{"id":"034f61e2-9045-4f82-bf58-85f3ba2662a2","arxiv_id":"2504.17613","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"TarDiff guides diffusion-based synthetic EHR generation with a gradient-alignment signal computed from a guidance set, reporting improved downstream mortality and ICU-stay classification versus prior generative models.","lead":"This paper presents TarDiff, a diffusion framework that generates synthetic medical time-series data by steering the denoising process with an influence-based signal, so the generated samples improve a downstream clinical prediction model. On six medical datasets, TarDiff reports AUPRC gains up to 20.4% and AUROC gains up to 18.4% over prior generative baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The implemented guidance J in Eq. (17) is not the gradient of the influence ΔL_T defined in Eq. (21), because the per-sample denominator ||∇ℓ(z)||^2 is z-dependent and is dropped in Algorithm 1; no experiment shows the simplified score tracks actual retraining effects.","rationale":"I agree with the reader's weakest_assumption: the cached-G simplification is the point most likely to break the central claim. The six-dataset consistency, the class-imbalance ablations in Section 4.4, and the complexity analysis are real supporting evidence, so I would not move to reject; the fix is a validation experiment plus a corrected derivation. I therefore keep the conditional verdict unchanged. The secondary absence of error bars strengthens the need for the correlation check but is not the reason for this verdict.","tokens_in":17403,"tokens_out":7697,"duration_ms":83306,"concrete_test":"On MIMIC-III mortality, sample 50 candidates from the unguided diffusion model; for each, compute (a) the implemented score S = G·∇φℓ(x,y;φ*) with cached G, (b) the exact Eq. (21) score including the denominator, and (c) the actual decrease in guidance-set loss after adding that sample and retraining TimesNet to convergence with fixed seeds. If the Spearman rank correlation between S and (c) is not significantly positive, the guidance signal in Eq. (17) is not the claimed influence; if (b) correlates with (c) much better than S, the dropped denominator is the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 defines the objective as ΔL_T(z) = -Σ_{Dguide} ε ∇ℓ(x',y';φ*)·∇ℓ(z;φ*) / ||∇ℓ(z;φ*)||^2 (Eq. 21). The denominator depends on the candidate z, so it cannot be folded into a constant vector. Eq. (22) writes G = -Σ ε∇ℓ(x',y';φ*)/||∇ℓ(z;φ*)||^2, but this G is z-dependent; Algorithm 1 nevertheless precomputes G from Dguide alone and then uses J = ∇_{x_t}[G·∇φℓ(x_t,y;φ*)] in Eq. (17). Dropping ||∇ℓ(z)||^2 changes the guidance direction, because the true gradient of Eq. (21) includes a term from differentiating the denominator. The paper provides no experiment showing that the simplified inner product G·∇ℓ(x_t) correlates with the actual reduction in downstream loss from adding the sample; Figure 3 only re-plots the same cached-G score, so it cannot validate the mechanism. If the simplified direction does not rank candidates by true influence, the reported gains cannot be attributed to the influence objective, even if the empirical improvements are real. This mismatch is load-bearing because the paper's central contribution—steering diffusion with the gradient of task-specific influence—rests on Eq. (17) being that gradient.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TarDiff, a diffusion-based generative framework for synthetic electronic health record (EHR) time series that aims to generate samples optimized for a downstream clinical task rather than merely for distributional fidelity. The central idea is to use influence functions: the influence of a candidate synthetic sample is defined as the expected reduction in task loss on a guidance set if that sample were added to the training set, and this influence is computed via the gradient inner product between the candidate sample and the cached gradients of the guidance set. This influence gradient is then added to the reverse diffusion mean at each denoising step. The method is evaluated on six datasets (MIMIC-III, eICU, APAVA, ADFD, PTB, TDBrain) in train-on-synthetic/test-on-real (TSTR) and train-on-synthetic-and-real/test-on-real (TSRTR) protocols, reporting AUPRC/AUROC improvements over several GAN, VAE, and diffusion baselines, plus additional analyses on class imbalance, computational overhead, fidelity, and privacy.","tokens_in":17747,"tokens_out":3332,"duration_ms":30850,"significance":"If the central mechanism is valid, the paper makes a valuable contribution: it explicitly targets synthetic data generation toward downstream task utility, which is a recognized weakness of fidelity-only generative models, and it demonstrates consistent gains across multiple clinical datasets and tasks. The paper also provides a useful experimental breadth, including minority-class analysis, guidance-set scale sensitivity, runtime overhead, and privacy metrics. However, the significance is contingent on resolving the mismatch between the influence derivation and the implemented guidance signal, and on providing statistically grounded comparisons; as presented, the evidence does not yet establish that the reported gains arise from the claimed influence objective.","major_comments":[{"comment":"The implemented guidance signal is not the gradient of the influence function defined in Eq. (21). In Eq. (21), the per-sample denominator ||∇φℓ(ˆz;φ)||^2 depends on the candidate sample ˆz, so it cannot be absorbed into a constant vector G as written in Eq. (22). Algorithm 1 precomputes G from Dguide alone (only accumulating per-sample gradients, then normalizing by |D0|) and then uses J = ∇_{x_t}(G·∇φℓ(x_t,y;φ*)) in Eq. (17). Dropping the denominator changes the guidance direction, because the true gradient of Eq. (21) would include a term from differentiating ||∇φℓ(ˆz;φ)||^2 with respect to x_t. Since Eq. (17) is the core contribution, this mismatch must be fixed or explicitly justified (e.g., by showing the simplified direction is a valid approximation of the true influence gradient).","section":"Section 3.3 and Algorithm 1"},{"comment":"The guidance scale w is selected on the Evaluation-Val subset, and then the final performance is reported on the entire validation set, which includes Evaluation-Val. This constitutes a form of validation-set overfitting: the reported figures in Figure 3 and the downstream tables may be optimistically biased because the same data used to choose w are retained in the final evaluation. The paper claims this is 'unbiased,' but the overlap between model selection and evaluation invalidates that claim. The final evaluation should be performed on a held-out test set (or at least on a split disjoint from both Guidance-Val and Evaluation-Val).","section":"Section 4.6 and evaluation protocol"},{"comment":"No error bars, confidence intervals, or seeds are reported for any of the main quantitative results. Several claimed improvements are small (e.g., MIMIC-III mortality AUPRC 0.1799 vs. Real Data 0.1736; PTB AUPRC 0.95435 vs. TimeVAE 0.95092), and without variance estimates it is not possible to assess whether these differences are meaningful. The paper should report mean and standard deviation over at least three independent runs for the primary comparisons.","section":"Tables 1, 2, 4, 5 and Figure 2"}],"minor_comments":[{"comment":"The symbol T is used both for the length of the time series and the number of diffusion steps in Section 2.1; this dual use is confusing and should be disambiguated (e.g., L for series length).","section":"Section 2.2"},{"comment":"The symbol ε in Eqs. (18)-(22) is used as a small perturbation magnitude but is never defined; it should be clarified whether it is a fixed scalar, a learning rate, or an infinitesimal, since the magnitude of G in Eq. (22) depends on it.","section":"Section 3.3"},{"comment":"The text says the right panel reports AUROC on the Evaluation-Val subset, but the Figure 3 caption says 'assessing AUROC performance on the Guidance-Val subset.' These are inconsistent and should be reconciled.","section":"Section 4.6 / Figure 3 caption"},{"comment":"Table 7 lists MIMIC-III as having 26,150 samples, while Section 4.1 and Table 6 report 20,920; one of these is a typo and should be corrected.","section":"Appendix A.3, Table 7"},{"comment":"The text refers to a 'PTBrain dataset,' which appears to be a typo for TDBrain.","section":"Appendix A.4"}],"recommendation":"major_revision","confidential_remarks":"The core idea is timely and the empirical scope is broad, but the paper currently has a load-bearing gap between its influence derivation and the implemented guidance, and the validation protocol in §4.6 risks overfitting the guidance scale to the evaluation set. Both are fixable with additional analysis and a cleaner split, but as submitted the evidence does not yet support the claimed mechanism or the reported margins."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on TarDiff. The core idea is genuinely new: instead of generating EHR time series that look like the training data, they use an influence-style objective to steer diffusion toward samples that improve a specific downstream classifier. That's a sensible and worth-pursuing direction, and it produces consistent reported gains over TimeGAN/TimeVAE/DiffusionTS/BioDiffusion across six datasets. The class-imbalance analysis—showing that minority examples have larger gradient norms and that guidance from minority-only samples helps minority F1—is a nice, credible extension.\n\nThe soft spots are real but fixable.\n\nFirst, the implementation doesn't match the derivation. Equation (21) defines influence as the per-sample inner product divided by ||∇ℓ(z)||^2, which depends on z. In Algorithm 1 they drop that denominator and precompute G from the guidance set alone. That makes the guidance signal a gradient-alignment score, not literally the gradient of the claimed influence. It may still be a useful heuristic—and the empirical results suggest it is—but the paper needs an experiment showing the simplified score tracks true leave-one-out influence (or a revised derivation). As written, the attribution is overstated.\n\nSecond, no error bars or multiple seeds appear in the main tables. Given the reported margins over the best baseline are meaningful in places but the absolute values are close for some tasks, we can't tell which differences are stable. This is a standard requirement and easy to fix.\n\nThird, the validation story in Section 4.6 is confusing. They partition the validation set, tune the guidance scale on one subset, then report final performance on the \"entire validation set.\" That sounds like they are using the tuning subset for final evaluation. The main TSTR tables appear to evaluate on a held-out test set, so the headline numbers might be clean, but the paper should say this explicitly and avoid reporting on the combined validation set.\n\nI don't think these are fatal. The central hypothesis—utility-guided generation can beat pure distribution matching—survives reading. The concerns are about precision of the claim and reproducibility. With a clearer derivation or validation experiment, plus error bars, this would be a solid contribution to the synthetic-EHR literature.\n\nMy recommendation: send it to peer review. A good referee can push for the missing validation and the derivation/implementation reconciliation. I'd bring it to reading group if you want a debate about validation hygiene, but I wouldn't cite it as-is until the reporting issues are cleaned up.","headline":"A genuinely new idea—influence-guided diffusion for EHR time series—with consistent reported gains, but the implemented guidance term doesn't match the derived influence and the empirical reporting needs error bars and a cleaner validation story.","tokens_in":18203,"tokens_out":3292,"would_cite":false,"duration_ms":31797,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Synthetic EHR time series generated by influence-guided diffusion are claimed to improve downstream clinical classifiers beyond distribution-matching baselines.","keywords":["synthetic EHR time series","diffusion models","influence functions","target-oriented generation","downstream task utility","class imbalance","TSTR evaluation","clinical time series"],"falsifier":"Compute the true influence of a batch of TarDiff-generated samples by adding each sample to the training set, retraining the downstream classifier, and measuring the loss change on a fixed held-out guidance set; then rank-correlate those true changes with the paper's predicted $\\Delta L_T(\\hat{z})$ from Eq. (23). If the correlation is not clearly positive on MIMIC-III mortality, the guidance direction is not optimizing the stated objective.","tokens_in":17222,"feed_emoji":"🏥","tokens_out":13052,"duration_ms":108160,"temperature":0.7,"pith_summary":"The paper proposes TarDiff, a diffusion-based generator for synthetic electronic health record time series whose samples are optimized for their expected effect on a downstream clinical classifier rather than for resemblance to the training data. The central claim is that embedding an influence-function gradient into the reverse diffusion process steers generation toward samples that reduce task-specific loss, and that this yields better mortality and ICU-stay classifiers than samples produced by distribution-matching generators. In Train-on-Synthetic-Test-on-Real experiments, a classifier trained only on TarDiff samples reaches MIMIC-III mortality AUPRC of 0.1799, above the 0.1736 achieved by a classifier trained on real data; across six datasets TarDiff reports gains of up to 20.4% in AUPRC and 18.4% in AUROC over generative baselines. The paper also reports that minority-class samples carry much larger gradient norms and that minority-only guidance more than doubles minority F1 on MIMIC-III (0.163 vs 0.056 real-only), indicating the mechanism can act as a counter to class imbalance. If the claim holds, clinical machine learning gains a way to synthesize training data for rare or underrepresented conditions, not merely a way to imitate the training distribution.","feed_headline":"Influence-guided synthetic EHRs can beat real-data training","feed_subtitle":"Influence-guided diffusion lifts mortality and length-of-stay AUPRC by up to 20.4 percent over generative baselines.","key_machinery":"The central object is the influence-gradient term $J = \\nabla_{x_t}(G\\cdot\\nabla_\\phi \\ell(x_t,y;\\phi^*))$, inserted into the reverse diffusion mean as $\\tilde{\\mu}_t = \\mu_t + w J$. Here $G$ is a single cached vector, the normalized sum of downstream-loss gradients over the guidance set, and $\\ell$ is the task loss. The design mirrors classifier guidance but replaces the class log-likelihood gradient with a task-utility signal; because $G$ is independent of $x_t$, the per-step guidance reduces to a dot product plus one gradient computation, which is why sampling remains cheap. The intended effect is to push the denoising trajectory toward regions where a synthetic sample's loss gradient aligns with the guidance-set gradient, i.e., toward samples whose addition to the training set would reduce expected task loss.","core_discovery":"TarDiff's central claim is that the reverse diffusion update can be modified to $\\tilde{\\mu}_\\theta(x_t,y,t) = \\mu_\\theta(x_t,y,t) + \\alpha \\nabla_{x_t}\\Delta L_T(\\hat{z}_t)$, where $\\Delta L_T(\\hat{z})$ is the expected reduction in downstream task loss caused by adding synthetic sample $\\hat{z}=(x,y)$ to the training set. The paper approximates this influence as $\\Delta L_T(\\hat{z}) = \\nabla_\\phi \\ell(\\hat{z};\\phi^*)\\cdot G$, with $G$ the accumulated (negative) gradient of the downstream loss over a guidance set drawn i.i.d. from the task distribution. In the implemented pipeline, $G$ is cached once from a pretrained downstream model, and at each denoising step the mean update is $\\tilde{\\mu}_t = \\mu_t + w J$, where $J = \\nabla_{x_t}(G\\cdot\\nabla_\\phi \\ell(x_t,y;\\phi^*))$. The paper reports that this produces synthetic time series that, in TSTR mode, beat a real-data-trained classifier on MIMIC-III mortality (AUPRC 0.1799 vs 0.1736) and outperform five generative baselines across six datasets by up to 20.4% AUPRC and 18.4% AUROC, with one-time gradient-caching overhead of 10-167 seconds.","pith_inferences":["My inference: if the reported gains are real, the same cached-gradient guidance could be ported to non-temporal tabular and image data with a well-defined downstream loss, since nothing in the update rule is specific to time series.","My inference: the minority-only guidance results suggest an explicit, tunable selection rule for the guidance set (for instance, choosing the hardest minority examples), which the paper stops short of specifying.","My inference: the paper's evidence implies that distributional fidelity alone is an insufficient benchmark for synthetic EHR generation; a utility-first evaluation protocol would report downstream AUPRC/AUROC gains after augmentation, not just distance-to-real-data scores."],"forward_implications":["Synthetic-only training with TarDiff can outperform a classifier trained on real data in at least one TSTR setting: MIMIC-III mortality AUPRC 0.1799 vs 0.1736.","Augmenting real data with TarDiff samples improves downstream AUROC across most tasks and mix ratios, with the gains generally growing as the synthetic proportion rises from 0.2 to 1.0.","Influence guidance naturally shifts generation toward minority-class patterns; TarDiff raises minority F1 from 0.056 to 0.108 on MIMIC-III and from 0.013 to 0.018 on eICU without explicit class weighting.","Restricting the guidance set to minority-only samples further raises minority F1 (0.163 on MIMIC-III), while majority-only guidance degrades it (0.066), showing that the guidance direction is controllable.","The added cost is modest: one-time downstream training plus gradient caching (10-167 s across datasets) and a per-step overhead ratio $g(L,D)/h(L,D)$ that is small, making the task-oriented guidance practical."],"supporting_citations":[{"why":"Supplies the gradient-based parameter-shift approximation from which the influence estimate in Eq. (21) is derived.","marker":"[Charpiat et al., 2019]"},{"why":"Provides the influence-function foundation for measuring how a training point changes model predictions and loss.","marker":"[Koh and Liang, 2017]"},{"why":"Contributes the classifier-guided diffusion update that TarDiff adapts by replacing the class-likelihood gradient with the influence gradient.","marker":"[Dhariwal and Nichol, 2021]"},{"why":"Provides the conditional diffusion formulation used as the base generative model for label-conditioned EHR sampling.","marker":"[Ho and Salimans, 2022]"},{"why":"Supplies the MIMIC-III critical-care dataset used for the mortality and ICU-stay TSTR and TSRTR experiments.","marker":"[Johnson et al., 2016]"},{"why":"Supplies the eICU critical-care dataset used for the second set of ICU mortality and length-of-stay evaluations.","marker":"[Pollard et al., 2018]"},{"why":"Serves as the TimeGAN baseline whose TSTR/TSRTR results TarDiff is compared against.","marker":"[Yoon et al., 2019]"},{"why":"Provides the TimesNet architecture used as the downstream classifier in all task evaluations.","marker":"[Wu et al., 2022]"}],"fun_headline_variants":["Influence-guided diffusion: synthetic EHRs that beat real-data models","Synthetic EHRs guided by influence functions improve downstream AUPRC","TarDiff: target-oriented diffusion for utility-driven synthetic EHRs","Diffusion that optimizes for task loss yields better synthetic EHRs","Up to 20.4% AUPRC gain: influence-guided synthetic EHR diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the simplified guidance signal actually points toward synthetic samples that improve the downstream model: the exact influence formula in Eq. (21) has a sample-dependent denominator $\\|\\nabla_\\phi \\ell(\\hat{z};\\phi)\\|^2$ that Algorithm 1 omits as a constant factor, and the paper provides no experiment verifying that the simplified direction correlates with true leave-one-out retraining effects.","fun_headline_variants_meta":{"raw":{"variants":["Influence-guided diffusion: synthetic EHRs that beat real-data models","Synthetic EHRs guided by influence functions improve downstream AUPRC","TarDiff: target-oriented diffusion for utility-driven synthetic EHRs","Diffusion that optimizes for task loss yields better synthetic EHRs","Up to 20.4% AUPRC gain: influence-guided synthetic EHR diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000353,"raw_usage":{"total_tokens":2000,"prompt_tokens":1099,"completion_tokens":901,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":715,"completion_tokens_details":{"reasoning_tokens":807}},"tokens_in":715,"tokens_out":901,"duration_ms":8647,"temperature":1.0,"reasoning_tokens":807,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:36:02.988248+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the true influence of a batch of TarDiff-generated samples by adding each sample to the training set, retraining the downstream classifier, and measuring the loss change on a fixed held-out guidance set; then rank-correlate those true changes with the paper's predicted $\\Delta L_T(\\hat{z})$ from Eq. (23). If the correlation is not clearly positive on MIMIC-III mortality, the guidance direction is not optimizing the stated objective.","supporting_citations":[],"review_version":1}