{"id":"b5b98455-4256-49a2-a46f-f55d0ef31ccd","arxiv_id":"2501.11715","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A glass-box CNN-EBM hybrid diagnoses Alzheimer's disease and predicts progression with accuracy on par with black-box models while exposing the brain regions behind each decision.","lead":"This paper introduces GL-ICNN, a model that combines a convolutional neural network with an explainable boosting machine to diagnose Alzheimer's disease from MRI scans while showing which brain regions influenced each decision. The model matches the accuracy of black-box deep learning models on ADNI and an external cohort.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's CNN update requires gradients through tree-ensemble EBM outputs, which are piecewise constant; absent a surrogate, the central 'end-to-end' training claim is underspecified.","rationale":"The reader's weakest assumption identifies exactly the point on which the paper's central claim depends. The paper says the model is end-to-end and alternatingly optimizes CNN and EBM weights, but the EBM is an ensemble of decision trees with zero gradients almost everywhere. The text offers no gradient surrogate or alternative learning rule for the CNN update. This is not a disagreement with external consensus; it is an internal gap in the method description. If the code contains a working surrogate, the concern is resolved and should be documented in the paper. If not, the model cannot be trained as written, and the performance results are not reproducible from the textual description alone. The concern does not necessarily falsify the empirical results, but it does undermine the methodological claim that the proposed training strategy produces an end-to-end interpretable model. The reader's conditional verdict is appropriate: the paper should be accepted only after the training mechanism is clarified or corrected, and the external validation results should be reported with the same detail as the ADNI results. No harsher verdict is warranted because the available code may resolve the issue, and the empirical comparison is otherwise plausible.","tokens_in":5753,"tokens_out":2634,"duration_ms":30542,"concrete_test":"Inspect the publicly released code linked in the abstract, specifically the optimizer or update routine corresponding to Algorithm 1 step 3. Determine how the loss for the CNN is computed while EBM weights are frozen: if it backpropagates through the EBM, verify that a differentiable approximation of the tree ensemble is implemented; if it instead uses EBM predictions as fixed targets or a separate distillation loss, then the model is not trained end-to-end through the glass-box output block as claimed. Run the update on one batch and check whether the gradient norm with respect to the CNN parameters is nonzero.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that GL-ICNN is an end-to-end, glass-box model performing comparably to black-box models. The load-bearing condition for that claim is that Algorithm 1 is actually executable: in each epoch, after training the EBM with CNN features fixed (step 2), the CNN is trained with EBM weights fixed (step 3). This step requires a gradient of the loss with respect to the CNN parameters. Because the EBM's shape functions are ensembles of bagged decision trees, they are piecewise constant functions of their inputs; the gradient of the EBM output with respect to the CNN feature vector is zero almost everywhere. Standard backpropagation through the EBM therefore supplies no learning signal to the CNN. The paper does not describe a differentiable surrogate, a straight-through estimator, a soft-tree relaxation, finite-difference approximation, or any alternate objective for step 3. The only loss mentioned, the weighted cross entropy on the validation set (Algorithm 1, line 4), is used for early stopping, not for updating the CNN. Without a specified mechanism for step 3, the end-to-end training procedure is incomplete as stated, and the central claim that the model is trained end-to-end is not supported by the text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GL-ICNN, an interpretable model that replaces the fully connected output block of a CNN with an Explainable Boosting Machine (EBM), aiming to provide end-to-end learning from raw MRI while retaining glass-box interpretability. The authors propose a block-coordinate alternating training scheme (Algorithm 1) in which the CNN acts as a feature extractor and the EBM as the output block. Validation is performed on the ADNI cohort (AD vs. CN classification and MCI-to-AD conversion prediction) and on the PND cohort as external testing. The reported AUCs are 0.956 for AD/CN and 0.694 for MCI conversion on ADNI, with feature importance maps highlighting temporal lobe regions. The model is compared against several black-box baselines and two interpretable baselines, and the authors report a large reduction in training time relative to their earlier non-end-to-end Glo&Loc-EBM.","tokens_in":5982,"tokens_out":6319,"duration_ms":59046,"significance":"If the proposed training procedure is well-defined and executable, GL-ICNN would be a valuable contribution to interpretable medical image analysis, offering both decisions and feature importance measures without sacrificing accuracy relative to black-box models. The strengths of the paper include the use of an external validation cohort, the public availability of code, and the clear experimental protocol with bootstrap confidence intervals and DeLong tests. However, the central methodological claim of end-to-end training is currently undermined by an incomplete specification of the gradient path from the EBM output to the CNN parameters. The paper's significance is therefore conditional on resolving this algorithmic gap.","major_comments":[{"comment":"The instruction to 'train the CNN component of GL-ICNN while keeping EBM weights fixed' is not executable as stated. The EBM shape functions in Eq. (1) are gradient-boosted ensembles of bagged decision trees, which are piecewise constant functions of their inputs. Consequently, the gradient of the EBM output with respect to the CNN output vector x_j is zero almost everywhere, so standard backpropagation supplies no learning signal to the CNN. The paper does not describe a differentiable surrogate (e.g., soft-tree relaxation, straight-through estimator, or finite-difference approximation) or any alternative loss for updating the CNN in step 3. Please specify the exact update rule used and justify how gradients are obtained; without this, the central claim of an end-to-end model trained by alternating optimization is unsupported.","section":"II-B (Algorithm 1, step 3)"},{"comment":"Related to the previous point, the alternating scheme as described is degenerate: step 2 trains the EBM on the current CNN features, but if step 3 cannot propagate gradients through the EBM, the CNN features never change in response to the EBM loss. In that case the CNN parameters are only updated during the warm-up phase, and the EBM is trained once on fixed features, which is not end-to-end learning. Please clarify whether the CNN is actually updated during the alternating loop and, if so, by what mechanism. If the CNN is updated using a surrogate or a separate objective, describe it explicitly and justify its use.","section":"II-B (block coordinate descent)"},{"comment":"The abstract and conclusion claim that GL-ICNN achieves 'comparable performance with other state-of-the-art black-box models.' However, the comparison study only includes the authors' own GL-CNN, VGG, DenseNet, and GL-ICNN-L. No published state-of-the-art results on ADNI or PND are cited or compared, and the DeLong test only establishes non-significant differences relative to these baselines. Please either add comparisons with recently published state-of-the-art models on the same cohorts or temper the claim to 'comparable performance with the included black-box baselines.'","section":"IV-A (performance comparison)"}],"minor_comments":[{"comment":"The early stopping condition compares loss_i to loss_{i-1} rather than to the best loss encountered so far. This can save a model that is worse than a previously seen model. Consider storing a loss_best variable and comparing to it.","section":"II-B (Algorithm 1, lines 5-7)"},{"comment":"The dataset notation D = {(X_i, y_i)}_1^N uses N for the number of subjects, while N is also used for the number of features in the same line ('i ∈ [1, N]' and 'x_j is the jth feature'). Use distinct symbols to avoid ambiguity.","section":"II-A (Eq. 1 and surrounding text)"},{"comment":"The model named 'GL-ICNN-L' is not an EBM-based model; it is a CNN with a linear output block. The name may confuse readers because it suggests a variant of GL-ICNN rather than a different architecture. Consider renaming it (e.g., GL-CNN-L or GL-Linear).","section":"III-B"},{"comment":"The description of the patch-based local inputs is vague: 'The entire brain image is divided into non-overlapping patches that collectively cover the whole image.' Please specify the number of patches, patch sizes, and how the global and local CNNs are combined.","section":"II-A"},{"comment":"The text states that error bars represent confidence intervals but does not state the type (e.g., 95% bootstrap CIs). Also, the PND AUC values are not reported numerically; please include them in the text or a table.","section":"Fig. 2 and Section III-C"},{"comment":"The phrase 'the GL-ICNN based the diagnosis mostly on regions of the temporal lobes, especially all regions containing the hippocampus and amygdala appeared' is awkward and slightly unclear; please rephrase for clarity.","section":"IV-B"},{"comment":"Reference [8] is to an arXiv preprint; if it has been published in the meantime, please cite the published version.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper aims at something useful: an interpretable model that combines a CNN feature extractor with an Explainable Boosting Machine (EBM) output, trained alternately, and it includes external validation on an independent cohort plus a public code link. If it worked as described, it would lower a real barrier to clinical adoption of deep learning for AD.\n\nThe catch: Algorithm 1, step 3, says to train the CNN while keeping the EBM fixed. The EBM shape functions are ensembles of decision trees, which are piecewise constant—gradients with respect to the CNN features are zero almost everywhere. The paper does not describe a differentiable surrogate, a straight-through estimator, or any alternate loss that would supply a learning signal. The only loss mentioned, a weighted cross entropy on the validation set, is used for early stopping, not for updating the CNN. So the central claim of end-to-end training is not supported by the text. This is a load-bearing gap, not a cosmetic one.\n\nThat said, the paper deserves a serious referee. The architecture is sensible, the problem is important, and the results are plausible: AUC 0.956 for AD/CN on ADNI, comparable performance to black-box models, and a 10x training time reduction over the authors' earlier non-end-to-end pipeline. The group-level feature importance pointing to hippocampus and amygdala is a good sanity check. The public code means a reviewer can verify whether the implementation actually matches the description.\n\nMinor issues: hyperparameters (Nmax, Ntolerate, warm-up epochs, patch sizes) are said to be tuned but not reported. External PND results are in the figures but not in the text. The comparison with black-box models uses only five repetitions; the 'not significantly different' claim is reasonable but not definitive.\n\nMy verdict: send it to peer review, but insist on a clear explanation of how the CNN update works. If the authors can provide that, the contribution is solid. If not, the model reduces to a staged feature-extractor-plus-EBM pipeline, which is much closer to their prior work. I wouldn't cite it until the training mechanism is clarified, but it's a useful paper to discuss in a reading group precisely because of that gap.","headline":"Worth a careful peer review, but the end-to-end training claim rests on an unspecified gradient path through a tree ensemble.","tokens_in":6562,"tokens_out":3285,"would_cite":false,"duration_ms":31467,"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 proposes GL-ICNN, an end-to-end interpretable convolutional network that combines CNN feature extraction with an explainable boosting machine, and claims it matches black-box models on Alzheimer's diagnosis while providing…","keywords":["Alzheimer's disease","MRI","convolutional neural network","explainable boosting machine","interpretable machine learning","MCI conversion prediction","feature importance","end-to-end training"],"falsifier":"Run Algorithm 1 on a small labeled MRI set, record the gradient norm of the CNN weights during step 3; if every gradient is zero or the CNN weights never change while the EBM is fixed, then the end-to-end training claim is false as stated and the reported performance must come from the warm-start or some unstated surrogate.","tokens_in":5554,"feed_emoji":"🧠","tokens_out":6784,"duration_ms":62069,"temperature":0.7,"pith_summary":"The paper claims that a convolutional neural network for MRI can be made fully interpretable without sacrificing accuracy by mounting an Explainable Boosting Machine (EBM) as its output layer. On Alzheimer's disease versus control classification it reports an AUC of 0.956 on ADNI, and for predicting conversion from mild cognitive impairment to Alzheimer's disease an AUC of 0.694, with no statistically significant difference from black-box CNN baselines on any of the four tasks tested. The model also produces per-patient and group-level feature importance for brain regions, and the regions it weights most heavily, temporal lobe structures including hippocampus and amygdala, match established clinical knowledge. If these claims hold, the practical barrier that interpretability forces a drop in diagnostic performance would be removed for this imaging task.","feed_headline":"Glass-box MRI model matches black-box Alzheimer's accuracy","feed_subtitle":"End-to-end CNN plus explainable boosting shows which brain regions drive each diagnosis and forecast.","key_machinery":"The load-bearing object is the EBM shape-function sum, $g(y)=\\beta+\\sum_j f_j(x_j)$, where each $f_j$ is a gradient-boosted ensemble of bagged decision trees operating on a single CNN-extracted feature $x_j$. Because the shape functions are univariate and additive, each subject's prediction decomposes into a vector of per-feature contributions, giving individual-level importance, and averaging absolute contributions over a cohort gives group-level importance. The training strategy is block-coordinate descent: in each epoch the EBM is retrained on the current CNN features, then the CNN is updated while the EBM stays fixed, with early stopping on validation loss.","core_discovery":"The central discovery is that an EBM can act as a transparent output block for a CNN and that the whole stack can be trained end-to-end by alternating the optimization: freeze the CNN and fit the EBM to its features, then freeze the EBM and update the CNN, repeating each epoch. The model is called GL-ICNN; it takes whole-image and region-level patches, learns image features with convolutional backbones, and expresses each prediction as an additive sum of per-feature contribution functions. The paper reports that this glass-box model performs statistically on par with the black-box GL-CNN, VGG, and DenseNet baselines, and that its training time is about ten times shorter than the earlier non-end-to-end version because the separate feature-selection stage is removed.","pith_inferences":["A natural extension: the shape functions explain features produced by the CNN, not the gray-matter voxels directly; whether this counts as a clinical explanation depends on how stably the CNN maps anatomy into those features.","Because the EBM is additive, the comparable accuracy suggests that pairwise and higher-order interactions among CNN features are not critical for these tasks; one could test this by adding interaction terms or comparing with a non-additive glass-box model.","The alternating training as written requires gradients to flow through the EBM, whose decision trees have zero gradients almost everywhere; if backpropagation is used unmodified, step 3 cannot change the CNN, so an implementable version must rely on a surrogate, a different update rule, or a detail not stated in the paper.","A testable practical claim: if the training procedure is implemented as described, the CNN weights will be frozen during the EBM phase, making reported performance reproducible from the code; checking the released code would settle this."],"forward_implications":["A clinician can see exactly which brain regions pushed a given patient's score, because the EBM output is an additive sum of per-patch contributions.","End-to-end training removes the expensive feature-selection stage, cutting training time from 81.3 hours to 7.8 hours on the AD-CN ADNI task.","The interpretable model need not cost accuracy: its AUC was statistically indistinguishable from black-box baselines on all four diagnosis and prediction tasks.","External-cohort testing on PND suggests the interpretable model generalizes beyond the training population.","The top-ranked regions, including temporal lobe, hippocampus, and amygdala, provide a clinically plausible anatomical check on model behavior."],"supporting_citations":[{"why":"Defines the Explainable Boosting Machine, the glass-box output block whose additive shape functions make predictions interpretable.","marker":"[7]"},{"why":"Describes the authors' earlier non-end-to-end CNN-EBM model whose separate feature-selection and extraction steps GL-ICNN removes.","marker":"[8]"},{"why":"Supplies the DenseNet backbone used for the global whole-image stream.","marker":"[9]"},{"why":"Supplies the VGG-based local backbones used for brain-region patches.","marker":"[10]"},{"why":"Establishes the use of EBM feature-importance measures in healthcare, supporting the paper's interpretability claim.","marker":"[11]"},{"why":"Provides the EBM training algorithm that motivates the alternating block-coordinate optimization in Algorithm 1.","marker":"[12]"},{"why":"Documents the Parelsnoer biobank cohort used as an external testing set.","marker":"[13]"},{"why":"Provides the image-processing pipeline and cross-cohort performance reference used to validate the model.","marker":"[14]"},{"why":"Names image patches after brain regions via the Hammers atlas, linking feature importance to anatomy.","marker":"[15]"},{"why":"Supplies clinical knowledge about Alzheimer's disease that the top-ranked temporal-lobe regions are said to align with.","marker":"[16]"}],"fun_headline_variants":["Glass-box CNN equals black-box accuracy for Alzheimer's detection","End-to-end transparent model predicts Alzheimer's and shows why","Explainable AI pinpoints brain regions in Alzheimer's diagnoses","Fast interpretable Alzheimer's model matches opaque neural networks","CNN plus explainable boosting diagnoses Alzheimer's with clarity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training loop assumes gradients can flow from the EBM's loss back into the CNN through the EBM, but the EBM's shape functions are ensembles of decision trees, and tree outputs are piecewise constant with zero gradient almost everywhere, so the CNN-update step as written may be impossible without a surrogate.","fun_headline_variants_meta":{"raw":{"variants":["Glass-box CNN equals black-box accuracy for Alzheimer's detection","End-to-end transparent model predicts Alzheimer's and shows why","Explainable AI pinpoints brain regions in Alzheimer's diagnoses","Fast interpretable Alzheimer's model matches opaque neural networks","CNN plus explainable boosting diagnoses Alzheimer's with clarity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1707,"prompt_tokens":983,"completion_tokens":724,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":645}},"tokens_in":599,"tokens_out":724,"duration_ms":7792,"temperature":1.0,"reasoning_tokens":645,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:55:44.645471+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a small labeled MRI set, record the gradient norm of the CNN weights during step 3; if every gradient is zero or the CNN weights never change while the EBM is fixed, then the end-to-end training claim is false as stated and the reported performance must come from the warm-start or some unstated surrogate.","supporting_citations":[{"cited_title":"Intelligible models for classification and regression,","cited_arxiv_id":null,"evidence_quote":"Defines the Explainable Boosting Machine, the glass-box output block whose additive shape functions make predictions interpretable."},{"cited_title":"An Interpretable Machine Learning Model with Deep Learning-based Imaging Biomarkers for Diagnosis of Alzheimer's Disease","cited_arxiv_id":"2308.07778","evidence_quote":"Describes the authors' earlier non-end-to-end CNN-EBM model whose separate feature-selection and extraction steps GL-ICNN removes."},{"cited_title":"Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission,","cited_arxiv_id":null,"evidence_quote":"Establishes the use of EBM feature-importance measures in healthcare, supporting the paper's interpretability claim."},{"cited_title":"Accuracy, interpretability, and differential privacy via explainable boosting,","cited_arxiv_id":null,"evidence_quote":"Provides the EBM training algorithm that motivates the alternating block-coordinate optimization in Algorithm 1."},{"cited_title":"The Parelsnoer Institute: a national network of standardized clinical biobanks in the Netherlands,","cited_arxiv_id":null,"evidence_quote":"Documents the Parelsnoer biobank cohort used as an external testing set."},{"cited_title":"Cross-cohort generalizability of deep and conventional machine learn- ing for MRI-based diagnosis and prediction of Alzheimer’s disease,","cited_arxiv_id":null,"evidence_quote":"Provides the image-processing pipeline and cross-cohort performance reference used to validate the model."},{"cited_title":"Three-dimensional maximum probability atlas of the human brain, with particular reference to the temporal lobe,","cited_arxiv_id":null,"evidence_quote":"Names image patches after brain regions via the Hammers atlas, linking feature importance to anatomy."},{"cited_title":"Towards a biological definition of Alzheimer disease,","cited_arxiv_id":null,"evidence_quote":"Supplies clinical knowledge about Alzheimer's disease that the top-ranked temporal-lobe regions are said to align with."}],"review_version":1}