An acute pain assessment method and system based on EDA signals
By decoupling common pain characteristics from individual-specific features through a conditional variational autoencoder, a standardized pain response template is constructed, which solves the problems of accuracy and real-time performance in cross-individual pain assessment and enables precise classification and real-time display of pain intensity.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGDONG UNIV OF TECH
- Filing Date
- 2026-03-16
- Publication Date
- 2026-05-29
AI Technical Summary
Existing pain assessment methods based on EDA signals suffer from accuracy and reliability issues in cross-individual applications, making it difficult to effectively distinguish between individual physiological differences and changes in pain intensity, resulting in insufficient model generalization ability.
A method based on conditional variational autoencoder (PCVAE) is adopted to decouple common pain features from individual-specific features through a dual-stream encoder and a pain-specific cross-attention module, construct a standardized sympathetic neural template for pain response, and generate a standardized SPainSymp curve for pain assessment.
It enables precise classification and real-time assessment of pain intensity, effectively removing interference from individual differences and improving the accuracy and real-time nature of pain assessment, making it suitable for clinical applications.
Smart Images
Figure CN121867707B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of signal analysis technology, and in particular to an acute pain assessment method and system based on EDA signals. Background Technology
[0002] Pain is defined in the medical field as "an unpleasant sensory and emotional experience associated with, or in relation to, actual or potential tissue damage." In clinical medicine, pain is primarily classified into acute pain and chronic pain based on its duration. Acute pain typically serves as an adaptive signal to prevent danger and ensure survival, prompting the body to avoid potential tissue damage. However, if acute pain is not managed promptly and effectively, it can easily become chronic pain, causing long-term physical and psychological suffering, severely reducing patients' quality of life, and constituting a heavy social, medical, and economic burden. Therefore, timely and accurate assessment of acute pain is a core element of clinical pain management.
[0003] Currently, the most commonly used pain assessment methods in clinical practice rely primarily on patient self-report measurements, such as the Visual Analogue Scale (VAS), Numerical Rating Scale (NRS), and Verbal Rating Scale (VRS). However, these subjective assessment methods have significant limitations in practical application: on the one hand, these methods are highly dependent on the patient's cognitive and expressive abilities, and are essentially ineffective when applied to special populations with limited communication abilities (such as patients with impaired consciousness, infants, intubated patients, or those with language expression disorders); on the other hand, subjective ratings are easily influenced by subjective factors such as the patient's current psychological state, emotional fluctuations, and past experiences, resulting in a lack of objective comparability of rating responses to noxious stimuli of the same intensity among different patients.
[0004] To overcome the limitations of subjective assessment, exploring objective physiological signals to quantify pain intensity has become a research hotspot in this field in recent years. Among these, the intrinsic relationship between pain and electrical skin activity (EDA) has received widespread attention. EDA, as a simple, readily available, and non-invasive quantitative assessment method for autonomic nervous system (ANS) function, can non-invasively capture changes in sweat gland activity caused by sympathetic nervous system activation. Pain, as a strong stressor, can significantly activate the sympathetic nervous system, triggering the body's "fight or flight" response, thereby causing drastic changes in EDA signals. Existing research shows that EDA signals include slowly changing skin conductance levels (SCL) and rapid transient skin conductance responses (SCR). The former reflects baseline sympathetic activity, while the latter is directly triggered by specific stimuli (such as pain). With further research, the time-varying sympathetic activity index (TVSymp), derived from EDA signals, has been applied to the objective assessment of acute pain. TVSymp can reflect the temporal patterns of sympathetic activity induced by noxious stimuli and has shown a certain sensitivity in reflecting sympathetic activity.
[0005] Despite some progress in objective pain assessment techniques based on EDA signals and TVSymp indices, serious shortcomings remain in practical clinical applications and cross-individual scenarios, specifically:
[0006] Due to the high heterogeneity of human physiological characteristics, differences in baseline physiological characteristics between individuals (such as baseline conductivity, sweat gland density, skin impedance, and emotional state) have a significant impact on EDA signals. Existing pain assessment methods based on TVSymp curves often show that the EDA response amplitude of the same intensity of pain stimulus varies by several times among different individuals.
[0007] Existing signal processing and feature extraction methods merely map or classify the entire EDA signal or TVSymp curve, making it difficult to effectively distinguish, from both mathematical and physiological perspectives, whether these signal differences stem from inherent individual physiological variations or from actual changes in pain intensity. This coupling and confusion at the feature level severely restricts the cross-individual generalization ability of existing assessment models, leading to a significant drop in accuracy when faced with new patients outside the training set, making it difficult to meet the stringent accuracy and reliability requirements of clinical applications. Summary of the Invention
[0008] Therefore, it is necessary to provide a pain assessment method and system that can effectively decouple and separate pain-induced physiological responses from individual baseline variability from mixed EDA signals, without being affected by individual differences. This method and system are characterized by high accuracy and strong real-time performance.
[0009] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows:
[0010] An acute pain assessment method based on EDA signals includes the following steps:
[0011] Collect EDA signals of the subject's skin conductance activity;
[0012] The EDA signal is preprocessed, and the individual time-varying sympathetic activity TVSymp curve is obtained based on the preprocessed EDA signal.
[0013] A prior knowledge base containing standardized pain response sympathetic nerve SPainSymp templates is constructed, wherein the SPainSymp templates represent typical sympathetic nerve activity response patterns under different pain intensities;
[0014] The EDA signal and the individual TVSymp curve are input into a pre-trained Conditional Variational Autoencoder (PCVAE) model. The PCVAE model includes a two-stream encoder, a pain-specific cross-attention module, and a decoder. The two-stream encoder extracts the encoded features of the EDA signal and the conditional embedding vectors of the individual TVSymp curves. The pain-specific cross-attention module fuses the encoded features and the conditional embedding vectors, decoupling the fused features into common pain features and individual-specific features in the latent space.
[0015] Pain intensity is classified using the common pain features. At the same time, the decoder of the PCVAE model is used to reconstruct a standardized SpainSymp curve with the individual-specific features removed based on the common pain features, the conditional embedding vector of the TVSymp curve, and the individual-specific features.
[0016] The pain intensity classification results and the dynamic trend of the standardized SPainSymp curve are output in real time.
[0017] Preferably, the specific process for preprocessing the EDA signal includes:
[0018] The EDA signal is downsampled to a preset frequency using multiphase filtering resampling technology; lost or constant value segments in the signal are identified, and reconstructed using conformal piecewise cubic Hermitian interpolation.
[0019] Motion artifacts are identified based on the first derivative of the signal, data points with instantaneous change rates exceeding a threshold are removed and then filled by linear interpolation.
[0020] A fourth-order Butterworth low-pass filter was used to filter the signal to preserve the main energy of the EDA signal; the Z-score normalization method was used to eliminate baseline variability in the subjects.
[0021] Furthermore, the specific process for calculating the individual TVSymp curve is as follows:
[0022] A variable frequency complex demodulation bandpass filter is used to filter the preprocessed EDA signal, preserving the characteristic frequency band of sympathetic nerve activity;
[0023] The Hilbert transform is applied to the filtered signal to obtain the analytic signal;
[0024] The modulus of the analyzed signal is calculated to obtain the instantaneous amplitude envelope reflecting the intensity of sympathetic nerve activity, which is the individual TVSymp curve.
[0025] Furthermore, the method for constructing the SPainSymp template is as follows:
[0026] Multiple pain EDA signal datasets are aggregated, and the TVSymp curves corresponding to each sample are calculated.
[0027] Group all TVSymp curves according to the corresponding pain stimulus intensity; within each pain intensity group, calculate the point-by-point mean of all TVSymp curves;
[0028] The point-by-point mean is normalized based on the global mean variance to obtain SPainSymp curve templates corresponding to different pain intensities.
[0029] Furthermore, the dual-stream encoder of the PCVAE model includes:
[0030] The EDA signal encoder consists of a convolutional layer, a residual module, and a bottleneck layer, and is used to extract deep features of EDA signals.
[0031] The TVSymp conditional encoder is used to process the individual TVSymp curves to generate conditional embedding vectors; the features extracted by the EDA signal encoder and the conditional embedding vectors generated by the TVSymp conditional encoder are jointly input into the pain-specific cross-attention module.
[0032] Furthermore, the pain-specific cross-attention module utilizes a query-key-value mechanism to promote feature interaction, specifically through the following calculation methods:
[0033] The conditional embedding vector is passed through the weight matrix Linear projection generates query vector ;
[0034] The features extracted by the EDA signal encoder are processed through a weight matrix. and Generate key vectors respectively Sum value vector ;
[0035] According to the formula:
[0036]
[0037] Calculate the fused features, where Scaling factor Represents the transpose symbol. This represents the normalized exponential function.
[0038] Furthermore, the specific mechanism of the decoupling is as follows:
[0039] The fused features are mapped to two independent latent variable distributions;
[0040] The latent variables representing common pain features are reparameterized and input into the classifier for supervised pain classification training.
[0041] The latent variables representing individual-specific characteristics are reparameterized and differencing is performed with the conditional embedding vector of the individual TVSymp curve to explicitly remove individual patterns contained in the prior information.
[0042] Furthermore, the decoder integrates the common pain features and optimized SPainSymp features through a latent variable projection layer, and reconstructs a standardized SPainSymp curve via a decoding module and an end convolutional layer.
[0043] Furthermore, an incremental update strategy is adopted to achieve real-time output: a fixed time window is set to capture the EDA signal, and the time window is updated incrementally at a preset time interval. During each update, the process returns to the step of calculating the individual time-varying sympathetic activity TVSymp curve, thereby dynamically refreshing the displayed pain intensity value and the standardized SPainSymp curve.
[0044] An acute pain assessment system based on EDA signals, comprising:
[0045] The signal acquisition module is used to acquire the EDA signal of the subject's skin conductance activity.
[0046] Memory, used to store computer programs and SPainSymp templates;
[0047] A processor for executing the computer program to implement the steps of the method;
[0048] The display module is used to display the patient's pain intensity classification results and the dynamic evolution trend of the standardized SPainSymp curve in real time.
[0049] The beneficial effects of this invention are as follows:
[0050] This patent starts with the characteristic changes of sympathetic nerve activity curves in pain and EDA signals, extracts and systematically analyzes the features of the TVSymp curve, and constructs a standardized pain response sympathetic nerve SPainSymp template as physiological prior knowledge. Based on this, an innovative method based on a pain conditional variational autoencoder (PCVAE) is proposed. Through a dual-stream encoder and a pain-specific cross-attention module, common pain features and individual-specific features are explicitly separated in the decoupled latent space to find standardized indicators related to pain changes. Furthermore, an acute pain assessment system based on EDA signal pain sympathetic nerve activity curves is developed, enabling real-time analysis and display of pain intensity, accurate pain intensity classification, and real-time generation and dynamic display of SPainSymp curves. This assists doctors in objectively, accurately, and in real-time quantifying the patient's pain level. Attached Figure Description
[0051] Figure 1 This is a flowchart illustrating the acute pain assessment method based on EDA signals in one embodiment.
[0052] Figure 2 This is a template diagram of the SPainSymp curve for standardized pain sympathetic nerve activity in one embodiment.
[0053] Figure 3 This is an algorithmic framework diagram of the acute pain assessment method based on EDA signals in one embodiment.
[0054] Figure 4 This is a schematic diagram of a pain-specific cross-attention module in one embodiment.
[0055] Figure 5 In one embodiment, a comparison chart of the consistency between the SPainSymp curves and the SPainSymp template curves of different pain intensity samples generated based on the Biovid and PMED datasets is shown. Detailed Implementation
[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0057] Example 1
[0058] like Figure 1As shown, this embodiment provides an acute pain assessment method based on EDA signals. The main process includes data preprocessing, template extraction, feature decoupling and standardization generation, including the following specific steps:
[0059] Collect EDA signals of the subject's skin conductance activity;
[0060] The EDA signal is preprocessed, and the individual time-varying sympathetic activity TVSymp curve is obtained based on the preprocessed EDA signal.
[0061] A prior knowledge base containing standardized pain response sympathetic nerve SPainSymp templates is constructed, wherein the SPainSymp templates represent typical sympathetic nerve activity response patterns under different pain intensities;
[0062] The EDA signal and the individual TVSymp curve are input into a pre-trained Conditional Variational Autoencoder (PCVAE) model. The PCVAE model includes a two-stream encoder, a pain-specific cross-attention module, and a decoder. The two-stream encoder extracts the encoded features of the EDA signal and the conditional embedding vectors of the individual TVSymp curves. The pain-specific cross-attention module fuses the encoded features and the conditional embedding vectors, decoupling the fused features into common pain features and individual-specific features in the latent space.
[0063] Pain intensity is classified using the common pain features. At the same time, the decoder of the PCVAE model is used to reconstruct a standardized SpainSymp curve with the individual-specific features removed based on the common pain features, the conditional embedding vector of the TVSymp curve, and the individual-specific features.
[0064] The pain intensity classification results and the dynamic trend of the standardized SPainSymp curve are output in real time.
[0065] This embodiment uses the publicly available Biovid dataset (part A) and the PainMonit dataset (PMED) for model training and analysis. To ensure data quality, an anomaly detection mechanism based on signal flatness was developed to address issues such as sensor detachment: the raw EDA data is processed by a fourth-order Butterworth low-pass filter with a cutoff frequency of 2.0Hz to suppress high-frequency noise, and then the first-order difference sequence of the signal is calculated. If the percentage of sampling points with an absolute value lower than a preset threshold exceeds 50%, they are identified as abnormal samples and removed. After rigorous screening, the final dataset distribution is shown in Table 1, containing samples at different levels, including baseline, painless (NP), threshold (T1), and tolerable temperature (T4).
[0066] Table 1
[0067]
[0068] To reduce the impact of individual baseline differences, the EDA signal underwent in-depth preprocessing. First, a multiphase filtering resampling technique was used to uniformly downsample the signal to 16Hz. Second, lost signal segments were identified and reconstructed using conformal piecewise cubic Hermitian interpolation, while motion artifacts with abnormal instantaneous rate of change were eliminated by analyzing the first derivative. Then, a fourth-order Butterworth low-pass filter with a cutoff frequency of 2.0Hz was used to filter out EMG artifacts and power line noise. Finally, Z-score normalization was applied to eliminate baseline variability among subjects. Subsequently, individual TVSymp curves reflecting sympathetic nerve activity were extracted: a variable-frequency demodulation bandpass filter was used to extract the characteristic frequency bands of the sympathetic nervous system, and Hilbert transform was applied to obtain the analytic signal. After calculating the modulus, the instantaneous amplitude envelope (i.e., the individual TVSymp curve) was obtained.
[0069] To provide prior physiological knowledge, the TVSymp curves of all samples were grouped according to pain level, the pointwise mean of each group was calculated, and normalization was performed based on the global mean variance. For example... Figure 2 As shown, the obtained SPainSymp curve template clearly reflects that the higher the pain intensity, the shorter the sympathetic nerve activation latency and the larger the peak amplitude. This template provides a benchmark for feature decoupling in subsequent models.
[0070] like Figure 3 As shown in the algorithm framework diagram, this invention innovatively adopts a dual-input structure. The preprocessed EDA signal is used for feature extraction by an encoder consisting of a convolutional layer, a residual module, and a bottleneck layer; the corresponding individual TVSymp curves are used to generate conditional embedding vectors through a conditional encoder.
[0071] The two information streams enter the pain-specific cross-attention module (PCAM) for fusion. For example... Figure 4 As shown, PCAM utilizes a query-key-value mechanism: the conditional embedding vector is projected onto a weight matrix to generate a query vector Q, and the EDA-encoded features are projected to generate a key vector K and a value vector V. The feature interaction calculation formula is as follows:
[0072]
[0073] The fused features are explicitly decoupled into two latent variable distributions: common pain features and individual-specific features.
[0074] The reparameterized common pain features are input into the classifier to perform a supervised pain classification task, outputting the pain intensity level. Simultaneously, the reparameterized individual-specific features are differentially analyzed with the conditional embedding vector of the individual's TVSymp curve to remove individual-specific patterns. These removed features, along with the common pain features, are input into the decoder and reconstructed through three decoding modules and a terminal convolutional layer, ultimately generating a standardized SPainSymp curve that eliminates individual differences.
[0075] Example 2
[0076] To verify the effectiveness of this invention, a comprehensive comparison with existing technologies was conducted under the same experimental configuration:
[0077] In this embodiment, the classification performance evaluation is shown in Table 2: In the binary classification task for T0 / T4, the algorithm of this invention achieved an accuracy of 89.94% and 93.15% on the Biovid and PMED datasets, respectively, which is 3.18% and 3.61% higher than the current best benchmark methods (AttenLSTM and CrossMod), demonstrating excellent cross-dataset generalization ability.
[0078] Table 2
[0079]
[0080] In this embodiment, the curve generation quality evaluation is shown in Table 3: This invention achieves optimal performance in curve reconstruction through an explicit decoupling mechanism. On the Biovid dataset, the mean squared error (MSE) of this method is as low as 0.0202, and the Pearson correlation coefficient (PCC) is as high as 0.9960, reducing the reconstruction error by 42.3% compared to the best-performing baseline model (CAE).
[0081] Table 3
[0082]
[0083] In this embodiment, interpretability analysis is as follows: Figure 5As shown: a) Comparison of the consistency between the SPainSymp predicted curve and the template curve for the Biovid dataset under T0 condition; b) Comparison of the consistency between the SPainSymp predicted curve and the template curve for the Biovid dataset under T4 condition; c) Comparison of the consistency between the SPainSymp predicted curve and the template curve for the PMED dataset under T0 condition; d) Comparison of the consistency between the SPainSymp predicted curve and the template curve for the PMED dataset under T4 condition. Under T0 and T4 conditions on different datasets, the SPainSymp predicted curves generated by the model show a high degree of consistency with the actual SPainSymp template curves, effectively eliminating misjudgments caused by individual baseline variations and providing highly transparent physiological indicators for clinical applications.
[0084] Example 3
[0085] Based on the above method, this embodiment also provides an acute pain assessment system. The present invention develops a prototype monitoring system that synchronously displays the original EDA signal, real-time TVSymp curve, and generated standardized SPainSymp curve, as well as the current pain assessment level, through a visual interface, including:
[0086] The signal acquisition module is used to acquire the EDA signal of the subject's skin conductance activity.
[0087] Memory, used to store computer programs and SPainSymp templates;
[0088] A processor for executing the computer program to implement the steps of the method;
[0089] The display module is used to display the patient's pain intensity classification results and the dynamic evolution trend of the standardized SPainSymp curve in real time.
[0090] To meet the needs of real-time clinical monitoring, this system adopts a fixed sliding time window of 5.5 seconds and supports a data update strategy with 0.5-second increments. As shown in Table 4, on a conventional desktop platform (CPU-only environment), the average processing time for a single sample on the Biovid and PMED datasets is only 6.46 milliseconds and 8.24 milliseconds, respectively, and the peak processing time is also far below the system-defined threshold of 500 milliseconds. This is further enhanced by the model's small parameter count of only 2.27M and extremely low computational complexity of 25.6MFLOPs.
[0091] Table 4
[0092]
[0093] In summary, this embodiment fully demonstrates the feasibility of achieving near real-time dynamic visualization on resource-constrained clinical equipment from an engineering deployment perspective.
Claims
1. An acute pain assessment method based on EDA signals, characterized in that, Includes the following steps: Collect EDA signals of the subject's skin conductance activity; The EDA signal is preprocessed, and the individual time-varying sympathetic activity TVSymp curve is obtained based on the preprocessed EDA signal. The specific preprocessing steps include: The EDA signal is downsampled to a preset frequency using multiphase filtering resampling technology; lost or constant value segments in the signal are identified, and reconstructed using conformal piecewise cubic Hermitian interpolation. Motion artifacts are identified based on the first derivative of the signal, data points with instantaneous change rates exceeding a threshold are removed and then filled by linear interpolation. A fourth-order Butterworth low-pass filter was used to filter the signal to preserve the main energy of the EDA signal; Z-score normalization was used to eliminate baseline variability in the subjects. The specific process for calculating the individual TVSymp curve is as follows: A variable frequency complex demodulation bandpass filter is used to filter the preprocessed EDA signal, preserving the characteristic frequency band of sympathetic nerve activity; The Hilbert transform is applied to the filtered signal to obtain the analytic signal; The modulus of the analyzed signal is calculated to obtain the instantaneous amplitude envelope reflecting the intensity of sympathetic nerve activity, which is the individual TVSymp curve. A prior knowledge base containing standardized pain response sympathetic nerve SPainSymp templates is constructed, wherein the SPainSymp templates represent typical sympathetic nerve activity response patterns under different pain intensities; The EDA signal and the individual TVSymp curve are input into a pre-trained Conditional Variational Autoencoder (PCVAE) model. The PCVAE model includes a two-stream encoder, a pain-specific cross-attention module, and a decoder. The two-stream encoder extracts the encoded features of the EDA signal and the conditional embedding vectors of the individual TVSymp curves. The pain-specific cross-attention module fuses the encoded features and the conditional embedding vectors, decoupling the fused features into common pain features and individual-specific features in the latent space. The specific mechanism of the decoupling is as follows: The fused features are mapped to two independent latent variable distributions; The latent variables representing common pain features are reparameterized and input into the classifier for supervised pain classification training. The latent variables that characterize individual-specific features are reparameterized and differiated with the conditional embedding vector of the individual TVSymp curve to explicitly remove individual patterns contained in the prior information. Pain intensity is classified using the common pain features. At the same time, the decoder of the PCVAE model is used to reconstruct and generate a standardized SPainSymp curve with the individual-specific features removed, based on the common pain features, the conditional embedding vector of the TVSymp curve, and the individual-specific features. The pain intensity classification results and the dynamic trend of the standardized SPainSymp curve are output in real time.
2. The method according to claim 1, characterized in that, The method for constructing the SPainSymp template is as follows: Multiple pain EDA signal datasets are aggregated, and the TVSymp curves corresponding to each sample are calculated. Group all TVSymp curves according to the corresponding pain stimulus intensity; within each pain intensity group, calculate the point-by-point mean of all TVSymp curves; The point-by-point mean is normalized based on the global mean variance to obtain SPainSymp curve templates corresponding to different pain intensities.
3. The method according to claim 1, characterized in that, The dual-stream encoder of the PCVAE model includes: The EDA signal encoder consists of a convolutional layer, a residual module, and a bottleneck layer, and is used to extract deep features of EDA signals. The TVSymp conditional encoder is used to process the individual TVSymp curves to generate conditional embedding vectors; the features extracted by the EDA signal encoder and the conditional embedding vectors generated by the TVSymp conditional encoder are jointly input into the pain-specific cross-attention module.
4. The method according to claim 3, characterized in that, The pain-specific cross-attention module utilizes a query-key-value mechanism to promote feature interaction, and the specific calculation method includes: The conditional embedding vector is passed through the weight matrix Linear projection generates query vector ; The features extracted by the EDA signal encoder are processed through a weight matrix. and Generate key vectors respectively Sum value vector ; According to the formula: Calculate the fused features, where Scaling factor Represents the transpose symbol. This represents the normalized exponential function.
5. The method according to claim 1, characterized in that, The decoder integrates the common pain features and the optimized SPainSymp features through a latent variable projection layer, and reconstructs a standardized SPainSymp curve through a decoding module and an end convolutional layer.
6. The method according to claim 1, characterized in that, Real-time output is achieved by using an incremental update strategy: a fixed time window is set to capture the EDA signal, and the time window is updated incrementally at preset time intervals. During each update, the process returns to the step of calculating the individual time-varying sympathetic activity TVSymp curve, thereby dynamically refreshing the displayed pain intensity value and the standardized SPainSymp curve.
7. An acute pain assessment system based on EDA signals, characterized in that, include: The signal acquisition module is used to acquire the EDA signal of the subject's skin conductance activity. Memory, used to store computer programs and SPainSymp templates; A processor for executing the computer program to implement the steps of the method as described in any one of claims 1 to 6; The display module is used to display the patient's pain intensity classification results and the dynamic evolution trend of the standardized SPainSymp curve in real time.