Balancing Performance and Stability in Motor Imagery EEG Decoding: A Linear-First Approach

Abstract

Motor Imagery (MI) EEG decoding lacks a principled way to decide when a complex nonlinear classifier is warranted over strong, data-efficient linear baselines. While literature often chases marginal accuracy gains, it rarely balances these against computational burden or stability under neurophysiological non-stationarity.

This work operationalizes a “Linear-First” rule using statistical inference and drift-aware neuroscience. We argue that a complex alternative should only be chosen if it is statistically superior to a simpler baseline; otherwise, non-inferiority defaults to the linear option for the sake of interpretability and efficiency.

Methodological Framework

1. The Linear-First Decision Rule

We implement a Paired Non-Inferiority Test (TOST) within a pre-registered, nested cross-validation framework.

• Protocol: Every learnable step (scaling, subspaces, shrinkage) is fit strictly inside the inner loop to prevent data leakage.
• Logic: We escalate to nonlinear models (e.g., Deep Learning or Riemannian MDM) only when they conquer a statistically robust niche defined by a pre-specified non-inferiority margin.

2. Computational Pareto Fronts

To ensure fair comparison, we construct Pareto fronts in two dimensions to adjudicate near-ties between models:

• Accuracy vs. Latency: Measured using a single-core inference protocol.
• Accuracy vs. FLOPs: Auditing budget alignment to ensure gains are not artifacts of increased search effort.

3. Neurophysiological Drift Analysis

We quantify within-session non-stationarity to connect decoding performance to sensorimotor physiology:

• Spectral Drift: Phase-wise changes in $\mu/\beta$ band power (ERD/ERS carrier).
• Geometric Drift: Affine-invariant distances on the Symmetric Positive Definite (SPD) manifold between phase-wise covariance means.

Key Insights

We utilize mixed-effects models to analyze the interaction between signal drift and model family. Our findings suggest that pipelines that remain stable are those most robust to the inherent dynamics of the sensorimotor code. This confirms that for clinical applications like closed-loop rehabilitation, stability and resource predictability are as critical as point accuracy.

For inquiries regarding this research, please contact the author directly.