{"id":"5ebc8a7f-04da-475f-8d1f-2ecb5869d626","arxiv_id":"1907.05496","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"Contextual linear bandits outperform fixed dosing and match clinical algorithms for initial Warfarin dose prediction on PharmGKB patient data.","lead":"This paper applies contextual linear bandit algorithms to predict initial Warfarin doses from patient features in real PharmGKB data. A smart generalist might read it to understand how online learning methods can support personalized dosing and reduce trial-and-error risks in common medical treatments.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Offline replay on PharmGKB historical outcomes assumes unbiased proxy for online policy performance without correcting for distribution shift or exploration harm","rationale":"The reader's weakest_assumption directly identifies the same evaluation-validity gap that the abstract's performance claims depend on. Because the full text is not supplied here, no additional internal inconsistency can be checked, so the reader's UNVERDICTED stance is left unchanged.","tokens_in":1689,"tokens_out":322,"duration_ms":15995,"concrete_test":"Re-run the reported linear bandit algorithms on the PharmGKB data using inverse-propensity scoring (or a doubly-robust variant) with the historical dosing policy as behavior policy; if the relative ranking versus fixed-dose and clinical baselines changes by more than one standard error or loses statistical significance, the original offline results are unreliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (algorithms outperform fixed-dose and match clinical algorithm) rests on evaluating linear bandit variants via offline replay of the PharmGKB dataset. This requires that observed patient outcomes under historical (non-bandit) policies serve as valid counterfactual rewards for the learned online policy. The setup does not appear to apply importance sampling, doubly robust estimators, or any correction for the fact that the data-generating policy differs from the bandit exploration policy; nor does it bound regret or safety violations that would arise during actual online deployment. If the historical dosing distribution is concentrated away from the regions the bandit would explore, performance estimates are biased and the outperformance claim does not transfer to the online regime.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper applies variants of contextual linear bandit algorithms to predict initial warfarin doses using features from the PharmGKB dataset. It claims that all proposed algorithms outperform a fixed-dose baseline and that some match the performance of the Warfarin Clinical Dosing Algorithm, with evaluation performed via offline replay of historical patient outcomes.","tokens_in":1836,"tokens_out":477,"duration_ms":24692,"significance":"If the offline evaluation is shown to be unbiased, the work would illustrate a practical use of linear bandits for dose personalization in a clinically relevant setting with real patient data, offering a potential improvement over fixed dosing. The empirical comparison to established baselines is a strength when properly validated.","major_comments":[{"comment":"The central empirical claim (outperformance over fixed-dose and parity with the clinical algorithm) rests on offline replay of PharmGKB historical outcomes, yet the manuscript provides no description of the evaluation protocol, importance sampling weights, doubly robust estimators, or any correction for the mismatch between the historical data-generating policy and the bandit exploration policy. This is load-bearing for the results reported in the abstract.","section":"Abstract and Experiments section"},{"comment":"No error bars, statistical significance tests, or handling of censored/missing outcomes are reported for the performance comparisons, making it impossible to assess whether the claimed parity with the clinical algorithm is robust or an artifact of the replay procedure.","section":"Experiments section"},{"comment":"The evaluation assumes historical outcomes under non-bandit policies serve as valid counterfactual rewards for the learned online policy without bounding distribution shift or exploration harm; this assumption is not justified or tested, directly affecting transferability of the outperformance claim to an actual online deployment setting.","section":"Methods and Evaluation"}],"minor_comments":[{"comment":"The abstract and introduction should explicitly state the number of patients, feature dimensionality, and the exact linear bandit variants (e.g., LinUCB, Thompson sampling) used.","section":"Abstract"},{"comment":"Notation for the contextual linear model and reward function should be introduced with a clear equation early in the methods section.","section":"Methods"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which highlight important gaps in the description and validation of our offline evaluation. We agree that these elements are central to the claims and will revise the manuscript accordingly. Our point-by-point responses follow.","responses":[{"response":"We acknowledge that the original manuscript omitted a clear description of the offline replay procedure. In the revised version we will insert a dedicated 'Evaluation Protocol' subsection that specifies how each bandit policy is simulated on the fixed PharmGKB dataset: at each step the policy selects an action for the current patient context, the historical outcome for that patient is used as the observed reward, and the process continues sequentially. We did not apply importance sampling or doubly robust corrections; we will explicitly state this choice and its limitations, noting that the historical dosing policy is treated as fixed and that any mismatch with the bandit exploration policy is not corrected. We will also add a short paragraph discussing the implications for the abstract claims.","revision_made":"yes","referee_comment":"[Abstract and Experiments section] The central empirical claim (outperformance over fixed-dose and parity with the clinical algorithm) rests on offline replay of PharmGKB historical outcomes, yet the manuscript provides no description of the evaluation protocol, importance sampling weights, doubly robust estimators, or any correction for the mismatch between the historical data-generating policy and the bandit exploration policy. This is load-bearing for the results reported in the abstract."},{"response":"We agree that the absence of variability measures weakens the empirical claims. The revision will report standard errors (computed via bootstrap resampling of the patient sequence) for all reported metrics and will include paired statistical tests (e.g., Wilcoxon signed-rank) comparing each bandit variant against the fixed-dose and clinical baselines. For missing outcomes in PharmGKB we will document the exact imputation or exclusion rule used and add a sensitivity table showing results under alternative handling strategies.","revision_made":"yes","referee_comment":"[Experiments section] No error bars, statistical significance tests, or handling of censored/missing outcomes are reported for the performance comparisons, making it impossible to assess whether the claimed parity with the clinical algorithm is robust or an artifact of the replay procedure."},{"response":"This is a substantive limitation of the current offline replay approach. The revised manuscript will contain an expanded 'Limitations' paragraph that states the untested assumption, notes the lack of distribution-shift bounds, and cautions that the reported gains may not translate directly to prospective online use. We will also outline a possible future direction using conservative policy evaluation techniques, but we cannot retroactively apply such bounds to the existing experiments without additional data or modeling assumptions not present in the PharmGKB release.","revision_made":"partial","referee_comment":"[Methods and Evaluation] The evaluation assumes historical outcomes under non-bandit policies serve as valid counterfactual rewards for the learned online policy without bounding distribution shift or exploration harm; this assumption is not justified or tested, directly affecting transferability of the outperformance claim to an actual online deployment setting."}],"tokens_in":1342,"tokens_out":649,"duration_ms":13537,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper takes existing contextual linear bandit methods and runs them on the warfarin initial dosing task with the public PharmGKB dataset. It reports that the bandit variants beat a fixed-dose baseline and that some reach parity with the clinical dosing algorithm in the offline results. That is the core contribution: a straightforward application to a real medical dataset rather than new algorithmic machinery. The authors also sketch a few future directions at the end. The work is honest about using established methods on an established dataset, and the medical framing is clear enough that a reader can see why dosing variability matters. The evaluation uses historical patient outcomes, which at least grounds the comparison in actual data instead of simulation. The soft spot is the evaluation itself. The claims rest on replaying past outcomes under the assumption that they serve as valid counterfactuals for a bandit policy. The abstract supplies no mention of importance sampling, doubly robust estimators, or any adjustment for the fact that the data came from non-bandit policies. Without those corrections, distribution shift between the historical dosing distribution and the regions the bandit would explore remains unaddressed, so the reported gains may not carry over to actual online deployment where exploration occurs. No error bars, statistical tests, or protocol details appear in the abstract either. This paper is mainly for people already working on bandit applications in clinical settings who want to see one more dataset tried. It shows clear engagement with the literature on both the medical side and the bandit side, so it is not incoherent on its own terms. I would send it to peer review so referees can check whether the full manuscript adds the missing evaluation safeguards or at least bounds the offline-to-online gap.","headline":"Applies standard contextual linear bandits to warfarin dosing on PharmGKB data but the offline replay evaluation leaves the transfer to online use unaddressed.","tokens_in":2288,"tokens_out":403,"would_cite":false,"duration_ms":15393,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"Contextual linear bandits for warfarin dosing – orthogonal to RS framework","alignment":"orthogonal","rationale":"Paper applies standard LinUCB / pseudo-reward variants to a 3-arm contextual bandit on PharmGKB patient features for dose classification. No J-cost, φ-ladder, 8-tick periodicity, ratio-symmetric forcing, or any element of the reality_from_one_distinction chain appears; domain is clinical RL with no overlap to RS theorems.","tokens_in":46298,"confidence":"high","tokens_out":115,"duration_ms":27214,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Contextual linear bandits can select initial Warfarin doses that match clinical algorithms on historical patient data.","keywords":["warfarin dosing","contextual linear bandits","online learning","personalized medicine","clinical decision support","pharmgkb data"],"falsifier":"A prospective study randomizing patients to bandit-recommended doses versus standard care and tracking the rate of correct initial dosing without adverse events.","tokens_in":2584,"feed_emoji":"💊","tokens_out":474,"duration_ms":19904,"temperature":0.7,"pith_summary":"The paper applies contextual linear bandit algorithms to predict the correct starting dose of Warfarin for patients. It evaluates these methods on real data from the PharmGKB database by simulating online learning and comparing against fixed-dose and clinical dosing baselines. The algorithms use patient features as context to choose among dose levels and learn from whether the chosen dose was appropriate. This matters because better initial dosing can minimize the risks associated with incorrect anticoagulant levels. Results indicate that the bandit approaches surpass the fixed baseline and some perform on par with the clinical algorithm.","feed_headline":"Bandits match clinical rules for Warfarin dosing","feed_subtitle":"Contextual linear algorithms from patient records beat fixed dosing and equal standard clinical predictions for initial dose.","key_machinery":"Contextual linear bandits that treat patient covariates as context and discrete dose categories as actions, estimating linear reward functions to guide dose selection.","core_discovery":"The authors show that contextual linear bandit algorithms, evaluated through offline replay on the PharmGKB Warfarin dataset, produce initial dose recommendations that yield a higher proportion of patients within the therapeutic range than a fixed-dose strategy, with multiple variants achieving performance comparable to the Warfarin Clinical Dosing Algorithm.","pith_inferences":["Live deployment would need mechanisms to limit exposure to suboptimal doses during learning.","The framework could extend to other drugs requiring individualized dosing.","Historical replay may not fully account for how dosing policies affect the patient population over time.","Integration with electronic health records could enable real-time adaptation."],"forward_implications":["Bandit-based dosing can improve upon fixed prescriptions using only clinical features.","Online updates enable continuous improvement as new patient responses are observed.","The methods achieve clinical-level accuracy without genetic testing.","Different bandit variants offer trade-offs in exploration suitable for medical use."],"fun_headline_variants":["Bandits match clinical Warfarin dosing","Contextual bandits match clinical dose rules","Linear bandits match Warfarin clinical algorithm","Bandits equal clinical Warfarin predictions"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Historical outcomes in the dataset serve as a valid proxy for the results that would occur if the learned policy selected doses for new patients.","fun_headline_variants_meta":{"raw":{"variants":["Bandits match clinical Warfarin dosing","Contextual bandits match clinical dose rules","Linear bandits match Warfarin clinical algorithm","Bandits equal clinical Warfarin predictions"]},"model":"grok-4.3","cost_usd":0.003792,"raw_usage":{"total_tokens":1940,"prompt_tokens":631,"num_sources_used":0,"completion_tokens":49,"cost_in_usd_ticks":37924500,"prompt_tokens_details":{"text_tokens":631,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1260,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":631,"tokens_out":49,"duration_ms":7448,"temperature":1.0,"reasoning_tokens":1260,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-24T22:56:13.106116+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A prospective study randomizing patients to bandit-recommended doses versus standard care and tracking the rate of correct initial dosing without adverse events.","supporting_citations":[],"review_version":1}