{"id":"f9a35d61-03f9-4045-a854-9a5b4139a1b5","arxiv_id":"2412.15818","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Combining T1 MRI latents with clinical data via the DAFT fusion model yields a small F1 improvement (0.37 to 0.41) for ICU admission prediction after brain surgery, though statistical significance is not established.","lead":"This paper tests whether adding MRI images to clinical records improves predictions of which brain surgery patients need intensive care. It reports a small F1 improvement, but the differences are tiny and no uncertainty is reported.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing DAFT tabular-only ablation confounds modality effect with architecture; imaging's contribution is unidentified.","rationale":"The reader's weakest assumption was that the F1 differences are within sampling noise. That is a valid concern, but it targets precision of the estimate, not the identification of the causal effect. The comparison in the abstract and Section 5 contrasts DAFT (clinical+3D SSL latents) with an XGBoost (clinical-only) baseline. Even if the 0.41 vs 0.37 difference is statistically significant, it could be explained by the change in model family. The DAFT experiments section states that they \"limited these experiments to combining clinical and imaging data,\" which means no tabular-only DAFT was evaluated. The absence of this ablation is the most load-bearing gap because the paper's title and abstract claim modality benefit, not just a better model. I would therefore require the authors to add the tabular-only DAFT control (and, ideally, an imaging-only DAFT control) with confidence intervals before the multimodal claim can be accepted. This does not change the reader's conditional verdict, but it sharpens the condition: the claim requires more than error bars; it requires an architecture-matched comparison.","tokens_in":5464,"tokens_out":4610,"duration_ms":38091,"concrete_test":"Run a DAFT model with clinical tabular data only (removing the image-latent branch) under the same 5-fold CV, data splits, undersampling, and 0.5 threshold; also run DAFT with 3D SSL latents only (no clinical data). If DAFT+clinical-only achieves F1 approximately 0.41 or higher, or DAFT+imaging-only achieves F1 approximately 0.37, the reported gain is not evidence for a multimodal fusion benefit. Report point estimates with bootstrap 95% CIs over the five folds for all four configurations: XGBoost tabular, DAFT tabular, DAFT imaging, and DAFT combined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that adding imaging to clinical data improves ICU admission prediction. The supporting comparison in Section 5 is between a DAFT model trained on tabular clinical data plus 3D SSL latents (F1 0.41, AUC 0.76) and an XGBoost model trained on tabular data alone (F1 0.37, AUC 0.77). These models differ in architecture, optimization, and input representation, and the DAFT experiments (Section 4) deliberately \"limited these experiments to combining clinical and imaging data.\" No DAFT model was trained on tabular data only, so the observed 0.04 F1 improvement cannot be attributed to the imaging modality; it may be an effect of switching from gradient-boosted trees to a deep neural network with DAFT feature modulation. Without a tabular-only DAFT ablation, the multimodal benefit is unidentified. The reader's concern about missing error bars and significance tests is real but secondary: even with tight confidence intervals, the comparison would not isolate the imaging contribution.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multimodal approach for predicting ICU admission after brain surgery, combining clinical tabular data with T1-weighted MRI. The authors compare XGBoost (gradient-boosted trees), ResNet, and the DAFT architecture, using three imaging representations: 2D autoencoder latents, a 3D ROI through DAFT's integrated encoder, and latents from a pretrained 3D masked autoencoder (the 'brain foundation model'). The central claim, stated in the abstract and Section 5, is that the best multimodal DAFT model using 3D SSL latents improves the F1 score from 0.37 (clinical-only GBT) to 0.41 on pre- and post-operative data, and from 0.29 to 0.30 on pre-operative data only. The paper concludes that multimodal data fusion, particularly with a foundation-model feature extractor, benefits ICU admission prediction despite severe class imbalance (59 positive out of 611 subjects).","tokens_in":5612,"tokens_out":4201,"duration_ms":34639,"significance":"If the result holds, the contribution is a clinically relevant, first-of-its-kind application of multimodal deep learning to ICU admission prediction after brain surgery. The paper leverages a real clinical cohort, a public pretrained foundation model, and compares several fusion strategies, which are commendable. However, the central quantitative claim currently rests on point estimates without uncertainty quantification and on a comparison that confounds architecture with modality. The reported improvements (0.01 and 0.04 F1) are small and may not be clinically meaningful. Because the authors do not provide code or detailed experimental setup for reproducibility, the study's practical impact is limited until the statistical and ablation concerns are addressed.","major_comments":[{"comment":"The central claim that adding imaging improves over clinical-only prediction is not isolated by the experiments. In Section 4, the DAFT experiments are explicitly 'limited to combining clinical and imaging data'; no DAFT model is trained on tabular data alone. Yet the headline comparison in Section 5 contrasts DAFT with clinical plus 3D SSL latents (F1 0.41) against a GBT trained on tabular data only (F1 0.37). These two models differ in architecture, optimization, and input representation, so the observed difference cannot be attributed to the imaging modality. The authors should include a tabular-only DAFT ablation (and ideally a tabular-only deep model with the same training setup) to identify the imaging contribution.","section":"Section 4 (DAFT experiments) and Section 5 (Results)"},{"comment":"All reported F1 and ROC-AUC values are point estimates without error bars, confidence intervals, or significance tests. With only 59 positive samples, a 0.04 F1 difference (0.37 vs. 0.41) is within the range of sampling noise. The authors should provide bootstrap confidence intervals or paired significance tests (e.g., a paired bootstrap over the stratified folds) for the key comparisons, and explicitly state whether the differences are statistically distinguishable.","section":"Section 5, Fig. 2"},{"comment":"The paper evaluates a large number of configurations (two feature extractors, three input scenarios for XGBoost and ResNet, three DAFT variants, two data settings) and then highlights the best-performing one. Without any correction for multiple comparisons or a pre-specified analysis plan, the probability of at least one spuriously large difference inflates. The total number of comparisons and the model-selection rule should be disclosed, and the final claim should account for this selection.","section":"Section 5, Results and Fig. 2"}],"minor_comments":[{"comment":"The sentence 'For all three scenarios, we used the same hyperparameters as [6]' likely intends to cite the baseline GBT study [5] rather than the DAFT paper [6]; please correct the reference.","section":"Section 4, XGBoost experiments"},{"comment":"The phrase 'plane combinations' should be 'plain combinations'.","section":"Section 3, Multimodal DAFT Model"},{"comment":"The ROI description 'sized at 160 mm³ with dimensions (320×5×5×5)' is dimensionally inconsistent; please clarify whether the side length is 160 mm and how the tensor dimensions are interpreted.","section":"Section 2, Dataset"},{"comment":"The right panel is described as a 'ROC-AUC curve'; it should be an ROC curve, since AUC is the area under that curve, not a curve itself.","section":"Section 5, Fig. 2"},{"comment":"The abstract's claim of improvement 'from 0.29 [F1] to 0.30 [F1]' is a gain of 0.01, which may not be clinically meaningful; the authors should discuss the clinical relevance of the effect sizes, or temper the wording if the gain is not significant.","section":"Abstract and Section 5"}],"recommendation":"major_revision","confidential_remarks":"The authors' best-performing feature extractor [8] and the baseline method [5] both involve the same research group (DKFZ/Heidelberg), which creates a self-referential dependence in the benchmarking. This is not inherently problematic, but the authors should declare any potential conflicts of interest and ensure the pretrained model is publicly available and not custom-tuned for this dataset. The manuscript would also benefit from a clearer statement of data and code availability, as reproducibility is currently limited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a modest but honest empirical paper. The genuinely new bit is applying DAFT multimodal fusion, with a 3D SSL foundation-model feature extractor, to ICU admission prediction after brain surgery. The authors also report negative results for naive fusion, which is useful. On the main claim: adding imaging to clinical data raises F1 from 0.37 to 0.41 (pre+post-op) and from 0.29 to 0.30 (pre-op only). Those are small effects, and the discussion is appropriately cautious, though the abstract states the outperformance claim too flatly.\n\nWhat is good: the paper is transparent. It uses stratified five-fold CV for feature extraction, reports F1 and ROC-AUC, which suit severe class imbalance, and works with a real cohort of 611 subjects (59 positive). The finding that a simple concatenation of latents with tabular data can hurt, while a dynamic fusion method helps, is a useful data point for the multimodal community. The negative results are not buried.\n\nThe soft spot is load-bearing. The headline comparison is between DAFT trained on clinical plus imaging and XGBoost trained on clinical only. Those differ in architecture, optimizer, and input representation. The authors explicitly limited DAFT experiments to combined inputs and state they skipped exhaustive ablations. Without a DAFT model trained on tabular data alone, the observed 0.04 F1 gain cannot be attributed to imaging; it could just be the switch from gradient-boosted trees to a neural network with dynamic modulation. The stress-test note is right: error bars and significance tests are secondary, because even with tight confidence intervals the comparison would not isolate the imaging contribution.\n\nTwo smaller concerns: all reported metrics are point estimates without variance or significance testing, and with 59 positive cases the differences could be sampling noise. Also, the best imaging extractor (ref 8) comes from the same group; that is not a flaw by itself, but an independent extractor would make the comparison more convincing.\n\nWho is this for: people working on multimodal fusion in medical imaging, and clinical informatics teams interested in ICU resource planning. It is not a methods paper; the architecture is from prior work. It deserves peer review because it is a legitimate new application with honestly reported experiments, but the reviewers should require the tabular-only DAFT ablation and some measure of uncertainty before the claim is accepted. I would not cite it in my own work yet, but I would keep an eye on a revised version.","headline":"Honest, modest multimodal ICU prediction paper whose central imaging benefit is undermined by a missing tabular-only DAFT ablation.","tokens_in":6174,"tokens_out":2659,"would_cite":false,"duration_ms":21915,"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":"The paper claims that fusing MRI image features with clinical data via DAFT predicts ICU admission after brain surgery better than clinical data alone, raising F1 from 0.37 to 0.41.","keywords":["ICU admission prediction","multimodal learning","brain surgery outcomes","dynamic affine feature map transform","gradient boosted trees","class imbalance","MRI feature extraction","autoencoder"],"falsifier":"Repeated resampling of the same five-fold cross-validation, such as 100 bootstrap iterations, that places a 95% confidence interval around the F1 difference between DAFT-with-3D-SSL and clinical-only GBT and finds it contains zero would falsify the claim that the multimodal model outperforms the baseline; a similarly sized external cohort showing no F1 improvement would also falsify it.","tokens_in":5236,"feed_emoji":"🧠","tokens_out":7082,"duration_ms":58583,"temperature":0.7,"pith_summary":"This paper tries to establish that ICU admission after brain surgery is better predicted by fusing clinical tabular data with MRI-derived image features than by clinical data alone. On a cohort of 611 patients, 59 of whom truly required ICU care, the authors compare an XGBoost clinical-only baseline with multimodal models that add imaging latents. They report that the best multimodal configuration, a DAFT network using 3D latents from a pretrained masked autoencoder, raises F1 from 0.37 to 0.41 when pre- and post-operative data are available, and from 0.29 to 0.30 with pre-operative data alone. The stakes are practical: most patients are routinely transferred to the ICU after brain surgery at high cost, and a trustworthy predictor could cut unnecessary admissions while still flagging the patients who genuinely need intensive care.","feed_headline":"Adding MRI features lifts ICU prediction from 0.37 to 0.41 F1","feed_subtitle":"Tumor MRI and patient data fused by DAFT beat clinical-only XGBoost in forecasting who needs intensive care after brain surgery.","key_machinery":"The load-bearing mechanism is the Dynamic Affine Feature Map Transform (DAFT), a fusion layer that predicts per-channel scale and shift parameters from the low-dimensional clinical vector and applies them to the image feature maps, so the imaging signal is reweighted by the patient's clinical state rather than merely concatenated with it. Image features come from a pretrained 3D masked autoencoder that produces compact latents from a tumor-centered region of interest, the configuration the paper calls the brain foundation model; a smaller 2D autoencoder and plain ResNet variants are used as comparisons. The clinical baseline is an XGBoost gradient-boosted tree, the same model class as the prior clinical-data study this work extends.","core_discovery":"The central claim is that multimodal fusion of clinical and imaging data improves ICU admission prediction over the clinical-only baseline, provided the fusion is dynamic rather than naive. The authors show that concatenating tabular data with 2D autoencoder latents and feeding them into XGBoost or ResNet does not beat the clinical-only gradient-boosted tree. The result that does beat it comes from the Dynamic Affine Feature Map Transform (DAFT), which scales and shifts image feature maps according to the patient's clinical data, using latents extracted by a pretrained 3D masked autoencoder from the tumor-centered ROI of a T1 scan. With pre- and post-operative data this configuration reaches F1 0.41 and ROC-AUC 0.76, compared with F1 0.37 and ROC-AUC 0.77 for the clinical-only baseline; the authors also report gains in the pre-operative-only setting. They describe this as the first multimodal approach to postoperative ICU admission prediction.","pith_inferences":["The paper leaves untested whether the reported F1 differences are statistically reliable; a bootstrap or repeated cross-validation estimate of the confidence interval around the 0.04 gain would show whether the multimodal advantage is distinguishable from sampling noise.","A natural extension is external validation on multi-center data or on other MRI sequences, which would test whether the DAFT-plus-foundation-model benefit generalizes beyond this single-institution cohort and T1 protocol.","The same conditioning recipe, tabular features modulating image feature maps, could be tested in other rare-outcome medical prediction tasks such as sepsis, stroke, or ICU readmission, where clinical variables are strong but imaging adds complementary risk information."],"forward_implications":["If the gain holds, post-operative triage can be personalized: patients the multimodal model scores as low-risk could be monitored outside the ICU, lowering cost and freeing beds.","The result implies that how modalities are fused matters as much as whether imaging is added, since concatenation made both XGBoost and ResNet worse than the clinical-only baseline.","The best pipeline depends on tumor segmentation and a large pretrained 3D image model, so segmentation quality and feature-extractor transferability become practical prerequisites.","Given the severe class imbalance and a best F1 of 0.41, the model still leaves substantial room for missed true ICU patients; clinical review remains necessary."],"supporting_citations":[{"why":"Defines the clinical cohort, the ICU-admission label based on nine post-operative events, and the XGBoost clinical-data baseline that this study extends.","marker":"[5]"},{"why":"Introduces the Dynamic Affine Feature Map Transform (DAFT) used to fuse tabular clinical data with imaging data.","marker":"[6]"},{"why":"Supplies the pretrained 3D masked autoencoder whose latents produce the best multimodal results.","marker":"[8]"},{"why":"Provides the nnU-Net tumor segmentation masks used to define the 2D slice and 3D ROI for image feature extraction.","marker":"[4]"}],"fun_headline_variants":["Dynamic fusion of MRI and clinical data boosts ICU prediction","Multimodal model with DAFT improves brain surgery ICU forecasts","Imaging plus clinical data lifts ICU admission F1 to 0.41","DAFT fuses MRI and tabular data for better ICU triage"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes that the reported F1 differences (0.01 with pre-operative data, 0.04 with pre- and post-operative data) reflect real population differences rather than sampling noise, because the paper presents point estimates without confidence intervals or significance tests.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic fusion of MRI and clinical data boosts ICU prediction","Multimodal model with DAFT improves brain surgery ICU forecasts","Imaging plus clinical data lifts ICU admission F1 to 0.41","DAFT fuses MRI and tabular data for better ICU triage"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1241,"prompt_tokens":903,"completion_tokens":338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":263}},"tokens_in":519,"tokens_out":338,"duration_ms":3640,"temperature":1.0,"reasoning_tokens":263,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:03:21.977896+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeated resampling of the same five-fold cross-validation, such as 100 bootstrap iterations, that places a 95% confidence interval around the F1 difference between DAFT-with-3D-SSL and clinical-only GBT and finds it contains zero would falsify the claim that the multimodal model outperforms the baseline; a similarly sized external cohort showing no F1 improvement would also falsify it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the clinical cohort, the ICU-admission label based on nine post-operative events, and the XGBoost clinical-data baseline that this study extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Dynamic Affine Feature Map Transform (DAFT) used to fuse tabular clinical data with imaging data."},{"cited_title":"iScience 27(2), 109023 (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the nnU-Net tumor segmentation masks used to define the 2D slice and 3D ROI for image feature extraction."}],"review_version":1}