File size: 1,838 Bytes
fe55bfe
 
 
 
d698733
 
 
 
 
fe55bfe
 
d698733
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
---
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: **[[email protected]](mailto:[email protected])**.