Separating Voice from Age in COPD Screening

George P. Kafentzis; Nikoletta Arvaniti

eess.ASAug 21, 2026

Separating Voice from Age in COPD Screening

Authors: George P. Kafentzis, Nikoletta Arvaniti

Organizations: Department of Computer Science, University of Crete

Abstract

Voice has been proposed as a low-cost screening signal for chronic obstructive pulmonary disease (COPD). COPD is strongly age-associated and voice changes with age, thus such results admit a trivial alternative explanation. We re-evaluate a public sustained-phonation corpus ( $1246$ recordings, $68$ participants) under a strictly participant-level protocol. We therefore evaluate on repeatedly drawn age-matched cohorts and report the discrimination achieved by the confounders themselves on those same cohorts. Where raw (unmodelled) age ( $0.510$ $[0.469, 0.551]$ ) and raw gender ( $0.479$ ) are both measured at chance, acoustic models excluding age retain ROC-AUC $0.717$ $[0.552, 0.859]$ and average precision $0.747$ $[0.581, 0.892]$ against a one-to-one baseline of $0.5$ , whereas models containing age fall to $0.531$ -- $0.679$ . The separation is reproduced by two further learners with fixed hyperparameters. Two findings have broader methodological implications: models trained with age transfer less effectively to an age-balanced target cohort than otherwise identical models trained without age, and fourteen classical voice-quality and perturbation measures achieve comparable discrimination to a $55$ -dimensional combined representation. We conclude that a non-age acoustic signal is present, that confounding by recording conditions cannot be excluded from the released features, and that the evaluation protocol in standard use cannot distinguish these possibilities.

Figures & tables

Feature group	Feature	No.
Demographic	Age	1
Demographic	Gender	1
Health	Cold (Cold/Flu)	1
	Pain (Sore throat)	1
	Slimy (Mucus in throat)	1
	Other	1

TABLE I: Features used for COPD classification. Refer to [ 15 ] for further details.

Fig. 1: Age structure of the cohort. (a) Participant-level age distributions; the shaded band marks the COPD age range, outside which ten controls fall. (b) For each control, the age distance to the nearest COPD participant; eight of

37

exceed the five-year matching criterion stated in the source publication and are shown in orange. (c) Group mean age computed over participants and over recordings, on the full

68

-participant corpus. The sign of the group difference reverses between the two weightings.

Fig. 2: COPD proportion by age band (

n

above each bar). Prevalence peaks in the

70

–

75

band and declines thereafter, because the COPD group is bounded above at

81

years. A non-monotone score in age alone attains ROC-AUC

0.721

, against

0.677

for a monotone score.

Fig. 3: Association with age against univariate COPD discrimination, for every acoustic descriptor at participant level. The strongest discriminators (perturbation measures, orange diamonds) are among the least age-correlated; the most age-correlated features (MFCC standard deviations) discriminate weakly. Univariate AUCs are in-sample and serve to order features, not to estimate performance.

Abbrev.	$p$	Composition
A $\dagger$	1	age
D $\dagger$	2	age, gender
DH $\dagger$	6	age, gender, and symptom items
AC	49	static acoustics
AC-Dur	48	static acoustics (no duration)
PTRB	14	[jitter, shimmer, HNR, $F_{0}$ ] stats

TABLE II: Feature configurations. D denotes demographics (age and gender), H the four symptom items, AC the

49

static acoustic descriptors, and

\Delta

the

52

MFCC derivative summaries. A suffix

-\texttt{X}

denotes removal of X and

+\texttt{X}

denotes addition of X , so predictor counts are additive.

\dagger

marks configurations containing age.

Hyperparameter	Values
Iterations	$\{300,600,800,1000\}$
Tree depth	$\{2,4,6\}$
Learning rate	$\{0.001,0.01,0.03,0.10\}$
$L_{2}$ leaf regularization	$\{1,3,6,10\}$

TABLE III: CatBoost hyperparameter search space.

Configuration	$p$	ROC-AUC	PR-AUC
DAC+ $\Delta$ $\dagger$	103	0.741 [0.622, 0.855]	0.697 [0.540, 0.864]
ALL+ $\Delta$ $\dagger$	107	0.717 [0.593, 0.840]	0.669 [0.509, 0.858]
D $\dagger$	2	0.716 [0.580, 0.844]	0.657 [0.498, 0.859]
DAC $\dagger$	51	0.712 [0.576, 0.835]	0.692 [0.535, 0.865]
A $\dagger$	1	0.709 [0.571, 0.840]	0.596 [0.436, 0.803] ∗
DH $\dagger$	6	0.704 [0.565, 0.833]	0.618 [0.462, 0.830]

TABLE IV: Pooled subject-level discrimination, all

67

participants, with participant bootstrap intervals (

2000

resamples). Chance is

0.5

for ROC-AUC and the prevalence

0.448

for PR-AUC.

\dagger

marks configurations containing the explicit age predictor.

Contrast	Isolates	$\Delta$ ROC-AUC	width
ALL-A $-$ DAC-A	health items	$+0.001$ $[-0.040,+0.037]$	0.077
AC $-$ AC-Dur	duration	$+0.005$ $[-0.040,+0.049]$	0.088
ALL $-$ ALL+ $\Delta$	MFCC $\Delta$	$-0.041$ $[-0.103,+0.014]$	0.116
DH $-$ D	health items	$-0.013$ $[-0.077,+0.044]$	0.121
DAC $-$ DAC+ $\Delta$	MFCC $\Delta$	$-0.030$ $[-0.104,+0.038]$	0.141
DAC-A $-$ AC	gender	$+0.005$ $[-0.068,+0.087]$	0.156

TABLE V: Paired comparisons on pooled scores, class-stratified participant bootstrap (

20{,}000

resamples). No interval excludes zero. Ordered by interval width; only the first two are narrow enough to constitute resolved nulls.

Configuration	$p$	Matched ROC-AUC	Matched PR-AUC	$\delta$
AC	49	0.717 [ 0.552 , 0.859]	0.747 [ 0.581 , 0.892]	$+0.021$
DAC-A	50	0.682 [0.498, 0.845]	0.734 [ 0.566 , 0.888]	$-0.020$
AC-Dur	48	0.681 [ 0.505 , 0.835]	0.740 [ 0.586 , 0.869]	$-0.010$
DAC+ $\Delta$ $\dagger$	103	0.679 [ 0.524 , 0.818]	0.703 [ 0.541 , 0.851]	$-0.062$
ALL-A	54	0.679 [0.495, 0.842]	0.741 [ 0.567 , 0.887]	$-0.024$
PTRB	14	0.665 [0.488, 0.823]	0.755 [ 0.612 , 0.873]	$-0.006$

TABLE VI: Matched-cohort discrimination, caliper

2

years,

2000

repetitions with a paired bootstrap over matched pairs. Chance is

0.5

for both metrics under one-to-one matching. Raw age scores

0.510

[0.469,0.551]

and raw gender

0.479

on these cohorts. Bold lower bounds exclude chance.

\delta

is the matched-subset AUC minus full-cohort pooled AUC.

\dagger

contains the explicit age predictor.

Configuration	CatBoost	HistGB	LogReg
AC	0.717	0.717	0.583
DAC-A	0.682	0.710	0.549
AC-Dur	0.681	0.724	0.606
DAC+ $\Delta$ $\dagger$	0.679	0.595	0.597
ALL-A	0.679	0.693	0.545
PTRB	0.665	0.668	0.628

TABLE VII: Matched ROC-AUC (

2

-year caliper, paired bootstrap over matched pairs) under CatBoost with tuned hyperparameters, histogram-based gradient boosting with fixed hyperparameters, and

L_{2}

-regularised logistic regression. Bold indicates the lower bound excludes chance.

	$\alpha=0.10$				$\alpha=0.20$
Configuration	Coverage	Efficiency	Correctness	E $\times$ C	Coverage	Efficiency	Correctness	E $\times$ C
PTRB	0.913	0.387	0.798	0.309	0.808	0.737	0.752	0.554
AC	0.920	0.261	0.705	0.184	0.814	0.675	0.737	0.498
AC-Dur	0.925	0.252	0.706	0.178	0.805	0.697	0.727	0.507
DAC-A	0.928	0.269	0.748	0.201	0.821	0.693	0.756	0.524
ALL-A	0.925	0.284	0.754	0.214	0.823	0.656	0.742	0.487

TABLE VIII: Exploratory cross-conformal prediction sets, averaged over

25

outer splits. Observed coverage met or exceeded the nominal level

1-\alpha

throughout; the construction carries no finite-sample guarantee (Section IV-F ). No prediction set was empty, so abstention consists entirely of sets containing both labels. “Efficiency” is the proportion of participants receiving a single-label set, “Correctness” the accuracy among those, and

\mathrm{E}\!\times\!\mathrm{C}

their product—the proportion of all participants receiving a single label and correct prediction.

\dagger

contains the explicit age predictor.

Clinical Speech Processing

Clinical speech processing uses speech recognition and acoustic analysis to support health assessment, monitoring, and communication.

See all 57 papers

Beyond Classification: A Cough Regression Benchmark for Respiratory Acoustic Foundation Models
Jun 13, 2026Mayur Sanap, Prasanna Desikan, Edgar LobatonAudio Understanding Clinical Prediction
Centific Global Solutions Inc. · Department of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC, USA.
Speaker-Disentangled Remote Speech Detection of Asthma and COPD Exacerbations
May 16, 2026Yuyang Yan, Sami O. Simons, Visara UroviSpeech Processing Clinical Prediction
Institute of Data Science, Maastricht University, Paul-Henri Spaaklaan 1, Maastricht, 6229 EN, the Netherlands · Department of Respiratory Medicine, NUTRIM Research Institute of Nutrition and Translational Research in Metabolism, Faculty of Health Medicine and Life Sciences, Maastricht University, P. Debyelaan 25, Maastricht, 6229 HX, the Netherlands · Department of Respiratory Medicine, Maastricht University Medical Centre, P. Debyelaan 25, Maastricht, 6229 HX, the Netherlands
SpiroPhonia: Non-Invasive Respiratory Health Assessment from Spontaneous Speech
Sep 15, 2026Roksana Khanom, Shafia Supty, Nirupam Roy +1
University of Maryland, College Park, USA · DR. M R Khan Shishu Hospital & Institute of Child Health, Bangladesh

Jun 13, 2026J/K move · Enter open · S save

Beyond Classification: A Cough Regression Benchmark for Respiratory Acoustic Foundation Models

Mayur Sanap, Prasanna Desikan, Edgar Lobaton

Centific Global Solutions Inc. · Department of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC, USA.

Audio Understanding Clinical Prediction

Open paper AskPDF arXiv

Respiratory acoustic foundation models (FMs) excel at cough classification, yet their ability to predict continuous health quantities from cough audio remains largely unexplored, despite the clinical value of passive age, BMI, and disease probability estimation in settings where physical measurements are unavailable. We introduce the multi-model, multi-target cough regression benchmark evaluating five FMs (OPERA-CT, OPERA-CE, OPERA-GT, HeAR, M2D+Resp) across six targets on three datasets under subject-disjoint protocols, comparing linear, MLP-small, and full MLP regression heads. MLP-small beats the mean-predictor baseline on all tasks and linear probing in 23 of 30 model x task cases, with full MLP overfitting on small clinical data but recovering on larger sets, revealing a dataset size x head-capacity trade-off. HeAR leads within-dataset age regression on Coswara (9.12 yr MAE); its CIDRZ result is excluded from headline claims owing to possible HeAR-CIDRZ pretraining overlap. OPERA-GT is favored over OPERA-CT on age in all three datasets, with the CIDRZ margin within seed variance, extending a generative-pretraining advantage from breath to cough. HeAR and M2D+Resp reach near-full performance at N = 50 samples while OPERA models require N = 400. Cross-dataset transfer is strongly asymmetric as large diverse data generalises to small clinical populations (CoughVID to CIDRZ: -0.17 yr) but not vice versa (CIDRZ to Coswara: +2.43 yr, +26.6%).

Abstract

Clinical Speech Processing

Explore similar work

Beyond Classification: A Cough Regression Benchmark for Respiratory Acoustic Foundation Models

Speaker-Disentangled Remote Speech Detection of Asthma and COPD Exacerbations

SpiroPhonia: Non-Invasive Respiratory Health Assessment from Spontaneous Speech