Update README.md

---
license: mit
language:
- id
- en
tags:
- audio-classification
- heart-sound
- medical
- cnn-lstm
- health
pipeline_tag: audio-classification
---

# SIRAJA Heart Sound Classifier

**CNN-LSTM model for heart sound classification** · 98.3% accuracy

Classifies heart sounds into three categories: Normal, Murmur, or Artifact.

---

## 🎯 Model Description

SIRAJA is a hybrid CNN-LSTM deep learning model trained to classify heart sounds from auscultation recordings. The model analyzes MFCC features extracted from audio signals to detect:

- **Normal** — Regular heart rhythm with clear S1/S2 sounds
- **Murmur** — Turbulent blood flow indicating potential valve issues
- **Artifact** — Noise/interference in the recording

---

## 📊 Performance

| Metric | Score |
|--------|-------|
| Test Accuracy | **98.30%** |
| Validation Accuracy | **98.86%** |

**Per-Class Results:**

| Class | Precision | Recall | F1-Score |
|-------|-----------|--------|----------|
| Artifact | 100.00% | 100.00% | 100.00% |
| Murmur | 97.37% | 94.87% | 96.10% |
| Normal | 98.41% | 99.20% | 98.80% |

---

## 🏗️ Architecture

```
Audio → MFCC (52 coefficients) → CNN → LSTM → Classification
```

**Specifications:**
- Input: 10-second audio clips
- Sample Rate: 22050 Hz
- Features: 52 MFCC coefficients
- Parameters: 14.1M
- Framework: TensorFlow/Keras

---

## 📦 Usage

### Installation

```bash
pip install tensorflow librosa numpy
```

### Inference

```python
import tensorflow as tf
import librosa
import numpy as np

# Load model
model = tf.keras.models.load_model('heart_sound_classifier.h5')

# Load and preprocess audio
audio, sr = librosa.load('heart_sound.wav', sr=22050, duration=10)

# Extract MFCC features
mfcc = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=52, n_fft=2048, hop_length=512)
mfcc_mean = np.mean(mfcc.T, axis=0).reshape(1, -1, 1)

# Predict
prediction = model.predict(mfcc_mean)
classes = ['Artifact', 'Murmur', 'Normal']
result = classes[np.argmax(prediction)]

print(f"Prediction: {result} ({prediction[0][np.argmax(prediction)]:.2%})")
```

### Web Application

Full-featured Streamlit app available at: [github.com/mdprana/SIRAJA](https://github.com/mdprana/SIRAJA)

```bash
pip install streamlit huggingface_hub
streamlit run app.py
```

---

## 🎵 Training Data

**Dataset:** PASCAL Challenge 2011 Heart Sound Classification

- Source: [Kaggle - Heartbeat Sounds](https://www.kaggle.com/datasets/kinguistics/heartbeat-sounds)
- Samples: 1,755 (after 3x augmentation)
- Split: 80% train / 10% validation / 10% test
- Augmentation: Time stretching (0.8x, 1.2x)

---

## 🔧 Training Details

- **Optimizer:** Adam (lr=0.0001)
- **Loss:** Categorical Crossentropy
- **Class Weighting:** Artifact (4.88x), Murmur (1.51x), Normal (0.47x)
- **Regularization:** Dropout 50%
- **Callbacks:** Early stopping, LR scheduling
- **Epochs:** 118

---

## ⚕️ Medical Disclaimer

**Important:** This model is for screening purposes only and is **not a diagnostic tool**.

- AI predictions are not medical diagnoses
- Always consult qualified healthcare professionals
- Not a replacement for clinical examination
- For research and educational use

---

## 📄 License

MIT License — Free for research and commercial use

---

## 👥 Authors

**Kelompok 3 - Universitas Udayana**

- Made Pranajaya Dibyacita (2208561122)
- I Made Sastra Wiguna (2208561121)
- Intara Pratama Harahap (2208561104)

---

## 🔗 Links

- **GitHub:** [mdprana/SIRAJA](https://github.com/mdprana/SIRAJA)
- **Dataset:** [Kaggle - Heartbeat Sounds](https://www.kaggle.com/datasets/kinguistics/heartbeat-sounds)
- **Paper:** PASCAL Challenge 2011

---

## 📊 Model Files

- `heart_sound_classifier.h5` (161.76 MB) — Trained model weights
- `model_config.json` — Model configuration

---

## 🙏 Citation

If you use this model in your research, please cite:

```bibtex
@misc {siraja2025,
title={SIRAJA: Heart Sound Classification using CNN-LSTM},
author={Dibyacita, Made Pranajaya and Wiguna, I Made Sastra and Harahap, Intara Pratama},
year={2025},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/mdprana/siraja}}
}
```

README.md

@@ -1,3 +1,18 @@
----
-license: mit
----
+---
+license: mit
+language:
+- id
+- en
+metrics:
+- accuracy
+base_model:
+- mdprana/siraja
+pipeline_tag: audio-classification
+library_name: keras
+tags:
+- audio-classification
+- heart-sound
+- medical
+- cnn-lstm
+- health
+---