README.md

---
language: en
tags:
  - audio
  - health
  - cough
  - tuberculosis
  - hear
  - medgemma
  - medical
license: apache-2.0
---

# HeAR-TB Domain-Aware Dual Heads

**Domain-Aware Adaptation of Google HeAR for TB Cough Screening**

This model adapts Google's **HeAR** (Health Acoustic Representations) foundation model for **tuberculosis cough screening** using a novel domain-aware dual-head approach.

## Model Description

- **Base Model**: Google HeAR (frozen)
- **Method**: Domain-Aware Dual Heads
  - One XGBoost head specialized on **Passive (natural) coughs**
  - One XGBoost head specialized on **Forced (voluntary) coughs**
  - Final patient score = average of both domain predictions (patient-level aggregation)
- **Embedding Weighting**: Embedding norm used as quality weight per cough
- **Dataset**: Nairobi TBscreen (Sharma et al., Science Advances 2024)
  - 27,343 clean coughs (Passive + Forced)

## Performance (Patient-Level, Repeated 5-Fold CV)

| Metric                  | Value                  |
|-------------------------|------------------------|
| Patient AUC             | 0.7476 ± 0.0932        |
| Sensitivity             | ~0.78–0.83 (tunable)   |
| Accuracy                | 0.728 ± 0.060          |

## Intended Use

- Research and demonstration of acoustic TB screening using foundation models
- Part of the **MedGemma Impact Challenge** (HAI-DEF model usage)
- Educational / exploratory use on cough audio

## Limitations & Important Disclaimer

**This is NOT a medical device or diagnostic tool.**

- Trained on a small number of patients (123)
- Performance is research-level only
- Must not be used for clinical decision-making
- Requires proper clinical validation before any real-world use
- May contain biases from the training data (hospital population in Nairobi)

**Always consult a qualified healthcare professional for medical advice.**

## CITE
@misc{hear-tb-2026,
  author = Sachiv.C,
  title = {HeAR-TB: Domain-Aware Dual Heads for Tuberculosis Cough Screening},
  year = {2026},
  dataset =https://zenodo.org/records/10431329,
  howpublished = (https://huggingface.co/sach3v/Domain_aware_dual_head_HEar),
  note = {Entry for MedGemma Impact Challenge}
}

## How to Use

```python
import joblib
import numpy as np
import librosa
from huggingface_hub import hf_hub_download

# Download and load
package = joblib.load(hf_hub_download("sach3v/hear-tb-domain-aware-dualheads", "hear_tb_prize_domain_aware.joblib"))

model_p = package["model_p"]
scaler_p = package["scaler_p"]
model_f = package["model_f"]
scaler_f = package["scaler_f"]

def predict_cough(audio_path):
    y, sr = librosa.load(audio_path, sr=16000)
    # ... (pad to 2s and get HeAR embedding)
    emb = get_hear_embedding(y)   # your HeAR extraction function
    
    p_passive = model_p.predict_proba(scaler_p.transform(emb))[0,1]
    p_forced = model_f.predict_proba(scaler_f.transform(emb))[0,1]
    
    return (p_passive + p_forced) / 2.0