SPDIM: Source-Free Unsupervised Conditional and Label Shift Adaptation in EEG

This paper proposes SPDIM, a parameter-efficient geometric deep learning framework that leverages information maximization on the symmetric positive definite (SPD) manifold to effectively address source-free unsupervised domain adaptation in EEG data, specifically overcoming the generalization limitations of prior methods when faced with label shifts.

The Gist

Here is an explanation of the paper SPDIM using simple language, everyday analogies, and creative metaphors.

The Big Problem: The "Weather" of Brain Signals

Imagine you are trying to teach a robot to recognize your thoughts using an EEG cap (a helmet with sensors that read brain waves). You train the robot on Monday. On Monday, you are well-rested, drinking coffee, and sitting in a quiet room. The robot learns your brain patterns perfectly.

But then, on Tuesday, you are tired, it's raining outside, and you are in a noisy office. Your brain waves change slightly. The robot, which was trained only on Monday's "weather," gets confused. It thinks your tired brain signals are a different thought entirely.

In the world of Brain-Computer Interfaces (BCI), this is called Distribution Shift. Your brain is non-stationary; it changes based on the day, your mood, your health, and who you are.

The Catch-22:
To fix this, usually, you need to re-train the robot with new data from Tuesday. But in real life, you often don't have labeled data for Tuesday (you don't know what the user is thinking at that exact moment). This is called a Source-Free Unsupervised Domain Adaptation (SFUDA) problem. The robot has to adapt to the new "weather" without a teacher.

The Old Solution: The "One-Size-Fits-All" Map

Previously, scientists used a clever mathematical trick involving Riemannian Geometry. Think of brain data not as a flat map, but as a curved surface (like the Earth).

The old method (called RCT+TSM) tried to align the "average" brain patterns of Monday and Tuesday. It was like taking a map of Monday and stretching it until it looked like Tuesday's map.

• What it did well: It handled changes in how the signal traveled (like if the sensor moved slightly).
• What it failed at: It failed when the balance of thoughts changed.

The "Label Shift" Problem:
Imagine on Monday, you thought "Left" 50% of the time and "Right" 50% of the time. On Tuesday, you were tired and thought "Left" 90% of the time and "Right" only 10%.
The old method tried to force the Tuesday map to look exactly like Monday's map. Because Tuesday was so unbalanced, the old method got confused. It tried to stretch the "Left" signals too much and squish the "Right" signals, making the robot perform worse than if it hadn't tried to adapt at all. It was like trying to fit a square peg into a round hole by hammering the peg until it broke.

The New Solution: SPDIM (The "Smart Compass")

The authors propose a new method called SPDIM. Instead of trying to force the whole map to match, they introduce a "Smart Compass" that adjusts for the specific imbalance of the day.

Here is how it works, step-by-step:

1. The Generative Model (The Recipe)

First, the authors created a mathematical "recipe" to simulate how brain signals are made. They realized that brain signals are a mix of:

• The Source: The actual thought (e.g., "Move Left").
• The Mixer: The body and environment (e.g., muscle tension, electrode placement).
• The Bias: The probability of thinking "Left" vs. "Right" (the Label Shift).

2. The "Over-Correction"

When the old method tried to align the days, it accidentally over-corrected for the "Label Shift." It thought, "Oh, Tuesday has too many 'Left' thoughts, so I need to shift everything to the right to balance it out." But because the nature of the thoughts changed, this shift made the data worse.

3. The Fix: Information Maximization

SPDIM uses a principle called Information Maximization. Imagine you are a detective trying to solve a mystery with very few clues.

• Goal A (Certainty): The detective wants to be very sure about each specific clue (e.g., "This signal definitely means 'Left'").
• Goal B (Diversity): The detective also wants to make sure they aren't guessing "Left" for everything. They want to use all the possible answers available.

SPDIM tweaks a single, special parameter (a "bias knob") for the target day. It turns this knob until the robot is confident in its guesses but also diverse enough to cover all possibilities. It essentially asks the robot: "If you guess 'Left' 90% of the time, are you sure you aren't just ignoring the 'Right' signals? Adjust your internal compass so you can see both clearly."

The Analogy: The Chameleon and the Mirror

• The Old Method (RCT+TSM): Imagine a chameleon trying to match a background. If the background is 90% green and 10% red, the chameleon tries to turn 90% green. But if the lighting changes (the "Label Shift"), the chameleon gets confused and turns the wrong shade of green, making it stand out.
• SPDIM: Imagine the chameleon has a smart mirror. It looks at the background, realizes, "Hey, there's a lot of green today, but I know red exists too." It adjusts its internal settings (the bias parameter) to ensure it can still see and react to the red patches, even if they are rare. It doesn't just copy the background; it adapts its perception of the background.

The Results: Does it Work?

The authors tested this on two things:

1. Simulations: They created fake brain data with known problems. SPDIM fixed the "Label Shift" problems where the old methods failed.
2. Real EEG Data:
   • Motor Imagery: People imagining moving their hands.
   • Sleep Staging: Classifying sleep stages (Deep Sleep, REM, etc.). Sleep naturally has "Label Shifts" (you spend more time in Deep Sleep than in REM).

The Outcome:
SPDIM consistently beat the old methods. In sleep staging, it improved accuracy significantly, especially for patients (who have more irregular sleep patterns). It proved that by using a "Smart Compass" (the bias parameter) and a "Detective's Logic" (Information Maximization), we can make brain-computer interfaces work better across different days and different people without needing to re-train them from scratch.

Summary

• The Problem: Brain signals change, and sometimes the types of thoughts change (Label Shift), confusing old AI models.
• The Old Fix: Tried to force old and new data to look identical, which broke things when the balance of thoughts changed.
• The New Fix (SPDIM): Uses a smart, math-based "compass" to adjust the model's internal bias. It ensures the model stays confident in its guesses while remembering to look for rare thoughts, allowing it to adapt to new "weather" without a teacher.

Technical Summary

Here is a detailed technical summary of the paper "SPDIM: Source-Free Unsupervised Conditional and Label Shift Adaptation in EEG", published as a conference paper at ICLR 2025.

1. Problem Statement

The paper addresses the challenge of Source-Free Unsupervised Domain Adaptation (SFUDA) in Electroencephalography (EEG) applications.

• Context: EEG signals are non-stationary, leading to distribution shifts across different domains (e.g., different recording sessions or subjects).
• The Gap: Traditional SFUDA methods for EEG, particularly those based on Riemannian geometry (operating on Symmetric Positive Definite (SPD) matrices), are state-of-the-art for handling conditional distribution shifts (changes in how features relate to labels). However, they often fail or degrade performance when label shifts (changes in the prior distribution of classes, e.g., sleep stages occurring with different frequencies) are present.
• Goal: Develop a framework that can adapt to target domains without access to source data or target labels, specifically compensating for both conditional shifts and label shifts simultaneously.

2. Methodology

The authors propose SPDIM (SPD Manifold Information Maximization), a geometric deep learning framework. The approach is built upon a novel generative model and a two-stage adaptation strategy.

A. Generative Model

The authors introduce a realistic generative model for EEG data where:

• Observed EEG signals $x$ are linear mixtures of latent sources $z$ via domain-specific forward models $A_j$.
• The covariance of latent sources $E_i$ has a log-linear relationship with the labels $y$.
• Crucially, the model accounts for label shifts by introducing domain-specific class priors $\pi_j$ that affect the latent log-space features.
• Theoretical Insight: The authors prove (Proposition 1 & 2) that standard Riemannian alignment methods (like RCT+TSM) successfully compensate for conditional shifts (scaling/rotation) but fail under label shifts because aligning the Fréchet mean (centering) inadvertently over-corrects for the shifted class priors, hurting generalization.

B. The SPDIM Framework

To remedy the over-correction caused by label shifts, SPDIM introduces a parameter-efficient optimization strategy:

1. Latent Alignment (Step 1):

   • A shared feature extractor $f_\theta$ (a neural network) maps raw EEG data to SPD matrices.
   • Standard Riemannian alignment (Tangent Space Mapping) is applied to align marginal distributions, projecting data to the tangent space of the Fréchet mean.

2. Bias Correction via Information Maximization (Step 2):

   • Instead of retraining the whole network, SPDIM learns a single domain-specific bias parameter $\Phi_j$ constrained to the SPD manifold for each target domain.
   • This parameter is optimized using an Information Maximization (IM) loss, which consists of:
     • Conditional Entropy Minimization ($L_{CEM}$): Encourages the model to make confident predictions for individual samples.
     • Marginal Entropy Maximization ($L_{MEM}$): Encourages the model's predictions across the target domain to match the expected class distribution (preventing collapse to a single class).
   • The bias parameter $\Phi_j$ acts to counteract the "over-correction" introduced by the initial Fréchet mean alignment, effectively compensating for the label shift.

C. Implementation Variants

• SPDIM(bias): Learns a general SPD bias matrix $\Phi_j$ (Eq. 19).
• SPDIM(geodesic): Constrains $\Phi_j$ to lie on the geodesic between the Fréchet mean and the identity matrix, reducing the parameter to a single scalar step-size $\phi_j$ (Eq. 20).

3. Key Contributions

1. Theoretical Analysis: The paper provides a rigorous theoretical proof demonstrating why standard Riemannian statistical alignment methods fail under label shifts in EEG data, identifying the mechanism of "over-correction."
2. Novel Generative Model: Introduction of a log-linear generative model for EEG that explicitly incorporates label shifts and domain-specific forward models, serving as a foundation for theoretical analysis.
3. SPDIM Algorithm: A parameter-efficient SFUDA method that learns a single SPD-manifold-constrained bias parameter per target domain using Information Maximization, avoiding the need for source data or target labels.
4. Comprehensive Evaluation: Extensive validation on both simulated data and real-world public datasets (Motor Imagery and Sleep Staging), demonstrating superiority over existing SFUDA baselines.

4. Experimental Results

The authors evaluated SPDIM against several baselines (RCT, Euclidean Alignment, STMA, SPDDSBN, and classic IM approaches) on two main tasks:

A. Simulations

• Setup: Binary classification with varying degrees of label shift (controlled by the ratio of minority to majority classes) and class separability.
• Result: SPDIM consistently outperformed RCT (the standard Riemannian method). While RCT performance dropped significantly as label shifts increased, SPDIM maintained high balanced accuracy, confirming the theoretical analysis.

B. Real-World EEG Datasets

1. Motor Imagery (BCI):

   • Datasets: BNCI2014001, BNCI2015001, Zhou2016, BNCI2014004.
   • Scenario: Cross-session and cross-subject transfer with artificially induced label shifts (0.2 ratio).
   • Result: SPDIM(bias) achieved the highest balanced accuracy, significantly outperforming SPDDSBN and other IM variants in cross-subject settings.

2. Sleep Staging:

   • Datasets: CAP, Dreem, HMC, ISRUC (426 subjects total).
   • Scenario: Inherent label shifts due to natural sleep variability across subjects.
   • Result: TSMNet+SPDIM(bias) significantly outperformed all baselines (including specialized sleep models like USleep and DeepSleepNet) across patient and healthy groups.
   • Ablation Study: Confirmed that the combination of IM loss and the manifold-constrained bias parameter is critical. Removing the bias parameter or using global alignment without per-domain adaptation resulted in significant performance drops (3-5% or more).

5. Significance and Impact

• Bridging the Gap: SPDIM addresses a critical limitation in current EEG neurotechnology: the inability to generalize to new users or sessions where the distribution of classes (e.g., sleep stages or mental states) is unknown and unbalanced.
• Clinical Relevance: The significant improvement in sleep staging accuracy (approx. 5% margin over the next best method) suggests high potential for clinical deployment, where robust, calibration-free models are essential.
• Efficiency: By learning only a single bias parameter per domain rather than retraining the entire network, SPDIM is computationally efficient and suitable for real-time or resource-constrained applications.
• Theoretical Foundation: The work establishes a new theoretical understanding of how Riemannian geometry interacts with label shifts, guiding future research in geometric deep learning for neuroimaging.

In summary, SPDIM represents a state-of-the-art advancement in EEG domain adaptation, offering a robust, theoretically grounded solution for the challenging scenario of source-free adaptation under both conditional and label distribution shifts.