Speed-enhanced Subdomain Alignment for Long-term Stable Neural Decoding in Brain-computer Interfaces

🔶 Recalibration of Neural Decoders in Brain-Computer Interfaces: The SSAR Framework 🔶

1. 🔍 Background and Motivation

Brain-computer interfaces (BCIs) enable communication between neural activity and external devices by decoding neural signals into intentional commands.

BCIs show great potential in:

• Restoration
• Rehabilitation
• Assistive technologies for individuals with disabilities

💡 Advancements

• Various neural decoding algorithms have achieved impressive accuracy:
  • Population Vector (PV)
  • Kalman Filter
  • KRSRL
  • DyEnsemble
  • EvoEnsemble
• Evolution in neuron population recording enables more complex interactions with the external world

⚠️ Core Challenge: Neural Signal Variability

Despite advancements, BCIs face a persistent challenge:
Neural signal instability over time, caused by:

• Electrode drift
• Neuronal turnover
• Synaptic plasticity

This results in:

• Unstable decoding performance
• Limited long-term usability of BCIs

2. 🛠️ Existing Recalibration Strategies

✅ Supervised Recalibration

• Common approach: daily retraining of decoders
• Limitation: high cost of labeled data acquisition in real-world settings

✅ Semi-Supervised Learning (SSL)

• Introduced to reduce labeling effort
• Limitation: rarely leverages historical data, missing cross-day patterns

✅ Unsupervised Domain Adaptation (UDA)

• Techniques: CCA, ADAN, CycleGAN, WDGRL, DA-DCF
• Limitation:
  • Ignores available few labeled data
  • Fails to achieve optimal performance

3. ⚡ Opportunity: Semi-Supervised Domain Adaptation (SSDA)

SSDA has recently gained attention due to its low-cost and high-performance characteristics.

🧱 Limitation in BCI Applications

• Most SSDA methods are for classification tasks
• Not suitable for regression-based neural decoding
• Currently, no SSDA-based method exists for BCI recalibration

4. 🧪 Observations from Neural Signal Patterns

To better understand neural signal changes across days, we conducted preliminary experiments based on:

• Speed
• Direction of ground-truth velocity

🧭 Key Findings

• Ideal feature distributions should be:
  • Radial in shape
  • Consistent between features and labels
• Cross-day decoding results:
  • Direction decoding: relatively stable
  • Speed decoding: significant deviation

🔄 Implications for UDA

• UDA methods using global alignment correct direction shifts reasonably well
• But fail to handle speed drift, as similar-speed samples may be far apart in feature space
• Existing conditional alignment methods are classification-specific and not compatible with regression tasks

5. 🚀 Proposed Framework: SSAR

We propose three key factors for effective regression-based BCI recalibration:

1. 🌐 Global alignment
2. ⚙️ Conditional speed alignment
3. 🔗 Feature-label consistency

🧠 SSAR: Speed-Enhanced Subdomain Adaptation Regression

A novel framework that integrates domain adaptation and semi-supervised learning for regression tasks.

🧩 Core Components

• Speed-enhanced Subdomain Alignment (SeSA)

  • Performs both global and conditional speed alignment
  • Aligns data points with similar labels

• Contrastive Consistency Constraint (CCC)

  • Enhances SeSA using contrastive learning
  • Enforces consistency between features and labels

6. 🎯 Contributions

• 🔍 Empirical Insight:
  We reveal that lacking conditional speed alignment limits UDA performance, and propose three key factors for effective recalibration.

• 🧠 Novel Method:
  We introduce SSAR, the first framework combining semi-supervised learning and subdomain adaptation for regression-based BCI recalibration.

• 📊 State-of-the-Art Results:
  Extensive qualitative and quantitative experiments demonstrate that SSAR achieves superior recalibration performance compared to existing methods.