SAFER: A Calibrated Risk-Aware Multimodal Recommendation Model for Dynamic Treatment Regimes

Hui Xiong; Yishan Shen; Yong Chen; Yuyang Ye

REVIEW 5 major objections 4 minor 16 references

SAFER: A Calibrated Risk-Aware Multimodal Recommendation Model for Dynamic Treatment Regimes

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A sepsis treatment recommender can keep unreliable recommendations below a user-chosen error rate, with formal guarantees.

desk verdict The paper's empirical idea is plausible, but the conformal FDR guarantee at its core is not supported by the proof. read the letter →

arxiv 2506.06649 v1 pith:66FSXSUL submitted 2025-06-07 cs.LG stat.ML

Yishan Shen , Yuyang Ye , Hui Xiong , Yong Chen This is my paper · ORCID

classification cs.LGstat.ML

keywords dynamictreatmentregimesconformalpredictionfalsediscoveryratecontrolmultimodalEHRclinicalnotesuncertaintyquantificationsepsisrecommendation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SAFER is a framework for dynamic treatment regimes (DTRs) that combines structured electronic health records with clinical notes and adds a statistical safety mechanism. The paper's central thesis is that label uncertainty is systematic: deceased patients' treatment labels are treated as ambiguous, while survivors' labels are considered reliable. An uncertainty score, defined as the KL divergence between a main predictor and a survivor-trained student predictor, is meant to capture each patient's label-mismatch risk. That score is fed into a conformal prediction procedure with Benjamini–Hochberg selection, and Theorem 5.1 states that the expected proportion of unreliable recommendations stays below a user-chosen $\alpha$. Experiments on two sepsis cohorts report higher recommendation accuracy and lower counterfactual mortality than several baselines, along with realized FDR below the target level.

What carries the argument

The load-bearing object is the uncertainty score $\kappa_i = D_{\mathrm{KL}}(p_\theta(y|h_i) \| p_\phi(y|h_i))$, the KL divergence between the main teacher predictor and a Lipschitz-regularized student predictor trained on survivors. The paper uses this score as a label-mismatch risk: high $\kappa$ means the model is internally inconsistent about the right treatment. That score defines the null hypotheses $H_j : \kappa_j \ge c$, is mapped through the conformal p-value in Eq. (8) with uniform tie-breaking, and is passed to the Benjamini–Hochberg procedure; Theorem 5.1 uses the selective super-uniformity of these p-values to bound the expected fraction of false recommendations.

What would settle it

On a held-out sepsis cohort, compute the empirical KL score for every patient and stratify by survival: if the mean $\kappa$ for deceased patients is not strictly above that for survivors, or if the teacher–student $\ell^1$ gap on any measurable region is zero, then the premise of Theorem 4.1 fails and the FDR guarantee no longer protects against unreliable labels.

Watch

Extended reading notes

Core claim

The paper's central claim is that a DTR model can be made risk-aware without giving up accuracy: SAFER learns a unified patient representation from EHR tables and clinical notes, trains a refined student predictor only on survivors, and uses the KL divergence between the teacher and student as a per-patient uncertainty score $\kappa_i$. That score is treated as a label-reliability measure, forming null hypotheses $H_j : \kappa_j \ge c$, and converted into conformal p-values with random tie-breaking. Applying the Benjamini–Hochberg procedure to those p-values yields a selection set $S$, and Theorem 5.1 states that under i.i.d. sampling and a bounded uncertainty score, $\mathrm{FDR} = \mathbb{E}[V/\max\{1,R\}] \le \alpha$ for any user-specified $\alpha$. On MIMIC-III and MIMIC-IV sepsis cohorts, the paper reports that SAFER outperforms LSTM-, attention-, and RL-based baselines on ranking and AUC metrics and achieves larger reductions in estimated counterfactual mortality than the baselines.

Load-bearing premise

The load-bearing premise is that the KL divergence between the teacher and the survivor-trained student really separates deceased from surviving patients with a positive margin; if that gap does not exist, the risk-aware loss and the conformal FDR interpretation lose their foundation.

Editorial extensions

If this is right

At decision time, a clinician can set $\alpha$ and the model will only return recommendations whose expected share of unreliable items is at most $\alpha$, without knowing the true optimal treatments for the test patients.
Including clinical notes and risk-aware fine-tuning improves measured recommendation quality on both sepsis cohorts across ranking, AUC, and counterfactual mortality metrics.
The FDR guarantee is architectural-agnostic: any uncertainty score satisfying the boundedness and exchangeability conditions yields the same bound, so the conformal selection layer can be reused with other DTR predictors.
Because the conformal guarantee does not need test labels, it remains valid in deployment when outcomes are not yet observed, provided the calibration set is exchangeable with the deployed population.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural stress test is to replace $\kappa$ with an outcome-independent score or a permuted version and check whether the realized FDR still stays below $\alpha$; if it does, $\kappa$ may be adding little beyond rank information, and the guarantee is doing all the work.
The same conformal wrapper would transfer to other high-stakes recommendation settings—drug combinations, triage, or discharge planning—as long as one can define a label-reliability score and an exchangeable calibration set.
The counterfactual mortality claims rest on ignorability and positivity assumptions; an external evaluation with randomized treatment assignments or a design that relaxes no-unmeasured-confounding would be the strong test of whether SAFER's recommendations are causally safer.
The theorem's i.i.d. condition is stronger than necessary; the authors note exchangeability suffices, and weighted conformal inference could extend SAFER to new hospitals with shifted covariate distributions, turning the guarantee into a per-site calibration procedure.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The headline: the paper's central safety guarantee is not supported by its own proof. The conformal p-value in Eq. (8) is defined as the fraction of calibration scores that fall below the test score after restricting to calibration samples with κ_i ≥ c, divided by n+1. That is not a valid conformal p-value: under the null H_j: κ_{n+j} ≥ c, a test score just above c can have p_j near zero because the numerator only counts scores in the conditional subset. So the BH procedure on these p-values does not control FDR. The attempt in Appendix A.2 to fix this with J(x,y)=κ(x)+2M·1{y≥c} compares a categorical treatment label to a scalar threshold, which is undefined. The reduction to Jin & Candès (2023b) does not go through.

That said, there is something here. The combination of tabular EHR and clinical notes with a teacher–student KL uncertainty score for DTR is new as far as I know. The experimental setup on MIMIC-III and IV is reasonable, the ablations are informative, and the reported gains over baselines are nontrivial. The code and dataset links are a plus.

But the soft spots are load-bearing. Theorem 4.1 assumes a δ-gap between teacher and student on a set G with probability gap π; the proof asserts this from P^- ≠ P^+ without justification. Two different latent distributions can lead to identical label distributions. And the uncertainty score predictor is fitted on calibration and test features in Section 6, which breaks the exchangeability that any conformal guarantee needs. On top of that, the headline metrics are evaluated only on survivors—the subset with reliable labels by the paper's own assumption—and the counterfactual mortality reduction relies on an LSTM outcome model with no uncertainty intervals.

None of this means the empirical work is worthless. As a predictive model, SAFER may be competitive. But the paper frames itself as providing formal assurances, and those assurances are not established. That's a serious issue for a high-stakes clinical application.

I'd send it to a serious referee—not to desk reject it—but the likely outcome is major revision or reject unless the theory is repaired. For a reading group, it could be an instructive example of why p-values need to be super-uniform before you apply BH.

Referee Report

5 major / 4 minor

Summary. The paper proposes SAFER, a multimodal framework for dynamic treatment regime (DTR) recommendation that fuses structured EHR data with clinical notes, models label uncertainty through a KL-divergence score between a teacher and a refined student module, and applies conformal inference with Benjamini–Hochberg (BH) selection to control the false discovery rate (FDR) of unreliable treatment recommendations. The main theoretical claim is Theorem 5.1, which asserts that the conformal selection set S satisfies FDR <= alpha under i.i.d. sampling and a bounded uncertainty score. The paper also reports experiments on two sepsis cohorts from MIMIC-III and MIMIC-IV, showing improved recommendation metrics and lower counterfactual mortality relative to several baselines, together with FDR and power curves for different thresholds.

Significance. If the theoretical guarantees were correct, the paper would be a meaningful contribution: it is, to my knowledge, the first DTR method to combine multimodal EHR/text modeling with conformal FDR control, and it ships code and data. The empirical study is broad, includes ablations, sensitivity analyses, and a case study, and the authors are candid about the i.i.d. assumption. However, the central safety claim rests on the conformal p-value construction and its FDR proof, and these are not valid as written. The uncertainty-score interpretation also depends on an unproved separation condition. Because the headline guarantee and the notion of 'risk-aware' recommendations are both undermined, the contribution's significance is conditional on a substantial reworking of the statistical formalism.

major comments (5)

[Section 5, Eq. (8)] The conformal p-value in Eq. (8) is not super-uniform under the null hypothesis H_j: kappa_{n+j} >= c. The numerator counts only calibration scores with kappa_i >= c, so a test point whose score is just above c will have p_j near 0 even though the null is true; for example, with kappa_i iid Unif(0,1), c=0.5 and kappa_{n+j}=0.5+epsilon, the numerator is empty with high probability. A valid conformal p-value must be super-uniform under the null to attach the Jin and Candès (2023b) guarantee, so Theorem 5.1's FDR bound is not established by the submitted argument.
[Appendix A.2] The reduction to Jin and Candès (2023b) is invalid as written: J(x,y)=kappa(x)+2M 1{y>=c} tests an event {y>=c} on a categorical treatment label y with no ordering relative to the uncertainty threshold c, so monotonicity of J in y is undefined. The proof asserts 'selective super-uniformity' for this statistic rather than deriving it; the displayed p-value in the restatement also contains a typo (a lone '+1/(n+1)' term after the sum), so the proof cannot be checked. The central FDR guarantee therefore rests on an unproved assertion.
[Appendix A.1] Theorem 4.1's key step is the asserted existence of a measurable set G with ||p_theta(.|h)-p_phi(.|h)||_1 >= delta for all h in G, derived from P^- != P^+ and Lipschitzness of f_phi. Distinct marginal latent distributions do not imply distinct conditional label distributions, so the delta-gap is an extra assumption, not a consequence. Since the risk-aware loss in Eq. (7) and the interpretation of kappa as label-mismatch risk both rely on this gap, the theoretical foundation of the uncertainty score is unsupported.
[Section 6] The experimental protocol for the FDR curves is not a valid test of the conformal procedure: the paper states that 'we train the uncertainty score predictor bkappa on the calibration and test sets using standard machine learning models leveraging the full embedding feature space X.' Fitting bkappa on test features breaks the exchangeability of calibration and test scores that the conformal p-values require, and it also means the reported FDR can reflect leakage rather than the guarantee in Theorem 5.1.
[Sections 3.2 and 5] The null hypothesis H_j: kappa_j >= c makes 'unreliable' a property of the model's own score; the FDR bound is then a rank and super-uniformity property of kappa, not an independent guarantee about treatment safety. The paper should state this limitation explicitly and validate the score against an external notion, such as mortality or clinical adjudication, before claiming that the expected fraction of incorrect recommendations is safely bounded.

minor comments (4)

[Table 1] The column header 'MRR@3↓Mortality' is ambiguous: it combines a ranking metric and a mortality reduction, and the text does not define whether lower or higher values are better for the mortality component.
[Section 6, preprocessing] The preprocessing description says variables were rescaled to [0,1] 'using z-score normalization,' which is inconsistent; standard z-score normalization does not map to [0,1].
[Throughout] Several citation keys are formatted with stray spaces (e.g., 'V ovk' instead of 'Vovk'), and there are typos such as 'avaliable' in the footnote; a careful proofreading pass is needed.
[Section 6.3] The reported 'p < 0.05' is not accompanied by any description of the significance test, multiple-comparison correction, or variability measure; please specify the test and the error bars.

Circularity Check

3 steps flagged · score 6.0 of 10

The FDR 'safety' guarantee is defined over the paper's own uncertainty score (κ≥c), Theorem 4.1's proof assumes the δ-gap that is its conclusion, and the test set is used to fit the score predictor.

self definitional [Section 3.2, Eq. (1); Section 5, Theorem 5.1]
"Concretely, we define a null hypothesis Hj :κ j ≥c, j= 1, . . . , m, for each test sample j, and reject (i.e., recommend) a subset S ⊆ {1, . . . , m} of hypotheses while ensuring FDR≤α ... V=(# of false rejections)."

A false rejection is, by Eq. (1), a recommended sample whose uncertainty score κ_j ≥ c. Thus the guaranteed quantity FDR = E[V/max(1,R)] is the expected fraction of recommended treatments whose own κ exceeds c. 'Unreliable treatment' in the abstract and Section 5 is therefore defined as κ≥c, and the 'safety guarantee' is a claim about the ranking of the paper's own score, not about an independent notion of wrong treatment. The conclusion FDR≤α reduces to a property of the score's calibration relative to its own threshold.
other [Appendix A.1, proof of Theorem 4.1]
"Since P − ̸=P +, there is a region G in latent space where pϕ cannot 'annihilate' the mismatch in pθ. That is on a measurable set G ⊂ H, the teacher predictions differ from the student by at least a fixed amount: ∥pθ(· |h)−pϕ(· |h)∥1 ≥δfor all h∈ G, with constantsδ > 0."

The theorem's conclusion is E_{h∼P−}[κ(h)] − E_{h∼P+}[κ(h)] ≥ c > 0, i.e., that κ separates deceased from survivor latents. The proof's only route to this is the asserted uniform gap δ between teacher and student label distributions on a set G. But P−≠P+ does not imply such a label-distribution gap; the two latent distributions could differ while pθ and pφ agree. The δ-gap is exactly the separation the theorem is supposed to establish, so Theorem 4.1—the result that makes κ a 'valid measure of uncertainty'—assumes its own conclusion.

1 more flagged steps

fitted input called prediction [Section 6, FDR Control experimental setup]
"Once the uncertainty scores are computed for the training set, we train the uncertainty score predictor bκ on the calibration and test sets using standard machine learning models leveraging the full embedding feature space X ."

The conformal p-values in Eq. (8) are computed from bκn+j for test samples. Here the same test samples' features are used to fit bκ before those p-values are produced, so the 'predicted' test uncertainties are in-sample outputs of a model fit on the test set. Theorem 5.1's assumption (1) requires calibration and test data to be i.i.d. and mutually independent; fitting the scoring function on the test set violates the exchangeability that the conformal guarantee relies on, so the claimed FDR bound is a self-evaluation rather than an out-of-sample guarantee.

full rationale

The paper's headline guarantee—'safe treatment recommendations' with FDR≤α—is not an independent safety claim. In Eq. (1), 'unreliable' is defined as κ_j≥c, and V counts exactly those rejections, so Theorem 5.1's FDR is the expected fraction of selected treatments whose own uncertainty score exceeds c. This is a self-referential calibration statement unless κ is independently shown to track label reliability. The only such showing, Theorem 4.1, is itself circular: the proof inserts a uniform δ-gap between teacher and student label distributions on a set G, asserted from P−≠P+ and Lipschitzness, but P−≠P+ does not imply that gap; the gap is precisely the separation that the theorem is meant to prove. Finally, the experimental protocol trains the uncertainty-score predictor on the combined calibration+test set before computing p-values for those same test points, making the test uncertainties in-sample and violating the i.i.d./exchangeability assumption the theorem requires. These three features make the central FDR claim reduce, in part, to the paper's own definitions and fitting choices rather than to an external measure of treatment safety. The paper also has independent proof defects (e.g., Eq. (8)'s p-value is not super-uniform under H_j, and the J(x,y)=κ(x)+2M·1{y≥c} reduction to Jin & Candès is ill-posed because c is an uncertainty threshold, not a label value), but those are correctness concerns, not circularity. Since the paper reports substantial external benchmark experiments and does not lean on a self-citation chain, the circularity score is 6 rather than 8.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on four tuned hyperparameters and six stated assumptions. The most fragile are the ad hoc 'delta gap' in Theorem 4.1's proof, the assumption that bkappa can be fit on combined calibration/test data, and the causal identification assumption for counterfactual mortality. No new physical or ontological entities are introduced.

free parameters (4)

gamma = not reported (grid-selected)
Regularization strength in Eq (7); sensitivity analysis in Fig 10 shows an optimal gamma chosen on the validation set.
uncertainty threshold c = evaluated at 0.1, 0.2, 0.3, 0.4
Defines null hypotheses H_j: kappa >= c. The paper gives an empirical guideline for choosing c based on FDR/power curves rather than a principled criterion.
hidden dimensionality h_d = 128
Selected after grid search over {128,256,512} based on Macro-AUC and mortality reduction on the validation set.
historical sequence length L = 8
Selected after sensitivity analysis (Fig 8); the same length is used for both datasets.

assumptions (6)

domain assumption Patient trajectories are i.i.d. sampled from a common distribution
Assumed for the conformal guarantee (Theorem 5.1, Remark A.5); real ICU cohorts have temporal drift and non-stationarity.
domain assumption No unmeasured confounding, positivity, and SUTVA for counterfactual mortality
Appendix B, conditions B1-B4; sepsis treatment is affected by many unrecorded variables, so the mortality reduction is not causally identified.
ad hoc to paper Treatment labels for deceased patients are ambiguous; survivor labels are reliable
Section 3.1; this is the central modeling assumption that drives risk-aware fine-tuning and the decision to evaluate metrics only on survivors.
ad hoc to paper There exists a region G where teacher and student label distributions differ by at least delta
Appendix A.1 Eq (2); asserted but not derived from P- != P+.
domain assumption The uncertainty predictor bkappa is fitted on calibration and test features without breaking exchangeability
Section 6; the score predictor is trained on calibration and test sets, yet the theorem assumes a fixed score function; this extra data dependence is not analyzed.
standard math The score function kappa is bounded by M
Theorem 5.1 assumption; boundedness is used to construct the auxiliary nonconformity score J.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SAFER: A Calibrated Risk-Aware Multimodal Recommendation Model for Dynamic Treatment Regimes." pith.science (2026). https://pith.science/paper/66FSXSUL

@misc{pith2026250606649,
  author       = {Pith},
  title        = {Pith review of: SAFER: A Calibrated Risk-Aware Multimodal Recommendation Model for Dynamic Treatment Regimes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66FSXSUL}},
  note         = {Machine review of arXiv:2506.06649}
}

read the original abstract

Dynamic treatment regimes (DTRs) are critical to precision medicine, optimizing long-term outcomes through personalized, real-time decision-making in evolving clinical contexts, but require careful supervision for unsafe treatment risks. Existing efforts rely primarily on clinician-prescribed gold standards despite the absence of a known optimal strategy, and predominantly using structured EHR data without extracting valuable insights from clinical notes, limiting their reliability for treatment recommendations. In this work, we introduce SAFER, a calibrated risk-aware tabular-language recommendation framework for DTR that integrates both structured EHR and clinical notes, enabling them to learn from each other, and addresses inherent label uncertainty by assuming ambiguous optimal treatment solution for deceased patients. Moreover, SAFER employs conformal prediction to provide statistical guarantees, ensuring safe treatment recommendations while filtering out uncertain predictions. Experiments on two publicly available sepsis datasets demonstrate that SAFER outperforms state-of-the-art baselines across multiple recommendation metrics and counterfactual mortality rate, while offering robust formal assurances. These findings underscore SAFER potential as a trustworthy and theoretically grounded solution for high-stakes DTR applications.

Figures

Figures reproduced from arXiv: 2506.06649 by the authors.

**Figure 1.** The overall framework of SAFER. 2024), to develop an end-to-end safe DTR framework with formal assurances. Clinical notes combined with EHR. Clinical notes, central to patient care, capture physicians’ thought processes, observations, and treatment rationale often absent from structured EHR data (Sheikhalishahi et al., 2019; Rosenbloom et al., 2011). Integrating clinical notes with structured EHRs has demonstrated … view at source ↗

**Figure 2.** Comparative visualization of the treatment frequency matrix in log scale from two datasets. Panel (A) represents MIMICIII, while Panel (B) corresponds to MIMIC-IV. For each patient, we extract 5 types of static demographic variables and 44 types of time-series variables from the tabular data. We set the historical sequence length to 8. All 6 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** FDR and power curves across different target α level and varies uncertainty threshold c on MIMIC-III with Ridge Regression [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: FDR and power curves across different target α level and varies uncertainty threshold c on MIMIC-IV with Ridge Regression. due to high patient heterogeneity in treatment responses. Second, while sequence embedding-based methods excel in classification tasks, their effe…

**Figure 5.** Figure 5: Comparative visualization of the treatment frequency matrix in log scale from the survivor subset of two datasets. Panel (A) represents MIMIC-III, while Panel (B) corresponds to MIMIC-IV. C.2. FDR Control In this section, we present the FDR control results using Linear…

**Figure 6.** Figure 6: FDR and power curves across different target α level with Linear Regression on MIMIC-III. (a) c = 0.1 (b) c = 0.2 (c) c = 0.3 (d) c = 0.4 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

**Figure 7.** Figure 7: FDR and power curves across different target α level with Linear Regression on MIMIC-IV. and the reduction in counterfactual mortality rate, two core evaluation metrics that reflect recommendation accuracy and treatment effectiveness. To ensure fair comparison and robu…

**Figure 8.** Figure 8: The performance of SAFER under different historical information seqeunce length L Hidden dimensionality hd [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

**Figure 9.** Figure 9: The performance of SAFER under different hidden dimensionality hd (a) AUC (b) ↓ Mortality Rate [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

**Figure 10.** Figure 10: The performance of SAFER under different γ value in loss function and improve computational efficiency. γ in the Loss Function [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]

**Figure 11.** Figure 11: Trends in uncertainty scores over time. To further illustrate the interpretability of SAFER’s uncertainty estimates, we select a subset of 10 surviving and 10 deceased 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 8 canonical work pages

[1]

collapse

This establishes that deceased-patient latents (drawn from P −) systematically lead to higher KL-based uncertainty κ(·) than do survivor latents from P +, giving a strict positive gap on average. Hence the theorem’s statement follows. RemarkA.1.By forcing pϕ to remain continuous with respect to h, we guarantee that if latent embeddings of deceased patient...

work page 2023
[6]

A multimodal transformer: Fus- ing clinical notes with structured ehr data for interpretable in-hospital mortality prediction

Lyu, W., Dong, X., Wong, R., Zheng, S., Abell-Hart, K., Wang, F., and Chen, C. A multimodal transformer: Fus- ing clinical notes with structured ehr data for interpretable in-hospital mortality prediction. InAMIA Annual Sympo- sium Proceedings, volume 2022, pp. 719,

work page 2022
[7]

A., Szolovits, P., and Ghassemi, M

Raghu, A., Komorowski, M., Celi, L. A., Szolovits, P., and Ghassemi, M. Continuous state-space models for optimal sepsis treatment—a deep reinforcement learning approach.arXiv preprint arXiv:1705.08422,

arXiv
[11]

R., Wang, W., Zong, B., Chen, H., and Zha, H

Wang, L., Yu, W., He, X., Cheng, W., Ren, M. R., Wang, W., Zong, B., Chen, H., and Zha, H. Adversarial cooperative imitation learning for dynamic treatment regimes. InPro- ceedings of The Web Conference 2020, pp. 1785–1795,

work page 2020
[13]

Safe- drug: Dual molecular graph encoders for recommending effective and safe drug combinations.arXiv preprint arXiv:2105.02711,

Yang, C., Xiao, C., Ma, F., Glass, L., and Sun, J. Safe- drug: Dual molecular graph encoders for recommending effective and safe drug combinations.arXiv preprint arXiv:2105.02711,

arXiv
[14]

teacher” module’s distribution and pϕ is the “student

13 A. Technical Proofs A.1. Lower-Bound Proof for Theorem 4.1. Restate of Theorem 4.1.We have two latent-representation distributions P −(h) and P +(h) over the same latent space H, corresponding todeceasedandsurvivingpatients, respectively. Let h− ∼P −(h) and h+ ∼P +(h). We define the predictive uncertaintyathas κ(h) =D KL pθ(y|h) pϕ(y|h) , where pθ is t...

work page 2009
[1980]

(B3) is a standard assumption on the ignorability of treatment assignment (Shi et al., 2023)

that no interference between units in (B1) and no hidden variations of treatments occur in (B2), If patient i actually receives treatment y, then the observed mortality Mi coincides with the potential outcome Mi(y), which allows us to assume that M= P y Y M(y) almost surely. (B3) is a standard assumption on the ignorability of treatment assignment (Shi et...

work page 2023
[1997]

Accurate Medical Named Entity Recognition Through Specialized NLP Models

Hu, J., Bao, R., Lin, Y ., Zhang, H., and Xiang, Y . Accurate medical named entity recognition through specialized nlp models.arXiv preprint arXiv:2412.08255,

work page Pith review arXiv

Show all 16 references

[2008]

Pre-training of graph augmented transformers for medication recommendation

Shang, J., Ma, T., Xiao, C., and Sun, J. Pre-training of graph augmented transformers for medication recommendation. arXiv preprint arXiv:1906.00346, 2019a. Shang, J., Xiao, C., Ma, T., Li, H., and Sun, J. Gamenet: Graph augmented memory networks for recommending medication co...

1906 arXiv
[2015]

Medical treatment migra- tion behavior prediction and recommendation based on health insurance data.World Wide Web, 23:2023–2042,

Cheng, L., Shi, Y ., and Zhang, K. Medical treatment migra- tion behavior prediction and recommendation based on health insurance data.World Wide Web, 23:2023–2042,

2023
[2016]

K., Fukumizu, K., Gretton, A., Sch¨olkopf, B., and Lanckriet, G

12 Sriperumbudur, B. K., Fukumizu, K., Gretton, A., Sch¨olkopf, B., and Lanckriet, G. R. On integral proba- bility metrics,\phi-divergences and binary classification. arXiv preprint arXiv:0901.2698,

arXiv
[2017]

Z., Dixit, A., Bodrova, A., Singh, S., Tu, S., Brown, N., Xu, P., Takayama, L., Xia, F., Varley, J., et al

Ren, A. Z., Dixit, A., Bodrova, A., Singh, S., Tu, S., Brown, N., Xu, P., Takayama, L., Xia, F., Varley, J., et al. Robots that ask for help: Uncertainty alignment for large lan- guage model planners.arXiv preprint arXiv:2307.01928,

arXiv
[2018]

and Cand`es, E

Jin, Y . and Cand`es, E. J. Model-free selective inference under covariate shift via weighted conformal p-values. arXiv preprint arXiv:2307.09291, 2023a. Jin, Y . and Cand`es, E. J. Selection by prediction with con- formal p-values.Journal of Machine Learning Research, 24(244)...

arXiv
[2020]

Adversarially reg- ularized medication recommendation model with multi- hop memory network.Knowledge and Information Sys- tems, 63:125–142, 2021a

Wang, Y ., Chen, W., Pi, D., and Yue, L. Adversarially reg- ularized medication recommendation model with multi- hop memory network.Knowledge and Information Sys- tems, 63:125–142, 2021a. Wang, Y ., Chen, W., Pi, D., Yue, L., Wang, S., and Xu, M. Self-supervised adversarial di...

2022
[2021]

Conformal alignment: Knowing when to trust foundation models with guarantees.arXiv preprint arXiv:2405.10301,

Gui, Y ., Jin, Y ., and Ren, Z. Conformal alignment: Knowing when to trust foundation models with guarantees.arXiv preprint arXiv:2405.10301,

arXiv
[2024]

Generating with confidence: Uncertainty quantification for black-box large language models.arXiv preprint arXiv:2305.19187,

Lin, Z., Trivedi, S., and Sun, J. Generating with confidence: Uncertainty quantification for black-box large language models.arXiv preprint arXiv:2305.19187,

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

collapse

This establishes that deceased-patient latents (drawn from P −) systematically lead to higher KL-based uncertainty κ(·) than do survivor latents from P +, giving a strict positive gap on average. Hence the theorem’s statement follows. RemarkA.1.By forcing pϕ to remain continuous with respect to h, we guarantee that if latent embeddings of deceased patient...

work page 2023

[2] [6]

A multimodal transformer: Fus- ing clinical notes with structured ehr data for interpretable in-hospital mortality prediction

Lyu, W., Dong, X., Wong, R., Zheng, S., Abell-Hart, K., Wang, F., and Chen, C. A multimodal transformer: Fus- ing clinical notes with structured ehr data for interpretable in-hospital mortality prediction. InAMIA Annual Sympo- sium Proceedings, volume 2022, pp. 719,

work page 2022

[3] [7]

A., Szolovits, P., and Ghassemi, M

Raghu, A., Komorowski, M., Celi, L. A., Szolovits, P., and Ghassemi, M. Continuous state-space models for optimal sepsis treatment—a deep reinforcement learning approach.arXiv preprint arXiv:1705.08422,

arXiv

[4] [11]

R., Wang, W., Zong, B., Chen, H., and Zha, H

Wang, L., Yu, W., He, X., Cheng, W., Ren, M. R., Wang, W., Zong, B., Chen, H., and Zha, H. Adversarial cooperative imitation learning for dynamic treatment regimes. InPro- ceedings of The Web Conference 2020, pp. 1785–1795,

work page 2020

[5] [13]

Safe- drug: Dual molecular graph encoders for recommending effective and safe drug combinations.arXiv preprint arXiv:2105.02711,

Yang, C., Xiao, C., Ma, F., Glass, L., and Sun, J. Safe- drug: Dual molecular graph encoders for recommending effective and safe drug combinations.arXiv preprint arXiv:2105.02711,

arXiv

[6] [14]

teacher” module’s distribution and pϕ is the “student

13 A. Technical Proofs A.1. Lower-Bound Proof for Theorem 4.1. Restate of Theorem 4.1.We have two latent-representation distributions P −(h) and P +(h) over the same latent space H, corresponding todeceasedandsurvivingpatients, respectively. Let h− ∼P −(h) and h+ ∼P +(h). We define the predictive uncertaintyathas κ(h) =D KL pθ(y|h) pϕ(y|h) , where pθ is t...

work page 2009

[7] [1980]

(B3) is a standard assumption on the ignorability of treatment assignment (Shi et al., 2023)

that no interference between units in (B1) and no hidden variations of treatments occur in (B2), If patient i actually receives treatment y, then the observed mortality Mi coincides with the potential outcome Mi(y), which allows us to assume that M= P y Y M(y) almost surely. (B3) is a standard assumption on the ignorability of treatment assignment (Shi et...

work page 2023

[8] [1997]

Accurate Medical Named Entity Recognition Through Specialized NLP Models

Hu, J., Bao, R., Lin, Y ., Zhang, H., and Xiang, Y . Accurate medical named entity recognition through specialized nlp models.arXiv preprint arXiv:2412.08255,

work page Pith review arXiv