Clean envelopes ➜ whole-cycle nonparametric SPM ➜ tv-VAR IRFs (EMG↔KIN) ➜ Proximal vs Distal EMG directionality
High accuracy + full interpretability: from black-box AI to muscle–phase therapeutic maps.
This repository supports the upcoming manuscript “From EMG to Insight: Explainable Deep Learning Identifies Muscle Drivers of Stiff-Knee Gait in Hereditary Spastic Paraparesis”, which will be submitted shortly to a peer-reviewed journal.
Large surface-EMG datasets in neurological gait disorders hold untapped clinical information — but which muscle, when, and how much?
We developed a fully explainable deep learning pipeline to map muscle×phase determinants of stiff-knee gait in Hereditary Spastic Paraparesis (HSP).
- Subjects: 43 total (26 HSP, 17 HS)
- Clinical subgroups: HSP1 (severe), HSP2 (moderate), HSP3 (mild)
- sEMG: 16 bilateral muscles (32 channels @1000 Hz)
- 3D Kinematics: 100 Hz
- Gait cycles: 3 valid cycles per subject
- Normalization: 0–100% gait cycle; 7 gait sub-phases
# 1. Install dependencies
pip install numpy pandas matplotlib scipy statsmodels scikit-learn spm1d joblib openpyxl xlsxwriter
# 2. Place your raw CSV (Italian field names supported)
# sEMG data HSP/Timeseries completo.csv
# 3. Run the script (Jupyter or CLI)
python EMG_HSP_Explainability.py
Outputs appear automatically in the `outputs/` folder:
- `Parameters4.xlsx` — preprocessing features
- `spm_clusters_corrected.xlsx` — SPM clusters (Holm/FDR)
- `results_irf.xlsx` — EMG↔KIN (XCorr/HAC/Granger/FDR + IRFs)
- `emg2emg_DI_heatmap.png` + per-phase bar charts| Section | Goal | Key Outputs |
|---|---|---|
| 1. Pre-processing | Clean envelopes, normalize EMG, resample 0–100% gait cycle | Parameters4.xlsx, overview plots |
| 2. Whole-cycle SPM (nonparam) | Compare full time-series between groups (Holm+FDR) | spm_clusters_corrected.xlsx, SPM plots with q annotations |
| 3. EMG↔Kinematics | tv-VAR IRFs, Granger causality, HAC OLS | results_irf.xlsx, controller flowmaps |
| 4. Proximal↔Distal EMG | Directionality Index (DI), phase-specific drivers | emg2emg_DI_heatmap.png, per-phase bar charts |
| Task | Accuracy | AUC / F1 |
|---|---|---|
| Binary (HS vs HSP) | 91.2% | AUC 0.95, F1 0.94 |
| Multiclass (HS, HSP1–3) | 91.4% | F1 ≥0.86 in all 28 cells |
- Quadriceps dominance: RF, VL, VM emerge as main drivers across phases.
- MER metric: High MER → stereotyped activation; Low MER → adaptive complexity.
- Proximal–distal interactions: Controls show alternation; HSP1 shows monochromatic blocks (loss of alternation).
- Clinical bridge: From deep learning to interpretable, muscle–phase therapeutic maps.
If you use this code or dataset, please cite:
Trabassi D. et al. From EMG to Insight: Explainable Deep Learning Identifies Muscle Drivers of Stiff-Knee Gait in Hereditary Spastic Paraparesis. SIAMOC 2025.
├── EMG_HSP_Explainability.py # Main pipeline ├── outputs/ # Generated figures and tables ├── docs/ # Images (pipeline, figures) ├── data/ # (Optional) input CSVs (add to .gitignore) └── README.md
This repository does not contain raw patient data.
Placeholders / synthetic examples are provided for reproducibility.
Please respect privacy regulations (GDPR) when using your own data.
- Dante Trabassi — main pipeline, preprocessing, complementary code modules
- Stefano Filippo Castiglia — explainability pipeline, analyses
- Prof. Mariano Serrao (supervision)
- Ing. Alberto Ranavolo
- Ing. Irene Gennarelli
- Ing. Tiwana Varrecchia