{"id":"d52aff8f-74b0-4975-a9d0-b93472c7e251","arxiv_id":"2607.09982","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Explicit unimodal, directional-bimodal and trimodal routes plus inference-time masking yield higher AUROC/F1 than fusion baselines on MIMIC-IV mortality and 25-phenotype tasks while exposing modality reliance.","lead":"The paper introduces a routing model that keeps separate pathways for structured labs, clinical notes, and chest X-rays (plus their directed combinations) so clinicians can see which sources drive each ICU prediction. It also lets the model drop missing modalities at test time and reweight the rest, giving an audit trail of how decisions change.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Route weights may encode architectural or shortcut artifacts rather than clinically meaningful asymmetric interactions, undercutting the interpretability/audit claims.","rationale":"The reader correctly isolates the weakest link: the leap from learned route weights to clinically meaningful asymmetric interactions. The manuscript supplies clean engineering, competitive numbers, and useful missing-modality audits, yet never directly tests whether directionality or the routing layer itself is doing the claimed clinical work versus merely re-weighting the strongest unimodal signals. The cerebrovascular shortcut already surfaces inside the paper, so the concern is not external. A single controlled symmetric-attention retrain would settle whether the interpretability story is load-bearing or ornamental; until that (or an equivalent clinician-alignment check) is done, CONDITIONAL remains the appropriate verdict. No stronger objection (e.g., data leakage, non-reproducibility, or outright numerical error) is evident from the full text.","tokens_in":18207,"tokens_out":538,"duration_ms":5731,"concrete_test":"Retrain an otherwise identical model in which the six directional bimodal routes are replaced by three undirected (symmetric) cross-attention blocks of matched parameter count, keep the same routing activations/coefficients and inference-time masking, and re-evaluate Table 1 AUROC/F1 plus the cerebrovascular missing-N drop. If performance and the shortcut pattern remain essentially unchanged, the directional-route interpretability claim is not supported by the architecture.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that the ten explicit routes plus patient/label-specific coefficients deliver interpretable, auditable multimodal reasoning rests on the premise that directional cross-attention (Eq. 3 and the six R2 routes) and the resulting effective weights W = α·R isolate genuine asymmetric clinical evidence rather than encoder-capacity or dataset-shortcut effects. The paper itself flags one such shortcut for cerebrovascular disease (high weight on I←N / N←I yet near-total collapse when notes are masked, §5.4). No controlled experiment distinguishes clinical asymmetry from architectural bias: the reverse directions are never ablated while holding capacity fixed, nor are route weights validated against clinician-annotated evidence sources or against a symmetric-attention control that still uses the same routing layer. Missing-modality performance drops and heatmaps (Figs. 2–3, Table 3) are consistent with the desired story but equally consistent with notes simply being the strongest unimodal encoder; they therefore supply only indirect support for the load-bearing interpretability claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes an explicit multimodal routing architecture for EHR prediction that constructs ten discrete routes—three unimodal, six directional bimodal (via cross-attention), and one hierarchical trimodal—over structured longitudinal data (L), clinical notes (N), and chest X-rays (I). Patient-specific route activations α and label-specific routing coefficients R yield effective weights W=α·R that drive prediction and are intended to support local, global, and group-level interpretability. Robustness and auditability are assessed by inference-time route masking (no retraining) that disables routes involving a missing modality and renormalizes the rest. On trimodal MIMIC-IV ICU stays the method reports higher AUROC/F1 than joint fusion, late fusion, TriMF, and MoE-Health for binary mortality and 25-phenotype multi-label prediction, with routing heatmaps and missing-modality breakdowns used to characterize modality reliance by clinical group.","tokens_in":18531,"tokens_out":1219,"duration_ms":17524,"significance":"If the directional routes and learned weights isolate clinically meaningful asymmetric evidence rather than encoder-capacity or dataset artifacts, the work would be a practical contribution to trustworthy multimodal clinical ML: it couples competitive accuracy with an auditable pathway structure and a no-retrain missing-modality protocol that can surface shortcut reliance (as the authors themselves note for cerebrovascular disease). Strengths that should be credited include matched-encoder baselines and ablations (Tables 1–2), open-source code, systematic per-phenotype and group-level analyses (Table 3, Figures 2–4), and the explicit separation of activations from coefficients. Absolute F1 remains modest under class imbalance, so the primary value is the interpretability/audit framing rather than a large accuracy leap.","major_comments":[{"comment":"The central interpretability/audit claim rests on the premise that the six directional bimodal routes (Eq. 3; R2 in §3.3) and effective weights W=α·R isolate clinically meaningful asymmetric interactions. Support is only indirect (missing-modality drops and heatmaps in §5.2–5.4). The paper itself flags a notes–image shortcut for acute cerebrovascular disease (§5.4: high I←N/N←I weight yet near-collapse when N is masked). There is no capacity-matched control that replaces directional cross-attention with symmetric attention (or undirected concatenation) while retaining the same routing layer, nor any external check (e.g., clinician-annotated evidence sources or route-level ablation of reverse directions only). Without such a control, the claim that routes are clinically asymmetric rather than architectural or shortcut artifacts remains under-supported for a paper whose title and abstract","section":null},{"comment":"Robustness is evaluated exclusively by training on fully trimodal stays and masking routes only at inference (§3.5, §4.1, §4.3.3; Table 4). Structured data L is never masked because it is ‘always retained.’ This design is a valid stress test of post-deployment missingness for a trimodal-trained model, but it does not establish robustness under the more realistic regime in which the model is trained with missing modalities (or with a missing-modality bank). The manuscript should either (i) add a train-with-missing comparison against methods designed for that setting, or (ii) clearly scope the robustness claim to inference-time masking of a trimodal-trained model and avoid broader deployment language.","section":null},{"comment":"Table 1 reports gains of roughly 3–6% over MoE-Health, but no statistical significance tests, confidence intervals beyond ±std, or multiple-run protocol details are given for the primary comparison. Given overlapping stds on some metrics and modest absolute F1 (0.45 mortality, 0.50 phenotypes), the performance superiority claim needs either paired significance testing across seeds or a clearer statement that gains are small and secondary to the routing analysis.","section":null}],"minor_comments":[{"comment":"Figure 1 is dense; route labels (e.g., L⟶N vs N⟶L) are hard to parse at print size. Consider a simplified schematic plus a table enumerating the ten routes.","section":null},{"comment":"Notation mixes R for the route set and R_{p,r,c} for routing coefficients; a distinct symbol for coefficients would reduce confusion in §3.4.","section":null},{"comment":"Appendix routing-coefficient heatmaps (Figures 5–6) largely duplicate effective-weight patterns; a short quantitative comparison (e.g., rank correlation of R vs W) would justify relegating them to the appendix.","section":null},{"comment":"Related-work placement of FAME and DrFuse is brief; a one-sentence contrast on directional vs scalar/fairness weighting would help readers place the contribution.","section":null},{"comment":"Typos/clarity: ‘dimen-sion’ line break in §4.2; ‘primary activation (α_r)’ in figure captions should match α_{p,r}; ACM reference placeholder URLs should be cleaned for camera-ready.","section":null}],"recommendation":"major_revision","confidential_remarks":"The empirical core is solid enough for a methods-oriented venue if the authors add a capacity-matched symmetric-attention control and tighten the robustness claim. The cerebrovascular shortcut finding is actually a selling point of the audit method; they should lean into it rather than treat it only as a caveat. Fit for CHASE/health-informatics is good; novelty relative to Tsai et al. multimodal routing and MoE-Health should be stated more sharply in the cover letter."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful part of this paper is the explicit ten-route design (3 unimodal + 6 directional bimodal + 1 hierarchical trimodal) kept separate until a patient- and label-specific weighted sum, plus inference-time route masking that reweights without retraining. That combination is cleaner than TriMF (which collapses pathways) and the MoE baselines they cite, and it actually lets them show how weights redistribute when notes or images disappear.\n\nThey do the experimental work carefully. Same encoders and training protocol for joint/late fusion, TriMF, and MoE-Health; ablations that knock out activations, coefficients, or directional cross-attention each hurt; missing-modality tables and per-phenotype breakdowns with standard deviations. Gains over the next-best model are real (roughly 4 points AUROC, 5–6 points F1). Absolute numbers stay modest, as expected under class imbalance, and they own that. The group-level heatmaps and the three patient examples make the routing patterns readable, and they themselves flag the cerebrovascular notes-image shortcut rather than burying it.\n\nThe soft spot is exactly the one the stress-test names, but it is not fatal. The claim that directional routes isolate clinically meaningful asymmetry is supported only indirectly by performance drops and weight shifts. They never ablate reverse directions while holding capacity fixed, nor do they check route weights against clinician-annotated evidence sources. So the interpretability story is plausible and useful for auditing, not proven. That is a limitation of the current evidence, not a circularity or a math error. Free parameters (d, a, optimizer settings) are ordinary; nothing is invented that is not defined in the methods.\n\nThis is for people building multimodal EHR models who care about missing-modality robustness and post-hoc inspection of evidence pathways. It will not reorganize the field, but it is a practical, reproducible step that a serious referee should see. I would send it out.","headline":"Clean multimodal routing with real audit machinery; interpretability claims rest on indirect evidence but the engineering and experiments are solid enough for peer review.","tokens_in":19096,"tokens_out":473,"would_cite":true,"duration_ms":5140,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Explicit multimodal routes for labs, notes, and X-rays make ICU predictions both stronger and auditable by showing which evidence pathways drive each decision.","keywords":["Multimodal learning","EHR","routing module","missing modalities","ICU outcome prediction","interpretability","clinical notes","chest X-rays"],"falsifier":"On a held-out trimodal ICU cohort, ablate or reverse the directional cross-attention (or replace learned coefficients with uniform weights) and check whether the claimed accuracy gains and clinically coherent missing-modality weight shifts both disappear.","tokens_in":19129,"feed_emoji":"🏥","tokens_out":807,"duration_ms":7312,"temperature":0.7,"pith_summary":"Most multimodal clinical models fuse labs, notes, and images into one opaque representation, so clinicians cannot tell which source supported a prediction. This paper argues that building ten discrete routes—unimodal, directional bimodal, and one trimodal—then learning patient- and label-specific weights for each route yields both higher accuracy and transparent reasoning on MIMIC-IV ICU mortality and 25-phenotype tasks. Inference-time route masking lets the model drop any modality without retraining and redistributes weight over the remaining routes, turning missing-data scenarios into an audit of how the system reweights evidence. The resulting routing patterns differ systematically by disease group: metabolic phenotypes lean on structured labs, respiratory phenotypes lean on imaging-plus-notes, and some low-prevalence conditions reveal shortcut reliance on notes. If the framework holds, hospitals gain a practical way to demand explanations and robustness checks for multimodal AI before deployment.","feed_headline":"Ten clinical routes make ICU AI both stronger and auditable","feed_subtitle":"Labs, notes and X-rays keep separate pathways so missing data and shortcut reliance become visible","key_machinery":"The multimodal routing layer: each of the ten routes produces an embedding; route activations measure expression strength and routing coefficients measure task relevance; their product forms an effective weight that is used both for prediction and for post-hoc audit under route masking.","core_discovery":"An explicit set of ten routes (three unimodal, six directional bimodal via cross-attention, one hierarchical trimodal) together with patient-specific activations and label-specific routing coefficients, plus inference-time route masking, produces higher AUROC and F1 than joint fusion, late fusion, TriMF and MoE-Health on trimodal MIMIC-IV stays while exposing clinically coherent modality-reliance patterns across phenotypes and mortality.","pith_inferences":["The notes–image shortcut observed for cerebrovascular disease suggests that route-level auditing could become a standard pre-deployment safety check for any multimodal clinical model.","Quadratic growth of directional bimodal routes implies sparse or dynamic routing will be needed before the method scales past three modalities.","Because structured data is treated as the always-available clinical anchor, the framework may understate fragility in settings where vitals themselves are missing or delayed."],"forward_implications":["Clinicians can inspect, for any patient and phenotype, which unimodal or cross-modal route carried the decision.","Missing-modality performance can be audited without retraining by simply zeroing routes that involve the absent source.","Disease groups that collapse when one modality is removed become candidates for more data, regularization, or architectural fixes.","The same route set can be reused for other multi-label ICU tasks that share the same three modalities."],"fun_headline_variants":["Ten routes make multimodal ICU AI auditable and robust","Explicit EHR routes expose modality reliance in phenotype prediction","Route masking audits missing-modality effects in clinical AI","Discrete unimodal-to-trimodal paths beat fusion on MIMIC-IV","Label-specific routes reveal coherent clinical modality patterns"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The directional routes and learned coefficients isolate genuine clinical interactions rather than architectural artifacts or dataset shortcuts that only look meaningful after the fact.","fun_headline_variants_meta":{"raw":{"variants":["Ten routes make multimodal ICU AI auditable and robust","Explicit EHR routes expose modality reliance in phenotype prediction","Route masking audits missing-modality effects in clinical AI","Discrete unimodal-to-trimodal paths beat fusion on MIMIC-IV","Label-specific routes reveal coherent clinical modality patterns"]},"model":"grok-4.5","effort":"low","cost_usd":0.002842,"raw_usage":{"total_tokens":1059,"prompt_tokens":782,"num_sources_used":0,"completion_tokens":63,"cost_in_usd_ticks":28420000,"prompt_tokens_details":{"text_tokens":782,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":214,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":782,"tokens_out":63,"duration_ms":2583,"temperature":1.0,"reasoning_tokens":214,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T01:15:51.418484+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out trimodal ICU cohort, ablate or reverse the directional cross-attention (or replace learned coefficients with uniform weights) and check whether the claimed accuracy gains and clinically coherent missing-modality weight shifts both disappear.","supporting_citations":[],"review_version":1}