- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper argues that offline-RL treatment recommendation folds heterogeneous evidence (routine practice, guidelines, learned policies) into one scalar objective and loses each source's provenance, strength, coverage, and ability to stay silent. PRUDENT anchors each decision at routine practice (a behavior-cloned policy) and represents other sources (rules compiled from SSC 2021 with their GRADE strength, and an IPW-IQL policy) as distributions over the same action space with per-dimension silence. It diagnoses the decision context by atypicality a(s) (a rank-calibrated Mahalanobis distance) and disagreement b(s) among the sources that speak, moves from the anchor toward each source along a conditional-flow-matching route by a distance a(s)κ_j (or, when a value signal is available, by the value-maximizing distance capped at a(s)κ_j), combines the endpoints with grade-dependent softmax weights, and projects onto hard constraints. On MIMIC-IV, the a-b plane isolates a "consensus extrapolation" region (22% of windows) whose 90-day mortality is 1.76 times that of the routine region. In a PK/PD simulator with five regimes, the value-guided gate has lower regret than the best fixed transport distance in every regime.
Critical strengths
S1. The position is thoughtful and relevant. Keeping evidence separate by source, strength, coverage, and silence matches how evidence-based medicine and GRADE are meant to work, and the requirement that a silent source must change neither the decision nor the diagnosed context is a clean formal statement.
S2. The real-data methodology is careful: the paper declines to report off-policy value on MIMIC-IV and explains why (effective sample size 0.27), uses stay-clustered and dyadic bootstraps, freezes reference samples, and documents the guideline compilation, including a human gold standard frozen before LLM compilation, inter-annotator agreement, and the two recorded departures from a mechanical reading of the guideline.
S3. The simulator includes regimes where the premise fails (F and G), and the paper reports oracle comparators that beat the method. This is honest.
S4. The consensus-extrapolation observation is an interesting hypothesis-generating finding, and it is checked across partition thresholds and reference seeds.
Critical weaknesses
W1. The headline simulation result depends on an oracle value signal. The gate beats the best fixed transport distance (chosen with hindsight per regime) only in its value-guided form, which uses the true three-step return of the known dynamics; Appendix E.5 says the claim is therefore conditional on a usable value signal, and that this branch is not used on MIMIC-IV.
W2. Several validations hold by construction. Silence invariance and exact reconstruction (Section 5.6, residual 2e-16) follow from masking and the deterministic arbitration formula, so they verify the implementation rather than a scientific claim. The sign of the transport-versus-mixture result (Section 5.5) follows from Jensen's inequality for a concave reward with two sources placed on opposite sides of the optimum, as Appendix E.1 acknowledges, so only its size is informative. The paper also concedes that the learned routes are nearly straight (Section 4.3, Appendix A.2) and keeps conditional flow matching for multimodal or constrained sources, but none of the sources tested is multimodal or constrained, so in the reported experiments it adds little over linear interpolation between paired draws. Please either show a case that needs a learned velocity field or simplify this component.
W3. The routine-practice anchor is mis-specified for the decision that matters most. 73% of windows have no vasopressor and 19% have no fluid (Appendix B.5), yet the anchor is a homoscedastic Gaussian with a softplus mean (Appendix D.1). Appendix D.2 concedes that a zero-inflated anchor would represent the give-or-withhold decision explicitly, and defends the choice because the anchor is never inverted. But every route starts from a draw of this anchor, so the give-versus-withhold decision is poorly represented in the actions PRUDENT outputs. Please show the effect of this approximation, or use the zero-inflated model already built for the importance weights as the anchor.
Questions for authors
Q1. In the simulator, what is PRUDENT's regret when the value-guided branch uses an estimated value function? Q2. Does the consensus-extrapolation region persist after excluding windows that follow a DNR or comfort-measures order? Q3. How sensitive are the decisions to the pre-specified κ_j values? Q4. How would PRUDENT behave with a second learned policy (for example one trained with a different reward), so that two sources speak on vasopressors?