This repository contains the official implementation and comparative analysis presented in our paper: "Efficiency Over Complexity: Optimal Neural Architectures for Near-Term Forecasting of Parkinsonian Freezing of Gait".
Freezing of Gait (FOG) is a debilitating Parkinson's symptom linked to falls and loss of independence. While deep learning (DL) has improved detection, near-term forecasting—predicting an onset before it happens—remains a challenge.
We benchmark eight different architectures (CNN, LSTM, Transformer, and Hybrids) using a unified preprocessing pipeline and a rigorous Leave-One-Subject-Out Cross-Validation (LOSO-CV) protocol. Our findings reveal that lightweight 1D-CNNs often outperform more complex, high-capacity models in generalizing to unseen patients.
We utilize high-fidelity gait dynamics from 16 individuals, captured via synchronized 3D acceleration and angular velocity sensors (200 Hz) mounted on the feet.
(Consider adding an image of the sensor placement here: )
- Input: 4-second sliding window of IMU data.
- Horizon: 2-second future prediction window.
- Labeling: "Any-event" strategy—if a freeze occurs within the 2s horizon, the input is labeled as Impending FOG.
(Consider adding an image of the LOSO-CV process here: )
We organized our models into three blocks to isolate the benefits of spatial vs. temporal modeling:
| Category | Architectures | Key Characteristics |
|---|---|---|
| Simple/Baselines | FD, LSTM (64), CNN (64) | Fundamental single-modality processing. |
| Spatiotemporal Hybrids | CNN-LSTM, LSTM-CNN | Synergy between feature extraction and temporal modeling. |
| High-Capacity | Deeper CNN, Transformer, LSTM-TCN | Investigating if increased complexity improves generalization. |
Our study highlights a significant "Generalization Gap" between standard random splits and the more realistic LOSO-CV protocol.
The CNN-64 and Deeper CNN (128) emerged as the most robust, achieving the best balance for clinical use.
| Architecture | Recall (Sensitivity) | F1-Score |
|---|---|---|
| CNN (64) | 0.9159 | 0.8415 |
| Deeper CNN (128) | 0.8920 | 0.8415 |
| LSTM-CNN | 0.8433 | 0.8183 |
| Transformer | 0.8176 | 0.7870 |
| FD (Baseline) | 0.0588 | 0.1110 |
The Takeaway: For short-horizon forecasting (4s windows), 1D convolutions capture the essential local pre-freeze kinematics better than heavy sequential models, which are more prone to overfitting subject-specific noise.
- Python 3.10+
- PyTorch / TensorFlow (update based on your framework)
- NumPy, Pandas, Scikit-learn
- Clone the repo: ```bash git clone https://github.com/ethan527w/FOG_Algorithm_Comparison_LOSOCV.git