PaperlayerView exact revision on arXiv

Source-linked AI summary

Correspondence Is Not Transfer: A Pairing-Controlled Study of Clinical Audio-Metadata Alignment

Liu He, Xi Wang, Jiahong Yuan, Yuanchao Li

arXiv:2610.35747v2cs.SDeess.AS

TL;DR

Clinical audio–metadata alignment can reward profile-identifying signals that are not disease signals, so its disease-prediction value is uncertain. The paper varies only metadata pairing while keeping audio backbones fixed, then evaluates retrieval, attribute probing, and matched disease prediction. Correct pairing improves correspondence and sex decodability, but disease AUROC differences show no consistent pattern, motivating separate reporting of these outcomes.

  • Problem

    Participant profiles contain disease information alongside symptoms, demographics, missingness, and recruitment patterns, so it is unclear whether alignment learns disease-relevant information or other profile-identifying signals.

    Source: §1 · p. 1

  • Method

    The study varies only audio–metadata pairing across correct and shuffled controls, trains a projector on frozen audio features, and evaluates retrieval, attribute probing, and matched disease classification.

    Source: §1 · p. 1

  • Results

    Correct pairing improves profile retrieval and sex decodability, but matched disease AUROC shows no consistent change and intervals across 12 dataset–backbone settings include zero.

    Source: §1 · p. 1, §3.3 · p. 3

  • Takeaways & Limitations

    The authors recommend reporting profile correspondence, attribute decodability, and matched disease prediction separately, with raw-audio references and explicit uncertainty.

    Source: §4 · p. 4

    Matching balances measured covariates only, external targets have limited precision, and pointwise bootstrap intervals omit choices made during UKCOVID discovery.

    Source: §4 · p. 4

Abstract

from arXiv · show

Clinical audio models are often trained by aligning recordings with participant metadata or text derived from it. This metadata is rarely a clean disease signal, since a single profile also contains symptoms, age, sex, missing responses, and recruitment information. We examine what alignment adds to frozen audio features by varying only the audio-meta data pairing. We keep the audio backbones fixed and train only a projector under correct pairing and three shuffled conditions that progressively preserve the disease label and recorded sex. We then evaluate each projector on profile retrieval, attribute probing, and disease classification on covariate-matched test sets. On UKCOVID, correct pairing clearly improves retrieval and makes sex easier to decode, confirming that the pairing manipulation works. Most of the sex-probe gap disappears when the control also preserves recorded sex, while a small retrieval gap remains. Matched disease AUROC differences are small, and every interval includes zero, so modest benefits or harms are not ruled out. The same pattern appears when we rerun the frozen pipeline on CODA TB, Cambridge COVID-19 Sounds, and Coswara, where retrieval and sex-probe contrasts are consistently positive but disease contrasts are not. We find no consistent evidence that better profile correspondence translates into better disease prediction and suggest reporting these outcomes separately.

1. INTRODUCTION

Clinical audio–metadata alignment may reward profile-identifying signals unrelated to disease, making it unclear whether alignment improves disease information. The study varies only pairing and separates profile retrieval, attribute decodability, and matched disease prediction.

1. INTRODUCTION

  • Metadata profiles combine disease-related information with symptoms, age, recorded sex, missing responses, and recruitment patterns that may identify profiles for non-disease reasons.Thus, contrastive alignment can encode disease information, demographic shortcuts, or metadata-form fingerprints while producing similar training curves.Source: §1 · p. 1
  • Prior findings report reduced COVID screening performance after symptom and demographic adjustment, alignment effects depending on shared modality information, and shortcut use by contrastive objectives.The paper frames these findings as reasons to test whether profile correspondence translates into disease prediction.Source: §1 · p. 1
  • The study varies only audio–metadata pairing, comparing correct, within-label, global, and label-and-sex-preserving shuffles with all other training factors fixed.It evaluates frozen projectors on profile retrieval, attribute probing, and disease prediction using covariate-matched test sets.Source: §1 · p. 1
  • Correct pairing improves profile retrieval and sex decodability, but matched disease AUROC shows no consistent change across 12 dataset–backbone settings.Every disease interval includes zero; modest benefits or harms remain possible.Source: §1 · p. 1

2. DATA AND METHODS

The study fixes its analysis pipeline on UKCOVID, then applies it unchanged to three external datasets. Frozen audio and text encoders support controlled pairing comparisons, with projector evaluation spanning retrieval, probing, matched disease prediction, and uncertainty estimation.

2.1. Discovery dataset and external panel

  • UKCOVID fixes the pairing arms, backbones, matching variables, probes, calibration, bootstrap, primary contrast and metric before external results are inspected.The discovery evaluation includes Standard, covariate-matched, and larger participant-disjoint matched test sets.Source: §2.1 · pp. 1–2
  • UKCOVID is the discovery dataset; the external panel comprises CODA TB, Cambridge COVID-19 Sounds, and Coswara, each with a matched test set of 200 participants.CODA uses match-first TB analysis, Cambridge uses a strict-COVID reconstruction of the Task-2 cough subset, and Coswara is matched before modeling; the external protocol remains unchanged.Source: §2.1 · pp. 1–2

2.2. Pairing-controlled training

  • The four arms pair audio with its own metadata (C), another participant’s same-label metadata (W), unrestricted shuffled metadata (G), or metadata preserving label and recorded sex (Wy,s).All arms use identical audio examples and training settings; only the metadata pairing changes.Source: §2.2 · p. 2
  • C − W measures effects of exact pairing beyond label-matched pairing, while the prespecified C − Wy,s contrast tests matched disease AUROC when recorded-sex consistency is also preserved.These are contrasts between pairing constraints, not causal estimands or decompositions into disjoint information sources.Source: §2.2 · p. 2

2.3. Training and evaluation

  • Frozen audio and text encoders feed a one-directional InfoNCE training objective; only an audio-side projector is trained for 500 epochs and then frozen for evaluation.The final checkpoint is used without epoch selection; this reproduces RespiraMFM Stage-1 alignment, not its full downstream system.Source: §2.3 · pp. 2–3
  • Evaluation measures macroaveraged profile-retrieval MRR, attribute-probe performance, and matched-set disease AUROC, NLL, and Brier score against raw-audio or metadata references.Probes measure attribute accessibility to a simple classifier, not whether a disease classifier uses that attribute.Source: §2.3 · p. 3

2.4. Uncertainty

  • The study reports pointwise 95% bootstrap intervals, jointly resampling profiles and seeds for retrieval and participants and seeds for probes and disease metrics.Intervals including zero do not establish equivalence; locking the external protocol does not remove analysis-selection freedom from the UKCOVID discovery stage.Source: §2.4 · p. 3

3. RESULTS

Label-preserving shuffling changes what probes decode beyond global shuffling, while correct pairing improves profile retrieval and sex decodability. Across matched disease evaluations, however, AUROC differences are inconsistent and sensitivity checks do not change that interpretation.

3.1. A label-preserving shuffle differs from a global shuffle

  • For AST, source-split W−G probe AUROC differences were 6.46 (COVID), 7.58 (cough), and 6.47 (recruitment); on the matched split COVID was 1.26 and cough was 8.03.Because W preserves disease labels while G does not, a direct C−G comparison combines label-level effects with any exact-pairing effect.Source: §3.1 · p. 3, Table 1 · p. 3

3.2. Correct pairing changes retrieval and sex decodability

  • For AST-6L, sex-probe AUROC C−W fell from 19.1 points to approximately zero under C−Wy,s; OPERA-CT and HeAR showed the same pattern, while Wy,s−W remained large and positive.Age probes improved under C for AST but not OPERA; age was not isolated because no age-preserving control was built.Source: §3.2 · p. 3
  • C−Wy,s retrieval MRR differences were positive for all three backbones, with intervals above zero for HeAR and AST (AST’s lower bound was zero), while OPERA’s interval included zero.The residual persisted when candidate pools were restricted by sex or by both label and sex.Source: §3.2 · p. 3

3.3. No consistent change in matched disease prediction

  • C−Wy,s matched disease AUROC differences were −0.03 (AST), 0.09 (OPERA), and 1.30 (HeAR), with all intervals covering zero; external disease intervals also included zero.External disease estimates were positive in five of 12 settings and negative in seven; no pooled claim was made.Source: §3.3 · pp. 3–4, Figure 2 · p. 4

3.4. Sensitivity analyses

  • Sensitivity analyses left the interpretation unchanged: aligned representations had no consistent advantage over metadata plus raw audio, and target recalibration reduced NLL differences while AUROC remained unchanged.This pattern points to probability-scale miscalibration rather than a difference in ranking.Source: §3.4 · p. 4

4. DISCUSSION AND LIMITATIONS

Correct pairing improves profile retrieval and attribute decodability, but the paper finds no consistent improvement in matched disease prediction. The authors recommend separating these outcomes in reporting and note limitations in matching, data scope, and analysis.

4. DISCUSSION AND LIMITATIONS

  • Matched disease prediction shows no consistent improvement, and intervals do not rule out modest benefits or harms despite improved profile retrieval and attribute decodability.The authors say alignment is not established to fail; disease prediction should be measured separately from alignment quality against raw-audio references.Source: §4 · p. 4
  • The label mix, recorded sex, and participant identity each shift representations, so correct-versus-global-shuffle comparisons conflate distinct pairing effects.Source: §4 · p. 4
  • Reports should separate profile correspondence, attribute decodability, and matched disease prediction, each with a raw-audio reference and explicit uncertainty.The authors also suggest label-preserving controls instead of global shuffling when metadata correlate with the outcome.Source: §4 · p. 4
  • Matching balances measured covariates only; external targets have limited precision, and the study tests one projector design and one deterministic schema rendering.Pointwise bootstrap intervals omit UKCOVID discovery-stage choices; the sex-preserving control may retain other attribute structure, and upstream pretraining exposure is not ruled out.Source: §4 · p. 4

5. CONCLUSION

Across the discovery dataset and an unchanged external rerun, alignment improves paired-profile retrieval, but the authors find no consistent evidence of better matched disease prediction. They conclude that clinical audio alignment reports should distinguish profile correspondence from disease transfer.

5. CONCLUSION

  • Alignment improves retrieval of the paired profile relative to within-label shuffling and, in most comparisons, relative to sex-preserving shuffling.Source: §5 · p. 4
  • The authors find no consistent evidence that improved profile correspondence leads to better disease prediction on matched test sets.Source: §5 · p. 4
  • Clinical audio alignment reports should distinguish correspondence from disease transfer.Source: §5 · p. 4

Compliance with Ethical Standards and Acknowledgment

The work is a retrospective analysis of de-identified datasets used under their respective data-use terms; GPT-5.5 and Claude assisted with language polishing.

Compliance with Ethical Standards and Acknowledgment

Related papers

Loading 2610.35747v2…