LNN-FoGNet

LNN-FoGNet – Exploring Liquid Neural Networks for Real-Time Freezing-of-Gait Detection

✨ Project summary

Freezing-of-Gait (FoG) is one of the most disabling symptoms of Parkinson’s disease.
LNN-FoGNet investigates whether Liquid Neural Networks (LNNs)—recurrent models whose neurons adapt their own time-constants—can match or exceed the accuracy of both LSTM and continuous-time RNN (CTRNN) while remaining smaller, faster, and more energy-efficient for round-the-clock wearables.

• Dataset: Daphnet FoG
• Models compared: Liquid Neural Network (LNN / LTC) | Long Short-Term Memory (LSTM) | Continuous-Time RNN (CTRNN).
• Key results:
  • Mean F1 ≈ 0.95 on 5-fold subject-wise Cross-Validation (CV).
  • LNN converges in ≈½ the epochs and 1/10th the training time of LSTM.
  • Inference latency per step is tens-of-times faster than LSTM on the same hardware.

---

🗂️ Repository layout

├── fog.py # Main training / evaluation script

├── ltc_model.py # LNN / Liquid-Time-Constant cell

├── ctrnn_model.py # CTRNN cell and helpers

├── vis.py # Training & ROC/PR visualisations

├── vis_fog_events.py # Optional: example raw‐signal plots

├── fog_data.zip # ⇣ Unzip to ./fog_data/ before training

└── results/ # Created automatically

├── fog/ # Metrics, pickles, CSVs

└── figures/ # Plots generated by vis.py

---

📋 Requirements

Package	Tested version
Python	3.9.13
TensorFlow / Keras	2.18 / 3.6
numpy, scikit-learn, matplotlib	recent
Optional	seaborn, tqdm

GPU: Any recent NVIDIA card with CUDA 11+ gives a big speed-up, but the code also runs on CPU.

Quick install

git clone https://github.com/Jonadler1/LNN-FoGNet.git
cd LNN-FoGNet

# Optional: create an isolated env
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# Install core deps
pip install "tensorflow~=2.18" keras==3.6 numpy scikit-learn matplotlib seaborn tqdm

---

📦 Dataset

Inside the repo root: unzip fog_data.zip -d fog_data

fog.py expects ./fog_data/ to contain the raw .txt files exactly as provided by Daphnet.

---

🚀 Training & evaluation

The fog.py driver covers the entire pipeline: normalization → windowing (+ micro-segmentation) → k-fold subject-wise CV → metrics aggregation.

For example, to train an LNN with 32 hidden units for 20 epochs and 5-fold CV, use the following command in the terminal:

python fog.py --model ltc --size 32 --epochs 20 --k 5

Other options:

--model lstm | ctrnn | ltc

--size hidden units (default 32)

--epochs max epochs (default 50)

--k CV folds (default 5)

Outputs (per fold and averaged) are written to results/fog/.

---

📊 Visualisation

Generate publication-quality plots once training is complete:

python vis.py

Those figures are saved under results/figures/.

vis_fog_events.py optionally displays raw sensor traces of FoG vs non-FoG windows.

---

📄 License

Released under the MIT License.

---