Web App: ShiftDx
Published:
Application: ShiftDx Title: Drift Diagnostics for Multi-Session MI-EEG BCI Authors: Yiming Shen, David Degras
Overview
ShiftDx is an interactive Streamlit dashboard for diagnosing cross-session nonstationarity in motor-imagery EEG. It supports the paper Drift Diagnostics, Adaptation, and Recalibration in Multi-Session Motor-Imagery EEG by linking session-level distribution shift to decoding loss, domain-adaptation benefit, retraining gap, and online drift triggers.
The dashboard follows a fixed-reference deploy-and-monitor protocol: session 0 is treated as calibration, while later sessions are monitored under No DA, DA, and Retrain strategies. This makes the app focus on the practical BCI question of what happens after a single calibration session is deployed unchanged.
What It Shows
- Dataset overview and per-subject drift trajectories
- Five drift metrics: MMD, Energy, Wasserstein, Mahalanobis, and Euclidean distance
- Four claim-explorer pages for drift-loss, DA decomposition, retraining gap, and feature robustness
- Subject-level deep dives across No DA, DA, and Retrain strategies
- DA Lab pages for live synthetic shift experiments, multi-metric comparison, method sweep, and Page-Hinkley drift detection
Experiment Scope
| Component | Coverage |
|---|---|
| Feature families | CSP, log-variance, tangent-space covariance features |
| Classifiers | LDA, linear SVM, radial SVM |
| DA methods | 10 DA4BCI methods including SA, TCA, CORAL, PT, ART, OT, and M3D |
| Protocol | Fixed-reference within-subject cross-session monitoring |
Tech Stack
Python, Streamlit, Plotly, pandas, NumPy, SciPy, scikit-learn, statsmodels, POT, DA4BCI-Python, CrossPython, MOABB.