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

The article introduces a cough regression benchmark for evaluating five respiratory acoustic foundation models (FMs): OPERA-CT, OPERA-CE, OPERA-GT, HeAR, and M2D+Resp, focusing on predicting continuous health metrics such as age and BMI from cough audio. Key findings include that MLP-small regression heads outperform mean predictors across all tasks, while HeAR achieves the best age regression performance with a mean absolute error of 9.12 years on the Coswara dataset. The study highlights the importance of dataset size and model capacity, revealing that smaller clinical datasets require more samples for effective performance, which is critical for practitioners developing applications in remote health monitoring where physical measurements are limited.