Skip to content
peer-reviewed pilot · respiratory medicine

Hearing the difference between respiratory illnesses

Using 514,377 ambient-recorded primary care visits from 379,225 adult patients, we trained 11 acoustic classifiers aligned with a respiratory triage cascade — demonstrating that conversational speech from real clinic visits carries signal for differential-diagnosis-level triage, not just sick-versus-healthy detection.

Authors
Vijay Ravi, Camille Noufi
Journal
Frontiers in Medicine — Respiratory Medicine
Published
July 2026
Type
Peer-reviewed pilot study
514,377Visits analyzedAmbient primary care audio
379,225Adult patientsUS clinic network
11Binary classifiersAligned to a triage cascade
0.602–0.745Test-set AUC rangeMean ECE = 0.018
the question

Not sick versus healthy. Which illness.

Respiratory complaints drive a large share of adult ambulatory visits, and how they are triaged has direct downstream consequences: antibiotic stewardship, choice of pathogen-specific therapy, escalation to imaging or in-person evaluation, and the total cost of care.

Yet most published voice-based respiratory work has evaluated a narrower question — can we tell a sick patient from a healthy one? — using scripted speech (sustained vowels, read passages, cough recordings) collected in crowdsourced apps or controlled research clinics.

That leaves the clinically important question unanswered: can voice help distinguish between respiratory illnesses in patients who are already known to be sick? This pilot is designed to test exactly that, on the kind of audio a primary care clinic actually generates.

the dataset

Real clinic audio at scale

Each visit is paired with clinically assigned ICD-10 diagnosis codes and de-identified demographic and geographic metadata. Patient audio is isolated from the doctor-patient conversation, and spectral, voice quality, and prosodic features are computed per visit.

property 01

Conversational, ambient speech

Not scripted vowels or coughs — the audio is what a patient naturally says during a visit, captured by the recording infrastructure the clinic already runs.

property 02

Differential contrasts among sick patients

Controls in most binaries are themselves patients presenting for care, not asymptomatic volunteers. This is a materially harder problem than sick-vs-healthy detection.

methods

A triage cascade of 11 binary classifiers

Eleven binary classification tasks aligned with a respiratory triage cascade — for example, acute respiratory illness versus acute non-respiratory illness, and lower versus upper respiratory tract infection. Each task received its own acoustic model.

01

Patient-stratified training

Every task was trained with patient-stratified five-fold cross-validation and evaluated on a held-out test set, so no patient appears on both sides of the split.

5-fold CVheld-out test set
02

Confounder baselines

Each acoustic model was compared against six non-acoustic baselines built from a single demographic, geographic, or temporal variable. This tests whether the model is learning acoustic signal, or simply riding demographic or clinic-level patterns.

6 baselinesmultiple-testing correction
03

Hierarchical composition

The 11 trained classifiers are composed into a hierarchical cascade — all visits, then acute respiratory versus non-respiratory, then lower versus upper respiratory tract, then pneumonia and viral/bacterial splits. Upstream decisions constrain the downstream differential.

L1 → L2 → L3case-study cascades
results

Six of eleven beat every shortcut

Across the 11 binaries, test-set AUC ranged from 0.602 (95% CI: 0.588–0.614) to 0.745 (95% CI: 0.742–0.748), with a mean expected calibration error of 0.018 — meaning predicted probabilities were well-aligned with observed frequencies, an important property for any triage-adjacent tool.

Six of eleven binaries outperformed all six confounder baselines after multiple-testing correction. In another four, median within-stratum AUC held at 0.61–0.70 when the confounder was fixed, indicating that acoustic signal was discriminating beyond what the confounder alone could explain.

Five binaries failed at least one axis. The most instructive failure was pneumonia versus non-pneumonia lower respiratory tract infection, which was outperformed by a patient-city baseline. The likely explanation is not that voice carries no signal for pneumonia, but that clinic-level differences in ICD-10 coding create a shortcut the demographic baseline can exploit — a useful reminder that in real-world clinical data, confounders can outperform models for reasons that have nothing to do with biology.

Headline results

Test-set AUC, baseline comparisons, and calibration across the 11 binary tasks.

Lowest task AUC0.602
Highest task AUC0.745
Chance0.500

Mean expected calibration error 0.018. Six of eleven binaries outperformed all six confounder baselines after multiple-testing correction.

framing

Moderate AUCs are the right frame here

Prior voice-AI work has often reported AUCs above 0.90, but under conditions far friendlier than clinical reality: scripted speech, curated microphones, and healthy controls. This study takes the opposite approach — conversational ambient audio, ICD-10-coded diagnoses, and differential-diagnosis contrasts among patients who are all seeking care. In that setting, an AUC in the mid-0.7s that beats every demographic and geographic shortcut is a meaningful signal.

implications

What this means for teams shipping respiratory triage

For nurse-line assistants, ambient documentation tools, and virtual-first primary care platforms, the practical takeaways are concrete.

Conversational audio is a viable substrate

You do not need to add scripted voice tasks to a visit. The speech a patient already produces carries the signal.

Confounder baselines are non-negotiable

Without non-acoustic baselines in the evaluation, reported performance is not interpretable — the model may only be reading demographics.

Cascades place evidence where it is strongest

Hierarchical composition lets you deploy acoustic evidence at the decision points it performs on, and defer at the ones it doesn't.

Coding artifacts are a real threat to validity

Clinic-level ICD-10 patterns can create shortcuts that outperform the model. They need to be surfaced, not hidden.

read the paper

Open access at Frontiers in Medicine

The full peer-reviewed article — including all methods, per-binary results tables, confidence intervals, and the hierarchical cascade case studies — is available open access.

Ravi V, Noufi C. Conversational Speech for Respiratory Triage in Primary Care: A Pilot Study. Frontiers in Medicine, Respiratory Medicine section, 2026. doi:10.3389/fmed.2026.1895376

Read the full article

All methods, per-binary results tables, confidence intervals, and the hierarchical cascade case studies are open access at Frontiers in Medicine.

Keywords: acoustic features, ambient audio, confounder analysis, differential diagnosis, hierarchical classification, ICD-10 cohorts, real-world clinical audio, vocal biomarkers.