Cough Regression Benchmark Advances Respiratory Acoustic AI Models

While respiratory acoustic foundation models (FMs) have shown strong performance in classifying coughs, their capability to predict continuous health metrics from cough audio has remained largely unexplored. This is a significant gap, as estimating quantities like age, BMI, or disease probability passively from coughs holds substantial clinical value, particularly in settings where direct physical measurements are not feasible. Researchers have introduced a comprehensive multi-model, multi-target cough regression benchmark. This benchmark assesses five different FMs across six distinct health targets using three diverse datasets, employing subject-disjoint protocols. The evaluation compares the effectiveness of linear, small MLP, and full MLP regression heads attached to these foundation models. Key findings indicate that a small MLP regression head generally outperforms both a mean-predictor baseline and linear probing across most tasks. While a full MLP can overfit on smaller clinical datasets, its performance recovers with larger data, highlighting a trade-off between dataset size and model capacity. Notably, models like HeAR and OPERA-GT demonstrate strong performance in age regression, and cross-dataset transfer learning shows asymmetric results, with large, diverse datasets generalizing better to smaller clinical populations than vice versa.