{"id":"1515d13d-f8f3-4438-99d2-46bfab62b899","arxiv_id":"2509.07108","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ADHAM combines additive per-covariate hazard functions with latent subgroup mixtures, and a post-training refinement merges similar subgroups without retraining.","lead":"ADHAM is a survival analysis model that mixes per-patient subgroup weights with covariate-specific hazard curves, offering population, subgroup, and individual level explanations in one framework. It matches leading interpretable survival models in accuracy and adds a post-training step that merges redundant subgroups.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Univariate hazard fits are marginal associations, not additive components; with correlated covariates, ADHAM's population-level explanations can misattribute risk.","rationale":"The reader's weakest assumption matches the load-bearing point. I considered the other flagged issues (unspecified regularizer coefficients, abstract overclaim, approximate subgroup merging) and none threaten the central construction as directly. The subgroup-merging proof (Appendix C.1) is exact for identical β rows and Table 1 suggests threshold robustness; missing coefficients are reproducibility issues, not correctness. The abstract overclaim is a presentation fix. The core mathematical construction in Eq. 4 and the importance-sampling likelihood is coherent. The univariate training step is the one place where the model's interpretability claim can fail even when the code runs and metrics look good: it presupposes that fitting each covariate alone recovers that covariate's role in the joint hazard. That assumption is false in general, and the paper provides no synthetic ground-truth validation for it. The proposed synthetic experiment would settle it cheaply. Hence CONDITIONAL remains appropriate; no verdict change, but the condition should explicitly require synthetic validation of component recovery.","tokens_in":26461,"tokens_out":8414,"duration_ms":87427,"concrete_test":"Run a controlled synthetic experiment with the public ADHAM code. Let D=2, generate X=(X1,X2) with correlation ρ∈{0,0.3,0.7}, and draw T from a known Weibull hazard λ(t|x)=λ0(t)exp(f1(x1)+f2(x2)) in Setting A, and from λ(t|x)=λ0(t)exp(f1(x1)) in Setting B, where X2 is irrelevant but correlated with X1. After fitting ADHAM, compute (i) the fitted λ_d and (ii) the true marginal hazard h_d(t|x_d) by numerical integration over X_-d conditional on T≥t. If λ_d≈h_d but λ_d deviates from the true additive component f_d, and in Setting B the fitted λ_2 shows a material risk gradient for the irrelevant covariate, then the univariate fits are marginal associations and the population-level interpretation misattributes risk. If λ_d tracks f_d regardless of ρ, the decoupled training recovers additive components and the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ADHAM's three-level interpretability claim depends on treating each λ_d(t|x_d), trained by maximizing the univariate likelihood ℓ_d in Eq. 9, as the covariate-specific hazard contribution in the joint additive mixture (Eq. 4). This is the load-bearing step. In expectation, the optimum of Eq. 9 is the marginal conditional hazard h_d(t|x_d)=E_{X_-d|T≥t,X_d=x_d}[λ_true(t|X_d,X_-d)], not the additive component f_d in a joint hazard λ_true(t|x)=Σ_d f_d(t,x_d). h_d includes, through the survivor-conditional distribution of X_-d, the averaged contributions of every other covariate. When covariates are correlated (the regime the paper aims to address), h_d can exhibit a strong gradient for a covariate with no direct effect, purely through confounding by X_-d. Decoupled training therefore does not 'purely capture covariate-specific trends'; it replaces the shared identifiability problem known as concurvity with a marginal-association bias. A second consequence is that Φ is never optimized against the full likelihood (Eq. 8): Algorithm 1 fixes Φ at univariate optima and fits only θ,β to the joint objective, so ADHAM is not a maximum-likelihood estimator of Eq. 8. The paper's limitation note that ADHAM is not causal does not resolve this, because the concern is about the faithfulness of the model's own additive decomposition, not about causal identification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADHAM, a survival model that combines per-covariate hazard networks with a latent subgroup assignment network. The marginal hazard is written as a weighted sum of univariate hazard functions, with subgroup-specific weights β and patient-specific assignment probabilities from a network f_θ. Training proceeds in two stages within a single loop: each univariate hazard λ(t|x_d;φ_d) is fit by maximizing its own marginal log-likelihood (Eq. 9), while the subgroup parameters θ,β are fit to the joint log-likelihood (Eq. 8). A post-training refinement merges subgroups whose β rows are similar, based on a correlation threshold. The paper claims interpretability at population, subgroup, and individual levels, and reports experiments on SUPPORT, FLCHAIN, and a CKD EHR dataset against ten baselines, concluding that regularized ADHAM is competitive with state-of-the-art interpretable survival models while also providing multi-level explanations.","tokens_in":26717,"tokens_out":7373,"duration_ms":69873,"significance":"If its interpretability claims are valid, ADHAM offers a practically useful combination: a single trained model that yields population-level covariate-risk curves, subgroup-specific importance profiles, and patient-level hazard decompositions, with a computationally cheap model-selection step. The paper has concrete strengths: code is provided; the likelihood is handled with the unbiased Monte Carlo estimator of Ketenci et al. (2023); the exact subgroup-merging proposition in Appendix C.1 is a correct algebraic identity; and the experimental appendix gives detailed hyperparameter settings and per-fold standard errors. The main risk is that the interpretability claim rests on treating univariate hazard fits as true additive components, which is not established for correlated covariates. The manuscript therefore needs substantial revision of its claims and additional validation before the central contribution can be regarded as supported.","major_comments":[{"comment":"The claim that decoupled training 'overcomes concurvity' and yields population-level hazard shapes that 'purely capture covariate-specific trends in data' is not supported. Each λ(t|x_d;φ_d) is trained by maximizing the univariate likelihood ℓ_d in Eq. (9), so in the population limit it approximates the marginal conditional hazard E_{X_-d | T≥t, X_d=x_d}[λ_true(t|X_d,X_-d)], not an additive component of a joint hazard. When covariates are correlated, this marginal object can exhibit a strong gradient for a covariate with no direct effect, simply through confounding by an influential omitted covariate. Consequently, the population-level explanations in Figures 2 and 6 are marginal associations, and the decomposition in Eq. (4) does not isolate covariate-specific contributions in the usual additive-model sense. A second consequence is that Algorithm 1 fixes Φ at univariate optima and fits only θ and β to the joint objective, so ADHAM is not a maximum-likelihood estimator of Eq. (8). The limitation note in Section 6 that ADHAM is not causal does not resolve this issue, because the concern is about the faithfulness of the model's own additive decomposition, not about causal identification. I would recommend either explicitly reframing the interpretability claims as descriptive of the model's weighting of marginal hazard curves, or adding a simulation study with known additive components and correlated covariates to show whether and when the univariate fits recover the true components.","section":"Section 3.3, Eq. (9), Algorithm 1"},{"comment":"The abstract's statement that ADHAM 'remains on par with existing state-of-the-art survival baselines' is stronger than the results support. The body in Section 5 and contribution 4 appropriately restrict the claim to state-of-the-art interpretable survival models, and even there ADHAM(R) is not always the best (TIMENAM and TIMENA2M win on several Brier Score and AUROC entries). Against the black-box DHA baseline, ADHAM(R) trails on Brier Score and AUROC at all three quantiles on SUPPORT (Table 2: 25th quantile BS 0.144 vs 0.138, AUROC 0.666 vs 0.672; 50th quantile AUROC 0.644 vs 0.650) and similar patterns appear in Tables 3 and 4 for CKD and FLCHAIN. Please align the abstract and the contribution list with the 'interpretable baselines' framing, and report which of the differences against DHA are statistically significant given the standard errors in Appendix G.","section":"Abstract and Section 5, Tables 2–4"}],"minor_comments":[{"comment":"The two regularization terms are subtracted from the objective, but no regularization weights or schedule is specified; please state how the magnitude of these terms is chosen and how they are combined with the log-likelihood.","section":"Eq. (10)"},{"comment":"The pseudo-code is hard to follow: the while condition `while T_{c*∈C*} c* ≠ {}` appears to contain corrupted notation, and the instruction `ρ_cc* ← Update entry to −∞` is underspecified. Please rewrite the merging procedure so that the pair collection and transitive merging steps are unambiguous.","section":"Algorithm 2"},{"comment":"The caption of Figure 2 refers to 'survival functions' but displays quantities labeled λ(t|x_id;φ_d), which according to Eq. (2) are hazards; please make the caption consistent with Eq. (2) and the survival function definition in Eq. (13).","section":"Figure 2 and Section 3.2"},{"comment":"In the notation table, the entry for D currently reads 'DEmpirical dataset', which merges the symbol with its meaning; please separate the symbol and the definition, and define C consistently as the number of subgroups throughout.","section":"Appendix A"},{"comment":"Statistical significance is assessed with a two-sided Welch t-test over only five folds; please report paired differences or confidence intervals where the bold/non-bold annotations are used, since some SEM values (e.g., ADHAM (R) on CKD at the 25th quantile, C-Index SEM 0.016) are large relative to the differences from the best baseline.","section":"Section 5 and Appendix G"}],"recommendation":"major_revision","confidential_remarks":"The core methodological weakness is that the univariate hazard fits are marginal associations, not additive components, and the manuscript's 'overcoming concurvity' framing overstates what decoupled training achieves. This is fixable by reframing the model as a mixture of marginal hazard curves and by adding a synthetic correlated-covariate experiment; the current text, however, makes the stronger claim. The abstract's predictive-parity statement also needs to be aligned with the interpretable-baseline results in Section 5. The paper is otherwise well organized, reproducible, and worth further consideration after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ADHAM is a genuine combination of known pieces—additive per-covariate hazards, latent subgroup mixtures, and post-training subgroup merging—and the whole ends up being more than the sum of its parts: one model that gives population-, subgroup-, and individual-level explanations. The math is straightforward, the code is public, and the experiments are careful, with standard errors and multiple datasets. That is a legitimate contribution to interpretable survival analysis in healthcare.\n\nThe strongest new bit is the decoupled training scheme: fitting each covariate-specific hazard on its own univariate likelihood, then fitting the mixture weights and assignment network with those curves fixed. That design does keep the population-level curves from being reshaped by the subgroup machinery, and the post-training merging of redundant subgroups is a nice practical touch, with stability across thresholds shown empirically.\n\nThe soft spots are real but not fatal. The main one, which the stress-test note correctly identifies, is the interpretation of the population-level curves. Optimizing Eq. (9) with only X_d does not yield an additive component in the joint hazard; it yields the marginal conditional hazard h_d(t|x_d). With correlated covariates, that marginal association can pick up effects of other variables and misattribute risk. The paper's claim that decoupled training \"purely captures covariate-specific trends\" and \"breaks pairwise covariate correlations\" is therefore overstated. The model still gives a coherent decomposition of the marginal hazard, so the central method survives, but the interpretability claims need reframing as marginal associations, or a demonstration that they approximate joint additive effects under correlation.\n\nTwo smaller issues: the abstract says \"on par with state-of-the-art survival baselines,\" but the body limits the claim to interpretable models and ADHAM trails DHA on several metrics; and the regularizer coefficients in Eq. (10) are never specified, which blocks exact reproduction. The subgroup-merging proof is tautological (identical β rows merge without changing the hazard), but the empirical threshold analysis shows the procedure works regardless.\n\nWho is this for? Researchers and practitioners in clinical ML who want a single interpretable survival model with three levels of explanation. It is not a breakthrough, but it is a useful, workmanlike contribution. I would send it to a serious referee and expect a revision that addresses the marginal-association point and the missing regularization details.","headline":"A practical and partly novel three-level interpretable survival model; the decoupled training makes the population-level explanations marginal associations rather than additive components, so the paper overstates its case, but it is solid enough to merit a genuine referee.","tokens_in":27285,"tokens_out":2853,"would_cite":false,"duration_ms":28012,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62N01","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"ADHAM writes each patient's hazard as a weighted sum of per-covariate risk curves, so a single trained model yields population-, subgroup-, and patient-level explanations while matching existing interpretable survival models on predictive…","keywords":["survival analysis","interpretable machine learning","additive hazard models","mixture density networks","latent subgroups","censored time-to-event data","concurvity","post-training model selection"],"falsifier":"Fit ADHAM on synthetic data with two correlated covariates whose true joint effects oppose their marginal effects (for example, either covariate alone predicts shorter times, but both together predict longer times). If the recovered population-level curves match the univariate marginals rather than the joint effects, the decoupled curves misstate covariate risk and the decomposition is not faithful. A cheaper corroborating check on real data: compare ADHAM's per-covariate curves with the per-covariate partial-dependence profiles of a flexible jointly-fit model; agreement supports the decomposition, while visible disagreement marks where it breaks.","tokens_in":26187,"feed_emoji":"📈","tokens_out":15401,"duration_ms":120407,"temperature":0.7,"pith_summary":"The paper introduces ADHAM (Additive Deep Hazard Analysis Mixtures), a survival model that expresses each patient's hazard function as a weighted sum of one neural risk curve per covariate, with the weights determined by a latent subgroup assignment learned from the data. The claim is that this single decomposition delivers interpretability at three levels at once: a population-level hazard curve for each covariate, subgroup-level feature-importance profiles, and a patient-specific hazard breakdown over time. The authors argue that training each covariate curve in isolation sidesteps the concurvity problem of additive models, where correlated features mask each other's effects. Subgroups with near-identical importance vectors can be merged after training without changing the likelihood, which answers the usual mixture-model question of how many subgroups to use. On three datasets, ADHAM matches the predictive performance of existing interpretable survival models, so the interpretability does not come at an obvious accuracy cost.","feed_headline":"Additive hazard mixtures explain risk at three levels at once","feed_subtitle":"One trained model yields population trends, subgroup profiles, and per-patient risk timelines.","key_machinery":"The load-bearing object is the identity of Equation 4 together with the decoupled training scheme of Algorithm 1: each population-level hazard $\\lambda(t|x_d;\\varphi_d)$ is fit by maximizing its own marginal log-likelihood $\\tilde\\ell_d$ using only covariate $d$, and then the subgroup assignment network $f_{\\theta}(x)$ and the $C\\times D$ importance matrix $\\beta$ (each row on a simplex) are fit to the full likelihood with the curves frozen. This ordering is what holds the model to an additive, interpretable form, because the mixture weights cannot reshape the individual curves. Two mechanisms support the identity in practice: an unbiased Monte-Carlo estimator for the cumulative-hazard integral, and a post-training refinement (Algorithm 2) that merges subgroups whose $\\beta$ rows have correlation above a threshold $h$, which the authors prove leaves the likelihood unchanged when the rows are exactly equal. Orthogonal and entropy regularizers respectively spread patients across subgroups and keep the covariate weights from collapsing to a narrow set early in training.","core_discovery":"ADHAM's central object is the marginal-hazard identity $\\lambda(t|x;\\theta,\\beta,\\Phi)=\\sum_{d=1}^{D}\\sum_{c=1}^{C}\\beta_{dc} f_{\\theta c}(x)\\lambda(t|x_d;\\varphi_d)$, in which $\\lambda(t|x_d;\\varphi_d)$ is a neural hazard curve trained on covariate $d$ alone, $f_{\\theta c}(x)$ is the probability that patient $x$ belongs to latent subgroup $c$, and $\\beta_{dc}$ is the importance of covariate $d$ inside subgroup $c$. The paper's claim is that this compact decomposition gives all three levels of explanation from a single fitted model: the curves are population-level covariate effects, the rows of $\\beta$ characterize the latent subgroups, and the per-patient weighted combination explains an individual's risk over time. To keep the decomposition from degenerating into arbitrary functions of all covariates, training is decoupled: each covariate curve is first fit by maximizing a likelihood that uses only that covariate, and only afterwards are the subgroup network and the $\\beta$ weights fit to reweight the fixed curves. A post-training refinement merges subgroups whose $\\beta$ rows are nearly identical, with a proof that exactly equal rows leave the data log-likelihood unchanged. The paper presents the explanations as associations learned from data, not as causal effects.","pith_inferences":["Because the curves are fit one covariate at a time, they are marginal associations. In real data where a covariate's apparent effect flips or disappears once correlated covariates are adjusted for, ADHAM's population curves will show the univariate pattern; a clinician should read them as 'how risk tracks this measurement alone' rather than 'what this measurement does.'","The same mixture identity extends naturally to competing risks: give each event type its own set of covariate curves and importance rows, and the model would yield three-level explanations for every competing outcome, a direction the paper explicitly leaves open.","The refinement criterion measures redundancy in the explanation, not separability of the patients: merging rows of $\\beta$ does not test whether the surviving subgroups are distinct in covariate space, so a companion check would apply the same correlation logic to the assignment distributions $f_{\\theta}(x)$ within each subgroup.","A direct stress test of the decoupling assumption is available: on synthetic data with two correlated covariates whose marginal and joint effects point in opposite directions, the recovered population curves reveal immediately whether univariate fits misattribute risk."],"forward_implications":["A single ADHAM model replaces the multi-model workflow: population, subgroup, and individual explanations all come from one set of fitted parameters, so explanations stay tied to the same model that makes the predictions.","Because each covariate curve is learned from that covariate alone before weighting, the reported per-covariate hazard is stable in the presence of correlated features, directly addressing the concurvity limitation of additive models in survival analysis.","Practitioners can train with many subgroups and choose the final number afterwards by picking the merge threshold $h$; the paper shows performance stays nearly flat as $h$ moves from 1 down to 0.65 while the subgroup count shrinks from 100 to about 11 on the SUPPORT dataset.","The readouts are associative, not causal: the curves describe how risk tracks a covariate in the observed data, and should not be treated as the effect of intervening on that covariate.","Among the interpretable baselines in the paper's tables (CoxPH, TimeNAM, TimeNA2M), regularized ADHAM posts the top metrics on most evaluations, while staying close to the best black-box comparator."],"supporting_citations":[{"why":"supplies the unbiased importance-sampling log-likelihood estimator that ADHAM uses to train its hazard networks and to compute survival predictions","marker":"(Ketenci et al., 2023)"},{"why":"introduces neural additive models, the per-covariate functional form ADHAM adopts for its population-level hazard components","marker":"(Agarwal et al., 2021)"},{"why":"defines TimeNAM and TimeNA2M, the interpretable additive survival baselines that ADHAM extends and compares against on all datasets","marker":"(Peroni et al., 2022)"},{"why":"gives the mixture-density-network flexibility argument that motivates modeling the time-to-event density as a mixture of components","marker":"(Bishop and Nasrabadi, 2006)"},{"why":"provides the agglomerative clustering routines used in the post-training subgroup-refinement algorithm","marker":"(Virtanen et al., 2020)"},{"why":"Cox-Time is the time-dependent hazard framework and integration routine on which the additive baselines and ADHAM's likelihood computations build","marker":"(Kvamme et al., 2019)"},{"why":"Deep Survival Machines is the mixture-density survival baseline whose parametric mixture approach ADHAM reworks into per-covariate additive form","marker":"(Nagpal et al., 2021a)"},{"why":"documents the concurvity problem in additive models, the failure mode the decoupled training scheme is designed to avoid","marker":"(Siems et al., 2023)"}],"fun_headline_variants":["Three risk lenses from one survival model","ADHAM: interpretable survival with subgroup-specific hazards","Population, subgroup, patient: three risk views at once","Survival analysis that explains risk at every level","Merging subgroups keeps survival models interpretable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on treating each covariate's alone-fit curve as its true contribution inside the full model; if correlations with other covariates change how a covariate relates to risk, the population-level curves track only that covariate's crude association and can misattribute risk.","fun_headline_variants_meta":{"raw":{"variants":["Three risk lenses from one survival model","ADHAM: interpretable survival with subgroup-specific hazards","Population, subgroup, patient: three risk views at once","Survival analysis that explains risk at every level","Merging subgroups keeps survival models interpretable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000711,"raw_usage":{"total_tokens":3231,"prompt_tokens":1006,"completion_tokens":2225,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2153}},"tokens_in":622,"tokens_out":2225,"duration_ms":16434,"temperature":1.0,"reasoning_tokens":2153,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:13:49.848488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit ADHAM on synthetic data with two correlated covariates whose true joint effects oppose their marginal effects (for example, either covariate alone predicts shorter times, but both together predict longer times). If the recovered population-level curves match the univariate marginals rather than the joint effects, the decoupled curves misstate covariate risk and the decomposition is not faithful. A cheaper corroborating check on real data: compare ADHAM's per-covariate curves with the per-covariate partial-dependence profiles of a flexible jointly-fit model; agreement supports the decomposition, while visible disagreement marks where it breaks.","supporting_citations":[{"cited_title":"Maximum likelihood estimation of flexible survival densities with importance sampling","cited_arxiv_id":null,"evidence_quote":"supplies the unbiased importance-sampling log-likelihood estimator that ADHAM uses to train its hazard networks and to compute survival predictions"},{"cited_title":"Extending the Neural Additive Model for Survival Analysis with EHR Data","cited_arxiv_id":"2211.07814","evidence_quote":"defines TimeNAM and TimeNA2M, the interpretable additive survival baselines that ADHAM extends and compares against on all datasets"},{"cited_title":"Time-to-event prediction with neural networks and cox regression","cited_arxiv_id":null,"evidence_quote":"Cox-Time is the time-dependent hazard framework and integration routine on which the additive baselines and ADHAM's likelihood computations build"},{"cited_title":"Curve your enthusiasm: concurvity regularization in differentiable generalized additive models","cited_arxiv_id":null,"evidence_quote":"documents the concurvity problem in additive models, the failure mode the decoupled training scheme is designed to avoid"}],"review_version":1}