BCoughBench: Benchmarking Respiratory Acoustic Foundation Models Under Body-Coupled Wearable Sensor Conditions

Technical Summary of BCoughBench

Problem Statement
Respiratory acoustic foundation models (FMs) are currently benchmarked exclusively on smartphone (air-conducted, AC) recordings. However, clinical deployment increasingly targets body-coupled (BC) wearable sensors (e.g., accelerometers, throat microphones, in-ear devices). These BC sensors act as low-pass filters, attenuating high-frequency content through tissue and bone. This spectral degradation introduces a domain shift that has not been characterized for existing FMs. Consequently, the reliability of these models under wearable conditions remains unknown, and their performance at clinically relevant operating points (specifically sensitivity at high specificity) is unverified.

Methodology
The authors introduce BCoughBench, a framework designed to evaluate five respiratory FMs under simulated body-coupled conditions.

• Datasets and Tasks: The benchmark utilizes five labeled cough datasets (CoughVID, Coswara, CIDRZ, COPD-CC, CovidUK) covering 12 tasks: nine classification tasks (disease detection, demographic inference, acoustic characteristics) and three age regression tasks.
• AC-to-BC Simulation: Instead of requiring physical wearable hardware for initial evaluation, the authors employ pre-trained EBEN (Extreme Bandwidth Extension Network) reverse models. These models transform clean smartphone audio into body-coupled equivalents across five distinct sensor placements:
  1. Forehead accelerometer (≤8 kHz)
  2. Soft in-ear microphone (≤6 kHz)
  3. Rigid in-ear microphone (≤5 kHz)
  4. Temple vibration pickup (≤2 kHz)
  5. Throat microphone (≤1.5 kHz)
• Models Evaluated: Five frozen respiratory FMs are tested without any fine-tuning or adaptation to BC audio:
  • OPERA-CT, OPERA-CE, and OPERA-GT (trained on 136K respiratory clips).
  • HeAR (a ViT-L masked autoencoder trained on 313M health audio clips).
  • M2D+Resp (combining AudioSet masking with respiratory fine-tuning).
• Evaluation Protocol: The models are evaluated in a strictly zero-shot regime.
  • Classification: A linear probe is trained on top of frozen embeddings. Metrics include AUROC, Sensitivity at 95% Specificity (Se@Sp95), and Expected Calibration Error (ECE). A Se@Sp95 < 0.20 is defined as clinically unusable.
  • Regression: Age prediction uses an MLP-small head, evaluated via Mean Absolute Error (MAE) against a mean-predictor baseline (MAD).

Key Contributions

1. BC Simulation Pipeline: A reproducible method using EBEN reverse models to generate BC-equivalent audio from five datasets across five sensor types without physical hardware, demonstrating progressive high-frequency attenuation (e.g., forehead preserves up to 8 kHz, throat attenuates above 1.5 kHz).
2. Multi-Metric Evaluation Framework: The benchmark moves beyond AUROC to include Se@Sp95 and ECE for classification, and MAE vs. baseline for regression, revealing failures that AUROC alone obscures.
3. Sensor Degradation Characterization: Quantification of performance drops across sensor types, identifying temple vibration as the most severe degradation source and soft in-ear as the least.
4. Task-Dependent Findings: Systematic analysis showing that some tasks (e.g., sex classification on CIDRZ) collapse under BC conditions, while others (e.g., age regression) remain robust or even improve.

Results

• Overall Performance: Mean AUROC declines from 0.785 (smartphone baseline) to 0.689–0.723 across BC sensors. The most severe degradation occurs under temple vibration pickup ($\Delta = -0.096$), while soft in-ear shows the least ($\Delta = -0.062$).
• Clinical Usability: No FM meets the clinical sensitivity threshold (Se@Sp95 $\ge$ 0.20) on most disease tasks (TB, Symptomatic, COVID) under any BC sensor. Only COPD and Symptomatic (Coswara) tasks retain clinical sensitivity.
• Task-Specific Variance:
  • Sex Classification: Performance on the CIDRZ cohort collapses dramatically (AUROC drops from 0.954 to 0.596–0.628, $\Delta = -0.341$), suggesting sex-discriminative features rely on high frequencies lost in BC transduction. Conversely, sex classification on Coswara slightly improves under forehead sensors.
  • COVID Detection: Uniquely robust, showing near-zero degradation ($\Delta = -0.004$), though still clinically unusable due to low sensitivity.
  • Age Regression: Remains robust across all sensors. On CoughVID, age regression actually improves under the forehead accelerometer (MAE decreases from 9.61 to 8.97 years), suggesting low-frequency age correlates are partially masked by high-frequency noise in open-air recordings.
• Model Performance: HeAR leads in demographic and regression tasks, while M2D+Resp leads in disease and characteristic tasks. OPERA models generally underperform.

Significance
The paper argues that AUROC alone is insufficient for evaluating respiratory FMs intended for wearable deployment, as it fails to reveal critical failures at the operating points required for clinical screening (Se@Sp95). The results demonstrate that sensor selection is as consequential as model selection; the performance gap between the best and worst sensors (0.034 AUROC) is comparable to inter-model differences on smartphone data. BCoughBench provides a necessary framework for exposing these domain shifts, highlighting that while low-frequency physiological signals (like age) survive BC transduction, high-frequency discriminative cues (like certain disease or demographic markers) often do not. The authors note that their use of simulated data is a limitation, and future work must validate these findings on physical wearable hardware.