acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Sensing versus Every Fixed Schedule: How Much Is Adaptivity Worth?

Abstract

Hospital staff cannot watch a patient continuously, so they measure on a schedule fixed in advance — before meals, or every four to six hours. Any fixed schedule misses events between checks, and the obvious fix is to decide each measurement from what is known at the time. But the two options are not equally tractable. A fixed schedule is a handful of clock times: a six-hour night at fifteen-minute resolution has twenty-four slots, so at most eight measurements gives schedules and the best is found by checking every one. A policy that decides as it goes must act at every state it can reach — on our grid — giving policies. One side can be settled exactly and the other cannot — which is why a reported gain over a fixed schedule can always be blamed on the chosen baseline. We remove that objection: a constrained partially observed Markov decision process makes the adaptive side tractable, and enumeration makes the fixed side an exact empirical optimum within its class. What we measure. On 112 hospitalized patients whose blinded sensor recorded every slot, we compare a solved adaptive policy against the best of all fixed schedules at six pre-specified measurement rates. The adaptive policy detects more nocturnal hypoglycemia at every rate, by percentage points on average. What we do not claim is its size: a bootstrap that reruns selection gives , positive in 96 percent of replicates but wide, because held-out excursions carry the information of some independent events. The advantage also mixes when to measure within a night with spending more on some nights than others. What generalizes beyond this dataset. An evaluation protocol for selected policies: selection is rerun inside every bootstrap replicate and splits are taken on unique identifiers, the usual alternative silently mixing about one unit in six however large the sample. A comparator that scales: detection coverage is monotone submodular, so a greedy schedule attains of the exhaustive optimum wherever enumeration is infeasible, and to percent of it here. Conditions under which adaptivity is worth nothing: provably zero when a measurement carries no reward-relevant information, at most when that holds to within . Zero is the right thing to characterize, because correlated sensing admits no universal constant bound on what adaptivity is worth.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
adaptive sensing, POMDP, submodularity, benchmark, uncertainty quantification, clinical monitoring

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_GoVKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors compare adaptive glucose measurement with an optimized fixed schedule using blinded CGM recordings. Adaptive measurement improves event detection by an estimated 14.2%.

Critical strengths

The blinded recordings allow both schedules to be evaluated against the same observed trajectories. Repeating model fitting and schedule selection within the patient-level bootstrap provides a useful assessment of uncertainty.

Critical weaknesses
  1. The computational argument overlooks the interval structure of the objective. An exact dynamic program solves fixed-schedule coverage in O(k * T^2) after preprocessing (T is the num. slots); this changes the author's claimed limit on tractable comparisons.

  2. The reported gain combines changes in measurement timing with reallocation across nights. Appendix H.16 uses a different fitting protocol... this leaves the contribution of timing unclear.

Questions for authors

Which feature of the coverage objective specifically prevents the interval dynamic program? Please compare fixed and adaptive schedules with the same training data and per-night budget, so the analysis isolates the value of adaptive timing, specifically.

Reviewer_XQtXOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks how much an adaptive measurement policy gains over the best fixed schedule for detecting nocturnal hypoglycemia in hospitalized patients. Using a public cohort of 112 inpatients with blinded CGM, it replays candidate schedules against the recorded glucose trace. A constrained POMDP policy, solved on a reduced Gaussian belief, is compared with the best fixed schedule found by enumerating all schedules over a 24-slot night. Across alarm rates of 1.5 to 4.0 per night, the paper reports an integrated frontier gap of +14.2 points, with a bootstrap interval of [−2.0, +39.8]. It also proposes an evaluation protocol that reruns selection inside every bootstrap replicate and splits on unique patient IDs. The theory includes a construction showing that local belief diagnostics cannot certify schedule value, submodularity of coverage with a greedy guarantee, and conditions under which the value of adaptivity is zero.

Critical strengths

The blinded-sensor cohort makes it possible to score any schedule against the recorded glucose trace, and enumerating the fixed-schedule class removes the usual objection that the baseline was hand-picked.

The evaluation is transparent about uncertainty. Selection is rerun inside every bootstrap replicate, splits are taken on unique IDs, and sensitivity to model misspecification is reported rather than hidden.

Critical weaknesses

The main claim is not supported by the evidence. The adaptive arm is constrained only in expectation and spends 0 to 24 measurements per night, while the fixed arm spends exactly k. The paper’s own pathwise-budget results (H.16: 55.8 / 58.3 / 62.5% at k = 2, 3, 4) roughly match the exhaustive fixed schedule in Figure 3, so the reported +14.2 gap likely comes from reallocating budget across nights rather than from adaptive timing. The estimate is also unresolved: the interval [−2.0, +39.8] covers zero, rests on about 18 effective events, and moves from +3.8 to +23.0 under different modeling choices. The “exact optimum” baseline is weaker than claimed, too. Only k ≤ 4 (12,950 schedules) is used, and on held-out data it beats a non-selecting lattice by only about 3 points.

The theory is limited and internally inconsistent. Most propositions are standard results (weighted coverage submodularity, zero gap under uninformative observations, bootstrap overlap). Several statements contradict each other: γ = 1 versus “the problem is discounted” (Sec. 2 vs. F.3); two incompatible versions of Proposition 1 (Sec. 3 vs. G.5); and Proposition 9’s proof derives a shared transition kernel from Assumption 1, which does not imply it. The main text also depends on propositions stated only in a 51-page appendix, and the numbers and references disagree across sections.

Questions for authors

Under a per-night budget, with both arms fitted and selected on the same training nights, does the adaptive policy still beat the exhaustive fixed schedule?

How do you reconcile γ = 1 with the discounting argument in F.3, and the two versions of Proposition 1?

How does Assumption 1 give the shared transition kernel used in Proposition 9’s proof?

Why does the abstract claim comparison against all 1,271,625 schedules when the frontier only uses k ≤ 4?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.