Constrained online recursive source separation framework for Real-Time electrophysiological signal processing

Yao Li, Xu Zhang, Haowen Zhao, Yunfei Liu, Xiang Chen, Lai Jiang

TL;DR

This paper introduces a novel "Constrained Online Recursive Source Separation" (CORSS) framework for real-time processing of electrophysiological signals. By combining a recursive learning rule with prior knowledge about target signals, CORSS achieves high accuracy in separating mixed signals with minimal computational cost, making it suitable for applications like human-machine interfaces and clinical monitoring.

What Problem Does This Paper Address?

Electrophysiological signals, such as those from the brain (EEG) or muscles (EMG), are often mixtures of signals from multiple underlying sources. To extract meaningful information, techniques like Blind Source Separation (BSS) are needed. However, many existing BSS methods, particularly those based on Independent Component Analysis (ICA), are computationally intensive and designed for offline analysis. This makes them unsuitable for real-time applications where immediate processing is crucial, such as in prosthetics control or continuous patient monitoring. The challenge is to develop BSS methods that are both accurate and fast enough for real-time use, especially when dealing with noisy, non-invasive signals.

Methods

The core of this work is the Constrained Online Recursive Source Separation (CORSS) framework. It builds upon existing BSS techniques but introduces two key innovations:

  • Recursive Matrix Updating Rule: Instead of batch processing, CORSS uses a stepwise recursive unmixing matrix learning rule. This allows for real-time updates of the separation matrix with minimal computational overhead, making it efficient for continuous data streams.
  • Incorporation of Prior Knowledge (Constrained ICA): CORSS integrates prior information about the target signals by using a specific cost function. This cost function is designed to mimic the probability distribution (CDF) of the desired signal. By guiding the algorithm towards these known characteristics, CORSS helps to prevent local optima and improves the accuracy of source separation, especially for signals with known properties like pulse-like patterns.

The framework also includes an online recursive pre-whitening step to prepare the data for the main separation module. The CORSS framework is illustrated in Figure 1 of the paper, showing the flow from input data to output source signals.

Datasets

The paper evaluates the CORSS framework on two real-time biomedical signal processing tasks using experimental data:

  • Real-time Surface Electromyogram (sEMG) Decomposition: Data was collected from eight healthy subjects performing isometric contractions of the thumb. This task aims to decompose sEMG signals into individual motor unit action potentials (MUAPTs).
  • Real-time Respiratory Intent Monitoring (sEMGdi Extraction): Data was collected from eight healthy subjects breathing at various intensities and rates. This task focuses on extracting diaphragmatic EMG (sEMGdi) signals to monitor respiratory intent, which is crucial for applications like ventilator control.

The paper does not introduce new benchmark datasets but utilizes existing experimental setups and data collection protocols.

Key Insights and Ideas

  • Prior Knowledge is Key for Electrophysiology: The paper strongly suggests that incorporating prior knowledge about the expected signal characteristics (like pulse shapes in EMG) is crucial for achieving accurate BSS in electrophysiological applications. This is a significant departure from traditional ICA, which often assumes minimal prior information.
  • Balancing Speed and Accuracy: CORSS demonstrates that it's possible to achieve both high separation accuracy and low computational latency. This is a critical breakthrough for real-time applications where these two factors are often in conflict.
  • Recursive Updates for Real-time: The recursive learning rule is the backbone of the real-time capability, allowing the system to adapt and update its separation parameters continuously without the need for large data buffers or offline processing.
  • CORSS Outperforms Standard ICA: The results clearly show that CORSS provides significantly better separation quality and faster convergence than standard online ICA methods like ORICA, especially in complex real-world scenarios.

Key Results

  • sEMG Decomposition: CORSS achieved a matching rate of 96.00 ± 2.04% for sEMG decomposition, which was statistically superior to the ORICA method. It also revealed distinct MU source signals, unlike the unusable signals produced by ORICA.
  • sEMGdi Extraction: For respiratory intent monitoring, CORSS achieved a matching rate of 98.12 ± 1.21%, outperforming other methods. It effectively extracted clear sEMGdi signals with minimal ECG interference, producing a smooth envelope crucial for accurate respiratory intent detection.
  • Computational Efficiency: The method exhibited minimal time delay, with only 12.5 ms delay when using a 0.1s block size. This is significantly faster than some other online methods like Online PFP.
  • Convergence Speed: CORSS demonstrated faster convergence in terms of correlation and RMSE compared to ORICA, reaching near-optimal performance within approximately 7-10 seconds.

Limitations and Future Work

  • Empirical Parameter Tuning: A current limitation is the dependence on empirical judgment for adjusting prior parameters (e.g., hyperparameters for the cost function).
  • Future Work: The authors suggest exploring methods for automatic adaptation of prior knowledge based on signal characteristics. This would enable the framework to intelligently adapt to a broader range of tasks without manual parameter tuning.

Summary updated

BibTeX
@inproceedings{li2025constrained,
  author = {Yao Li and Xu Zhang and Haowen Zhao and Yunfei Liu and Xiang Chen and Lai Jiang},
  title = {{Constrained online recursive source separation framework for Real-Time electrophysiological signal processing}},
  booktitle = {Biomedical},
  year = {2025},
  doi = {10.1016/j.bspc.2025.108769},
}
Figure