When Your BCI Decoder Drifts: Handling Non-Stationarity with NimbusSTG

Brain-computer interfaces make a quiet assumption that most engineers don't notice until it bites them: the signal you trained on will look like the signal you see at inference time. For a 5-minute calibration session followed by a 6-minute task, that assumption holds. For a 45-minute clinical trial, a day-long assistive device session, or a multi-hour research recording, it almost never does.

This is the non-stationarity problem — and it is one of the most underappreciated failure modes in BCI engineering. In this post, we'll unpack why it happens, why classical decoders are structurally ill-equipped to handle it, and how NimbusSTG — the Bayesian Structural Time Series model in NimbusSDK — is purpose-built to adapt as the signal evolves.

Why EEG Signals Drift (and Why You Should Care)

EEG non-stationarity has multiple, simultaneous causes:

• Electrode impedance drift — gel dries, contact resistance rises, signal amplitude and noise floor shift
• Cognitive fatigue — sustained attention tasks modulate alpha and theta power over tens of minutes
• Adaptation effects — the brain itself changes strategy mid-session as the user becomes more familiar with the BCI paradigm
• Muscle artifact accumulation — posture changes introduce slow-wave contamination that preprocessing filters don't fully remove

For a static classifier — a fixed LDA boundary trained on the first 60 trials — any of these shifts pushes the true decision boundary away from the trained one. The decoder doesn't know it's wrong. It keeps outputting predictions with high confidence while accuracy quietly degrades. By minute 40 of a session, a decoder that opened at 85% accuracy can be sitting below chance on some users.

This isn't a data quality problem. It's a modeling problem: classical BCI decoders have no internal representation of time.

The Bayesian Structural Time Series Approach

Bayesian Structural Time Series (BSTS) is a class of state-space models that decompose a time-varying signal into interpretable components — trend, seasonal patterns, regression effects — and propagate uncertainty through those components over time.

The core idea is elegantly suited to the BCI drift problem. Instead of treating your decoder's internal parameters as fixed after calibration, NimbusSTG represents them as latent state variables that evolve according to a learned dynamics model. At each new trial, the model does two things:

1. Predicts the current state from the previous state plus a transition model (the "prior step")
2. Updates that prediction using the new observation via Extended Kalman Filter (EKF)-style inference

This is the same update loop that makes Active Inference tick — a generative model making predictions, then revising its beliefs when reality differs. NimbusSTG simply applies this loop at the level of decoder parameters rather than raw neural signals.

The practical outcome: the decoder's effective decision boundary tracks the true boundary as the signal drifts. No periodic retraining. No manual recalibration triggers. No session interruptions.

When to Use NimbusSTG vs. Other Nimbus Models

NimbusSDK provides four Bayesian classifiers. Choosing between them is a function of your session structure and signal characteristics:

The heuristic is straightforward: if your session exceeds 30 minutes or you've observed accuracy degradation over time in user testing, reach for NimbusSTG. For shorter sessions with stable recording conditions, the simpler static models will train faster and generalize well.

NimbusSTG is also the right call for online BCIs — paradigms where the user interacts with the system continuously rather than in discrete calibration + test phases. In those settings, there's no natural point to retrain, so adaptive inference is the only viable path.

Setting Up NimbusSTG in Nimbus Studio

Nimbus Studio makes it straightforward to drop NimbusSTG into an existing pipeline without touching your preprocessing or feature extraction nodes. Here's the setup pattern:

1. Build your standard pipeline

Start with the usual chain: EEG source → bandpass filter → epoch → spatial filter (CSP or FBCSP) → feature vector. This part is identical to any other Nimbus Studio pipeline.

2. Swap in NimbusSTG at the classification node

In the model selection panel, choose NimbusSTG instead of NimbusLDA or NimbusQDA. The node exposes two key configuration parameters:

• state_noise_scale — controls how aggressively the model allows parameters to drift between trials. Higher values track fast changes; lower values smooth over noise.
• obs_noise_scale — the assumed observation noise on each new trial. Tune this based on your expected signal quality.

For most motor imagery or SSVEP pipelines, the defaults are a reasonable starting point. The model is robust to moderate misspecification of these values.

3. Enable the live confidence overlay

NimbusSTG outputs a full posterior over class probabilities at each trial, not just a point prediction. Nimbus Studio's real-time dashboard can display this as a confidence band alongside the prediction trace — useful for spotting adaptation dynamics during a session and for setting threshold-based control policies in assistive applications.

4. Export and deploy

Once validated offline on a recorded dataset (Nimbus Studio ships with 40+ public BCI datasets, several of which include long multi-hour sessions ideal for testing drift correction), use the one-click deploy to stream from hardware. The same NimbusSTG pipeline that trained offline runs live inference with state propagation intact — no code rewrites required.

A Note on Interpretability and Clinical Use

One underappreciated advantage of the BSTS formulation is that the latent state trajectory is interpretable. You can inspect how the model's internal representation of each class evolved over a session — which directions drifted, how quickly, and whether the adaptation stabilized or continued monotonically.

For clinical and assistive BCI applications, this matters beyond engineering performance. Regulatory bodies and clinical teams increasingly require explainability — a record of why the system made each decision and how its behavior changed over time. NimbusSTG's probabilistic state history provides exactly that audit trail, distinguishing it from black-box adaptive methods.

This aligns with Nimbus BCI's broader positioning: a "white box" engine where every inference step is traceable, uncertainty-aware, and grounded in a principled generative model.

Conclusion

Non-stationarity is not an edge case — it is the default condition for any BCI deployed beyond a controlled short-session lab setting. Classical static decoders ignore it by design, and users pay the price in degrading accuracy.

NimbusSTG offers a principled, practical alternative: a Bayesian state-space model that treats decoder parameters as living quantities that evolve with the signal, updating beliefs at every trial via EKF-style inference. The result is a decoder that stays calibrated across long sessions without interruption, produces confidence-scored outputs, and generates an interpretable state history suitable for clinical audit.

If you're building a real-world assistive device, running extended research recordings, or simply tired of watching accuracy curves slope downward over time, NimbusSTG is the model to reach for — and Nimbus Studio makes the integration a matter of minutes, not weeks.

Explore NimbusSTG and the full NimbusSDK model suite in Nimbus Studio. Available in both Python and Julia SDKs.