{"id":"aa72e31a-0f7b-46df-bed6-b27e25d3fb91","arxiv_id":"2510.08350","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"An offline RL policy trained on ICU feeding records recommends personalized enteral nutrition targets and reports an estimated 18.8% vs 22.5% mortality advantage over clinicians.","lead":"DeepEN uses offline reinforcement learning on 11,000 MIMIC-IV ICU stays to recommend personalized 4-hourly calorie, protein, and water targets for tube-fed patients. It reports an estimated 3.7-percentage-point mortality advantage over observed clinician practice, but that estimate is produced by an off-policy evaluation loop that shares its reward with the training objective.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mortality estimate is calibrated on returns that include the terminal survival/death reward, so Table 2's 18.8% vs 22.5% may be circular and not a valid counterfactual.","rationale":"The paper's central claim is the 3.7-percentage-point mortality reduction (18.8% vs 22.5%). The reader identified the bin-mapping in Section 3.2.4(3) as the weakest assumption. My analysis agrees: the bin-mapping uses returns that contain the terminal mortality reward, making the mortality estimate partly circular. Additionally, the clinician baseline is an observed value while DeepEN's is estimated, so the comparison is not apples-to-apples. The proposed concrete test would directly assess whether the bin-mapping is calibrated. If the calibration fails, the headline claim collapses. This supports the reader's REJECT verdict: the current evidence does not establish the claimed mortality benefit.","tokens_in":13586,"tokens_out":4568,"duration_ms":96164,"concrete_test":"On held-out data (or via cross-validation), compute expected returns for the clinician policy, apply the bin-mapping from Section 3.2.4(3), and compare the predicted mortality to the observed 22.5%. If the predicted value deviates by more than ~1 percentage point, the bin-mapping is not calibrated and Table 2's mortality estimates for DeepEN are unreliable. Also rerun the binning using returns with the terminal reward term removed; if the DeepEN vs clinician mortality gap vanishes, the headline result is an artifact of circularity.","verdict_should_be":"REJECT","load_bearing_attack":"Section 3.2.4(3) converts expected returns into estimated mortality by binning returns and using observed mortality within bins. But returns include the terminal reward Rterm(m)=+15 for ICU survival and −15 for death (Section 3.1.3), the same outcome the policy is trained to maximize. The bin-mapping therefore learns the training signal itself; a policy with higher expected return will automatically be assigned lower 'estimated mortality' even if it does not truly reduce mortality. Worse, the clinician policy's mortality in Table 2 is the observed cohort mortality (22.5%), while DeepEN's 18.8% is a model-derived estimate. Comparing an observed value to a calibrated prediction is apples-to-oranges and can produce a spurious 3.7-point gap. No validation of the bin-mapping is provided (e.g., predicting the clinician policy's mortality on held-out data). Since the central claim rests entirely on this estimator, it is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DeepEN formulates enteral nutrition dosing in the ICU as an offline RL problem, using a D3QN with CQL on 11,378 MIMIC-IV patients. The state includes 102 variables, actions are four-level discretized calorie/protein/water targets, and the reward combines a terminal survival/death reward (R_term = +15/-15) with intermediate SOFA, lactate, glucose, and phosphate rewards. The paper reports that DeepEN achieves the lowest estimated mortality (18.8% vs. 22.5% for clinicians) and the highest CWPDIS return, and claims that deviations from the policy are independently associated with mortality and biomarker instability. The central evaluation, however, rests on a mortality estimator that is circular: it converts expected returns, which already include the terminal survival/death reward, into a mortality estimate via binning. This undermines the headline result.","tokens_in":13887,"tokens_out":2803,"duration_ms":37470,"significance":"If the central claim were valid, DeepEN would be a notable first application of offline RL to enteral nutrition, and the inclusion of a physiologically motivated reward and conservative Q-learning would be a useful contribution. The paper also makes code and data availability explicit, uses a substantial MIMIC-IV cohort, and compares against four baselines. However, the headline mortality reduction is not established: the estimated mortality in Table 2 is a model-derived calibration of the same return signal that the policy was trained to optimize, so the 18.8% vs. 22.5% result can be produced by construction even if the policy has no true mortality benefit. The 'independent association' claim in the abstract is also not supported by the unadjusted, binned analyses in Section 3.2.4. As presented, the paper is a proof-of-concept of an RL framework, but its main clinical performance claim requires a non-circular evaluation before it can be accepted.","major_comments":[{"comment":"The estimated-mortality procedure is circular. The terminal reward in Eq. (3.1.3) is +15 for ICU survival and -15 for death, so the expected return of a trajectory is a direct function of the outcome used to define mortality. Section 3.2.4(3) then bins expected returns and reads observed mortality in those bins to estimate a policy's mortality rate. A policy trained to maximize return will automatically be assigned a lower 'estimated mortality' even if it does not reduce mortality. Moreover, Table 2 compares the clinician policy's observed mortality (22.5%) with DeepEN's calibrated estimate (18.8%), which is an apples-to-oranges comparison. No validation of the bin-mapping is provided (for example, applying the same mapping to the clinician policy and checking that it reproduces 22.5%). Because this estimator is the sole basis for the headline 3.7 percentage-point reduction, the central","section":"Sec. 3.2.4(3), Sec. 3.1.3, Table 2"},{"comment":"The abstract states that deviation from the DeepEN policy was 'independently associated' with increased mortality and biomarker instability, but the main text presents only unadjusted, binned plots of dosage difference versus mortality (Figure 3) and versus biomarker deviation (Figure 4). There is no multivariable analysis, no adjustment for confounders such as severity of illness or concurrent treatments, and no formal statistical test for the U-shaped relationship. These plots can reflect confounding by indication (sicker patients receiving different dosages) rather than a causal effect of deviation. The word 'independently' is therefore unsupported by the evidence in the manuscript.","section":"Sec. 3.2.4(4), Fig. 3, Abstract"},{"comment":"The quantitative OPE results lack sufficient detail to be reproduced or interpreted. The definition of the expected-return bins is not given (e.g., number of bins, boundaries, whether bins are fit on the training or test set), and the reported confidence intervals (e.g., 3.7±0.17) do not state the resampling or uncertainty procedure. For CWPDIS, the behavior policy is a BC model fit to the same data, and policies like the expert-guidelines policy or random policy may fall outside the support of the behavior policy, making importance weights unstable. The absence of these details is particularly important because the entire performance comparison rests on these estimates.","section":"Table 2 and Sec. 3.2.4"}],"minor_comments":[{"comment":"There is a duplicated sentence: 'DeepEN also resulted the highest CWPDIS score among the policies, indicating superior expected returns under the learned value function (Table 2).' appears twice in the same paragraph.","section":"Sec. 4.1, Table 2"},{"comment":"The abstract reports a 4.0 percentage-point reduction (22.8% to 18.8%), while the main text and Table 2 report 22.5% and a 3.7-point reduction. Please harmonize these numbers.","section":"Abstract vs. Sec. 4.1"},{"comment":"The intermediate reward notation is inconsistent: the definition of R_t writes R_im(t), but the text and equations use R_phys and R_bio without defining R_im. Please clarify the relationship between these terms.","section":"Sec. 3.1.3, Eq. (1)"},{"comment":"The p-value for the return-mortality correlation is reported as p<0.001 without specifying the test or the correlation coefficient. Please provide the test name, the effect size, and the confidence interval.","section":"Sec. 3.2.4(2), Fig. 2"},{"comment":"The paper acknowledges the handcrafted nature of the reward and the single-center data. These limitations are appropriately stated, but they should be connected to the evaluation: a handcrafted reward that includes terminal survival/death makes the return-based mortality calibration particularly suspect, not merely a future direction.","section":"Sec. 3.2.3, Sec. 5"}],"recommendation":"reject","confidential_remarks":"The central result is not merely weak but circular: the estimated mortality is derived from the same return signal that includes the terminal reward used in training. This is not fixable by polishing the text; it requires a fundamentally different evaluation strategy (e.g., direct outcome models that separate the terminal reward from the return, or proper off-policy estimators that do not use the trained return to read out mortality). Because the abstract and title emphasize the mortality reduction, the paper cannot be accepted in its current form. The reviewer sees value in the RL formulation for nutrition and the availability of code, but the present evaluation does not support the claimed clinical benefit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nOne thing to know up front: this is the first RL framework aimed at enteral nutrition dosing in the ICU, and the MDP design is genuinely careful and clinically informed. The other thing: the headline mortality result — DeepEN 18.8% vs clinician 22.5%, a 3.7-point reduction — is circular and should not be taken as evidence of reduced mortality.\n\nThe circularity is concrete. Section 3.1.3 defines the terminal reward as +15 for survival, −15 for death; those values dominate the reward scale (intermediate shaping terms are an order of magnitude smaller). Section 3.2.4(3) then bins each trajectory's expected return, computes observed mortality within bins, and uses that calibration curve to convert any policy's expected return into an 'estimated mortality rate.' Since the return already contains the terminal survival/death reward, the calibration curve is essentially a monotone relabeling of the return. Any policy with a higher estimated return automatically gets assigned a lower mortality. Figure 2's strong return–mortality correlation (p<0.001) is guaranteed by construction, not evidence of clinical alignment. On top of that, the clinician baseline (22.5%) is the raw observed cohort mortality, while DeepEN's 18.8% is a model-derived estimate — an apples-to-oranges comparison. No held-out or leave-one-out validation of the bin-mapping is reported.\n\nCredit where it's due: the application is new and relevant — the ICU nutrition literature has conflicting trial evidence, and a personalized dosing policy could matter. The state space (102 variables with trends), the action space that counts IV fluids and propofol as part of total energy/fluid, and the glucose/phosphate target-range shaping are all sensible and well motivated. The baseline set (random, clinician, behavior cloning, ASPEN-guideline) is the right instinct, and the CWPDIS return comparisons (DeepEN 11.89 vs guideline 8.11 vs clinician 5.87) are a legitimate internal metric — still shaped by the same reward, but at least a true off-policy estimator. Code is on GitHub, data is open MIMIC-IV, and the limitations section is honest about single-center data and the handcrafted reward.\n\nOther soft spots, in proportion. The abstract's claim that deviation from DeepEN is 'independently associated' with mortality is not supported by any multivariable analysis in the main text; the deviation plots are unadjusted binned rates. Hyperparameters were selected on the same OPE metrics that are then reported as results. Minor: no uncertainty interval for the clinician baseline, and the cohort excludes non-feeding periods.\n\nThe paper deserves a serious referee — the framework and domain work are real and the evaluation pitfall is worth dissecting. But the central claim needs to change: either validate the calibration properly (e.g., returns excluding the terminal reward, on held-out data) or drop the mortality framing and report returns as returns. I'd send it to review with a clear note that the 18.8% vs 22.5% comparison, as currently constructed, is not a valid counterfactual mortality estimate.","headline":"First RL-for-enteral-nutrition framework, thoughtfully designed; but the headline 3.7-point mortality gap is a circular relabeling of returns that already contain the terminal survival reward — don't take it at face value.","tokens_in":14335,"tokens_out":12970,"would_cite":true,"duration_ms":130942,"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":"DeepEN, an offline reinforcement learning framework for ICU enteral nutrition, is estimated to cut mortality by 3.7 percentage points versus clinician practice.","keywords":["reinforcement learning","offline reinforcement learning","enteral nutrition","critical care","personalized medicine","ICU mortality","clinical decision support","biomarker stability"],"falsifier":"Compute the same score-to-mortality translation for a version of the policy trained without the terminal survival/death reward, keeping only the biomarker and physiology shaping rewards; if the estimated mortality advantage over clinicians disappears or shrinks significantly, the headline result would reflect the evaluation loop rather than the dosing decisions.","tokens_in":1348,"feed_emoji":"🥗","tokens_out":1692,"duration_ms":55542,"temperature":0.7,"pith_summary":"DeepEN asks whether a machine-learning policy can personalize calorie, protein, and fluid targets for tube-fed ICU patients every four hours, and it claims the answer is yes. Trained on more than 11,000 patients from a public intensive-care database, the policy is estimated to achieve 18.8% mortality versus 22.5% for observed clinician behavior, and it also outperforms an expert guideline policy in expected return. The paper argues that this gain comes from conditioning nutrition decisions on evolving organ function, labs, vital signs, and trends rather than static per-kilogram rules, and from balancing long-term survival with short-term metabolic stability. If true, this would mean data-driven personalization could complement—not replace—guideline-based nutrition care and address part of the persistent gap between guidelines and actual practice.","feed_headline":"RL feeding policy cuts estimated ICU mortality by 3.7 points","feed_subtitle":"A policy trained on ICU records recommends four-hourly calorie, protein, and fluid targets that beat both clinician and expert-guideline bas","key_machinery":"The argument is carried by a composite reward function paired with a conservative offline Q-learning update. The reward gives a terminal signal of +15 for survival and −15 for death, plus intermediate shaping rewards: a physiological term penalizing SOFA stagnation and worsening lactate, and a biomarker term rewarding maintenance of glucose within 140–180 mg/dL and phosphate within 2.5–4.5 mg/dL (with a bonus for improvement toward target). This reward is optimized by a dueling double DQN with conservative Q-learning regularization, which suppresses overestimation of rarely seen or unseen actions so the learned policy stays close to observed clinical practice. The evaluation then converts ex","core_discovery":"DeepEN's central claim is that a conservative offline reinforcement learning policy—specifically a dueling double deep Q-network with conservative Q-learning regularization—can learn safe, individualized enteral nutrition dosing from retrospective data. The policy selects 4-hourly combinations of weight-adjusted calories, protein, and water from a discretized space of 51 observed dose combinations, using a 102-variable state representation that includes demographics, diagnoses, vitals, labs, feeding history, and concurrent interventions like vasopressors and insulin. The paper reports the lowest calibrated mortality (18.8% ± 1.0%) among all tested policies—3.7 percentage points below the cli","pith_inferences":["A natural next test, beyond the paper's internal evaluation, would be to run the same return-to-mortality mapping on a version of the policy trained without the terminal survival/death reward; that would reveal how much of the headline mortality advantage depends on the reward's own mortality signal.","The paper's action space covers only periods where enteral nutrition is actively given, so the policy is conditional on a decision to feed; a fuller decision-support tool would also need to decide when to withhold or interrupt feeding, a direction the paper itself flags.","The same reward-design logic—terminal survival balanced with biomarker shaping—could plausibly transfer to other slow-effect ICU interventions such as sedation titration or glycemic control, where immediate physiological targets and long-term outcomes both matter."],"forward_implications":["If DeepEN is right, ICU nutrition could move from fixed weight-based targets to dynamic, four-hourly recommendations conditioned on the patient's organ function, labs, and recent feeding response.","Because the learned policy outperforms behavior cloning, the benefit does not come solely from imitating good clinician decisions; the reinforcement learning optimization itself adds value.","The inclusion of biomarker stability in the reward means following the policy should also yield tighter glucose and phosphate control, not only a mortality difference.","The U-shaped deviation curves imply that both underfeeding and overfeeding relative to the learned targets are harmful, supporting a middle-range rather than a maximal-or-minimal dosing philosophy for enteral nutrition.","A reduction of roughly 37 deaths per 1,000 patients is comparable to or larger than the modest effect sizes seen in recent nutrition randomized trials, making personalization a potentially meaningful clinical lever if confirmed externally."],"fun_headline_variants":["DeepEN reduces ICU mortality by 4 percentage points","RL-based enteral nutrition cuts ICU death rate by 4 points","Offline deep RL personalizes ICU feeding, lowers mortality 4%","Dueling DQN feeding policy beats clinicians in ICU survival","AI-driven nutrition targets improve ICU outcomes by 4%"],"cache_read_input_tokens":15744,"weakest_assumption_plain":"The headline mortality comparison assumes that the procedure translating a policy's expected score into a mortality rate produces a true counterfactual estimate, even though the score itself is built from the same survival/death signal the policy was trained on.","fun_headline_variants_meta":{"raw":{"variants":["DeepEN reduces ICU mortality by 4 percentage points","RL-based enteral nutrition cuts ICU death rate by 4 points","Offline deep RL personalizes ICU feeding, lowers mortality 4%","Dueling DQN feeding policy beats clinicians in ICU survival","AI-driven nutrition targets improve ICU outcomes by 4%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000303,"raw_usage":{"total_tokens":1618,"prompt_tokens":824,"completion_tokens":794,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":707}},"tokens_in":568,"tokens_out":794,"duration_ms":7003,"temperature":1.0,"reasoning_tokens":707,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T10:44:53.787755+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the same score-to-mortality translation for a version of the policy trained without the terminal survival/death reward, keeping only the biomarker and physiology shaping rewards; if the estimated mortality advantage over clinicians disappears or shrinks significantly, the headline result would reflect the evaluation loop rather than the dosing decisions.","supporting_citations":[],"review_version":1}