💬 NLP

Adaptive Test-Time Scaling for Zero-Shot Respiratory Audio Classification

The paper introduces TRIAGE, an adaptive zero-shot framework for respiratory audio classification that dynamically allocates computational resources across three reasoning tiers based on input uncertainty, achieving state-of-the-art performance without task-specific training while significantly reducing costs by resolving easy cases early.

Original authors: Tsai-Ning Wang, Herman Teun den Dekker, Lin-Lin Chen, Neil Zeghidour, Aaqib Saeed

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tsai-Ning Wang, Herman Teun den Dekker, Lin-Lin Chen, Neil Zeghidour, Aaqib Saeed

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor listening to a patient's breathing through a stethoscope. Sometimes, the sound is crystal clear: a loud, unmistakable wheeze that screams "asthma." Other times, the sound is faint, muffled by background noise, or so subtle that it requires years of experience to distinguish a "crackle" from a "rattle."

In the world of Artificial Intelligence, most computer programs treat every sound the same way. They run the exact same, heavy calculation on a clear sound as they do on a confusing, noisy one. It's like using a sledgehammer to crack a nut, or using a supercomputer to decide if a coin is heads or tails. It's wasteful, and for the tricky cases, it's often not smart enough.

This paper introduces TRIAGE, a new way for AI to listen to respiratory sounds. The name comes from the medical practice of "triage"—sorting patients by how urgent their needs are. TRIAGE doesn't just listen; it decides how hard to think based on how easy or hard the sound is to understand.

The Three-Tier System: A "Smart Clinic" Analogy

Think of TRIAGE as a three-level clinic for audio:

1. The Triage Nurse (Tier-L): The "Quick Glance"

What happens: When a sound comes in, the AI first takes a quick, cheap look. It compares the sound to simple labels like "Wheeze" or "Normal."
The Analogy: Imagine a nurse at a hospital entrance. If a patient walks in with a broken leg that is clearly visible, the nurse says, "Okay, we know what that is," and sends them straight to the orthopedist.
The Result: If the AI is very confident (like the nurse seeing the broken leg), it stops there. It saves time and money. About 46% of all sounds get solved here.

2. The Specialist (Tier-M): The "Detailed Check"

What happens: If the sound is a bit fuzzy or the nurse isn't 100% sure, the case moves up. The AI doesn't just look at the label; it breaks the sound down into specific medical features. Is the sound high-pitched? Does it happen when the patient breathes in or out? Where in the chest is it loudest?
The Analogy: This is like a specialist doctor who asks specific questions: "Does the pain hurt when you move your arm?" "Is it sharp or dull?" They use a checklist to figure it out.
The Result: If the checklist gives a clear answer, the AI stops here. This handles another chunk of the cases.

3. The Expert Panel (Tier-H): The "Deep Dive"

What happens: If the sound is still confusing after the checklist, the AI escalates to the "Big Gun." It goes to a massive library of past medical records, finds the most similar cases, and asks a super-smart AI (like a large language model) to read the reports and make a final, evidence-based decision.
The Analogy: This is like calling in a team of top specialists to review a rare, complex case. They look at similar past cases, discuss the evidence, and write a detailed report.
The Result: This is expensive and slow, so TRIAGE only uses it for the 19% of cases that are truly difficult. But for those hard cases, it boosts accuracy by nearly 20%.

Why This Matters

1. It's "Zero-Shot" (No New Training Needed)
Usually, to teach an AI a new disease, you need thousands of labeled examples (doctors saying "this is X, this is Y"). That takes years and costs a fortune. TRIAGE works without this. It uses a pre-trained brain that already knows how to match sounds to text. It just changes how it thinks, not what it knows.

2. It Saves Money and Time
By letting the easy cases leave early, TRIAGE saves a massive amount of computing power. It's like a smart traffic light that lets empty cars pass quickly but stops heavy trucks only when necessary.

3. It's Fair to the Hard Cases
The most important finding is that TRIAGE doesn't just get better on average; it specifically fixes the worst cases. The easy sounds stay just as easy, but the confusing, dangerous sounds get the extra attention they need to be diagnosed correctly.

The Bottom Line

TRIAGE is like giving an AI a "brain" that knows when to stop thinking. Instead of blindly crunching numbers on every single sound, it acts like a wise human doctor:

"This is obvious? Done."
"This is tricky? Let me check the details."
"This is a mystery? Let's call the experts and look at the history books."

This approach allows AI to screen for respiratory diseases (like COPD, asthma, or COVID-19) more accurately and efficiently, without needing to be retrained for every new hospital or every new disease. It's a step toward making high-quality medical diagnosis accessible to everyone, everywhere, instantly.

1. Problem Statement

Automated respiratory audio analysis (e.g., detecting wheezes, crackles, or coughs) holds promise for scalable disease screening but faces two major bottlenecks:

Data Scarcity: High-quality, labeled medical audio datasets are rare and expensive to annotate.
Inference Rigidity: Existing zero-shot methods (which do not require task-specific training) apply a uniform computational cost to every input. Whether a recording is a clear textbook wheeze or a faint sound buried in noise, the model performs a single forward pass. This fails to address the varying difficulty of clinical recordings, leading to suboptimal robustness on ambiguous cases without increasing costs for easy ones.

The authors propose that test-time compute should be a dynamic resource, allocated adaptively based on the difficulty of the input, rather than a fixed cost.

2. Methodology: The TRIAGE Framework

The paper introduces TRIAGE (Tiered Retrieval and Inference for Audio with Gated Escalation), a three-tier inference framework designed for zero-shot respiratory audio classification. The system uses a frozen audio-text encoder (specifically AcuLa) and scales computation by routing samples through progressively richer reasoning stages based on confidence.

Core Architecture

The framework consists of three tiers, escalating only when confidence is low:

Tier-L (Low Compute): Fast Label Cosine Scoring
- Mechanism: Computes cosine similarity between the audio embedding and natural language label prompts (e.g., "wheeze," "normal").
- Decision: If the margin between the top two scores exceeds a threshold ( $\tau_L$ ), the prediction is finalized immediately.
- Cost: Minimal (single embedding lookup).
Tier-M (Medium Compute): Structured Descriptor Matching
- Trigger: Activated if Tier-L confidence is low.
- Mechanism: Instead of raw labels, the audio is matched against a set of clinician-approved descriptor templates (e.g., timing, anatomical location, sound quality). A rule-based aggregator votes on the final label based on these attributes.
- Decision: If the margin between top rule-based scores exceeds a threshold ( $\tau_M$ ), the prediction is finalized.
- Cost: Moderate (multiple template comparisons and rule aggregation).
Tier-H (High Compute): Retrieval-Augmented LLM Reasoning
- Trigger: Activated if both Tier-L and Tier-M fail to reach confidence thresholds.
- Mechanism:
  1. Retrieval: Retrieves the $k$ -nearest neighbor audio-report pairs from an external corpus using the audio embedding.
  2. Reasoning: A Large Language Model (LLM, e.g., Gemini 3 Pro) is prompted with the audio descriptors, Tier-L scores, and the retrieved clinical reports to synthesize a final prediction with an evidence trail.
- Cost: High (retrieval + LLM inference).

Gated Escalation Policy

The system uses a confidence-based router to determine the exit point. By tuning thresholds $\tau_L$ and $\tau_M$ , practitioners can trade off between computational efficiency (finalizing more cases early) and accuracy (escalating more cases to Tier-H).

3. Key Contributions

Adaptive Test-Time Scaling for Medical Audio: Demonstrates that selectively increasing inference computation on uncertain recordings—without any task-specific training or fine-tuning—significantly improves zero-shot classification performance.
Compute-Aware Formulation: Casts auscultation inference as a staged decision process where compute is an explicit, tunable resource, enabling principled study of efficiency-accuracy tradeoffs.
Interpretable Pipeline: The tiered approach provides transparency. Tier-L gives a score, Tier-M provides an attribute profile (e.g., "late inspiratory crackle"), and Tier-H offers retrieved evidence and LLM reasoning, making the decision process clinically explainable.

4. Experimental Results

The authors evaluated TRIAGE on nine classification tasks across five public respiratory datasets (including COVID-19 detection, COPD screening, and smoking status).

Performance:
- Mean AUROC: TRIAGE (Adaptive) achieved a mean AUROC of 0.744.
- Comparison: It outperformed prior zero-shot baselines (CLAP and AcuLa) by substantial margins (e.g., +0.175 over CLAP on average).
- Supervised Baselines: It matched or exceeded supervised linear probing baselines (which require task-specific training) on multiple tasks, despite using zero task-specific training data.
Efficiency vs. Accuracy:
- Tier Distribution: Approximately 46% of samples were resolved at Tier-L (lowest cost), 35% at Tier-M, and only 19% required Tier-H.
- Targeted Gains: The adaptive routing concentrated gains where they mattered most.
  - High-confidence cases (Tier-L) saw no loss in accuracy compared to running them through the full pipeline.
  - Ambiguous cases escalated to Tier-H saw a 19% relative improvement in AUROC compared to a baseline Tier-L prediction.
Ablation Studies:
- Descriptor Robustness: Performance degraded gracefully when clinical descriptors were masked, validating the need for the escalation to Tier-H when evidence is weak.
- Retrieval Depth: Increasing the number of retrieved documents for the LLM (from 1 to 3) yielded significant gains, with diminishing returns beyond 3–5 documents.

5. Significance

Clinical Viability: TRIAGE offers a practical path to robust medical AI in settings where labeled data is unavailable or regulatory approval for fine-tuning is too slow. It allows systems to "think harder" only when necessary.
Resource Optimization: By resolving nearly half of all cases at the cheapest tier, the framework makes high-accuracy zero-shot inference economically feasible for large-scale screening.
Paradigm Shift: The paper shifts the focus from solely scaling model size or training data (training-time scaling) to scaling inference compute (test-time scaling) as a viable strategy for improving robustness in safety-critical domains.

In summary, TRIAGE proves that a structured, confidence-aware inference pipeline can unlock strong zero-shot performance in medical audio, providing a scalable, interpretable, and cost-effective alternative to traditional supervised learning.