RA-QA: A Benchmarking System for Respiratory Audio Question Answering Under Real-World Heterogeneity

RA-QA is a large-scale benchmark for Respiratory Audio Question Answering which covers real-world heterogeneity across modalities, devices, and question types

---

What is RA-QA?

RA-QA (Respiratory-Audio Question Answering) is a benchmarking framework and dataset release that turns heterogeneous public respiratory-audio datasets into a unified multimodal QA collection.

RA-QA provides:

• a standardized data generation pipeline ,
• format-diverse QA files (open-ended / multiple-choice / single-verify),
• a unified evaluation protocol for both discriminative (categorical) and regression (continuous) targets.

This benchmark is designed to stress-test models under real-world heterogeneity: different datasets, recording conditions/devices, audio modalities (e.g., cough/breath/speech/auscultation), and question intents.

---

What’s in this repository?

This repo is organized around two deliverables:

1. Standardized metadata
   Per-dataset harmonized tables exposing canonical attributes and provenance.

2. QA files
   Generated QA pairs in multiple formats (OE/MC/SV), linked to recordings and subjects, structured in a JSON format style.

QA JSON format (example)

Each split is stored as a single JSON file containing a list of QA records.
Below is an example entry for a single-verify question:

[
  {
    "id": 0,
    "question_id": 1,
    "version": 1,
    "patient_id": "01GtHP1FUbXKdWEUwApFdusuO773",
    "split": "train",
    "question_type": "single-verify",
    "question": "Does the patient suffer from asthma?",
    "answers_list": [
      "Yes, the patient suffers from asthma.",
      "Yes, the patient suffers from asthma.",
    ],
    "audio_filename_list": [
      "20220224_01GtHP1FUbXKdWEUwApFdusuO773_breathing-shallow.wav",
      "20220224_01GtHP1FUbXKdWEUwApFdusuO773_cough-heavy.wav",
    ],
    "attribute": "asthma"
  }
]

QA JSON fields

Each QA file is a JSON array of records, where each record has the following fields:

• id (int): Unique record index within the file.
• question_id (int): Identifier of the underlying question template / generation rule.
• version (int): Version tag for the generation pipeline/templates used to create this record.
• patient_id (str): Patient identifier used to group multiple recordings belonging to the same subject.
• split (str): Dataset split for this record. One of {train, val, test}.
• question_type (str): Question format. One of {open-ended, multiple-choice, single-verify}.
• question (str): Natural-language question text.
• answers_list (list[str]): List of natural-language reference answers (multiple references, one per audio recording).
• audio_filename_list (list[str]): List of audio filenames associated with the patient_id (e.g., cough/breath/speech recordings or longitudinal data).
• attribute (str): Canonical target attribute queried by the question (e.g., age, asthma, device, severity).