---
license: apache-2.0
language:
  - es
  - sk
tags:
  - audio
  - parkinsons
  - speech
  - health
  - classification
  - contrastive-learning
  - attention
  - adain
  - wavelets
  - self-supervised
datasets:
  - ewa-db
  - pc-gita
model-index:
  - name: BDHPD
    results:
      - task:
          type: audio-classification
          name: Parkinson's Disease Detection
        dataset:
          name: EWA-DB (Slovak)
          type: ewa-db
        metrics:
          - type: f1
            value: 69.03
          - type: accuracy
            value: 84.72
          - type: sensitivity
            value: 56.52
          - type: specificity
            value: 88.56
      - task:
          type: audio-classification
          name: Parkinson's Disease Detection
        dataset:
          name: PC-GITA (Spanish)
          type: pc-gita
        metrics:
          - type: f1
            value: 90.83
          - type: accuracy
            value: 90.83
          - type: sensitivity
            value: 93.33
          - type: specificity
            value: 88.33
---

# BDHPD: Bilingual Dual-Head Architecture for Parkinson's Disease Detection from Speech

This model implements **BDHPD**, a deep neural network designed to detect Parkinson's Disease (PD) from speech signals, with bilingual support for Slovak and Spanish datasets.

## Model Description

BDHPD combines several modern audio processing techniques:
- **Self-supervised learning (SSL)** with models like `microsoft/wavlm-base`
- **Wavelet-based spectrogram features**
- **Adaptive Instance Normalization (AdaIN)** for domain adaptation
- **Convolutional Bottleneck Layers** for feature recalibration
- **Dual-head classification architecture** to handle different speech types (e.g., diadochokinetic and continuous)
- **Contrastive learning** for embedding space refinement
- **Attention pooling** for better sequence summarization

The architecture supports bilingual inputs and has been evaluated on **EWA-DB** (Slovak) and **PC-GITA** (Spanish).

## Intended Use

- **Research** in pathological speech detection
- **Benchmarking** bilingual speech-based PD detection models
- **Development** of real-world diagnostic support tools in healthcare

## Training

Training was performed using:
- AdamW optimizer
- Linear learning rate scheduling with warmup
- Binary cross-entropy loss for classification
- Contrastive loss via `pytorch-metric-learning`
- 20 epochs with early stopping
- Balanced batch sampling for both datasets

## How to Use

You can find all information on the GitHub repository: [BDHPD GitHub](https://github.com/MorenoLaQuatra/BDHPD)

## Datasets

- [**EWA-DB**](https://zenodo.org/records/10952480): Slovak pathological and healthy speech
- [**PC-GITA**](https://aclanthology.org/L14-1549/): Spanish pathological speech

## Limitations

- The model is only trained on Slovak and Spanish speakers; cross-lingual generalization outside these languages is untested.
- Sensitive to audio quality-ensure audio is preprocessed with proper VAD and dereverberation.
- Should not be used as a standalone diagnostic tool.

## Citation

If you use this model or find useful this research work, please cite the following paper:

```bibtex
@inproceedings{laquatra2025bilingual,
  title={Bilingual Dual-Head Deep Model for Parkinson's Disease Detection from Speech},
  author={La Quatra, Moreno and Orozco-Arroyave, Juan Rafael and Siniscalchi, Marco Sabato},
  booktitle={ICASSP 2025 - IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2025}
}
```