running on a single scan would be advisable; I agree. A revision that re-runs the analysis on individual sessions (or with session as a covariate), includes surrogate nulls, and quantifies mean graph density would make the empirical case much stronger.\n\nBottom line: the paper is worth a serious referee, mainly because the subregion definition is genuinely useful and the application is clear. But as it stands, the empirical support is conditional on the concatenation assumptions, and the claims outrun the evidence. If I were the editor, I'd send it out, with the expectation that the authors address the iid violation and the missing null.","headline":"A simple, clearly-presented idea for defining voxel-level connectivity subregions, with real-data consistency results that are suggestive but undercut by concatenation artifacts and missing null controls.","tokens_in":17227,"tokens_out":3663,"would_cite":false,"duration_ms":36800,"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":"ROI averaging hides which voxels drive inter-regional communication, and two causal search algorithms recover those connectivity subregions directly from voxel-level resting-state fMRI.","keywords":["effective connectivity","voxel-level connectivity","connectivity subregions","FASK algorithm","Two-Step algorithm","resting-state fMRI","medial temporal lobe","causal graphical models"],"falsifier":"Run the same FASK and Two-Step searches on each of the ten sessions separately instead of on their concatenation; if the recovered subregions do not reappear in the single-session graphs, or if their spatial consistency across sessions is no better than a null distribution obtained by permuting voxel coordinates within each ROI, then the reported connectivity subregions are artifacts of the concatenation or of voxel-level autocorrelation rather than stable signaling.","tokens_in":16289,"feed_emoji":"🧠","tokens_out":9280,"duration_ms":84273,"temperature":0.7,"pith_summary":"Standard resting-state fMRI connectivity analysis averages voxel time series inside regions of interest, and this paper argues that the averaging discards the signal that distinguishes which voxels actually drive communication between regions. The paper's proposal is to infer a directed graph at the voxel level over all voxels in a set of ROIs, then define a connectivity subregion for each member of a pair of ROIs as the set of voxels in that ROI that have an adjacency to the other ROI. Applied to eighty longitudinal resting-state scans of one person's medial temporal lobe, two independent causal-search procedures, FASK and Two-Step, recover similar subregions that are spatially consistent across scanning sessions and across hemispheres. The intended payoffs are anatomical: subregions can coexist inside one ROI, they tend to sit near contact surfaces of adjacent regions, and their directed edges are roughly balanced, suggesting reciprocal communication. The methodological point is that correlation and partial-correlation analyses cannot distinguish direct signaling from common causes, indirect paths, or conditioning on common effects, while the two algorithms used here are designed to do so.","feed_headline":"Causal search finds which voxels drive region-to-region brain links","feed_subtitle":"Two algorithms recover stable medial-temporal subregions from one person's scans without voxel averaging.","key_machinery":"The machinery that carries the argument is a pair of graphical search algorithms that exploit the non-Gaussianity of BOLD signals instead of second-order correlations. FASK first runs FAS-stable, an order-independent conditional-independence search using BIC scores, to identify undirected adjacencies without mistaking common causes or indirect paths for direct connections; it then orients each adjacency as a two-way or one-way edge by comparing correlations conditioned on positive values of each variable, with a heuristic extra-adjacency rule for cases where a true 2-cycle cancels to near-zero correlation. Two-Step writes the linear causal system as $X=(I-B)^{-1}E$, uses adaptive lasso in a first step to restrict the free parameters of $(I-B)$, and then uses sparse ICA to estimate the connectivity matrix $B$ so that the residual components $E$ are as independent and non-Gaussian as possible. The two procedures are the source of the voxel-level adjacencies; the Connectivity Subregions Algorithm is simply the extraction rule that turns those adjacencies into subregions $S_P$ and $S_Q$ for each pair of ROIs.","core_discovery":"The central claim is that effective-connectivity subregions are identifiable from voxel-level BOLD time series without any anatomical prior about where within an ROI the communicating voxels lie. For a pair of ROIs $P$ and $Q$ with exclusive voxel sets $V_P$ and $V_Q$, a voxel $v_P\\in V_P$ belongs to the connectivity subregion $S_P$ exactly when the adjacency $v_P-v_Q$ appears in the voxel-level adjacency graph for some $v_Q\\in V_Q$, and $S_Q$ is defined symmetrically. To get those adjacencies, the authors first estimate a full directed, possibly cyclic graph over all 570 (left) or 530 (right) voxels spanning entorhinal cortex, CA32DG, CA1, and subiculum, using FASK or Two-Step on each of eight concatenated ten-session datasets. Across the eight temporally separated datasets, the undirected voxel graphs agree at Jaccard values around 0.7 for both algorithms, most subregions agree at 0.77–0.94, and the direction counts between subregion pairs are nearly symmetric. The authors conclude that connectivity subregions are a recoverable feature of resting-state data, that they can differ and overlap within a single ROI, and that most medial-temporal-lobe interregional signaling is reciprocal.","pith_inferences":["A natural next test is to compare recovered subregions against anatomical tracers or known axonal termination zones in an animal model; the paper's observation that subregions hug contact surfaces is a prediction that neuroanatomy could confirm.","Because session-to-session orientation estimates are unstable, direction-of-influence conclusions in single-subject resting-state studies may need pooling across sessions or subjects, rather than thresholding each graph independently.","The concatenation of ten sessions into one 4610-point sample treats the scans as exchangeable; checking whether the same subregions survive when each session is analyzed separately would directly test that assumption.","The same subregion-extraction rule could be applied to time-varying or task-based connectivity graphs, asking whether the voxels driving inter-regional communication shift with cognitive state."],"forward_implications":["A single ROI can contain several overlapping connectivity subregions, each tied to a different partner region, so ROI-level connectivity is not a homogeneous property of the region.","Adjacency structure at the voxel level is reproducible across sessions (Jaccard around 0.7), while edge orientations are much less so (around 0.3), so connectivity-subregion identification should be treated as stable but direction claims as provisional.","Medial temporal lobe communication inferred this way is largely reciprocal, with comparable numbers of directed edges in both directions between subregions.","If a single scanning session can support the same search, as the paper's simulations suggest, connectivity subregions could be mapped without the expense of longitudinal data.","Including all voxels in the search, not just the pair of interest, protects the inferred subgraphs against false adjacencies mediated by common causes or intermediate voxels in other regions."],"supporting_citations":[{"why":"Supplies both search algorithms (FASK and Two-Step), their orientation rules, sparsity penalties, and the cyclic simulation results that set parameter choices.","marker":"Sanchez-Romero et al. (2019)"},{"why":"Supplies the FAS-stable order-independent adjacency search that FASK uses in its first stage.","marker":"Colombo and Maathuis (2014)"},{"why":"Supplies the independent component analysis principle that Two-Step uses to estimate the connectivity matrix.","marker":"Hyvärinen and Oja (2000)"},{"why":"Supplies adaptive lasso, which Two-Step uses in step one to find undirected adjacencies.","marker":"Zou (2006)"},{"why":"Supplies the BIC score that FAS-stable uses for conditional-independence tests with a user-chosen penalty.","marker":"Schwarz (1978)"},{"why":"Provides the longitudinal single-subject resting-state dataset that the demonstration is built on.","marker":"Poldrack et al. (2015)"},{"why":"Provides the acquisition and preprocessing details for the same highly sampled individual dataset.","marker":"Laumann et al. (2015)"},{"why":"Supplies the standard medial-temporal-lobe connectivity model used to select the five region pairs for subgraph extraction.","marker":"Lavenex et al. (2000)"},{"why":"Supplies the manual ROI delineation procedure followed for the hippocampal subfields.","marker":"Liang et al. (2012)"}],"fun_headline_variants":["Voxel-level causal search finds brain subregions that drive ROI links","No ROI averaging: algorithms recover stable connectivity subregions","Medial-temporal subregions mapped via causal voxel graphs","Voxel subsets driving region links emerge without aggregation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 4610 BOLD time points obtained by concatenating ten separate sessions can be treated as independent, identically distributed samples for causal search, even though each session has its own baseline, drift, and motion and the BOLD signal is autocorrelated.","fun_headline_variants_meta":{"raw":{"variants":["Voxel-level causal search finds brain subregions that drive ROI links","No ROI averaging: algorithms recover stable connectivity subregions","Medial-temporal subregions mapped via causal voxel graphs","Voxel subsets driving region links emerge without aggregation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1507,"prompt_tokens":1036,"completion_tokens":471,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":400}},"tokens_in":652,"tokens_out":471,"duration_ms":4892,"temperature":1.0,"reasoning_tokens":400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:18:57.925950+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same FASK and Two-Step searches on each of the ten sessions separately instead of on their concatenation; if the recovered subregions do not reappear in the single-session graphs, or if their spatial consistency across sessions is no better than a null distribution obtained by permuting voxel coordinates within each ROI, then the reported connectivity subregions are artifacts of the concatenation or of voxel-level autocorrelation rather than stable signaling.","supporting_citations":[],"review_version":1}