TLDR: A new deep learning autoencoder effectively removes muscle artifact noise (EMG) from ECG recordings, which typically obscure Skin Nerve Activity (SKNA). This method significantly improves SKNA signal quality and allows for accurate detection of cognitive stress, even in noisy conditions, paving the way for more reliable non-invasive sympathetic nervous system monitoring in real-world settings.
The human body’s response to stress is intricately managed by the sympathetic nervous system (SNS). Understanding SNS dynamics can offer crucial insights into various health conditions, from heart disease to anxiety. A promising non-invasive method to observe this activity is through Skin Nerve Activity (SKNA), which can be extracted from high-frequency electrocardiogram (ECG) recordings.
However, measuring SKNA accurately has faced a significant hurdle: contamination from electromyographic (EMG) signals, essentially noise from muscle activity. Traditional methods, like simple bandpass filtering, often struggle because the frequency ranges of muscle noise and SKNA can overlap, making it difficult to isolate the true nerve signals, especially during movement.
A recent study introduces an innovative solution to this problem using a deep learning approach. Researchers developed a lightweight one-dimensional convolutional autoencoder, enhanced with a Long Short-Term Memory (LSTM) bottleneck, designed to reconstruct clean SKNA from ECG recordings that are heavily contaminated with muscle artifacts. This method moves beyond merely detecting and discarding noisy data segments, as previous approaches did, to actively restoring the valuable physiological information within them.
The team simulated realistic noise conditions by adding EMG signals from chaotic muscle stimulation recordings to clean ECG-derived SKNA data obtained during cognitive stress experiments. They tested their model at two challenging noise levels: -4 dB and -8 dB signal-to-noise ratio (SNR).
The results were highly encouraging. The deep learning autoencoder significantly improved the signal quality, boosting the SNR by up to 9.65 dB. It also dramatically increased the cross-correlation with clean SKNA, from 0.40 to 0.72 for SKNA and from 0.15 to 0.72 for integrated SKNA (iSKNA) at the more severe -8 dB SNR. This indicates that the model effectively suppressed muscle noise while preserving the crucial temporal patterns of SKNA.
Furthermore, the reconstructed signals maintained the ability to accurately reflect sympathetic activity patterns. Key SKNA features, such as burst count and burst duration, showed discriminability (AUROC) values exceeding 0.93 even under severe noise, closely matching the performance of clean, uncontaminated data. When it came to classifying baseline versus cognitive stress conditions, the reconstructed signals achieved impressive accuracies of 91–98% across severe noise levels, a stark contrast to the poor performance of traditional filtering methods (50–65%).
This breakthrough signifies a major step forward for SKNA monitoring. By enabling the preservation of physiologically relevant sympathetic bursts even amidst substantial muscle interference, this deep learning-based reconstruction method paves the way for more robust and continuous SKNA monitoring in everyday, movement-rich environments. This could lead to more practical, real-world applications in areas such as portable cardiovascular disease management, emotion recognition, cognitive fatigue monitoring, and sleep assessment. For more details, you can read the full research paper here.
Also Read:
- Enhanced Silent Speech Recognition: A Two-Stage AI Breakthrough
- i-Mask: Recognizing Human Activities Through Exhaled Breath Patterns
While the study focused on cognitive stress and used an additive noise model, the researchers acknowledge these limitations and plan for future work to include a wider range of physiological states and incorporate simultaneously collected EMG and ECG data for even more realistic noise modeling.
