TLDR: This research introduces a novel method for automatic pain recognition using electrodermal activity (EDA) signals. It transforms multiple representations of EDA (raw, phasic, tonic, etc.) into a single “multi-representation diagram” image. This image is then processed by a Vision Transformer for pain classification (No, Low, High Pain). The approach demonstrates comparable and often superior performance to traditional signal fusion methods, achieving high accuracy and showing the effectiveness of visualizing biosignals for leveraging powerful 2D vision models in pain assessment.
Pain is a complex and widespread issue affecting many people, and accurately assessing it is crucial for effective management. Traditional pain assessment often relies on subjective patient reports, which can be inconsistent and lead to challenges like over-prescription of medication. To address this, researchers are developing automatic pain assessment systems that use objective physiological signals.
A recent study introduces a novel approach to automatic pain recognition by integrating various electrodermal activity (EDA) signals into a single image. EDA, which measures changes in skin conductivity due to sweat gland activity, is a non-invasive indicator of sympathetic nervous system function and has long been recognized for its reliability in quantifying emotional responses and stress.
The proposed method, detailed in the research paper Multi-Representation Diagrams for Pain Recognition, involves a pipeline that takes raw EDA signals and processes them into multiple distinct representations. These representations include the raw signal itself, its phasic and tonic components (which capture rapid and slow sympathetic activity, respectively), detrended signals, a time-varying sympathetic index (TVSymp), and a set of handcrafted features derived from the EDA signal. Each of these six representations is then visualized as a waveform diagram.
What makes this approach unique is the creation of “multi-representation diagrams.” Instead of analyzing each signal representation separately or fusing them through traditional methods, these individual waveform diagrams are combined and jointly visualized within a single 224×224 pixel image. This transformation allows the system to leverage powerful 2D vision models, specifically a hierarchical Vision Transformer encoder, which are excellent at extracting complex patterns from images.
The study conducted extensive experiments using a dataset of EDA recordings from 65 participants, who experienced “No Pain,” “Low Pain,” and “High Pain” levels induced by electrical stimulation. The performance was evaluated using metrics like accuracy, precision, and F1 score.
The results showed that the multi-representation diagrams consistently yielded comparable, and in several cases superior, results to traditional signal fusion methods like addition and concatenation. For instance, combining all six EDA waveform representations into a single diagram achieved an impressive 80.67% accuracy, outperforming the best classic fusion method. Even combinations of fewer representations within a single diagram showed competitive performance.
Also Read:
- AI Breakthrough: Detecting Pain Through Breathing Patterns
- New AI Model Enhances Emotion Recognition from Imperfect Body Signals
This innovative method demonstrates the effectiveness of transforming physiological signal waveforms into image-based visualizations for pain assessment. It opens up new possibilities for utilizing advanced 2D vision models in analyzing biosignals, challenging the conventional reliance on 1D signal analysis. While the method shows great promise, the researchers note that scalability might be a consideration if a significantly larger number of representations were to be used, as excessive compression could degrade performance. Nevertheless, this work represents a robust alternative for integrating different signal representations or modalities in automatic pain recognition systems.
