acceptodds
Under review as a conference paper at ICLR 2027

Label-Efficient Risk Certification for Medical Question Answering and Triage via Predictive Calibration Ordering

Abstract

Approving medical AI policies for repeated use can require a high-confidence population-error certificate, even when expectation-based methods offer higher yield. With scarce calibration labels, however, these strict certificates often force low release rates. We introduce Predictive Calibration Ordering (PCO), which forecasts dependent certification outcomes from development data and uses exact dynamic programming to optimize a first-failure testing sequence, preserving the finite-sample selected-risk guarantee of Learn then Test. We evaluate PCO across medical question answering (HealthAdvice and 1,000 MedQA questions) and current-action triage (CARE-Bench). Across 80 target-averaged settings, Exact PCO achieves higher useful yield than Holm-certified adaptations of SCoRE, LEC and SCRC. At 20 labels and 90% certification confidence, MedQA useful yield reaches 51.52% and 43.12% for two models, versus 18.68% and 12.20% for SCoRE+Holm. Matched controls isolate joint planning, with small-budget gains across all three benchmarks of up to 2.4 percentage points over marginal ordering and up to 7.8 over Learn then Test. By increasing certified yield from scarce labels, PCO could help clinical AI systems release more answers to patient questions and triage recommendations without individual review.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
selective prediction, risk control, uncertainty quantification, calibration, multiple hypothesis testing, dynamic programming, medical question answering, medical triage

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xLFVOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

I found this paper extremely difficult to parse and understand, which may be due to my lack of prior knowledge in this area, or may be due to the writing quality, or both. The following is what I understood the problem setting for this paper to be:

  • You have a medical AI assistant that produces some output or decision f(X), in response to some question/prompt/scenario X. It also produces a risk score s(X), which is similar but not indentical to its confidence in its answer. This process may be done in 2 stages, by first having a model or multiple models produce candidate answers f(X)_1, ...., f(X)_n, and then having another model pick one and output a risk score.
  • We are told that deploying this assistant requires a threshold t that says "if s(X) > t" then do not release this assistant's response, route it to a human.
  • Where to set this threshold depends on the error rate alpha that you are willing to accept. The deployment should contain a mapping that says for a given alpha, here's where you set t.
  • This mapping should be certified, which is to say that if you follow its given t for a choice of alpha, with some probability delta the actual error rate should be less than alpha.
  • Ultimately, this means there is a policy T(alpha, delta) -> t, which is a mapping from a risk tolerance alpha and certification confidence delta, to a threshold t. This says that if you chose your threshold to be t, then with probability > 1 - delta the error rate of the system will be less than alpha.
  • We can rollout policy T on real data and systems, and evaluate the true error rate, and how often it witholds an assistants correct answer and so on.
  • This paper is about how to design the policy T and produce statistically sound certificates (proofs) that it has the desired property of "if you chose your threshold to be t, then with probability > 1 - delta the error rate of the system will be less than alpha. "
Critical strengths
  • The paper tackles a worthwhile aim, the safe deployment of medical AI
  • The paper evaluates over 3 different datasets and multiple cross-validation folds within datasets.
  • The authors compare to a few existing methods, but hard to assertain to what extent they are fair comparisons
  • For their chosen experimental setup, their method vastly outperforms existing methods
Critical weaknesses
  • The paper is extremely hard to follow
  • The problem setting is very complicated, and the paper does not give good evidence that it is a real setup (ie laws require medical AI to have a risk score classifier and a certificate etc) or a contrived one
  • Further analysis and weaknesses require a greater understanding of the paper and its method than i am able to give
Questions for authors

I believe this paper would benefit massively from well structured and seperated sections that tackle

  1. problem setup (a model produces answers, a classifier gives risk score, a rule decides when to release etc etc)
  2. background (what prior work is needed to understand this paper, motivate and validate the problem statement etc)
  3. related work (existing work, if it fits into your problem setup then should be described and compared in tables, if doesn't fit in problem setup, why not and why that doesn't make your setup wrong)

If the authors are able to restructure and improve the clarity of their writing, then I am happy to review the method etc in more depth and revise my scores.

Reviewer_jx5EOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper describe Predictive Calibration Ordering, a technique developed to certify a release rule for medical question answering and triage. PCO simulates how different decision thresholds would pass a fixed testing sequence and then uses dynamic programming to select the threshold that maximizes expected useful release.

Evaluations include HealthAdvice, CARE-Bench and 1,000 MedQA tasks, with Holm-certified adaptations of SCoRE, LEC and SCRC as baselines.

Critical strengths

The paper addresses a practical setting, which is defining a decision/release rule for repeated use of medical AI models label information that can be used for calibration is scarce.

Critical weaknesses

The paper often uses coined/invented terms instead of standard ML terminology. Examples include “clinical autonomy–risk frontier” for the risk–coverage curve, “predictive worlds” for Monte Carlo samples, and “useful yield” for the fraction of cases released and answered correctly. Others are “passage indicators”, “envelope”, “support”, “tie loss” and “pseudo-observations”. We need to translate each term to established terminology in selective prediction and multiple testing. This makes the method hard to follow and the paper hard to compare with prior work.

Section 6 combines related work with deployment rationale and limitations, so the paper lacks a discussion of related work and then relevant benchmarking. Related work section needs to state how PCO is different from existing methods, and from RCPS and conformal risk control, in objective, guarantee and use of the development data.

Questions for authors

Please address critical weaknesses raised in the field above.

Reviewer_N4MdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a label-efficient risk certification framework for medical question answering and triage named Predictive Calibration Ordering (PCO). Specifically, PCO consists of three core components: the joint certification outcome forecasting, the exact testing-sequence optimization, and the implementation-aware utility certification. The joint certification outcome forecasting framework predicts dependent certification outcomes and useful release mass. The exact testing-sequence optimization maximizes predicted useful yield under first-failure stopping via dynamic programming. Finally, the implementation-aware utility certification framework bounds operational utility regret on predictive simulations. The experimental evaluation in this paper assesses the effectiveness of PCO across three benchmark datasets. Results demonstrate that PCO achieves higher certified useful yield while preserving finite-sample selected-risk guarantees, outperforming several Holm-certified selective prediction baselines.

Critical strengths
  • The idea of certifying an automatic release rule for long-term use, allowing more correct answers or triage actions to be released without individual human review while controlling error risk, is considered meaningful.
  • Modeling certification efficiency with limited labeled data as a sequence planning problem that accounts for dependencies among tests is novel and effective.
Critical weaknesses
  • In my understanding, a key motivation of PCO is that knowing the passing probability of each threshold individually is insufficient, and the joint passage of an entire sequence of tests needs to be considered. Therefore, Eqn(5) involves the probability that all tests in a prefix pass, the probability that subsequent tests pass conditional on earlier tests passing, and the relationship between passage events and corresponding utility, among other factors. However, can a simulator constructed from limited development data reliably estimate the joint dependencies that can change the optimal testing sequence?
  • The paper uses 2,048 training worlds for planning and then independent check worlds for validation. This design can mitigate overfitting to a particular set of Monte Carlo samples, but both sets of worlds come from the same predictive distribution. I agree that sharing the same simulated world does preserve the dependencies induced by a given predictive distribution. However, this is not entirely equivalent to those dependencies accurately reflecting the actual calibration process. I am also convinced that increasing the number of worlds allows the utility under the predictive distribution to be estimated more precisely, and independent checks can assess whether a plan is suited only to the particular simulated samples used during planning. However, neither can automatically eliminate systematic discrepancies between the predictive distribution and the real process. In fact, the independent forecast checks in Appendix E already reflect this phenomenon, it seems the independent simulation checks still favor Exact, whereas the observed utility differences have the opposite sign.
  • In dynamic programming, the optimization objective is the utility of the last passing threshold, whereas the actual deployment selector chooses the passing threshold with the largest calibration support and selects the smaller threshold when support is tied. Since the largest certified threshold satisfies the same risk certification requirements and is consistent with the objective, why does the actual controller still select the smaller threshold when calibration support is tied?
  • The clinical objective is reduced to an average binary error rate, which does not fully capture the value of actual decisions. The current objective mainly encourages more correct actions to be released automatically, while a correct current action is not equivalent to completing a useful triage episode.
Questions for authors
  • Is the use of local fractional losses in the predictor reasonable? PCO relies on discrete, nonlinear joint passage events with small samples. An approximation to the mean does not seem to imply a reasonable approximation to the passing probabilities. I have read Appendix F.1, but the existing results still do not establish a clear rule for determining which predictive construction can more accurately reflect actual joint certification events.
  • Is exact global dynamic programming really necessary? Exact has exponential worst-case complexity. I understand that its overhead on the current small grids is low and does not constitute a deployment bottleneck, but the cost of exact optimization will increase as candidate thresholds become denser and the policy space becomes richer. If a simple joint greedy method can already achieve comparable utility, would the practical contribution of exact DP’s theoretical optimality then become limited?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.