PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluator-Dependent Patient-Adaptive ECG Lead-Channel Allocation

Xiaoyang Li, Zeyan Tao

arXiv:2608.26827v1cs.CV

TL;DR

The paper asks whether patient-adaptive ECG channel allocation remains superior to fixed protocols when the downstream diagnostic evaluator changes. It freezes policies trained with a controlled evaluator, scores their unchanged trajectories with a stronger evaluator, and finds a positive evaluator-dependent interaction across policies, budgets, and probabilistic metrics. The practical conclusion is to develop and validate acquisition policies with their intended diagnostic backbone.

  • Problem

    Channel marginal utilities are defined relative to a downstream evaluator, so utilities learned under one evaluator may not transfer when the diagnostic backbone is replaced.

  • Method

    The study freezes ECG-on-Demand and MGA trained with a controlled arbitrary-mask logistic evaluator, then scores their unchanged trajectories with a masked raw-waveform ResNet1D and evaluator-specific exhaustive fixed comparators.

  • Results

    All 30 interaction estimates across two policies, five budgets, and three probabilistic metrics are positive with paired confidence intervals excluding zero.

  • Takeaways & Limitations

    Adaptive ECG channel allocation should be developed and validated with its downstream diagnostic evaluator, while jointly optimized sensing-diagnosis systems remain future work.

  • Takeaways & Limitations

    The study is retrospective and restricted to PTB-XL; k counts channels rather than electrode actions, and fold-8 results are exploratory.

Abstract

from arXiv · show

Patient-conditioned acquisition policies for ECG lead-channel selection can outperform population-wide fixed protocols by tailoring the channel budget to each patient's observed cardiac state. However, the value of acquiring any given channel is defined relative to a downstream diagnostic evaluator, so marginal utilities learned under one evaluator need not transfer when the evaluator is replaced. We study this evaluator dependence empirically on PTB-XL by freezing two policies (ECG-on-Demand and MGA) trained with a controlled arbitrary-mask logistic evaluator, then scoring their unchanged acquisition trajectories with a more predictive masked raw-waveform ResNet1D. Exhaustive search provides metric-matched population-wide fixed comparators separately for each evaluator, enabling a clean interaction contrast. At budget $k=4$ on a held-out evaluation fold, ECG-on-Demand shifts from $D_\mathrm{C}=-0.011$ (favoring adaptive under the controlled evaluator) to $D_\mathrm{S}=+0.029$ (favoring fixed under the strong evaluator), yielding an NLL interaction of $+0.041$ (95% CI $[+0.030, +0.050]$). Across two policies, five budgets, and three probabilistic metrics, all 30 interaction estimates are positive with paired confidence intervals excluding zero. Three post-hoc sensitivity analyses -- common-reference scoring, training the strong evaluator on a mixture of policy-generated and random masks, and evaluator-aligned Strong-MGA policy training -- each preserve a positive interaction interval, making reference-choice and mask-distribution artifacts less plausible explanations. Evaluator-aligned training reduces but does not eliminate the gap. These results indicate that adaptive ECG channel allocation should be developed and validated jointly with its intended diagnostic backbone, and that jointly optimized sensing-diagnosis systems remain an open problem.

ORCID: 0009-0002-4863-5761

The paper frames patient-conditioned ECG channel allocation as a response to patient-level variation in channel utility, while emphasizing that utility depends on the downstream evaluator. It asks whether adaptive policies retain their advantage when that evaluator is replaced.

  • Motivation: Patient-conditioned policies allocate remaining ECG channels from a partial recording according to the current patient state.The channel budget k counts accessed lead channels, not electrodes.
  • Evaluator dependence: Task-driven sensing defines channel value through a downstream diagnostic objective rather than as an evaluator-free property.The paper formalizes this value as retrospective marginal loss reduction for evaluator f.
  • Evaluator dependence: In general, marginal utilities differ across evaluators, so policies developed with a lightweight surrogate may not transfer cleanly to a replacement diagnostic backbone.The paper poses this as the central evaluator-replacement question.

ORCID: 0009-0000-3149-9105

The study freezes policies trained with a controlled evaluator, evaluates their unchanged trajectories with a stronger evaluator, and adds post-hoc analyses targeting reference, mask-distribution, and policy-training explanations.

  • Study design: Frozen ECG-on-Demand and MGA trajectories are scored with a more predictive masked raw-waveform ResNet1D after training under a controlled arbitrary-mask logistic evaluator.Exhaustive search supplies evaluator-specific fixed comparators.
  • Study design: The paper quantifies evaluator dependence across policies, budgets, and probabilistic metrics using patient-disjoint roles without using fold 8 for training or tuning.The added analyses are interpreted as robustness evidence because they were motivated after the original analysis.
  • Study design: The transfer experiment replaces the evaluator while retaining the acquisition policy, distinguishing selection from prediction robustness.Related work identifies raw-waveform convolutional networks as high-capacity diagnostic models for 12-lead ECGs.

Active feature acquisition and task-driven sensing:

The paper situates adaptive ECG acquisition within task-driven sensing: policies select channels for downstream prediction, and evaluator replacement can alter those selections because models weight ECG representations differently.

  • Active feature acquisition and task-driven sensing:: Task-driven sensing treats the channel budget as a measurement budget and the acquisition policy as an adaptive measurement operator.The sensor is designed around a downstream inference objective, although this paper tests evaluator replacement rather than joint design.
  • Active feature acquisition and task-driven sensing:: An exact-budget policy selects a k-channel subset for each ECG record, while the evaluator maps the selected recording to superclass probabilities and a loss.The available leads are the standard 12-lead set.
  • Active feature acquisition and task-driven sensing:: Different evaluators can induce different acquisition rankings for the same observed ECG.This follows because channel marginal utility is defined relative to the evaluator’s loss.
  • Active feature acquisition and task-driven sensing:: ECG-on-Demand scores channels using acquired features, identities, masks, evaluator probabilities, and remaining budget, whereas MGA predicts one-step controlled-evaluator NLL reduction.Cohort-Conditional Greedy Acquisition serves only as a mechanism control.
  • Active feature acquisition and task-driven sensing:: The strong evaluator is a 2.77M-parameter masked ResNet1D using zeroed unavailable standardized channels and a 12-bit mask.During transfer, it scores the same policy-selected subsets rather than regenerated trajectories.

B. Evaluator-specific and common-reference contrasts

Evaluator-specific contrasts compare each adaptive trajectory with the best fixed subset for that evaluator, while common-reference analyses test whether the interaction depends on reference choice.

  • B. Evaluator-specific and common-reference contrasts: The minimizing fixed subset is selected on the selection split and then frozen before scoring the adaptive allocation.The fixed comparator is evaluator-specific and budget-specific.
  • B. Evaluator-specific and common-reference contrasts: Negative D favors adaptive allocation, whereas positive I indicates that evaluator replacement shifts the contrast less favorably.The interaction is defined relative to the evaluator-specific adaptive-versus-fixed comparisons.
  • B. Evaluator-specific and common-reference contrasts: Common-reference contrasts score both evaluators against the same frozen reference R drawn from the controlled or strong fixed subsets.These are post-hoc sensitivity estimands and do not replace the original evaluator-specific comparison.

C. Mask-robust evaluator and aligned policy

Mask-mixture training preserves the evaluator-dependent interaction, while evaluator-aligned Strong-MGA training provides only partial recovery and is not jointly optimized end to end.

  • Mask-robust evaluator: The mask-robust evaluator mixes random, ECG-on-Demand, and MGA masks while reserving fold 8 for final evaluation.The mixture uses 50% random exact-cardinality masks, 25% frozen ECG-on-Demand masks, and 25% frozen MGA masks.

A. Data roles and provenance

The study uses fold-specific patient-disjoint roles and treats fold-8 results as a held-out exploratory evaluation because fold 8 appeared in historical pre-reset training pools.

  • Data roles: PTB-XL v1.0.3 contains 21,799 recordings from 18,869 patients, and fold 8 contains 2,173 records from 1,881 patients.The recordings are ten-second, 100-Hz ECGs with five diagnostic superclasses.
  • Data roles: Training uses folds 1–6, validation uses fold 7, fixed subset selection uses fold 10, development auditing uses fold 9, and evaluation uses fold 8.Post-hoc additions use fold 8 only for final scoring.
  • Development audit: On the fold-9 development audit, ECG-on-Demand and MGA each beat their controlled-evaluator metric-matched fixed subsets in all 15 budget–metric comparisons.This is the development-fold result before the held-out exploratory evaluation.
  • Frozen-policy evaluation: At k = 4 on fold 8, ECG-on-Demand shifts from DC = −0.011120 under the controlled evaluator to DS = +0.029466 under the strong evaluator.The controlled-evaluator interval is [−0.018652, −0.002885], while the strong-evaluator interval is [+0.022848, +0.036799].
  • Interaction results: All 30 evaluator-specific interaction estimates are positive with paired intervals excluding zero across two policies, five budgets, and three probabilistic metrics.These contrasts share data and are not 30 independent tests.

C. Common-reference sensitivity

Common-reference scoring preserves a positive evaluator-dependent interaction, indicating that the primary result is not explained solely by changing the fixed reference.

  • Common-reference scoring: At k = 4 for ECG-on-Demand, the common-controlled-reference interaction is +0.020760 (95% CI [+0.010131, +0.031402]).Both evaluators use the controlled evaluator’s fixed reference in this contrast.
  • Common-reference scoring: The common-strong-reference interaction is +0.048098 ([+0.037788, +0.057694]) at the same budget.Both evaluators use the strong evaluator’s fixed reference in this contrast.
  • Common-reference scoring: The evaluator-specific interaction is +0.040585, and all three reference choices produce intervals excluding zero.The complete policy–budget–metric table is supplied in the Supplementary Material.
  • Interpretation: Common-reference results support an evaluator-dependence effect because each contrast subtracts the evaluator’s own exhaustive best fixed subset.This construction cancels additive evaluator-level bias in the interaction.

A Full-12 evaluator quality B Frozen ECG-on-Demand transfer C NLL evaluator interaction

Evaluator replacement makes adaptive ECG allocation less favorable relative to fixed references, and this interaction persists across robustness analyses, although Strong-MGA only partly reduces the gap.

  • Evaluator dependence and robustness: Positive interactions persist under common-reference scoring and a mask-mixture-trained strong evaluator, making reference-only and mask-exposure explanations less plausible.The robustness analysis reports an NLL of 0.037649, with all 30 intervals above zero.
  • Evaluator-aligned policy training: Strong-MGA reduces the gap by −0.016777 NLL but leaves a strong-evaluator contrast of DS = +0.012362.The residual may reflect myopic targets, finite MLP capacity, trajectory shift, or one-step versus multi-step mismatch.
  • Metric sensitivity: Strong-evaluator ECE interactions are negative at k ∈ {2, 3, 4}, whereas NLL and Brier interactions are positive at all budgets.ECE interactions still exclude zero, indicating that calibration is less uniformly affected than the proper scoring rules.
  • Practical implication and scope: Adaptive policies should be developed and validated with the downstream diagnostic evaluator intended for deployment.The study is retrospective, restricted to PTB-XL, and based on exploratory fold-8 evidence rather than a pristine confirmatory test.

VI. CONCLUSION

Patient-conditioned ECG allocation can outperform fixed protocols under its development evaluator, but that advantage is not evaluator-invariant. The findings support evaluator-specific development and validation, while jointly optimized sensing-diagnosis systems remain future work.

  • VI. CONCLUSION: Adaptive ECG allocation can outperform fixed protocols under its development evaluator, but the advantage is not evaluator-invariant.The evidence is specific to the studied policies, evaluators, and PTB-XL.
  • VI. CONCLUSION: Channel marginal utilities learned against one inference model do not transfer cleanly when diagnostic evaluators weight ECG representations differently.Common-reference and mask-mixture checks support this shift, while evaluator-aligned training provides only partial recovery.
  • VI. CONCLUSION: Policies should be developed and validated with the downstream diagnostic evaluator used at deployment.Jointly optimized sensing-diagnosis systems remain necessary future work.
Loading 2608.26827v1…