---
tags:
- model_hub_mixin
- pytorch_model_hub_mixin
metrics:
- F1
- Precision
- Recall
- ROC-AUC
---

# ARespiratory audio classification model

This model classifies respiratory audio recordings from the [ICBHI 2017 Challenge](https://bhichallenge.med.auth.gr/ICBHI_2017_Challenge) dataset into **crackles**, **wheezes**, **both**, or **none** (multi-label classification). It utilizes the [AST encoder](https://huggingface.co/MIT/ast-finetuned-audioset-14-14-0.443) (`MIT/ast-finetuned-audioset-14-14-0.443`) with a lightweight classification head.

The model has been pushed to the Hub using the [PyTorchModelHubMixin](https://huggingface.co/docs/huggingface_hub/package_reference/mixins#huggingface_hub.PyTorchModelHubMixin) integration.

---

## **Dataset**

- **Source:** [ICBHI 2017 Challenge](https://bhichallenge.med.auth.gr/ICBHI_2017_Challenge)

---

## **Performance metrics**

| **Label**    | **F1**  | **Precision** | **Recall** | **AUC**   |
|--------------|---------|---------------|------------|-----------|
| **Crackle**  | 0.6756  | 0.6147        | 0.7500     | 0.7033    |
| **Wheeze**   | 0.4853  | 0.6565        | 0.3849     | 0.8031    |
| **Macro Avg**| 0.5805  | 0.6356        | 0.5674     | 0.7532    |

---

## **Usage**

Run inference using [`s05_inference.py`](https://github.com/fabiocat93/icbhi_2017_challenge/blob/main/src/code/s05_inference.py).  
Ensure you install the necessary dependencies. For setup instructions, please see the [documentation](https://github.com/fabiocat93/icbhi_2017_challenge).

---

## **Notes**

For additional details, check the [documentation notes](https://github.com/fabiocat93/icbhi_2017_challenge/edit/main/docs/notes.md).

---

## **Contact**

For any questions or further information, feel free to reach out via email: **[fabiocat@mit.edu](mailto:fabiocat@mit.edu)**.