GeHirNet

GeHirNet turns a sustained /a/ vowel recording into a segment-level prediction for healthy voice or one of six voice-related conditions: ALS, COVID-19, dysphonia, laryngitis, Parkinson's disease, and vocal cord paresis.

This repository is useful if you want to:

• run the pretrained model on a WAV recording or prepared Mel spectrogram;
• evaluate the released weights on a labeled dataset;
• retrain or adapt the model for voice pathology research;
• reproduce the paper's preprocessing, hierarchical training, and augmentation experiments.

Ready-to-use weights are included in checkpoints/hierarchical/, so inference does not require training first. The model uses a sex-aware hierarchy of three ResNet-50 classifiers to account for physiological voice differences while producing the final health or pathology label.

GeHirNet is intended for research on sustained vowels and is not a clinical diagnostic tool.

Install

Python 3.12 is the locally validated version.

python -m venv .venv
source .venv/bin/activate
python -m pip install -e .

Use the model

Recommended:

gehirnet predict path/to/vowel.wav \
  --model-dir checkpoints/hierarchical \
  --output outputs/prediction.json

Using the three .pth files directly:

gehirnet predict path/to/vowel.wav \
  --pd checkpoints/hierarchical/pd.pth \
  --mp checkpoints/hierarchical/mp.pth \
  --fp checkpoints/hierarchical/fp.pth

See docs/USAGE.md for WAV/Mel requirements, Python usage, outputs, and evaluation.

Reproduce the research workflow

See docs/REPRODUCE.md for data preparation, training, evaluation, seeds, augmentation experiments, and known reproduction limits.

Repository layout

Path	Purpose
src/gehirnet/	Reusable preprocessing, training, evaluation, inference, and model-bundle code
checkpoints/hierarchical/	Complete PD/MP/FP model artifact
configs/, scripts/	Training configurations and experiment helpers
data/templates/	Metadata schema examples
experiments/	Original notebooks, preprocessing, augmentation, ablation, and analysis materials
docs/USAGE.md	How to call the pretrained model
docs/REPRODUCE.md	How to reproduce the experimental workflow

Validation

python -m unittest discover -s tests -v

The implementation uses the original class orders and hard routing. Model manifests verify preprocessing metadata, label order, and SHA-256 checksums before loading.

License

Code and the supplied model artifact use Apache-2.0. See LICENSE.txt and NOTICES.txt.

Citation

If you use this repository, please cite:

F. Wu, K. Zhao, E. Fleisch and F. Barata, "GeHirNet: A Sex-Aware Hierarchical Model for Fair and Accurate Voice Pathology Classification," 2026 IEEE 14th International Conference on Healthcare Informatics (ICHI), Minneapolis, MN, USA, 2026, pp. 262–270, doi: 10.1109/ICHI69079.2026.00043.