🤖 machine learning

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

This paper introduces a multi-target cough regression benchmark for respiratory acoustic foundation models, demonstrating that while MLP-based regression heads outperform linear probing and mean predictors, optimal performance depends on a trade-off between dataset size and model head capacity, with significant asymmetric transfer capabilities favoring general-to-specific adaptation.

Original authors: Mayur Sanap, Prasanna Desikan, Edgar Lobaton

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mayur Sanap, Prasanna Desikan, Edgar Lobaton

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant library of cough recordings. Scientists have built "super-smart" computer brains (called Foundation Models) that have read through millions of these coughs to learn what they sound like. Usually, these computers are only asked to play a simple game: "Is this a healthy cough or a sick cough?" (Classification).

This paper asks a much harder question: "Can these computers guess specific numbers just by listening to a cough?"

Think of it like this: Instead of just saying "This person is sick," can the computer guess, "This person is likely 45 years old," or "Their Body Mass Index is 22," or "There is a 60% chance they have Tuberculosis"?

Here is what the researchers found, broken down into simple stories:

1. The "Head" Matters (The Hat Analogy)

The computer brain has a "head" (a small extra layer) that takes the sound information and turns it into a number guess. The researchers tried three different types of heads:

The Simple Hat (Linear): Very basic, like a straight line.
The Medium Hat (MLP-small): A bit more flexible, with a small internal network.
The Giant Hat (Full MLP): A massive, complex network.

The Result: The Medium Hat was the winner. It beat the Simple Hat in most cases. The Giant Hat was too fancy for the small amount of medical data they had; it tried to memorize the specific patients instead of learning the general rules (this is called "overfitting"). However, when they gave the Giant Hat a huge pile of data, it finally started working well.

Takeaway: Don't use a sledgehammer to crack a nut, but if you have a mountain of nuts, the sledgehammer works great.

2. The "Training Style" Matters (The Teacher Analogy)

The researchers compared different ways the computers were originally trained:

Contrastive Training (OPERA-CT): Like a teacher who says, "This cough sounds like that cough, but different from this one."
Generative Training (OPERA-GT): Like a teacher who says, "Here is a cough with a missing piece; can you fill in the blank?"

The Result: The Generative teacher (the one who practices filling in blanks) was slightly better at guessing age than the Contrastive teacher. This held true across all three groups of people they tested. It suggests that learning to "reconstruct" a sound helps the computer understand the physical details of the body better.

3. The "Data Size" Problem (The Crowd vs. The Clinic)

This is where things get interesting regarding where the data came from.

The Big Crowd: They had huge datasets of coughs collected from the internet (like CoughVID).
The Small Clinic: They had smaller, very specific datasets from a hospital in Zambia (CIDRZ).

The Result: The transfer of knowledge was one-way street.

If you trained the computer on the Big Crowd and tested it on the Small Clinic, it worked surprisingly well. The computer learned general rules from the crowd that applied to the clinic.
If you tried to train on the Small Clinic and test on the Big Crowd, it failed miserably. The clinic data was too narrow and specific to teach the computer about the whole world.

4. The "Few-Shot" Miracle (The Quick Learner)

How much data does the computer need to learn to guess age?

Some models (like HEAR and M2D+RESP) are like genius toddlers. They can learn to guess age almost perfectly after seeing just 50 examples.
Other models (the OPERA family) are like hardworking students. They need to see about 400 examples before they get good at it.

This suggests that the variety of sounds the computer heard during its initial training (its "pretraining") matters more than the specific architecture of the computer itself.

5. The Reality Check (The "Good News, Bad News")

While the computer got better at guessing numbers than just guessing "sick" or "healthy," the results weren't perfect.

The Good News: On the large internet datasets, the computers could guess age with an error of about 9–10 years. This is better than just guessing the average age of everyone.
The Bad News: On the specific clinical dataset (CIDRZ), the computers barely did any better than just guessing the average age of the group. The "signal" (the actual useful information in the cough) was too weak for the computer to pick up specific details about individual patients in that specific group.

Summary

The paper introduces a new "test track" for cough-sounding computers. They found that:

A medium-sized prediction tool works best for small medical datasets.
Generative training (filling in blanks) is slightly better for guessing age than contrastive training.
Big data can teach computers to handle small, specific groups, but small data cannot teach them to handle big, diverse groups.
Some models are fast learners (needing only 50 samples), while others need more practice.

The authors are careful to say this is a benchmark study. They are showing how these models perform on these specific tasks, not claiming that doctors should start using coughs to diagnose patients yet. The technology is promising, but for now, it's mostly a tool for understanding what these AI models can and cannot do.

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

Problem Statement

Respiratory acoustic foundation models (FMs) have demonstrated proficiency in categorical cough classification. However, their capability to predict continuous health quantities—such as age, Body Mass Index (BMI), and disease probability—from cough audio remains largely unexplored. This gap is critical for low- and middle-income countries (LMICs) and settings where physical measurements (e.g., birth records, weighing scales) are unavailable. In these contexts, passive estimation of physiological states via cough acoustics could serve as actionable proxies for triage. While existing literature has benchmarked FMs on classification, their regression capabilities have not been systematically characterized across multiple models, targets, and datasets.

Methodology

The authors introduce a unified, multi-model, multi-target cough regression benchmark designed to evaluate the regression capabilities of five frozen respiratory acoustic foundation models.

Datasets and Tasks

The benchmark utilizes three distinct datasets under subject-disjoint protocols:

CIDRZ: Smartphone-recorded volitional coughs from a TB clinic in Zambia ( $N=1049$ ). Targets include age, BMI, chest X-ray abnormality scores, and TB probability.
Coswara: Remotely collected shallow coughs from India during the pandemic ( $N=2560$ ). Target: Self-reported age.
CoughVID: Globally submitted volitional coughs via smartphone ( $N=6858$ ). Target: Age.

Foundation Models (Encoders)

Five frozen encoders spanning three pretraining paradigms are evaluated:

OPERA Family (Respiratory-specific SSL):
- OPERA-CT: Contrastive Transformer (768-d).
- OPERA-CE: Contrastive CNN (1280-d).
- OPERA-GT: Generative Masked Autoencoder (MAE) with an 8.18s positional grid (384-d).
HEAR: A ViT-L MAE pretrained on 313M health audio clips (512-d), covering clinical audio beyond just respiratory sounds.
M2D+RESP: Combines masked spectrogram prediction on AudioSet with respiratory fine-tuning, mean-pooled to a 3840-d vector.

Regression Heads and Training

The study compares three regression heads applied to the frozen embeddings:

Linear: Standard linear probing ( $d_{feat} \to 1$ ).
MLP-small: A bottleneck architecture ( $d_{feat} \to 256 \to 1$ ) with ReLU and dropout, designed to be embedding-size agnostic.
Full MLP: A full-width architecture ( $d_{feat} \to d_{feat} \to 1$ ), expected to have high capacity (e.g., ~15M parameters for M2D+RESP).

All models are trained using Adam optimizer, MSE loss, and early stopping. Performance is measured via Mean Absolute Error (MAE) in natural units, compared against a naive mean-predictor baseline (Mean Absolute Deviation, MAD).

Evaluation Regimes

Within-Dataset Regression: Evaluating all targets on held-out test splits.
Head Comparison: Characterizing the interaction between dataset size and head capacity.
Cross-Dataset Transfer: Training on one dataset and evaluating on others (specifically for age, the only shared target across multiple datasets).
Low-Data Regime: Analyzing performance as a function of training set size ( $N$ ).

Key Contributions

The paper provides five primary contributions:

Comprehensive Benchmark: A multi-model, multi-target evaluation of five FMs across six targets and three datasets under subject-disjoint protocols.
Regression Head Analysis: Identification that a small MLP head generally outperforms linear probing and avoids the overfitting associated with full MLPs on small clinical datasets.
Generative Pretraining Advantage: Evidence that generative pretraining (OPERA-GT) outperforms contrastive pretraining (OPERA-CT) on age regression across all datasets, extending previous findings from breath audio to cough.
Asymmetric Transfer: Demonstration that cross-dataset transfer is strongly asymmetric; large web-collected data generalizes to small clinical populations, but the reverse is not true.
Low-Data Efficiency: Analysis showing that models with diverse pretraining corpora (HEAR, M2D+RESP) reach near-full performance with as few as 50 labeled samples, whereas models with smaller, domain-specific corpora (OPERA) require significantly more data ( $N=400$ ).

Results

Within-Dataset Performance

Signal Strength: The strongest signal was observed for age prediction on the Coswara dataset (Best/MAD ratio = 0.81). CoughVID age prediction showed a weaker signal (0.93).
Clinical Scores: For CIDRZ targets (BMI, X-ray abnormality, TB probability), all models performed near the chance floor (Best/MAD between 0.92 and 0.99). The minimal inter-model spread suggests a representational ceiling where embeddings do not capture usable patient-level signal for these specific clinical scores under linear probing.
Baseline Comparison: All models nominally exceeded the mean-predictor baseline, but margins were often marginal.

Head Architecture Comparison

MLP-small Dominance: The MLP-small head won in 23 of 30 model $\times$ task combinations, outperforming linear probing by up to 0.38 years (HEAR on Coswara).
Overfitting vs. Recovery: Full MLPs overfit on small datasets (e.g., M2D+RESP on CIDRZ degraded by +0.53 years compared to MLP-small) but recovered on larger datasets (CoughVID), revealing a dataset-size $\times$ head-capacity trade-off.

Pretraining Paradigms

Generative vs. Contrastive: OPERA-GT consistently outperformed OPERA-CT on age regression across all three datasets. While the margin on CIDRZ was within seed variance, the gaps on Coswara and CoughVID were significant, suggesting generative objectives capture more physiological variation.
Cross-Dataset Transfer: Transfer was successful only when large web-collected data (CoughVID, Coswara) was used to train models for small clinical populations (CIDRZ).
- CoughVID $\to$ CIDRZ: Improved performance by -0.17 years.
- CIDRZ $\to$ Coswara: Performance degraded significantly (+2.43 years, +26.6%).
- This asymmetry suggests clinical cohorts have narrower demographic ranges than global crowdsourced corpora.

Low-Data Regime

Data Efficiency: HEAR and M2D+RESP reached near-full performance at $N=50$ samples.
OPERA Models: Required $N=400$ samples to stabilize, attributed to their smaller, respiratory-only pretraining corpus compared to the diverse corpora of HEAR and M2D+RESP.

Significance and Claims

The paper claims that while cough acoustic FMs show promise for continuous health estimation, their utility is highly dependent on the specific task, dataset, and architectural choices.

Practical Head Selection: For frozen embeddings, a small MLP head is the preferred architecture, balancing non-linear capacity with robustness against overfitting in data-scarce clinical settings.
Pretraining Impact: Generative pretraining objectives appear superior to contrastive ones for age regression, and the diversity of the pretraining corpus is a stronger determinant of low-data performance than the model architecture itself.
Transfer Limitations: The benchmark reveals a critical limitation in current models: they generalize well from large, diverse web data to small clinical populations but fail to generalize from small clinical data to broader populations. This highlights the need for diverse pretraining data to enable effective clinical deployment.
Task Specificity: The results indicate that while age estimation shows learnable structure, predicting specific clinical scores (like TB probability or X-ray abnormalities) from cough audio alone remains a significant challenge, with current models performing near chance levels on these specific targets.

The authors explicitly note that these conclusions are scoped to the frozen-embedding, shallow-probe regime and do not represent the performance ceiling achievable with full fine-tuning. Future work is suggested to explore additional targets, fine-tuning regimes, and attention-pooling heads.