{"id":"abf023f9-0047-4158-a0e5-21c02c1328b9","arxiv_id":"2504.17277","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A bandit-based policy for ICU lab test orders, trained with clinical rule bounds and predicted patient status, outperforms physician ordering on a self-defined utility metric in retrospective MIMIC-IV and HiRID data.","lead":"ExOSITO learns a policy for ordering ICU blood tests by combining predictions of future patient status, clinical rules, and cost weights, and reports lower cost and better adherence to rules than the recorded physician policy on two critical care datasets. The method is a candidate clinical decision support tool, but its evaluation uses the same hand-crafted reward it is trained to maximize, with no external outcome validation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 12% GPS improvement is contradicted by Table 1: recomputing g = ΔX − Lb − C gives a 1.4% MIMIC gain and a large HiRID loss, and GPS increases out-of-bounds orders on both datasets.","rationale":"The reader's weakest assumption—that the hand-crafted reward g(t,x) is an unvalidated proxy for clinical utility—is a genuine and important limitation, explicitly acknowledged by the paper in Section 6 ('our dataset does not include true outcome measures'). That concern justifies a conditional verdict on its own. However, the single most load-bearing issue for the paper's central claim is internal: the abstract and results section assert that the GPS-based reliable policy achieves approximately 12% higher average total outcome than the non-GPS policy, but the paper's own tables contradict this when g is computed as specified. The GPS constraint also increases Lb (out-of-bounds orders) on both datasets, undermining the claim that it improves reliability. This is a concrete, checkable inconsistency rather than a question of external clinical validity, so it should be the focus of revision. The reader's Eq. (7) sign concern appears to be a misreading; the Lagrangian is correctly signed for λi ≥ 0. Because the issues are addressable—by correcting the quantitative claim, reporting the actual g values, and re-evaluating whether the GPS constraint helps—a conditional accept remains appropriate, with the required revisions centered on the numerical claim and the reward-validity qualification.","tokens_in":32313,"tokens_out":6014,"duration_ms":58791,"concrete_test":"Recompute g(t,x) = ΔX − Lb − C for every row of Tables 1, 5, 8, 9, and 11 using the reported test-set means and the Appendix F evaluation setting β1 = β2 = 1. Then compute the per-dataset and pooled average improvement of 'Ours (w GPS)' over 'Ours (w/o GPS)' and compare it to the claimed ~12%. Also verify the direction of Lb (out-of-bounds orders) across datasets. If no configuration yields an approximately 12% improvement, the claim should be removed or replaced with the actual computed values, and the GPS benefit needs to be re-established on a configuration where it does not degrade the utility or increase out-of-bounds orders.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central methodological claim is that the GPS overlap constraint makes learning more reliable, with 'the average total outcome of a reliable policy approximately 12% higher than that of a policy trained without GPS constraints.' This is not supported by the paper's own reported numbers. Using the evaluation setting stated in Appendix F (β1 = β2 = 1), the lab-order utility in Eq. (5) is g = ΔX − Lb − C for each test-set row. From Table 1: MIMIC-IV w/o GPS gives 0.44 − 1.06 − 0.80 = −1.42, while w/ GPS gives 0.42 − 1.16 − 0.66 = −1.40, a 1.4% improvement, not 12%. On HiRID, w/o GPS gives 1.08 − 0.62 − 0.57 = −0.11, while w/ GPS gives 1.01 − 0.89 − 0.52 = −0.40, a large degradation. Moreover, the out-of-bounds metric Lb, which the GPS constraint is supposed to reduce, increases on both datasets (MIMIC: 1.06 → 1.16; HiRID: 0.62 → 0.89). Thus the headline '12% higher' claim is either a miscalculation or based on an unstated hyperparameter setting, and the main technical contribution—overlap-guaranteed learning—appears to hurt rather than help the reliability objective. The reader's separate concern about the reward proxy remains valid, but the internal numerical contradiction is more decisive because it undercuts the paper's core advance even under its own evaluation metric. I also checked the reader's claimed sign error in Eq. (7): with λi ≥ 0, the term −λi[f̂(πθ(xi),xi) − ε] equals λi[ε − f̂(·)], which correctly penalizes overlap violations, so I do not find a sign inconsistency there.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ExOSITO, a contextual-bandit method for daily ICU lab test ordering. It augments each patient's 48-hour history with a PatchTST forecast of the next 24 hours, derives per-stay lower and upper bounds on ten test orders from clinician-curated rules (Algorithm 1), and defines a lab-order utility g = ΔX − β1 Lb − β2 C (Eq. 5) that rewards predicted changes in lab values, penalizes deviation from the bounds, and penalizes cost. A conditional normalizing flow estimates the propensity of the logging policy, and Algorithm 2 trains a policy by a Lagrangian objective intended to constrain the learned policy to actions with estimated propensity above a threshold ε. The method is evaluated on MIMIC-IV and HiRID, comparing against random, physician, bound, and prior RL policies, and the paper claims lower cost, better information, and fewer out-of-bound orders than the physician, with a 12% improvement from the GPS constraint.","tokens_in":32678,"tokens_out":11694,"duration_ms":110817,"significance":"Strengths: the problem framing as a clinician-facing assistive bandit is appropriate; the use of clinical rules as side information and forecasts as explainability is a useful idea; the manuscript includes code, two large real-world datasets, and extensive ablations. If the GPS constraint actually improved reliability and if the reward captured clinical utility, this would be a valuable contribution. As it stands, the numerical evidence in the paper's own tables contradicts the main GPS claim, the evaluation is circular with respect to the hand-crafted reward, and the paper acknowledges there are no true outcome measures, so the clinical significance is not yet established.","major_comments":[{"comment":"The claim that 'the average total outcome of a reliable policy is approximately 12% higher than that of a policy trained without GPS constraints' is not supported by Tables 1 and 11 under the evaluation specified in Appendix F. With β1 = β2 = 1, the test metric is g = ΔX − Lb − Cost. On MIMIC-IV, Ours(w/o GPS) gives 0.44 − 1.06 − 0.80 = −1.42 and Ours(w GPS) gives 0.42 − 1.16 − 0.66 = −1.40, a 1.4% change, not 12%. On HiRID, the corresponding values are 1.08 − 0.62 − 0.57 = −0.11 and 1.01 − 0.89 − 0.52 = −0.40, a large degradation. Moreover, the out-of-bounds metric Lb, which the GPS constraint is supposed to reduce, increases on both datasets (MIMIC: 1.06 → 1.16; HiRID: 0.62 → 0.89). Either the 12% figure was computed under different hyperparameters, which must be reported and used consistently, or it is a miscalculation; as written, the paper's central reliability claim is contradicted by its own results.","section":"§5.3, Table 1, Appendix F"},{"comment":"Equation (7) as printed does not enforce the overlap constraint f̂ ≥ ε. The objective is min_θ max_{λ_i ≥ 0} −(1/n)Σ [g − λ_i(f̂ − ε)]. Expanding the inner term gives −g + λ_i(f̂ − ε), so when f̂ > ε the inner maximization is unbounded above as λ_i → ∞, and when f̂ < ε it is maximized at λ_i = 0. The λ update in Algorithm 2, λ ← λ + η_λ(f̂ − ε), likewise increases the multiplier exactly when the constraint is satisfied. The printed optimization therefore encourages f̂ ≤ ε, the opposite of the intended feasible set Π_r. Only the final validation selection (the indicator 1[f̂ > ε]) applies the constraint as stated. This needs to be corrected, and the GPS experiments re-run or re-explained, before the overlap-guaranteeing contribution can be assessed.","section":"§4.4, Eq. (7), Algorithm 2"},{"comment":"The evaluation is circular with respect to the definition of clinical benefit. The policy is trained to maximize g(t,x) in Eq. (5), and the reported comparisons use the same g with β1 = β2 = 1 (Appendix F); the Lb and C terms are simultaneously training objectives and evaluation metrics, so the learned policy is being compared with the physician on its own reward. The paper states in Section 6 that the dataset has no 'true' outcome measures quantifying lab test utility. Consequently, statements such as 'reduces costs without omitting any vital lab orders' and 'outperforming both a physician's policy and prior approaches' are only claims about the hand-crafted proxy. Table 1 additionally shows nonzero Llow for both learned policies (MIMIC w GPS: 0.67; HiRID w GPS: 0.5), which contradicts the abstract's 'without omitting any vital lab orders.' The authors should either provide expert/outcome validation or substantially temper the clinical claims.","section":"§4.3, §5.2, §6"}],"minor_comments":[{"comment":"The bound loss only penalizes tests where tmax_j = tmin_j; for tests with tmax_j > tmin_j there is no penalty at all, so the sentence 'The indicator function ensures the policy includes all necessary labs suggested by rules while avoiding redundant ones' overstates the constraint that Lb actually imposes.","section":"Eq. (3), §4.3"},{"comment":"The notation for random restarts is inconsistent: Section 5.2 says m = 5 random restarts, while Appendix F refers to k = 5 experimental runs; please unify the notation.","section":"§5.2, Appendix F"},{"comment":"The rule 'If patient Phosphate is greater than 0.6 or greater than 1.8' is likely a typo for 'less than 0.6 or greater than 1.8'; please correct it.","section":"Appendix C"},{"comment":"The text says the MIMIC-IV database follows 'HIPPA'; the correct acronym is HIPAA.","section":"Appendix A"},{"comment":"The row '10, 5 (Main result)' is ambiguous: please clarify whether these numbers refer to tests per 24-hour window or per 48-hour window, and align them with the description in Section 5.1.","section":"Appendix G, Table 7"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant health-ML application and has a useful framework, but the two load-bearing problems need to be fixed before publication: the headline 12% GPS result is contradicted by the authors' own tables, and Eq. (7) inverts the intended overlap constraint. I would support a revised version that corrects the Lagrangian, re-reports the GPS comparison using the stated evaluation metric, and reframes the clinical claims in light of the acknowledged absence of true outcome measures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, ExOSITO is a genuine engineering contribution: it is a clinician-facing causal bandit for ICU lab test ordering that combines clinical rules as side information, forecasted patient trajectories, and a constrained off-policy learner, with code, two real ICU datasets, and careful ablations. Second, the headline claim about the GPS constraint—\"the average total outcome of a reliable policy is approximately 12% higher\"—does not survive arithmetic on the paper's own Table 1.\n\nWhat is actually good: the framing is honest about the clinician-in-the-loop problem, the reward design is transparent and clinically motivated, and the ablations in Appendices G and H show real effort to isolate the contribution of each component. The paper also states clearly in Section 6 that no true outcome measures exist for lab test utility, which is the right kind of caveat. The citation pattern is fine; the relevant off-policy RL and clinical utilization literature is covered.\n\nNow the soft spots, in proportion. The internal numerical contradiction is the load-bearing one. Using the evaluation setting in Appendix F (β1=β2=1), g = ΔX − Lb − C. On MIMIC-IV: w/o GPS gives 0.44−1.06−0.80 = −1.42; w/ GPS gives 0.42−1.16−0.66 = −1.40, a 1.4% improvement, not 12%. On HiRID: w/o GPS gives 1.08−0.62−0.57 = −0.11; w/ GPS gives 1.01−0.89−0.52 = −0.40, a large degradation. Worse, the GPS constraint is supposed to reduce out-of-bounds orders, but Lb increases on both datasets (1.06→1.16 and 0.62→0.89). Either the 12% figure is a miscalculation or it comes from an unstated hyperparameter setting; as written, the main technical contribution appears to hurt the reliability objective under its own metric.\n\nThe reward circularity is real but secondary. Training and evaluation use the same hand-crafted g, and the paper acknowledges the absence of true outcomes. That makes \"outperforms the physician\" relative to a self-defined utility, not a clinical validation. If reframed as optimizing a clinician-defined utility, the contribution still stands, but the abstract overstates it.\n\nOne reader concern I do not share: the claimed sign error in Eq. (7). With λi ≥ 0, the term −λi[f̂(πθ(xi),xi) − ε] correctly penalizes overlap violations. That particular worry does not hold up.\n\nWho this is for: people working on off-policy learning in healthcare and on clinical decision support. It deserves a serious referee—send it to peer review—but the authors need to remove or re-derive the 12% GPS claim, and the discussion should be reframed as optimizing a defined utility rather than demonstrating clinical benefit. I would cite it for the problem framing and the careful reward construction, not for the GPS reliability evidence.","headline":"A useful, reproducible clinician-facing bandit for ICU lab ordering whose core GPS reliability claim is contradicted by its own Table 1.","tokens_in":33312,"tokens_out":2701,"would_cite":true,"duration_ms":28729,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ExOSITO learns ICU lab orders that cut cost without missing vital tests.","keywords":["off-policy learning","contextual bandits","ICU lab test ordering","clinical decision support","global propensity score","time-series forecasting","MIMIC-IV","HiRID"],"falsifier":"Run a silent ICU deployment: log ExOSITO's daily recommended orders without acting on them, and check whether the tests it omits but physicians ordered are followed within 48 hours by a clinically significant event such as a new vasopressor start, transfusion, arrhythmia, or death. If omitted tests are systematically followed by such events, the paper's claim that no vital lab orders are omitted is contradicted.","tokens_in":32011,"feed_emoji":"🧪","tokens_out":10657,"duration_ms":86254,"temperature":0.7,"pith_summary":"The paper tries to establish that ICU lab test ordering can be improved without sacrificing essential tests by reframing it as an off-policy contextual bandit whose reward is built for clinicians rather than for mortality. It claims that a policy balancing how much a test is expected to change a patient's lab values, how well an order respects clinical rule-based bounds, and what each test costs beats the observed physician policy and prior RL methods on two large ICU datasets. The practical point is that such a tool could reduce over-ordering, with its costs, patient discomfort, and environmental waste, while keeping the clinician as the final decision maker and making each recommendation explainable.","feed_headline":"Learned policy cuts ICU lab test orders without missing vital tests","feed_subtitle":"Beats physician ordering on MIMIC-IV and HiRID while keeping vital labs in the daily set.","key_machinery":"The load-bearing object is the lab order utility function $g(t,x)=\\Delta X(t,x)-\\beta_1 L_b(t,x)-\\beta_2 C(t)$, where $\\Delta X$ scores the mean and range change between observed past and predicted future lab values, $L_b$ penalizes deviations from rule-derived minimal and maximal order bounds, and $C$ charges test costs. It carries the argument by turning whether an order is useful to a clinician into a differentiable objective. Around it, the method builds a PatchTST forecasting model to supply predicted future patient status for both the reward and explainability, and a conditional normalizing flow estimate of the global propensity score (the conditional density of test orders given patient status) that enforces overlap through a Lagrangian constraint, keeping the learned policy inside actions the data can support.","core_discovery":"On its own terms, ExOSITO's central claim is that a reliable, explainable lab-ordering policy can be learned offline from routine EHR data, and that this policy dominates the observed physician policy: it yields higher lab-test informativeness (measured by mean and extreme-value changes between predicted future and observed past lab values), lower cost, and fewer out-of-bounds orders relative to clinician-derived minimal and maximal order sets. The paper further claims that adding a learned global propensity score constraint raises the average total outcome by about 12% compared with training without the constraint, and that prior RL methods using mortality as reward are ill-suited because lab tests help clinicians decide rather than directly changing patient outcomes.","pith_inferences":["Editorial inference: the same template of predicted change, rule bounds, and cost could transfer to other recurring diagnostic-ordering decisions such as imaging or microbiology where the clinician stays the decision maker and overuse is driven by precaution.","A testable extension the paper does not run: compare ExOSITO against a purely rule-based ordering policy under the same utility function, which would isolate how much of the gain comes from learned patient-specific adaptation rather than from the rules themselves.","The paper's admission that the data contain no true utility outcomes implies the decisive validation would be a silent trial in which recommendations are logged but not enacted, and omitted tests are checked against downstream deterioration.","Left implicit: since the reward includes predicted future lab values, the policy should improve as forecasting accuracy improves; the ablations point that way but do not map the sensitivity curve."],"forward_implications":["Daily ICU lab ordering can be reduced in cost and volume without increasing the number of missed rule-required tests, because the learned policy holds orders inside clinician-derived bounds.","Mortality-based RL rewards are the wrong signal for lab-ordering decisions; an informativeness-plus-bounds-plus-cost objective is what lets a learned policy beat the physician policy.","Adding the GPS overlap constraint yields a more reliable policy; the paper reports roughly 12% higher average total outcome for GPS-constrained than unconstrained training.","Because each recommendation is tied to concrete patient data and a predicted future trajectory, clinicians can audit why a test was recommended, which is what makes clinician-facing deployment plausible."],"supporting_citations":[{"why":"Supplies the RL baseline whose mortality-based reward ExOSITO argues is unsuited to lab ordering, and the policy that ExOSITO must beat on cost and bound adherence.","marker":"Chang et al., 2019"},{"why":"Source of the GPS-estimation and overlap-guaranteed off-policy learning machinery that keeps the learned policy inside data-supported actions.","marker":"Schweisthal et al., 2023"},{"why":"Provides the PatchTST architecture used as the patient-status forecasting model that generates predicted future lab values for the reward.","marker":"Nie et al., 2022"},{"why":"Primary dataset (MIMIC-IV) used to train and evaluate the learned ordering policy.","marker":"Johnson et al., 2023"},{"why":"Second ICU dataset (HiRID) used to show the method transfers to an independent patient cohort.","marker":"Hyland et al., 2020"},{"why":"Grounds the consistency, ignorability, and overlap assumptions under which the potential-outcome reward is identifiable.","marker":"Rubin, 1974"},{"why":"Introduces the conditional normalizing flows used to estimate the global propensity score.","marker":"Trippe and Turner, 2018"},{"why":"Supplies real-world lab test cost values for the cost term in the reward function.","marker":"Kandalam et al., 2020"},{"why":"Supplies additional real-world cost data for normalizing the cost term in the reward function.","marker":"Spoyalo et al., 2023"}],"fun_headline_variants":["ExOSITO: Explainable off-policy learning trims ICU lab orders","AI learns to order fewer ICU labs, keeps vital ones","Interpretable AI beats clinicians on ICU lab test sets","Side info + off-policy = better ICU lab ordering","ExOSITO: ICU lab orders that explain themselves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pivotal assumption is that the paper's reward formula -- how much a test's values are predicted to change, how well an order matches the clinical rules, and what it costs -- actually captures what makes a lab order worth placing; if it does not, then outperforming the physician policy on this formula is not the same as improving care.","fun_headline_variants_meta":{"raw":{"variants":["ExOSITO: Explainable off-policy learning trims ICU lab orders","AI learns to order fewer ICU labs, keeps vital ones","Interpretable AI beats clinicians on ICU lab test sets","Side info + off-policy = better ICU lab ordering","ExOSITO: ICU lab orders that explain themselves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000731,"raw_usage":{"total_tokens":3245,"prompt_tokens":892,"completion_tokens":2353,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":2270}},"tokens_in":508,"tokens_out":2353,"duration_ms":17622,"temperature":1.0,"reasoning_tokens":2270,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:45:01.598757+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a silent ICU deployment: log ExOSITO's daily recommended orders without acting on them, and check whether the tests it omits but physicians ordered are followed within 48 hours by a clinically significant event such as a new vasopressor start, transfusion, arrhythmia, or death. If omitted tests are systematically followed by such events, the paper's claim that no vital lab orders are omitted is contradicted.","supporting_citations":[{"cited_title":"Dynamic measurement scheduling for event forecasting using deep rl","cited_arxiv_id":null,"evidence_quote":"Supplies the RL baseline whose mortality-based reward ExOSITO argues is unsuited to lab ordering, and the policy that ExOSITO must beat on cost and bound adherence."},{"cited_title":"Reliable Off-Policy Learning for Dosage Combinations","cited_arxiv_id":"2305.19742","evidence_quote":"Source of the GPS-estimation and overlap-guaranteed off-policy learning machinery that keeps the learned policy inside data-supported actions."},{"cited_title":"Inappropriate repeat testing of complete blood count (cbc) and electrolyte panels in inpatients from alberta, canada","cited_arxiv_id":null,"evidence_quote":"Supplies real-world lab test cost values for the cost term in the reward function."},{"cited_title":"Patient, hospital and environmental costs of unnecessary bloodwork: capturing the triple bottom line of inappropriate care in general surgery patients","cited_arxiv_id":null,"evidence_quote":"Supplies additional real-world cost data for normalizing the cost term in the reward function."}],"review_version":1}