Artificial Intelligence Respiratory Rate Estimation from Low-Cost Smartphone Audio Under Variable Ambient Clinic Noise: A Self-Supervised Domain Adaptation Evaluation of Healthcare Triage Sensitivity

Respiratory rate is a sensitive but frequently neglected vital sign in triage, and smartphone audio offers a low-cost route to its estimation. The obstacle is domain shift: clinic noise varies between rooms, devices and hours, and a model trained on clean or laboratory audio degrades exactly where triage decisions are made. The adaptation objective is posed as a minimax problem over an invariance penalty on unlabelled target audio, and a bound is derived linking target-domain error to source error, the invariance penalty and the divergence between source and target feature distributions. The bound implies that adaptation gains are capped by the invariance penalty and are not monotone in unlabelled target volume. A before-after-control-impact design with 8, 16, 24 and 32 clusters, effect sizes of 3, 5 and 8 percentage points, and 2000 replications per cell was characterised under a declared variance structure with between-cluster SD of 6.0, residual SD of 4.5 and intraclass correlation of 0.64. Power reached 80 per cent at 54 clusters for a 5 percentage point effect. The contribution is a design-level account of what self-supervised adaptation can and cannot buy a triage service, and a statement of the cluster counts required before a field evaluation is worth mounting.