BCoughBench exposes 12% accuracy drop for respiratory AI on wearables

BCoughBench, accepted to the KDD 2026 Workshop on Reliable Scientific Foundation Models, evaluates how five respiratory acoustic foundation models (OPERA-CT/CE/GT, HeAR, M2D+Resp) perform when using audio from body-coupled wearable sensors instead of standard smartphone recordings. Using five labelled cough datasets and five EBEN-simulated body-coupled conditions (e.g., temple vibration pickup, soft in-ear, forehead accelerometer), the benchmark reveals a systematic degradation in classification performance. Mean AUROC falls from 0.785 (smartphone baseline) to between 0.689 and 0.723, depending on sensor type — a drop of 6–12%. The worst sensor was temple vibration pickup (Δ = -0.096), while soft in-ear was most resilient (Δ = -0.062).

Critically, no foundation model reached the clinical sensitivity threshold (Se@Sp95 ≥ 0.20) on most disease-classification tasks under any body-coupled condition. The most dramatic failure occurred in sex classification on the CIDRZ cohort, where AUROC tumbled from 0.954 to 0.596–0.628 (Δ = -0.341). Conversely, COVID-19 detection was nearly unaffected (Δ = -0.004), and age regression proved robust — the forehead accelerometer even improved MAE on CoughVID from 9.61 to 8.97 years. Among models, HeAR led on regression and demographic tasks, while M2D+Resp excelled on disease and characteristic tasks. BCoughBench provides a reproducible framework for evaluating models before clinical deployment on wearable devices, highlighting that current respiratory acoustic FMs are not yet reliable for body-coupled sensor pipelines.