{"id":"5b76d3b0-8d37-4b5a-a54a-46b39f64bb7a","arxiv_id":"2502.05950","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SurvCBM and SurvRCM combine concept bottleneck learning with Cox and Beran survival models, and SurvCBM achieves the best C-index and concept F1 on synthetic MNIST and CIFAR experiments.","lead":"The authors propose two concept-based survival models, SurvCBM and SurvRCM, that predict censored event times using human-understandable concepts as intermediate variables. The paper shows that the bottleneck variant, SurvCBM, outperforms a concept-regularized variant and a non-concept baseline on synthetic image datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The experiments generate event times from the same concept labels used to supervise SurvCBM, so its large C-index gains over SurvBase may be a structural artifact; real-data evidence is needed before claiming concept information improves survival prediction.","rationale":"The reader identifies the same load-bearing weakness: the synthetic datasets are constructed so that event times depend directly on the concepts used to supervise SurvCBM, which structurally favors the proposed model. I agree with that assessment. The paper is a reasonable engineering contribution with public code, and the model descriptions are mostly clear, but the headline claim about the importance and advantage of concept information is not established by the present experiments. The MNIST-sin variant uses sin(b^T c), but it still ties the target to the same concept vector, so it does not break the structural advantage. I also note a secondary consistency issue: Section 5.1 lists MNIST generation coefficients (0.5, 1.5, −1, 0.001), while Section 6 uses (0.5, 1.5, 0.0001, 0.001) and different λ/ν for the interpretation examples; this complicates reproduction but does not by itself invalidate the performance comparison. The most decisive remedy is a real-data validation with standard survival baselines, which would test whether the advantage persists when concepts are not the exact generative covariates. Until that is provided, the conditional verdict is appropriate.","tokens_in":15238,"tokens_out":4838,"duration_ms":50613,"concrete_test":"Evaluate SurvCBM and SurvBase on a real censored dataset with human-interpretable concepts, e.g., GBSG2/Rotterdam breast cancer with clinical concepts (grade, ER status, tumor size) or MIMIC-IV with comorbidity concepts, adapting the concept predictor to a tabular or multimodal encoder as needed. Compare C-index on a held-out test set against Cox proportional hazards, DeepSurv, and random survival forests under the same hyperparameter search budget. If SurvCBM does not at least match the best baseline, the claim that concept information improves survival prediction accuracy is not supported outside the synthetic generative setup.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on Eqs. (13)–(14), where the synthetic event time T is a deterministic function of the same concept vector c that is supplied as supervision to the concept predictor. For the MNIST setting, T = (−ln u / (λ exp(b^T c)))^(1/ν), so the Cox variant of SurvCBM—whose risk score is a learned linear function of concept logits—can in principle recover the exact data-generating risk. SurvBase, by contrast, must discover the relevant information from raw pixels without concept labels. The large C-index gap in Figs. 6 and 8 therefore measures how much the experiment embeds the target in the concept labels, not whether concept bottlenecks improve survival predictions on data whose event times arise from arbitrary image features. No real censored dataset is used, and Section 7 nevertheless extrapolates to the broad 'crucial role' of concept information in improving accuracy. This is the load-bearing assumption: if real survival outcomes are not generated by predefined concept covariates, the reported advantage may vanish.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SurvCBM and SurvRCM, two neural architectures that combine concept-based learning with survival analysis. SurvCBM uses predicted concept logits as the input to a Cox proportional hazards head or a Beran estimator head; SurvRCM uses concept logits only as an auxiliary classification loss while the survival head is trained on a continuous embedding. Both models are trained end-to-end with a loss that balances a smoothed C-index term and a concept cross-entropy term. The paper also presents two interpretation methods: Cox-coefficient contributions and a nearest-instance concept-matching procedure for the Beran estimator. Experiments on synthetic MNIST, MNIST-sin, and CIFAR-10 image datasets compare C-index and concept F1 across sample sizes and censoring proportions, reporting that SurvCBM outperforms SurvRCM and the concept-free SurvBase baseline.","tokens_in":15474,"tokens_out":4743,"duration_ms":48504,"significance":"If the reported results hold, this is a useful and timely first integration of concept bottleneck models with censored survival data, and the public code is a concrete asset. The end-to-end training formulation and the two interpretability mechanisms are clearly relevant to medical and reliability applications. However, the central comparative claim is currently supported only by synthetic data in which the event time is generated from the same concept vector that supervises the concept predictor, and the survival loss is not fully specified because the expected event times are never defined. These two gaps are load-bearing for the abstract and Section 7 claims that concept information substantially improves survival prediction accuracy.","major_comments":[{"comment":"The smoothed C-index loss in Eq. (12) uses estimated expected event times \\hat T_i, but the paper never defines how \\hat T_i is computed from the predicted survival function for either the Cox or Beran head. Since L_surv is the only survival term in Eq. (10), the training objective is underspecified. Please provide the exact formula (e.g., numerical integration of S(t), median survival, or restricted mean survival time) and state how it is made differentiable for end-to-end training.","section":"4.2, Eq. (12)"},{"comment":"The synthetic event times are generated as deterministic functions of the same concept vector c that is supplied as supervision to the concept predictor. In the MNIST and CIFAR-10 settings, SurvCBM's Cox head is a linear function of concept logits and can in principle recover the exact data-generating risk, while SurvBase must discover the relevant information from raw pixels. The large C-index gaps in Figs. 6 and 8 may therefore reflect the experimental design rather than a general benefit of concept information. No real censored dataset is used, so the abstract and Section 7 claims about the crucial role of concept information are not established for data whose event times are not driven by predefined concepts; real-data validation or additional generative settings with event times depending on image features not fully captured by the concepts would be needed.","section":"5.1, Eqs. (13)–(14), and Section 7"},{"comment":"The text states that different hyperparameter values are tested and those leading to the best results are chosen, and all reported metrics are computed on the test set. Selecting hyperparameters on the test set and then reporting test metrics on the same set, even with 100 repetitions, can produce optimistic estimates and may favor SurvCBM if it has more tunable components. Please use validation-based selection or nested cross-validation, and report the selected hyperparameter values for reproducibility.","section":"5, hyperparameter selection"},{"comment":"The claim that LCE in SurvRCM 'does not directly influence the embeddings z' is questionable, because the concept logits p are produced from z by fully connected networks, so the gradient of LCE backpropagates through those networks into z. If a stop-gradient is applied before the concept heads, this should be stated explicitly; otherwise the description of SurvRCM as a regularized model with unaffected embeddings is inaccurate.","section":"4.3"}],"minor_comments":[{"comment":"The first Beran-based explanation method is described inconsistently: it first says neighbors are selected by closeness in concept space, then defines closeness by the distance between predicted survival functions, and Eq. (15) uses distances between logits. Please state exactly which distance is used to select the nearest instances.","section":"6"},{"comment":"The MNIST generation vector b is given as (0.5, 1.5, −1, 0.001) in the experimental setup, but Section 6 uses b = (b_1, ..., b_4) with each b_i ∈ R^10. Please clarify whether concepts enter Eq. (13) as scalar digit values or as one-hot vectors, since this changes the data-generating model.","section":"5.1 and 6"},{"comment":"The denominator of Eq. (12) sums pairs with δ_j = 1 and T_j < T_i; for small batches or certain censoring patterns this set can be empty, leading to division by zero. Please specify the handling of empty pair sets and ties in event times.","section":"4.2, Eq. (12)"},{"comment":"The abstract says the models outperform 'traditional survival models', but the experiments compare only SurvBase and SurvRCM. Either add standard survival baselines (e.g., DeepSurv, random survival forests, or Cox on raw features) or temper the claim to 'the compared baselines'.","section":"Abstract and 7"},{"comment":"There are several typos and inconsistencies to clean up: 'impove', 'significanlty', 'CIF AR-10' versus 'CIFAR-10', the phrase 'To implement We apply Lsurv' in Section 4.2, and the label 'Mixture model' used in Fig. 7 for what the text calls SurvRCM.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the proposed models are sensible, but the evaluation is entirely synthetic with event times generated from the same concepts used for supervision. I would recommend the authors add at least one real censored dataset and specify the expected-time computation before the comparative claim can be accepted. The priority claim of being the first survival concept-based model should also be checked against recent literature beyond Forest et al. and the cited CBM variants."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before citing it as evidence that concept bottlenecks help survival prediction: the experiments stack the deck. Event times are generated from the same concept labels that supervise the concept predictor, so SurvCBM is essentially reading off the generative mechanism. That doesn't mean the paper is worthless, but the strong claim in the abstract and conclusion is not supported by the reported evidence.\n\nWhat's actually new: this is the first combination of concept-based learning with censored survival analysis, using Cox and Beran heads, trained end-to-end with a smoothed C-index loss. The architectures are clearly described, the code is public, and the idea is natural. The two interpretation mechanisms (Cox coefficient contributions and instance-based SF proximity) are reasonable. Credit where due: the authors correctly identify Forest et al. as the closest CBM-for-prognostics work and note it doesn't handle censoring. The paper is readable and the experiments are extensive in sweep, if not in realism.\n\nSoft spots:\n\n1. The main quantitative claim is undercut by the data generation. In Eqs. (13) and (14), T is a deterministic function of the concept vector c. SurvCBM's Cox head sees concept logits trained to predict that exact c, so it can approximate the true risk up to noise. SurvBase must infer the hidden concept structure from raw pixels alone. The large C-index gap in Figs. 6 and 8 therefore measures how directly the simulation encodes the target in the supervised concepts, not whether concepts help in general. The paper extrapolates to a 'crucial role' of concepts without any real censored dataset. This is the load-bearing flaw.\n\n2. Technical gaps. The expected event times \\hat T_i used in the smoothed C-index loss are never defined. The hyperparameters are selected as the ones giving 'the best results' with all metrics computed on the test set; that invites overfitting. And the interpretation experiments use inconsistent coefficient descriptions: Section 5.1 says the MNIST generation vector is (0.5, 1.5, -1, 0.001), but Section 6 later says MNIST event times were generated with coefficients (0, 7, 10, 14). That discrepancy needs to be resolved.\n\n3. The interpretability section is illustrative only; no quantitative measure of explanation quality is reported.\n\nThe central idea is a reasonable engineering contribution, and with real-data validation, honest baselines, and a clear specification of the loss it could be a solid paper. As it stands, the evidence doesn't support the abstract's conclusion. I'd send it to peer review—the novelty is there and the problems are fixable—but I'd expect a major revision, not acceptance. For your own work, cite it if you need the first CBL+survival reference, but don't rely on its comparative claims.","headline":"The first concept-bottleneck-plus-survival model is a real novelty, but its headline accuracy claim is an artifact of generating event times from the very concepts the model is supervised to predict.","tokens_in":15973,"tokens_out":2957,"would_cite":true,"duration_ms":27228,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62N01","62N05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Survival prediction gains concept bottleneck models whose risk scores can be read off concept by concept.","keywords":["concept-based learning","survival analysis","concept bottleneck model","Cox proportional hazards model","Beran estimator","censored data","interpretability","neural networks"],"falsifier":"Run SurvCBM against SurvRCM and SurvBase on a real censored dataset with concept annotations, such as a medical cohort with expert-defined risk factors and follow-up times, and measure the C-index difference; if the advantage shrinks to zero or reverses when event times are not generated from the concepts, the paper's central claim of concept-driven accuracy improvement is falsified.","tokens_in":15060,"feed_emoji":"⏳","tokens_out":5727,"duration_ms":50827,"temperature":0.7,"pith_summary":"This paper tries to bring concept-based learning into survival analysis, where the target is a time to an event and some observations are censored. It proposes two neural architectures, SurvCBM and SurvRCM, that predict survival functions from high-level concepts and are trained end-to-end with a loss combining the smoothed C-index and concept classification cross-entropy. The paper's central assertion is that the bottleneck version, SurvCBM, in which the Cox or Beran survival model is trained on predicted concept logits, outperforms both the regularized SurvRCM and a concept-free baseline on synthetic composite image datasets. If true, this means that explicitly representing human-understandable concepts does not merely make survival predictions explainable; it can also improve their accuracy. The paper also proposes two ways to read concept relevance from the survival prediction, one based on Cox regression coefficients and one based on nearest neighbors in survival-function space.","feed_headline":"SurvCBM beats concept-free survival models on synthetic tests","feed_subtitle":"Concept bottlenecks let Cox and Beran survival heads predict event times and explain risk by concept.","key_machinery":"The load-bearing object is the concatenated concept-logit vector $p$ formed from the outputs of $m$ concept CNNs; this vector is the input to the survival head, which is either the Cox model with hazard $H(t \\mid x, b) = H_0(t)\\exp(b^T p)$ or the Beran estimator with Gaussian-kernel weights computed on $p$. Training minimizes $L = -\\alpha L_{\\text{surv}} + (1-\\alpha) L_{\\text{CE}}$, where $L_{\\text{surv}}$ is a differentiable smoothed C-index and $L_{\\text{CE}}$ is cross-entropy for concept classification. The bottleneck forces all survival information to flow through the concepts, which is why the model can both produce survival functions and explain them by concept contributions.","core_discovery":"The central discovery claimed is that the concept bottleneck model, previously used for classification, can be adapted to censored time-to-event data by placing a survival head after the concept bottleneck. In SurvCBM the input images are mapped by separate CNNs to concept logits; these logits are concatenated and passed to either the Cox proportional hazards model or the Beran estimator, which outputs a survival function. Each concept contributes a linear term in the Cox log-risk, and the Beran estimator weights training instances by similarity of concept logits, so predictions stay interpretable in terms of concepts. Numerical experiments on synthetic MNIST and CIFAR-10 composites show SurvCBM achieving higher C-index than SurvRCM and SurvBase, with the largest gains when the Cox model is used on Weibull-generated event times.","pith_inferences":["The reported superiority may depend on the generative match: in these synthetic datasets the event time is a function of the very concepts the model is trained to predict, so the advantage could shrink on real data where concepts are imperfectly correlated with event time.","Because the paper uses only synthetic composite images, the most direct testable extension is to apply SurvCBM to a real medical or reliability dataset with expert-annotated concepts and compare its C-index against concept-free deep survival models.","The Cox-based interpretation method inherits the proportional-hazards assumption; on datasets where hazards cross, the Beran-based nearest-neighbor explanation would likely be more dependable, a comparison the paper does not make.","Continuous or time-varying concepts, which the paper names as future work, might be needed for reliability applications where system components degrade gradually rather than being described by categorical attributes."],"forward_implications":["If SurvCBM's reported advantage holds, then collecting concept labels can improve the accuracy of survival models, not just their interpretability, on tasks where concepts align with the drivers of event times.","The Cox-based SurvCBM yields concept relevance scores as products of learned coefficients and softmax concept probabilities, so a risk prediction can be audited concept by concept.","The Beran-based SurvCBM offers example-based explanations: the training instances with survival functions closest to the explained instance show which concepts are doing the work.","The paper's negative result for SurvRCM indicates that using concepts only as a regularizer is weaker than a true bottleneck for survival tasks, so future concept-based survival models should route predictions through concepts.","The claim of being the first integration of concept-based learning with survival analysis establishes a baseline architecture against which later concept-based survival models can be compared."],"supporting_citations":[{"why":"Supplies the concept bottleneck model architecture that SurvCBM adapts by placing a survival head after the concept logits.","marker":"[12]"},{"why":"Defines the Cox proportional hazards model used as the linear-risk survival head and as the basis for coefficient-based concept interpretation.","marker":"[23]"},{"why":"Defines the Beran estimator used as the nonparametric survival head and as the basis for nearest-neighbor example-based explanations.","marker":"[25]"},{"why":"Provides the survival analysis definitions of censoring, survival functions, and the Nelson-Aalen estimator used throughout the paper.","marker":"[21]"},{"why":"Formalizes the C-index used as the survival accuracy measure and as the smoothed differentiable loss term in training.","marker":"[66]"},{"why":"Surveys survival machine learning and contextualizes the C-index and the comparison against traditional survival models.","marker":"[38]"},{"why":"Supplies the MNIST digit images from which the synthetic concept-based dataset is constructed.","marker":"[67]"},{"why":"Supplies the CIFAR-10 images from which the synthetic concept-based dataset with animal and vehicle concepts is constructed.","marker":"[68]"},{"why":"Represents the prior concept bottleneck model for remaining useful life that the paper distinguishes from its own because it does not handle censored data.","marker":"[22]"}],"fun_headline_variants":["SurvCBM brings concept bottlenecks to survival analysis","Concept-based survival models outperform on synthetic tests","SurvCBM: interpretable survival predictions via concept bottlenecks","Cox and Beran heads learn survival from concept bottlenecks","Survival analysis adapts concept-based learning for event times"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative case for SurvCBM depends on synthetic datasets in which the true event time is generated from a linear combination of the same predefined concepts the model is trained to predict; on real censored survival data, where event times may not be driven by those concepts, the claimed accuracy advantage may not hold.","fun_headline_variants_meta":{"raw":{"variants":["SurvCBM brings concept bottlenecks to survival analysis","Concept-based survival models outperform on synthetic tests","SurvCBM: interpretable survival predictions via concept bottlenecks","Cox and Beran heads learn survival from concept bottlenecks","Survival analysis adapts concept-based learning for event times"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000348,"raw_usage":{"total_tokens":1884,"prompt_tokens":909,"completion_tokens":975,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":897}},"tokens_in":525,"tokens_out":975,"duration_ms":7366,"temperature":1.0,"reasoning_tokens":897,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:16:10.954771+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SurvCBM against SurvRCM and SurvBase on a real censored dataset with concept annotations, such as a medical cohort with expert-defined risk factors and follow-up times, and measure the C-index difference; if the advantage shrinks to zero or reverses when event times are not generated from the concepts, the paper's central claim of concept-driven accuracy improvement is falsified.","supporting_citations":[{"cited_title":"Mussmann, E","cited_arxiv_id":null,"evidence_quote":"Supplies the concept bottleneck model architecture that SurvCBM adapts by placing a survival head after the concept logits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Beran estimator used as the nonparametric survival head and as the basis for nearest-neighbor example-based explanations."},{"cited_title":"Hosmer, S","cited_arxiv_id":null,"evidence_quote":"Provides the survival analysis definitions of censoring, survival functions, and the Nelson-Aalen estimator used throughout the paper."},{"cited_title":"Uno, Tianxi Cai, M.J","cited_arxiv_id":null,"evidence_quote":"Formalizes the C-index used as the survival accuracy measure and as the smoothed differentiable loss term in training."},{"cited_title":"Krizhevsky and G","cited_arxiv_id":null,"evidence_quote":"Supplies the CIFAR-10 images from which the synthetic concept-based dataset with animal and vehicle concepts is constructed."},{"cited_title":"Interpretable Prognostics with Concept Bottleneck Models","cited_arxiv_id":"2405.17575","evidence_quote":"Represents the prior concept bottleneck model for remaining useful life that the paper distinguishes from its own because it does not handle censored data."}],"review_version":1}