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

Whether adaptive ECG channel-selection policies retain an advantage when the downstream evaluator changes matters because channel utility depends on the evaluator. This study compares adaptive policies against evaluator-specific fixed references and tests alternative explanations. All 30 interaction estimates across two policies, five budgets, and three probabilistic metrics are positive, with paired confidence intervals excluding zero.

  • Problem

    Whether adaptive ECG channel-selection advantages persist when the downstream evaluator changes is uncertain because channel utility is evaluator-dependent.

    Source: §I · p. 1

  • Method

    The study compares adaptive-versus-fixed contrasts against each evaluator’s own exhaustive best fixed subset and tests reference, mask-exposure, and evaluator-aligned training alternatives.

    Source: ORCID: 0009-0000-3149-9105 · p. 1, §V · p. 4

  • Results

    All 30 evaluator-specific interaction estimates across five budgets and three probabilistic metrics are positive, with paired intervals excluding zero.

    Source: B. Frozen-policy evaluator replacement · p. 3

  • Takeaways & Limitations

    Adaptive policies should be developed and validated with the downstream diagnostic evaluator intended for deployment.

    Source: §VI · p. 4

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

    Source: §V · p. 4

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.

I. INTRODUCTION

The introduction frames patient-conditioned ECG channel acquisition as an alternative to fixed reduced-lead subsets, while emphasizing that channel value depends on the diagnostic evaluator and may change when the backbone is replaced.

I. INTRODUCTION

  • Reduced-lead systems commonly use one population-wide channel subset, whereas patient-conditioned policies allocate remaining channels based on the observed patient state.Here, k denotes the number of accessed lead channels, not the number of electrodes.Source: §I · p. 1
  • The standard 12-lead ECG provides complementary views of cardiac electrical activity.Reduced-lead channel subsets are used in prior systems –.Source: §I · p. 1
  • Channel utility is evaluator-relative: task-driven sensing defines measurement value by the downstream evaluator, so marginal channel utility need not transfer across evaluators.The introduction highlights the risk of developing an acquisition policy against a lightweight surrogate and later replacing the diagnostic backbone.Source: §I · p. 1
  • The study asks whether a policy that beats the best fixed protocol under its development evaluator retains that advantage after the diagnostic backbone is replaced.The motivating distinction is that marginal channel utility is not evaluator-free.Source: §I · p. 1

ORCID: 0009-0000-3149-9105 … Active feature acquisition and task-driven sensing:

The paper tests whether patient-adaptive ECG channel allocation remains advantageous when the diagnostic evaluator changes. Its contributions include evaluator-dependence analysis and post-hoc tests of reference choice, mask exposure, and evaluator-aligned policy training.

ORCID: 0009-0000-3149-9105

  • Two policies trained with a controlled arbitrary-mask logistic evaluator are scored unchanged with a masked raw-waveform ResNet1D, against metric-matched fixed comparators for each evaluator.The fixed comparators come from exhaustive search.Source: ORCID: 0009-0000-3149-9105 · p. 1
  • Three post-hoc sensitivity analyses test common-reference scoring, mixed-source mask exposure, and Strong-MGA training against strong-evaluator marginal targets.These tests address alternative explanations involving evaluator-specific fixed references, mask distributions, and policy training.Source: ORCID: 0009-0000-3149-9105 · p. 1
  • The study quantifies evaluator dependence of adaptive-versus-fixed contrasts across policies, budgets, and probabilistic metrics.The added analyses use patient-disjoint roles and exclude fold 8 from training and tuning; their intervals are robustness evidence, not prespecified confirmation.Source: ORCID: 0009-0000-3149-9105 · p. 1

Reduced and incomplete-lead ECG:

  • Reduced-ECG work includes fixed configurations, cohort-level selection, and reconstruction from predefined subsets –; lead-agnostic methods improve missing-channel robustness but predict from subsets rather than select them,.Lead switching is described as the closest ECG analogue to sequential observation.Source: Reduced and incomplete-lead ECG · p. 1

Active feature acquisition and task-driven sensing:

  • Cost-sensitive acquisition, generative conditional or information-gain methods, and dynamic feature selection address feature revelation –, while task-driven sensing jointly designs measurements and a decoder,.Here the channel budget is the measurement budget and the acquisition policy is the adaptive measurement operator; the contribution is a controlled test of evaluator replacement.Source: Active feature acquisition and… · pp. 1–2
  • Figure 1 presents fold-10 controlled-evaluator examples of fixed and adaptive channel sets, oracle first-channel choices, and fixed-versus-oracle NLL by budget.Its oracle quantities are descriptive, evaluator- and NLL-specific, label-aware, and non-deployable.Source: Figure 1 · p. 2, Active feature acquisition and… · p. 2
  • The paper distinguishes probability quality from ranking performance and limits its evidence to the studied policies, evaluators, and PTB-XL.Source: Active feature acquisition and… · p. 2

III. METHOD · A. Exact-budget allocation and evaluators

The method selects an exact number of ECG leads and scores each selection with an evaluator that maps the observed ECG and mask to diagnostic probabilities. Because evaluator choice can change lead rankings, the strong evaluator scores the policies’ unchanged acquisition trajectories.

A. Exact-budget allocation and evaluators

  • Each exact-budget policy selects a subset of exactly k leads from the 12-lead ECG record.The evaluator returns superclass probabilities for the selected set and computes loss against the record’s label.Source: A. Exact-budget allocation and… · p. 2
  • The same observed ECG can produce different lead-acquisition rankings under different evaluators.Source: A. Exact-budget allocation and… · p. 2
  • The controlled evaluator uses ten deterministic waveform statistics per lead and a mask-aware one-vs-rest logistic model trained on random exact-cardinality masks.Source: A. Exact-budget allocation and… · p. 2
  • ECG-on-Demand uses acquired-lead features, identities, the mask, evaluator probabilities, and remaining budget to select its next channel until reaching k.Source: A. Exact-budget allocation and… · p. 2
  • MGA predicts one-step reductions in controlled-evaluator NLL, providing a behaviorally distinct comparison policy.Source: A. Exact-budget allocation and… · p. 2
  • The strong evaluator is a 2.77M-parameter masked ResNet1D trained on 12-lead, 10-s waveforms; unavailable standardized channels are zeroed, and a 12-bit mask is provided.Source: A. Exact-budget allocation and… · p. 2
  • During evaluator transfer, the strong evaluator scores the same lead sets chosen by the controlled-evaluator policies rather than regenerating their trajectories.The strong evaluator’s original training used random exact-cardinality masks.Source: A. Exact-budget allocation and… · p. 2

B. Evaluator-specific and common-reference contrasts

The section defines how evaluator-specific fixed references are selected and introduces a post-hoc common-reference sensitivity contrast, which does not replace the original evaluator-specific comparison.

Evaluator-specific and common-reference contrasts

  • The minimizing fixed subset is selected on the selection split and then frozen.The notation indexes the fixed subset by metric and budget.Source: B. Evaluator-specific and… · p. 2
  • Negative D favors adaptive allocation, while positive I means the contrast shifts less favorably after evaluator replacement.Source: B. Evaluator-specific and… · p. 2
  • In a post-hoc sensitivity analysis, both evaluators are scored against the same frozen reference R in {FC, FS}; these estimands do not replace the original comparison.Source: B. Evaluator-specific and… · p. 2

C. Mask-robust evaluator and aligned policy

This section tests robustness to policy-generated mask distributions and evaluator-aligned policy training. The aligned policy is evaluated on held-out fold 8, but it is not a jointly optimized end-to-end system.

C. Mask-robust evaluator and aligned policy

  • The robustness evaluator trains on a fixed mixture of 50% random masks, 25% ECG-on-Demand masks, and 25% MGA masks, with fold 8 excluded from training and model selection.It uses cardinalities k ∈ {1, 2, 3, 4, 6, 8, 12}, trains on folds 1–6, selects the epoch on fold 7, and uses fold 10 only for a boundary check.Source: C. Mask-robust evaluator and aligned… · p. 2
  • Strong-MGA fits one-step marginal NLL reductions from the frozen original strong evaluator and is evaluated on fold 8 after fitting.Its target states come only from training records, and it uses the same MLP form and hyperparameters as controlled-trained MGA.Source: C. Mask-robust evaluator and aligned… · p. 2
  • Strong-MGA tests whether evaluator-aligned policy training can recover transfer loss, but it is not a jointly optimized end-to-end system.The passage frames recovery of transfer loss as the question being tested, not as a reported result.Source: C. Mask-robust evaluator and aligned… · pp. 2–3
  • The study reports mean multilabel NLL, Brier score, and macro label-wise 10-bin ECE, while AUROC and AUPRC are descriptive.Intervals use 1,000 paired patient-cluster bootstrap replicates with identical patient resamples across policies, evaluators, and fixed references.Source: C. Mask-robust evaluator and aligned… · p. 3

IV. EXPERIMENTS · A. Data roles and provenance · B. Frozen-policy evaluator replacement

On a fold-8 exploratory evaluation, frozen adaptive ECG policies that favored adaptation under the controlled evaluator instead favored fixed subsets under the stronger evaluator at k = 4. The experiments use official PTB-XL folds and report positive evaluator interactions across policies, budgets, and metrics.

A. Data roles and provenance

  • PTB-XL v1.0.3 contributes 21,799 ten-second, 100-Hz recordings from 18,869 patients, assigned to official folds 1–6 for training and folds 7, 10, 9, and 8 for validation, fixed subset selection, development audit, and evaluation, respectively.Source: A. Data roles and provenance · p. 3
  • Fold 8 contains 2,173 records from 1,881 patients but is an exploratory, not untouched confirmatory, evaluation because it appeared in historical pre-reset training pools.Reported models were retrained without fold 8; post-hoc additions use it only for final scoring, with budgets k ∈ {2, 3, 4, 6, 8}.Source: A. Data roles and provenance · p. 3

B. Frozen-policy evaluator replacement

  • At k = 4, ECG-on-Demand shifts from DC = −0.011120 under the controlled evaluator to DS = +0.029466 under the strong evaluator.The frozen trajectories therefore change from favoring adaptive acquisition to favoring the metric-matched fixed subset.Source: B. Frozen-policy evaluator replacement · p. 3
  • On the fold-9 development audit, ECG-on-Demand and MGA each beat controlled-evaluator metric-matched fixed subsets in all 15 budget–metric comparisons.Source: B. Frozen-policy evaluator replacement · p. 3
  • At k = 4, the Brier interaction is +0.013960 and the MGA NLL interaction is +0.047150.Source: B. Frozen-policy evaluator replacement · p. 3
  • Across two policies, five budgets, and three probabilistic metrics, all 30 evaluator-specific interaction estimates are positive with paired intervals excluding zero.The contrasts share data and are not 30 independent tests.Source: B. Frozen-policy evaluator replacement · p. 3

C. Common-reference sensitivity · D. Mask-distribution robustness · E. Evaluator-aligned Strong-MGA

Common-reference scoring and mask-distribution robustness preserve positive evaluator interactions, while evaluator-aligned Strong-MGA training improves performance but does not eliminate the gap to the strong fixed baseline.

C. Common-reference sensitivity

  • +0.020760 (95% CI [+0.010131, +0.031402]) with both evaluators using FC and +0.048098 (95% CI [+0.037788, +0.057694]) using FS.The evaluator-specific interaction is +0.040585; all three intervals exclude zero.Source: C. Common-reference sensitivity · p. 3

D. Mask-distribution robustness

  • 0.2584 NLL, 0.0775 Brier, 0.0263 ECE, 0.9345 AUROC, and 0.8425 AUPRC for the mask-robust evaluator compare with 0.5031, 0.1587, 0.1553, 0.8355, and 0.6283 for the controlled evaluator.These full-12 fold-7 scores show the mask-robust evaluator is more predictive on each reported metric.Source: D. Mask-distribution robustness · p. 3
  • +0.037649 NLL (95% CI [+0.026942, +0.047100]) is the ECG-on-Demand interaction on fold 8 at k = 4; Brier and ECE interactions are +0.012817 and +0.018932.All 30 robust interaction intervals exclude zero; policy-generated plus random-mask exposure does not remove the shift, though it modestly changes its magnitude.Source: D. Mask-distribution robustness · p. 3

E. Evaluator-aligned Strong-MGA

  • −0.016777 (95% CI [−0.025963, −0.008355]) is the paired k = 4 strong-evaluator NLL improvement from Strong-MGA over controlled-trained MGA.Strong-MGA changes NLL from 0.2952 to 0.2784.Source: E. Evaluator-aligned Strong-MGA · p. 3
  • +0.012362 ([+0.005769, +0.019204]) remains Strong-MGA’s DS gap relative to the exhaustive strong fixed baseline at k = 4.The aligned policy’s Brier gap is reduced but remains positive; ECE shows no clear pairwise change, so aligned training provides partial recovery rather than eliminating the gap.Source: E. Evaluator-aligned Strong-MGA · p. 3

A Full-12 evaluator quality B Frozen ECG-on-Demand transfer C NLL evaluator interaction · V. DISCUSSION

Adaptive-versus-fixed performance depends on the diagnostic evaluator, and robustness analyses leave a positive interaction. The discussion recommends developing and validating acquisition policies with their intended evaluator while noting limits to the evidence.

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

  • Figure 3 presents full-12 diagnostic quality, frozen ECG-on-Demand contrasts against evaluator-specific fixed references, and NLL interactions for ECG-on-Demand and MGA with paired patient-cluster 95% CIs.The contrasts and interactions are reported for fold 8; the full-12 quality audit is on fold-9 development data.Source: Figure 3 · p. 4
  • +0.037649 NLL was the interaction under the mask-robust evaluator, and all 30 intervals remained above zero, weakening reference-only and mask-exposure explanations.The common-reference analyses also preserved a positive interaction interval.Source: Figure 4 · p. 4, §V · p. 4

V. DISCUSSION

  • Every interaction subtracts each evaluator’s own exhaustive best fixed subset, cancelling additive evaluator-level bias; adaptive outputs diverged from the strong evaluator’s NLL-optimal leads, with mean fold-8 Jaccard 0.300.The strong-evaluator NLL-optimal set was {III, aVR, V2, V5}; adaptive outputs concentrated on leads favored by the controlled logistic evaluator.Source: §V · p. 4
  • Strong-MGA reduced the gap by −0.016777 NLL but did not close it, with DS = +0.012362.The residual may reflect myopic targets, finite MLP capacity, trajectory shift, or a one-step versus multi-step mismatch.Source: §V · p. 4
  • Strong-evaluator ECE interactions were negative at k ∈ {2, 3, 4}, while NLL and Brier interactions were positive at all budgets; ECE intervals still excluded zero.The calibration surface was less uniformly affected than the proper scoring rules.Source: §V · p. 4
  • The practical implication is to develop and validate adaptive acquisition policies with the downstream diagnostic evaluator intended for deployment.The study is retrospective and restricted to PTB-XL; k counts channels rather than electrode actions, and fold-8 findings are exploratory rather than pristine confirmatory evidence.Source: §V · p. 4

VI. CONCLUSION · COMPLIANCE WITH ETHICAL STANDARDS

Adaptive ECG channel allocation is not evaluator-invariant, supporting development and validation with the intended diagnostic evaluator; the study used public, de-identified PTB-XL data without intervention, enrollment, or new protected health information.

VI. CONCLUSION

  • Patient-conditioned ECG lead-channel allocation can outperform fixed protocols under its development evaluator, but that advantage is not evaluator-invariant.Source: §VI · p. 4
  • Channel marginal utilities learned against one inference model do not transfer cleanly when models weight the ECG representation differently.Source: §VI · p. 4
  • Common-reference and mask-mixture checks support the evaluator-dependent shift in adaptive allocation performance.Source: §VI · p. 4
  • Evaluator-aligned policy training gives only partial recovery of the allocation gap.Source: §VI · p. 4
  • Policies should be developed and validated with their downstream diagnostic evaluator, while jointly optimised sensing–diagnosis systems remain necessary future work.Source: §VI · p. 4

COMPLIANCE WITH ETHICAL STANDARDS

  • The study used only public, de-identified PTB-XL data under its released terms, with no intervention, enrollment, or collection of new protected health information.Source: COMPLIANCE WITH ETHICAL STANDARDS · p. 5

Related papers

Loading 2608.26827v1…