TRIAGE: Adaptive Test-Time Scaling for Zero-Shot Respiratory Audio Classification
A zero-shot respiratory audio classification framework that adaptively allocates test-time compute to uncertain cases.
This project corresponds to our paper (Wang et al., 2026).
TRIAGE (Tiered Retrieval and Inference for Audio with Gated Escalation) is a zero-shot respiratory audio classification framework that improves medical audio inference without task-specific training. Instead of applying the same inference procedure to every recording, TRIAGE adaptively routes each sample through progressively richer reasoning stages depending on prediction confidence.
The framework starts with a fast Tier-L stage that performs label-cosine scoring in a frozen audio–text embedding space. Confident cases are finalized immediately, while uncertain cases are escalated to Tier-M, where clinician-style descriptor templates are matched against the audio and converted into rule-based predictions. The most ambiguous recordings are further routed to Tier-H, which retrieves similar audio–report examples and uses an LLM to make an evidence-grounded final decision.
Across nine respiratory audio classification tasks spanning cough, exhalation, and lung-sound datasets, TRIAGE achieves strong zero-shot performance, reaching a mean AUROC of 0.744 without any task-specific training. The adaptive router concentrates additional computation where it matters most: high-confidence cases exit early, while uncertain cases benefit from descriptor-based reasoning and retrieval-augmented LLM inference.
