A method, system, device and medium for gynecological postoperative pain assessment

Abstract

The application relates to the medical information technology field. A gynecological postoperative pain evaluation method, system, device and medium are provided, wherein the method comprises the following steps: inputting a multi-modal feature vector into a pre-constructed pain component decoupling network, the pain component decoupling network comprising a somatic pain prediction head and a visceral pain prediction head; performing feature fusion on the multi-modal feature vector by the pain component decoupling network to obtain a multi-modal fusion feature vector; performing dynamic distribution on the multi-modal fusion feature vector by a decoupling attention mechanism; outputting a somatic pain intensity prediction value via the somatic pain prediction head and outputting a visceral pain intensity prediction value via the visceral pain prediction head; and generating a gynecological postoperative pain component decoupling evaluation result for a patient based on the somatic pain intensity prediction value and the visceral pain intensity prediction value, so that the objectivity and continuity of pain evaluation are improved, the distinguishing ability for somatic pain components and visceral pain components is enhanced, and the decoupling of the evaluation result is improved.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method, system, device and medium for assessing postoperative pain in gynecological surgery. Background Technology

[0002] With the deep integration of artificial intelligence and smart healthcare, objective and accurate pain assessment is playing an increasingly crucial role in the rapid recovery after gynecological surgery. Postoperative pain assessment is a core component guiding clinical analgesia decisions and accelerating patient recovery, and it is also a key focus and challenge in perioperative management research. Effective pain assessment aims to accurately identify the source and intensity of pain, thereby providing targeted intervention plans for different pain components.

[0003] In traditional techniques, patient-reported tools such as visual analog scales or numerical rating scales are commonly used in clinical practice for pain assessment. However, these methods are highly dependent on patient cooperation, have poor applicability to patients who are weak or confused after surgery, and cannot achieve continuous dynamic monitoring. Existing objective pain recognition technologies based on single-modal physiological signals can usually only output a comprehensive pain intensity index, making it difficult to distinguish between the two different components of postoperative gynecological surgery: incision somatic pain and uterine contraction visceral pain, resulting in a lack of targeted analgesia measures. In addition, some multimodal assessment schemes often simply splice various signals and directly map them to pain scores, ignoring the differential contributions of different modal signals in representing different pain components, resulting in insufficient decoupling ability and clinical interpretability of the assessment results. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, system, device, and medium for assessing postoperative pain in gynecology to address the aforementioned technical problems, so as to improve the objectivity and continuity of pain assessment, enhance the ability to distinguish between somatic pain components and visceral pain components, and improve the decoupling effect of assessment results.

[0005] Firstly, this application provides a method for assessing postoperative pain in gynecological surgery, the method comprising:

[0006] Simultaneously acquire multimodal physiological signals and facial video data of the patients to be evaluated during postoperative recovery. The multimodal physiological signals include electrocardiogram signals, skin conductance signals, abdominal surface electromyography signals, and respiratory signals.

[0007] Multimodal physiological signals and facial video data are preprocessed, and features are extracted from the preprocessed signals within a unified time window to generate multimodal feature vectors. The multimodal feature vectors include electrocardiogram features, skin conductance features, electromyography features, respiratory features, and facial expression dynamic features.

[0008] The multimodal feature vector is input into a pre-constructed pain component decoupling network, which includes a somatic pain prediction head and a visceral pain prediction head. The pain component decoupling network performs feature fusion on the multimodal feature vector to obtain a multimodal fused feature vector. The multimodal fused feature vector is dynamically allocated through a decoupling attention mechanism. The somatic pain intensity prediction value is output through the somatic pain prediction head, and the visceral pain intensity prediction value is output through the visceral pain prediction head.

[0009] Based on the predicted values ​​of somatic pain intensity and visceral pain intensity, a decoupled assessment result of the pain components after gynecological surgery is generated for the patient.

[0010] In one embodiment, a decoupled attention mechanism is used to dynamically allocate multimodal fusion feature vectors, and the predicted values ​​of somatic pain intensity are output through a somatic pain prediction head and the predicted values ​​of visceral pain intensity are output through a visceral pain prediction head, including:

[0011] A nonlinear transformation is performed on the multimodal fusion feature vector to generate query vector and key value vector respectively;

[0012] Calculate the similarity between the query vector and the key value vector to obtain the weight distribution. The weight distribution represents the proportion of information corresponding to somatic pain mode and visceral pain mode in the multimodal fusion feature vector.

[0013] The multimodal fusion feature vector is weighted and split using weight distribution to obtain somatic pain latent variables and visceral pain latent variables;

[0014] Input the hidden variables of somatic pain into the somatic pain prediction head for pain intensity mapping processing, and output the predicted value of somatic pain intensity.

[0015] The latent variable of visceral pain is input into the visceral pain prediction head for pain intensity mapping processing, and the predicted value of visceral pain intensity is output.

[0016] In one embodiment, a nonlinear transformation is performed on the multimodal fusion feature vector to generate a query vector and a key-value vector, including:

[0017] The multimodal fusion feature vector is subjected to a first linear projection to obtain a first linear projection vector, and a nonlinear activation is applied to the first linear projection vector to generate a query vector;

[0018] The multimodal fusion feature vector is subjected to a second linear projection to obtain a second linear projection vector, and a nonlinear activation is applied to the second linear projection vector to generate a key value vector.

[0019] The parameter matrices of the first linear projection and the second linear projection are learned independently during the training of the pain component decoupling network, enabling the query vector and key value vector distribution to capture discriminative information corresponding to different pain modes in the multimodal fusion feature vector.

[0020] In one embodiment, the pre-training process of the pain component decoupling network includes:

[0021] Obtain a multimodal sample dataset, which includes multimodal physiological signal samples and facial video data samples from multiple samples, as well as the corresponding somatic pain intensity annotation values ​​and visceral pain intensity annotation values ​​for each sample;

[0022] Each sample in the multimodal sample dataset is input into the pain component decoupling network to be trained, and the pain component decoupling network to be trained outputs the corresponding predicted values ​​of somatic pain intensity and visceral pain intensity.

[0023] Calculate the prediction loss, which represents the deviation between the predicted value of somatic pain intensity and the labeled value of somatic pain intensity, as well as the deviation between the predicted value of visceral pain intensity and the labeled value of visceral pain intensity.

[0024] Calculate the decoupling loss, which characterizes the correlation between the somatic pain latent variables and the visceral pain latent variables generated within the pain component decoupling network to be trained;

[0025] The parameters of the pain component decoupling network to be trained are updated based on the prediction loss and decoupling loss until the preset convergence condition is met, thus obtaining the trained pain component decoupling network.

[0026] In one embodiment, a decoupling loss is calculated, which characterizes the correlation between somatic pain latent variables and visceral pain latent variables generated within the pain component decoupling network to be trained, including:

[0027] Obtain somatic pain latent variables and visceral pain latent variables from the pain component decoupling network to be trained;

[0028] Nonlinear mapping is applied to the latent variables of somatic pain to obtain the projected representation of somatic pain.

[0029] Nonlinear mapping is applied to the latent variables of visceral pain to obtain the projected representation of visceral pain;

[0030] Calculate the distribution difference measure between the somatic pain projection representation and the visceral pain projection representation. The distribution difference measure characterizes the distance between the somatic pain projection representation and the visceral pain projection representation in the feature space.

[0031] The distribution difference metric is used as the decoupling loss.

[0032] In one embodiment, based on predicted values ​​of somatic pain intensity and predicted values ​​of visceral pain intensity, a decoupled assessment result of the components of post-gynecological postoperative pain for the patient is generated, including:

[0033] By comparing the predicted values ​​of somatic pain intensity and visceral pain intensity, the dominant pain component type was determined. The dominant pain component types include somatic pain dominant type, visceral pain dominant type, and mixed type.

[0034] Based on the dominant pain component type and the numerical relationship between the predicted values ​​of somatic pain intensity and visceral pain intensity, pain component composition description information is generated. This pain component composition description information represents the composition ratio of somatic pain component to visceral pain component in postoperative gynecological pain.

[0035] The results of the decoupling assessment of pain components after gynecological surgery include the type of dominant pain component and descriptive information on the composition of pain components.

[0036] Secondly, this application also provides a gynecological postoperative pain assessment system, the system comprising:

[0037] The multimodal data acquisition module is used to simultaneously acquire multimodal physiological signals and facial video data of the patient to be evaluated during postoperative recovery. The multimodal physiological signals include electrocardiogram signals, skin conductance signals, abdominal surface electromyography signals, and respiratory signals.

[0038] The multimodal feature extraction module is used to preprocess multimodal physiological signals and facial video data, and extract features from the preprocessed signals within a unified time window to generate multimodal feature vectors. The multimodal feature vectors include electrocardiogram features, skin conductance features, electromyography features, respiratory features, and facial expression dynamic features.

[0039] The pain component decoupling prediction module is used to input multimodal feature vectors into a pre-constructed pain component decoupling network. The pain component decoupling network includes a somatic pain prediction head and a visceral pain prediction head. The pain component decoupling network performs feature fusion on the multimodal feature vectors to obtain a multimodal fused feature vector. The multimodal fused feature vector is dynamically allocated through a decoupling attention mechanism. The somatic pain intensity prediction value is output through the somatic pain prediction head, and the visceral pain intensity prediction value is output through the visceral pain prediction head.

[0040] The assessment results generation module is used to generate decoupled assessment results of postoperative gynecological pain components for patients based on the predicted values ​​of somatic pain intensity and visceral pain intensity.

[0041] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect of this application.

[0042] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods in the first aspect of this application.

[0043] This application provides a method, system, device, and medium for assessing postoperative pain in gynecological surgery. By simultaneously acquiring multimodal physiological signals and facial video data, it replaces traditional subjective scale assessments, effectively reducing human reporting bias and thus improving the objectivity and continuity of pain assessment. Features are extracted from electrocardiogram signals, skin conductance signals, abdominal surface electromyography signals, respiratory signals, and facial video data within a unified time window to generate a multimodal feature vector that integrates multi-source physiological and behavioral information, providing rich and synchronous input representations for distinguishing different pain components.

[0044] By utilizing the decoupling attention mechanism in a pain component decoupling network, multimodal fusion feature vectors are dynamically allocated to generate latent variables for somatic pain and visceral pain. Intensity values ​​are then independently predicted via somatic pain prediction heads and visceral pain prediction heads, respectively, thereby enhancing the ability to distinguish between somatic and visceral pain components. The decoupling assessment results for pain components are generated based on these two independent prediction values, reducing the mutual interference between somatic and visceral pain information, improving the decoupling of the assessment results, and enabling more targeted interventions for different components in postoperative pain management. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a gynecological postoperative pain assessment method according to one embodiment of the present invention;

[0047] Figure 2 A flowchart for generating decoupled assessment results of post-gynecological pain components for patients based on predicted values ​​of somatic pain intensity and predicted values ​​of visceral pain intensity in one embodiment of the present invention.

[0048] Figure 3 This is a structural diagram of a gynecological postoperative pain assessment system according to one embodiment of the present invention. Detailed Implementation

[0049] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0050] First, the application scenarios of the embodiments of this application will be described. In the embodiments of this application, methods, systems, devices, and media for assessing post-gynecological post-operative pain are provided, applicable to, but not limited to, post-gynecological post-operative pain assessment scenarios.

[0051] In illustrative purposes, the method, system, device and medium for assessing postoperative pain in gynecology provided in this application embodiment can also be applied to other surgical postoperative pain assessment, obstetric labor pain monitoring, chronic pelvic pain management and tumor-related pain assessment, etc. This is only an example and does not limit the specific application scenarios.

[0052] like Figure 1 As shown, this application provides a method for assessing postoperative pain in gynecological surgery, the method comprising:

[0053] S101: Simultaneously acquire multimodal physiological signals and facial video data of the patient to be evaluated during postoperative recovery. The multimodal physiological signals include electrocardiogram signals, skin conductance signals, abdominal surface electromyography signals, and respiratory signals.

[0054] For example, the pain assessment terminal synchronously collects multimodal physiological signals through multiple sensors deployed on the patient's body surface. These sensors include electrocardiogram (ECG) electrodes, skin conduction electrodes, abdominal surface electromyography (EMG) electrodes, and a respiratory sensor. The ECG electrodes are attached to a designated location on the patient's chest to continuously collect ECG signals; the skin conduction electrodes are attached to the patient's fingers or wrists to continuously collect skin conduction signals; the abdominal surface EMG electrodes are attached to the area around the patient's abdominal incision to continuously collect abdominal surface EMG signals; and the respiratory sensor is fixed to the patient's chest or abdomen to continuously collect respiratory signals.

[0055] Simultaneously, the pain assessment terminal activates its camera to capture continuous facial video data of the patient. Through hardware synchronization or timestamp alignment, the terminal rigorously aligns the collected ECG signals, skin conductance signals, abdominal surface electromyography signals, respiratory signals, and facial video data along the time dimension, resulting in synchronously acquired multimodal physiological signals and facial video data.

[0056] S102: Preprocess the multimodal physiological signals and facial video data, and extract features from the preprocessed signals within a unified time window to generate multimodal feature vectors. The multimodal feature vectors include electrocardiogram features, skin conductance features, electromyography features, respiratory features, and facial expression dynamic features.

[0057] For example, the pain assessment terminal performs bandpass filtering on the synchronously acquired ECG signal to remove baseline drift and high-frequency noise interference, resulting in a filtered ECG signal. The filtered ECG signal is then subjected to motion artifact detection and abnormal segment removal to obtain a denoised ECG signal. Subsequently, the denoised ECG signal is subjected to QRS complex localization and R-wave peak detection, and heart rate variability time-domain parameters and frequency-domain parameters are extracted. These heart rate variability time-domain parameters and frequency-domain parameters are used as ECG domain features.

[0058] When processing the skin conductance signal, the pain assessment terminal performs low-pass filtering to remove high-frequency noise interference, resulting in a filtered skin conductance signal. The filtered skin conductance signal is then subjected to baseline drift correction and trend separation to obtain a denoised skin conductance signal. The denoised skin conductance signal is then decomposed into skin conductance level components and skin conductance response components, and skin conductance response amplitude and frequency are extracted. The skin conductance response amplitude and frequency are used as skin conductance domain features.

[0059] For abdominal surface electromyography (EMG) signals, the pain assessment terminal performs power frequency notch filtering to remove power frequency electromagnetic interference, obtaining notch EMG signals. Then, motion artifact identification and interference segment marking are performed on the notch EMG signals, and the interference segments are removed to obtain denoised EMG signals. Subsequently, full-wave rectification and envelope extraction are performed on the denoised EMG signals to obtain EMG envelope signals. The frequency band energy is divided and the root mean square value is calculated on the EMG envelope signals to extract EMG time-domain features and EMG frequency-domain features, which are then used as EMG domain features.

[0060] For the respiratory signal, the pain assessment terminal performs smoothing filtering to remove high-frequency spikes and interference, resulting in a filtered respiratory signal. The filtered respiratory signal is then subjected to abnormal segment detection and removal to obtain a denoised respiratory signal. The denoised respiratory signal is then subjected to peak and trough detection and respiratory cycle segmentation to extract respiratory frequency and respiratory depth, which are used as respiratory domain features.

[0061] In the facial video data processing stage, the pain assessment terminal performs frame-by-frame face detection on the synchronously acquired facial video data to determine the face region, and performs facial key point localization on the face region to obtain the facial key point coordinate sequence. Based on the facial key point coordinate sequence, the facial action unit intensity is calculated to obtain the facial action unit intensity sequence. Then, the facial action unit intensity sequence is subjected to temporal modeling and dynamic feature encoding to extract the dynamic features of facial expressions.

[0062] After extracting features from each modality, the pain assessment terminal combines features from the electrocardiogram domain, electrodermal domain, electromyography domain, respiratory domain, and facial expression dynamics to generate a multimodal feature vector.

[0063] S103: Input the multimodal feature vector into the pre-constructed pain component decoupling network. The pain component decoupling network includes a somatic pain prediction head and a visceral pain prediction head. The pain component decoupling network performs feature fusion on the multimodal feature vector to obtain a multimodal fused feature vector. The multimodal fused feature vector is dynamically allocated through a decoupling attention mechanism. The somatic pain intensity prediction value is output through the somatic pain prediction head, and the visceral pain intensity prediction value is output through the visceral pain prediction head.

[0064] For example, the pain assessment terminal inputs the multimodal feature vector into a pre-constructed pain component decoupling network. The pain component decoupling network first performs cross-modal interaction processing on each modality feature in the multimodal feature vector, and then performs adaptive weighting on each modality feature through a gating mechanism to suppress redundant information unrelated to pain and retain discriminative features related to pain, thus obtaining weighted modality features.

[0065] After completing cross-modal interaction processing, the pain component decoupling network performs hierarchical feature fusion on the weighted modal features. First, it fuses the electrocardiogram and electrodermal features into autonomic nervous system response features, and fuses the electromyography and respiratory features into somatic response features. Then, it deeply fuses the autonomic nervous system response features, somatic response features, and facial expression dynamic features, mapping them to a unified semantic space to generate a multimodal fusion feature vector.

[0066] The pain component decoupling network performs decoupling attention processing on the multimodal fusion feature vector. It generates query representation and key-value representation from the multimodal fusion feature vector through learnable projection transformation. It calculates the similarity weight between the query representation and the key-value representation and uses the similarity weight to perform weighted splitting of the multimodal fusion feature vector to obtain somatic pain latent variables and visceral pain latent variables.

[0067] The pain component decoupling network inputs the latent variables of somatic pain into the somatic pain prediction head. The somatic pain prediction head performs multi-level nonlinear mapping and pain intensity regression on the latent variables of somatic pain, and outputs the predicted value of somatic pain intensity. At the same time, the pain component decoupling network inputs the latent variables of visceral pain into the visceral pain prediction head. The visceral pain prediction head performs multi-level nonlinear mapping and pain intensity regression on the latent variables of visceral pain, and outputs the predicted value of visceral pain intensity.

[0068] S104: Based on the predicted values ​​of somatic pain intensity and visceral pain intensity, generate decoupled assessment results of postoperative gynecological pain components for patients.

[0069] For example, the pain assessment terminal compares the predicted values ​​of somatic pain intensity and visceral pain intensity, and determines the type of pain dominance based on the comparison results. When the predicted value of somatic pain intensity is significantly higher than the predicted value of visceral pain intensity, it is determined to be somatic pain dominance type. When the predicted value of visceral pain intensity is significantly higher than the predicted value of somatic pain intensity, it is determined to be visceral pain dominance type. When the difference between the two is within the range of a preset mixed threshold, it is determined to be mixed type.

[0070] The pain assessment terminal generates a description of pain components based on the dominant pain component type and the ratio of predicted somatic pain intensity to predicted visceral pain intensity. The terminal integrates the dominant pain component type and the description of pain component composition to generate a decoupled assessment result for the patient's post-gynecological surgery pain components.

[0071] One embodiment of this application provides a method for assessing postoperative pain in gynecological surgery. By simultaneously acquiring multimodal physiological signals and facial video data, it replaces traditional subjective scale assessments, effectively reducing human reporting bias and thus improving the objectivity and continuity of pain assessment. Feature extraction is performed on electrocardiogram signals, skin conductance signals, abdominal surface electromyography signals, respiratory signals, and facial video data within a unified time window to generate a multimodal feature vector that integrates multi-source physiological and behavioral information, providing rich and synchronous input representations for distinguishing different pain components.

[0072] By utilizing the decoupling attention mechanism in a pain component decoupling network, multimodal fusion feature vectors are dynamically allocated to generate latent variables for somatic pain and visceral pain. Intensity values ​​are then independently predicted via somatic pain prediction heads and visceral pain prediction heads, respectively, thereby enhancing the ability to distinguish between somatic and visceral pain components. The decoupling assessment results for pain components are generated based on these two independent prediction values, reducing the mutual interference between somatic and visceral pain information, improving the decoupling of the assessment results, and enabling more targeted interventions for different components in postoperative pain management.

[0073] In one embodiment, a decoupled attention mechanism is used to dynamically allocate multimodal fusion feature vectors, and the predicted values ​​of somatic pain intensity are output through a somatic pain prediction head and the predicted values ​​of visceral pain intensity are output through a visceral pain prediction head, including:

[0074] (1) Perform nonlinear transformation on the multimodal fusion feature vector to generate query vector and key value vector respectively.

[0075] For example, the pain assessment terminal obtains the multimodal fusion feature vector output by the pain component decoupling network, and inputs the multimodal fusion feature vector into the first linear projection layer and the second linear projection layer in the decoupling attention mechanism, respectively. The first linear projection layer performs a first linear projection process on the multimodal fusion feature vector to obtain a first linear projection vector, and the second linear projection layer performs a second linear projection process on the multimodal fusion feature vector to obtain a second linear projection vector. The first linear projection vector is then input into a first nonlinear activation function to obtain a query vector, and the second linear projection vector is input into a second nonlinear activation function to obtain a key value vector.

[0076] (2) Calculate the similarity between the query vector and the key value vector to obtain the weight distribution. The weight distribution represents the proportion of information corresponding to the somatic pain mode and the visceral pain mode in the multimodal fusion feature vector.

[0077] For example, the pain assessment terminal calculates the similarity between the query vector and the key-value vector, divides the similarity result by a scaling factor to obtain an unnormalized similarity score, and inputs the unnormalized similarity score into a normalized exponential function to obtain the weight distribution. The calculation formula for the above process is as follows:

[0078]

[0079] In the formula, Represents the query vector. Represents a key-value vector. This represents the transpose operator. Indicates the scaling factor. Represents the normalized exponential function, This represents the weight distribution.

[0080] The above formula is based on the following: the dot product between the query vector and the key vector measures the degree of matching between them in the feature space. The higher the degree of matching, the more discriminative information related to the pain pattern is contained in the current multimodal fusion feature vector. The physical basis of the scaling factor is that when the feature dimension is high, the variance of the dot product result increases with the increase of dimension, causing the subsequent normalization exponential function to enter the gradient saturation region. By dividing by... It can compress similarity scores to a reasonable range, ensuring gradient stability.

[0081] The normalization exponential function is based on mapping similarity scores to a positive real number interval and normalizing them by summing the components, so that the output weight distribution satisfies non-negativity and the sum of each component is a fixed constant. Thus, it can be interpreted as the proportion of information belonging to the somatic pain mode or visceral pain mode in the multimodal fusion feature vector, realizing dynamic attention allocation among pain components.

[0082] (3) Use weight distribution to perform weighted splitting of multimodal fusion feature vectors to obtain somatic pain latent variables and visceral pain latent variables.

[0083] For example, the pain assessment terminal performs sub-distribution partitioning on the weight distribution according to the somatic pain mode and the visceral pain mode. The sub-distribution partitioning process obtains the somatic pain weight sub-distribution and the visceral pain weight sub-distribution. Then, the somatic pain weight sub-distribution is used to perform element-wise weighting on the multimodal fusion feature vector to obtain the somatic pain latent variable. The visceral pain weight sub-distribution is used to perform element-wise weighting on the multimodal fusion feature vector to obtain the visceral pain latent variable.

[0084] The somatic pain latent variable includes discriminative information related to somatic pain from the multimodal fusion feature vector, while the visceral pain latent variable includes discriminative information related to visceral pain from the multimodal fusion feature vector. The multimodal fusion feature vector contains mixed information of both somatic and visceral pain. Weighted splitting allows for the decoupling of pain components, rather than simply using the same feature to predict both types of pain intensity.

[0085] (4) Input the hidden variables of somatic pain into the somatic pain prediction head for pain intensity mapping processing, and output the predicted value of somatic pain intensity.

[0086] For example, the pain assessment terminal inputs the hidden somatic pain variables into the somatic pain prediction head. The somatic pain prediction head then sequentially performs a first linear mapping process, a first nonlinear activation process, and a second linear mapping process on the hidden somatic pain variables. Finally, the regression mapping function of the somatic pain prediction head performs pain intensity mapping processing on the result of the second linear mapping process, and outputs the predicted somatic pain intensity value. The above calculation process is as follows:

[0087]

[0088] In the formula, This represents a hidden variable related to somatic pain. This represents the first weight matrix of the somatic pain prediction head. This represents the first bias vector of the somatic pain prediction head. This represents the first activation function of the somatic pain prediction head. This represents the second weight matrix of the somatic pain prediction head. This represents the second bias vector of the somatic pain prediction head. The regression mapping function representing the somatic pain prediction head. This represents the predicted value of somatic pain intensity.

[0089] The above formula is based on the following: the combination of the first linear mapping and subsequent nonlinear activation is used to learn the complex nonlinear mapping from somatic pain latent variables to intermediate semantic features, enabling the somatic pain prediction head to fit the high-order feature expressions corresponding to incision pain, abdominal wall electromyographic tension, and noxious stimuli on the body surface; the second linear mapping is used to compress the intermediate semantic features to a low-dimensional space; the regression mapping function is based on aligning the network output to the clinical pain intensity scale, so that the predicted somatic pain intensity can be directly compared with the labeled values ​​of the numerical rating scale or visual analog scale; the multi-layer mapping is based on the fact that a single-layer linear transformation cannot fully express the nonlinear coupling relationship between physiological signals, facial expressions, and pain intensity, while multi-layer nonlinear mapping can gradually approximate the true pain intensity through hierarchical feature abstraction.

[0090] The somatic pain prediction head includes a first fully connected layer, a first activation function, a second fully connected layer, and a regression mapping function. The first fully connected layer corresponds to the first weight matrix and the first bias vector, the second fully connected layer corresponds to the second weight matrix and the second bias vector, and the regression mapping function is used to output a somatic pain intensity prediction value aligned with the clinical pain rating scale.

[0091] (5) Input the latent variable of visceral pain into the visceral pain prediction head for pain intensity mapping processing, and output the predicted value of visceral pain intensity.

[0092] For example, the pain assessment terminal inputs the latent visceral pain variable into the visceral pain prediction head. The visceral pain prediction head then sequentially performs a third linear mapping process, a second nonlinear activation process, and a fourth linear mapping process on the latent visceral pain variable. Finally, the regression mapping function of the visceral pain prediction head performs pain intensity mapping processing on the result of the fourth linear mapping process, outputting a predicted visceral pain intensity value. The calculation formula for the above process is as follows:

[0093]

[0094] In the formula, This represents a latent variable indicating visceral pain. This represents the first weight matrix of the visceral pain prediction head. This represents the first bias vector of the visceral pain prediction head. This represents the first activation function of the visceral pain prediction head. This represents the second weight matrix of the visceral pain prediction head. This represents the second bias vector of the visceral pain prediction head. The regression mapping function represents the predictive head for visceral pain. This indicates the predicted intensity of visceral pain.

[0095] The above formula is based on the following: the parameters of the visceral pain prediction head and the somatic pain prediction head are independent. The physical basis for this is that the pathophysiological mechanisms, neural conduction pathways, and physiological signal expression patterns of visceral pain and somatic pain are different. Independent parameters can avoid mutual interference between the two types of pain features. The first layer of linear mapping and nonlinear activation is used to extract nonlinear features related to autonomic nerve response, diffuse pain, and visceral-somatic reflexes from the latent variables of visceral pain. The second layer of linear mapping and regression mapping function is used to map the above features to the visceral pain intensity scale. This structure enables the predicted value of visceral pain intensity to reflect the degree of pain caused by traction or inflammation of organs in the uterine cavity and pelvic cavity, rather than simply reusing the somatic pain prediction path.

[0096] In one embodiment, a nonlinear transformation is performed on the multimodal fusion feature vector to generate a query vector and a key-value vector, including:

[0097] (1) Perform a first linear projection on the multimodal fusion feature vector to obtain a first linear projection vector, and apply nonlinear activation to the first linear projection vector to generate a query vector.

[0098] For example, the pain assessment terminal obtains the multimodal fusion feature vector output by the pain component decoupling network, inputs the multimodal fusion feature vector into the first linear projection layer in the decoupling attention mechanism, and performs a first linear projection process on the multimodal fusion feature vector to obtain a first linear projection vector. The calculation formula for the above process is as follows:

[0099]

[0100] In the formula, This represents a multimodal fusion feature vector. Denotes the first linear projection parameter matrix. This represents the first bias vector. This represents the first linear projection vector.

[0101] The above formula is based on the following: the multimodal fusion feature vector is obtained by fusing features from the electrocardiogram domain, skin conductance domain, electromyography domain, respiratory domain, and facial expression dynamic features. Different modal features are highly overlapping in the original fusion space, making it impossible to directly distinguish between somatic pain patterns and visceral pain patterns. The first linear projection processing maps the multimodal fusion feature vector to the query subspace through the first linear projection parameter matrix, so that the subsequently generated query vector can express the degree to which the somatic pain pattern or visceral pain pattern needs to be focused on in the current mixed pain state, thereby providing a task-oriented query representation for decoupling the attention mechanism.

[0102] The pain assessment terminal inputs the first linear projection vector into the first nonlinear activation function, which performs nonlinear activation processing on the first linear projection vector to generate a query vector.

[0103] The above steps are based on the following: the first linear projection processing can only provide linear feature transformation and cannot characterize the pain pattern information of nonlinear superposition between physiological signals and facial video features; the first nonlinear activation function introduces a nonlinear discrimination boundary, enabling the query vector to distinguish between highly aliased somatic pain features and visceral pain features, thereby avoiding the inseparability of the two types of pain patterns caused by linear mapping; the query vector serves as the matching benchmark for decoupling the attention mechanism, and its physical meaning is the attention query representation of somatic pain patterns and visceral pain patterns in the current multimodal fusion feature vector.

[0104] (2) Perform a second linear projection on the multimodal fusion feature vector to obtain a second linear projection vector, and apply nonlinear activation to the second linear projection vector to generate a key value vector.

[0105] The parameter matrices of the first linear projection and the second linear projection are learned independently during the training of the pain component decoupling network, enabling the query vector and key value vector distribution to capture discriminative information corresponding to different pain modes in the multimodal fusion feature vector.

[0106] For example, the pain assessment terminal inputs the multimodal fusion feature vector into the second linear projection layer in the decoupled attention mechanism. The second linear projection layer performs a second linear projection process on the multimodal fusion feature vector to obtain a second linear projection vector. The calculation formula for the above process is as follows:

[0107]

[0108] In the formula, This represents a multimodal fusion feature vector. Denotes the second linear projection parameter matrix. This represents the second bias vector. This represents the second linear projection vector.

[0109] The above formula is based on the following: the second linear projection process maps the same multimodal fusion feature vector to the key-value subspace through a second linear projection parameter matrix that is independent of the first linear projection parameter matrix; the key-value subspace and the query subspace have different functions. The query subspace is used to express attention matching requirements, while the key-value subspace is used to express the feature content of somatic pain patterns and visceral pain patterns contained in the multimodal fusion feature vector; the two projection path parameters are independent, which can ensure that the query vector and the key-value vector capture discriminative information of different semantics respectively, rather than simply copying the same set of transformation results.

[0110] The pain assessment terminal inputs the second linear projection vector into a second nonlinear activation function, which performs nonlinear activation processing on the second linear projection vector to generate a key-value vector. The second nonlinear activation function enables the key-value vector to express the pain pattern information of nonlinear superposition in the multimodal fusion feature vector. The key-value vector and the query vector achieve attention matching through subsequent similarity calculation. If the key-value vector is only a linear projection result, it cannot fully express the complex nonlinear relationship between abdominal surface electromyography, facial expression dynamic features and visceral pain or somatic pain. The key-value vector generated after nonlinear activation processing can provide a richer feature expression, thereby improving the ability of the decoupled attention mechanism to distinguish pain components.

[0111] The reason why the parameter matrices of the first and second linear projections are learned independently during the training of the pain component decoupling network is as follows: If the parameter matrices of the first and second linear projections share parameters, the query vector and the key-value vector will degenerate into two representations of the same feature, and will not be able to assume the semantic roles of "query requirement" and "feature content" respectively. This will lead to the decoupling attention mechanism being unable to distinguish between somatic pain patterns and visceral pain patterns. Independent parameter learning enables the query vector to focus on expressing the matching requirement of "whether the current pain is more likely to be somatic pain or visceral pain", and enables the key-value vector to focus on expressing the content information of "which features in the multimodal fusion feature vector belong to somatic pain and which features belong to visceral pain", thereby achieving the semantic separation required for pain component decoupling.

[0112] In one embodiment, the pre-training process of the pain component decoupling network includes:

[0113] (1) Obtain a multimodal sample dataset, which includes multimodal physiological signal samples and facial video data samples from multiple samples, as well as the corresponding body pain intensity labeling value and visceral pain intensity labeling value for each sample.

[0114] For example, the multimodal physiological signal samples include electrocardiogram signal samples, skin conductance signal samples, abdominal surface electromyography signal samples, and respiratory signal samples; the facial video data samples include a sequence of continuous video frames of the face of patients in the postoperative recovery period; the somatic pain intensity labeled values ​​include the true values ​​of incision pain or abdominal wall pain intensity obtained by clinical assessment or patient self-assessment; the visceral pain intensity labeled values ​​include the true values ​​of uterine pain or pelvic and abdominal visceral pain intensity obtained by clinical assessment or patient self-assessment; and the multimodal sample dataset includes the supervised training data set required for the pain component decoupling network to be trained.

[0115] Multimodal physiological signal samples and facial video data samples together constitute the objective expression carrier of pain. ECG signal samples, skin conductance signal samples, and respiratory signal samples mainly reflect autonomic nerve responses, abdominal surface electromyography signal samples mainly reflect abdominal wall muscle tension and protective reflexes, and facial video data samples mainly reflect changes in pain-related facial movement units. Somatic pain intensity labels and visceral pain intensity labels provide regression targets for two types of pain components, enabling the pain component decoupling network to be trained to simultaneously learn pain intensity prediction ability and pain component decoupling ability in subsequent forward propagation and loss calculation.

[0116] The pain assessment terminal performs preprocessing operations on the multimodal physiological signal samples in the multimodal sample dataset to obtain standardized multimodal physiological signal samples. It also performs face detection and keypoint localization operations on the facial video data samples in the multimodal sample dataset to obtain standardized facial expression dynamic feature samples. The calculation formula for the above process is as follows:

[0117]

[0118]

[0119] In the formula, Indicates the first Each sample corresponds to a multimodal physiological signal sample. This represents the channel mean vector of a multimodal physiological signal sample. This represents the channel standard deviation vector of a multimodal physiological signal sample. This represents standardized multimodal physiological signal samples. Indicates the first Each sample corresponds to a facial video data sample. This represents a facial video preprocessing function. This represents a standardized sample of facial expression dynamic features.

[0120] Among them, the channel mean vector includes the set of mean values ​​of multimodal physiological signal samples in each signal channel, the channel standard deviation vector includes the set of standard deviations of multimodal physiological signal samples in each signal channel, the facial video preprocessing function includes face detection, key point localization, action unit intensity calculation and dynamic feature encoding operation, and the standardized facial expression dynamic feature sample includes facial behavior representations related to pain intensity.

[0121] The above formula is based on the following: the sampling range of each channel of the multimodal physiological signal sample is different, and the numerical magnitudes of the electrocardiogram signal sample, skin conductance signal sample, abdominal surface electromyography signal sample, and respiratory signal sample are different. If standardization is not performed, the signal channel with a larger numerical magnitude will dominate in the subsequent feature fusion, causing fine-grained pain component features such as abdominal surface electromyography signal sample to be submerged; facial video data samples need to undergo face detection and key point localization processing to obtain stable facial expression dynamic feature samples, thereby providing a consistent input representation for subsequent multimodal feature extraction.

[0122] (2) Input each sample in the multimodal sample dataset into the pain component decoupling network to be trained, and output the corresponding somatic pain intensity prediction value and visceral pain intensity prediction value from the pain component decoupling network to be trained.

[0123] For example, the pain assessment terminal inputs standardized multimodal physiological signal samples and standardized facial expression dynamic feature samples into the multimodal feature extraction part of the pain component decoupling network to be trained. The multimodal feature extraction part extracts electrocardiogram domain features, skin electric field features, electromyography domain features and respiratory domain features from the standardized multimodal physiological signal samples, and extracts facial expression dynamic features from the standardized facial expression dynamic feature samples, thus obtaining the multimodal feature vector corresponding to each sample.

[0124] The above processing is based on the following: different modal signals have different abilities to express pain. The electrocardiogram and electrodermal features mainly reflect the autonomic nervous system's stress response to pain stimuli, the electromyography features mainly reflect the electrical activity response of the abdominal wall muscles to incision pain or visceral-somatic reflexes, the respiratory features mainly reflect changes in respiratory rhythm caused by pain, and the facial expression dynamic features mainly reflect the behavioral expression of pain. The physical basis for concatenating multiple features into a multimodal feature vector is that single modal signals are easily affected by individual differences, changes in body position, or environmental noise, while multimodal joint representation can complement the information from different signal sources, thereby improving the robustness of subsequent pain component separation.

[0125] Among them, the electrocardiogram domain features include time-domain and frequency-domain parameters of heart rate variability; the skin conductance domain features include skin conductance response amplitude and skin conductance response frequency; the electromyography domain features include time-domain and frequency-domain features of abdominal surface electromyography signals; the respiratory domain features include respiratory rate and respiratory depth; the facial expression dynamic features include the time-series coding results of facial action unit intensity sequences; and the multimodal feature vector includes the joint representation of multimodal signals within a unified time window.

[0126] The pain assessment terminal inputs the multimodal feature vectors into the feature fusion part of the pain component decoupling network to be trained. The feature fusion part performs cross-modal interaction processing and hierarchical fusion processing on the multimodal feature vectors to obtain the multimodal fused feature vectors corresponding to each sample.

[0127] The pain assessment terminal inputs the multimodal fusion feature vector into the decoupling attention mechanism of the pain component decoupling network to be trained. The decoupling attention mechanism performs a first linear projection process, a first nonlinear activation process, a second linear projection process, and a second nonlinear activation process on the multimodal fusion feature vector, generating a query vector and a key-value vector respectively. The calculation formula for the above process is as follows:

[0128]

[0129]

[0130] In the formula, Indicates the first The multimodal fusion feature vector corresponding to each sample Denotes the first linear projection parameter matrix. This represents the first bias vector. Denotes the first nonlinear activation function. Indicates the first The query vector corresponding to each sample Denotes the second linear projection parameter matrix. This represents the second bias vector. This represents the second nonlinear activation function. Indicates the first The key-value vector corresponding to each sample.

[0131] The pain assessment terminal calculates the similarity between the query vector and the key-value vector. The similarity result is divided by a scaling factor and processed by a normalized exponential function to obtain the weight distribution for each sample. The calculation formula is as follows:

[0132]

[0133] In the formula, Indicates the first The query vector corresponding to each sample Indicates the first The key-value vector corresponding to each sample This represents the transpose operator. Indicates the scaling factor. Represents the normalized exponential function, Indicates the first The weight distribution corresponding to each sample.

[0134] The pain assessment terminal performs sub-distribution partitioning on the weight distribution according to the somatic pain mode and the visceral pain mode. This sub-distribution partitioning yields somatic pain weight sub-distributions and visceral pain weight sub-distributions. Then, the somatic pain weight sub-distributions and visceral pain weight sub-distributions are used to perform element-wise weighting on the multimodal fusion feature vector, respectively, to obtain the somatic pain latent variable and visceral pain latent variable corresponding to each sample. The calculation formula for the above process is as follows:

[0135]

[0136]

[0137] In the formula, Indicates the first The weighted quantum distribution of somatic pain corresponding to each sample. Indicates the first The visceral pain weighted sub-distribution corresponding to each sample This represents the element-wise multiplication operator. Indicates the first The multimodal fusion feature vector corresponding to each sample Indicates the first The hidden variables of physical pain corresponding to each sample. Indicates the first Latent variables of visceral pain corresponding to each sample.

[0138] The pain assessment terminal inputs the latent somatic pain variables into the somatic pain prediction head of the pain component decoupling network to be trained. The somatic pain prediction head then sequentially performs linear mapping, nonlinear activation, and regression mapping on the latent somatic pain variables, outputting the predicted somatic pain intensity value for each sample. The calculation formula for the above process is as follows:

[0139]

[0140] In the formula, Indicates the first The hidden variables of physical pain corresponding to each sample. This represents the first weight matrix of the somatic pain prediction head. This represents the first bias vector of the somatic pain prediction head. This represents the first activation function of the somatic pain prediction head. This represents the second weight matrix of the somatic pain prediction head. This represents the second bias vector of the somatic pain prediction head. The regression mapping function representing the somatic pain prediction head. Indicates the first Predicted values ​​of somatic pain intensity for each sample.

[0141] The pain assessment terminal inputs the latent visceral pain variables into the visceral pain prediction head of the pain component decoupling network to be trained. The visceral pain prediction head then sequentially performs linear mapping, nonlinear activation, and regression mapping on the latent visceral pain variables, outputting the predicted visceral pain intensity value for each sample. The calculation formula for the above process is as follows:

[0142]

[0143] In the formula, Indicates the first The latent variable of visceral pain corresponding to each sample This represents the first weight matrix of the visceral pain prediction head. This represents the first bias vector of the visceral pain prediction head. This represents the first activation function of the visceral pain prediction head. This represents the second weight matrix of the visceral pain prediction head. This represents the second bias vector of the visceral pain prediction head. The regression mapping function represents the predictive head for visceral pain. Indicates the first Predicted visceral pain intensity values ​​for each sample.

[0144] (3) Calculate the prediction loss, which represents the deviation between the predicted value of somatic pain intensity and the labeled value of somatic pain intensity, as well as the deviation between the predicted value of visceral pain intensity and the labeled value of visceral pain intensity.

[0145] For example, the pain assessment terminal calculates the single-sample prediction loss for each sample based on the predicted values ​​of somatic pain intensity, labeled somatic pain intensity, predicted visceral pain intensity, and labeled visceral pain intensity. Then, the prediction loss is obtained by batch averaging the single-sample prediction losses. The calculation formula for the above process is as follows:

[0146]

[0147]

[0148] In the formula, Indicates the first Predicted values ​​of somatic pain intensity for each sample Indicates the first The labeled value of somatic pain intensity for each sample Indicates the first Predicted visceral pain intensity for each sample Indicates the first The visceral pain intensity label value corresponding to each sample This represents the intensity prediction loss function. Indicates the first The single-sample prediction loss for each sample. Represents the total number of samples. This indicates the predicted loss.

[0149] The above formula is based on the following: the intensity prediction loss function measures the deviation between the network output and the clinical label, forcing the predicted value of somatic pain intensity to approach the labeled value of somatic pain intensity and the predicted value of visceral pain intensity to approach the labeled value of visceral pain intensity; the physical basis for calculating the single-sample prediction loss first and then the batch average is to reduce the impact of single-sample noise on parameter updates and make the gradient direction more stable.

[0150] (4) Calculate the decoupling loss, which represents the correlation between the somatic pain latent variables and the visceral pain latent variables generated within the pain component decoupling network to be trained.

[0151] For example, the pain assessment terminal obtains the somatic pain latent variables and visceral pain latent variables corresponding to each sample from the pain component decoupling network to be trained, performs nonlinear projection processing on the somatic pain latent variables to obtain the somatic pain projection representation corresponding to each sample, and performs nonlinear projection processing on the visceral pain latent variables to obtain the visceral pain projection representation corresponding to each sample.

[0152] The above processing is based on the following: Although the somatic pain latent variable and the visceral pain latent variable have been separated through the decoupled attention mechanism, they are still in the dedicated feature space of their respective predictors. Directly comparing the distribution differences between the two may produce bias due to spatial inconsistency. Projection nonlinear mapping is used to transform the somatic pain latent variable and the visceral pain latent variable to the same common projection space so as to make a fair comparison of the distribution differences between the two in the future.

[0153] The pain assessment terminal calculates the maximum mean difference between the distributions of somatic pain projection and visceral pain projection based on the somatic pain projection representation and visceral pain projection representation corresponding to each sample using a kernel function. This maximum mean difference is used as the decoupling loss. The calculation formula for the above process is as follows:

[0154]

[0155] In the formula, Indicates the first The somatic pain projection representation corresponding to each sample. Indicates the first The visceral pain projection representation corresponding to each sample Represents the kernel function. Represents the regenerating nucleus Hilbert space. The norm square operator is used to represent the Hilbert space of the reproducing kernel. This represents the decoupling loss.

[0156] The above formula is based on the core idea of ​​maximum mean difference: if two distributions are statistically identical, their mean embeddings in the regenerating kernel Hilbert space should be equal; after mapping the projected representation to the regenerating kernel Hilbert space through the kernel function, the distribution difference can be measured by comparing the distance between the mean embeddings of the two distributions; when the distribution difference is used as the decoupling loss, minimizing the decoupling loss is equivalent to forcing the distribution of somatic pain latent variables and the distribution of visceral pain latent variables to move away from each other, thereby reducing the correlation between the two.

[0157] (5) Update the parameters of the pain component decoupling network to be trained based on the prediction loss and decoupling loss until the preset convergence condition is met, and obtain the trained pain component decoupling network.

[0158] For example, the pain assessment terminal calculates the total loss based on the prediction loss and the decoupling loss. The total loss is then used to calculate the gradients of each trainable parameter in the pain component decoupling network to be trained via backpropagation. The calculation formula for the above process is as follows:

[0159]

[0160]

[0161] In the formula, Indicates the predicted loss. This represents the decoupling loss balance coefficient. This represents the decoupling loss. Indicates the total loss. This represents all trainable parameters of the pain component decoupling network to be trained. This represents the gradient of the total loss with respect to the trainable parameters. This represents the gradient of the prediction loss with respect to the trainable parameters. This represents the gradient of the decoupling loss with respect to the trainable parameters.

[0162] The above formula is based on the following: the prediction loss ensures that the predicted pain intensity value is close to the clinical standard; the decoupling loss ensures that the somatic pain latent variable and the visceral pain latent variable are independent of each other in the feature space. Since there is a target conflict between the two, overemphasizing the decoupling loss may reduce the prediction accuracy, while overemphasizing the prediction loss may lead to the re-overlapping of pain components; the decoupling loss balance coefficient dynamically adjusts the relative weights of the two losses in parameter updates, so that the network can achieve pain component decoupling while maintaining prediction accuracy; the gradient calculation is based on the fact that the gradient of the total loss with respect to the trainable parameters indicates the direction of the influence of parameter changes on the total loss, and updating the parameters along the negative gradient direction can make the total loss gradually decrease.

[0163] The pain assessment terminal updates the trainable parameters of the pain component decoupling network under training based on the gradient corresponding to the total loss, using either a stochastic gradient descent algorithm or an adaptive learning rate optimization algorithm. The formula is as follows:

[0164]

[0165] In the formula, Indicates the first In the next iteration, the pain component to be trained decouples all trainable parameters of the network. Indicates the learning rate. Indicates the first The gradient of the total loss with respect to the trainable parameters at the nth iteration. Indicates the first The trainable parameters are updated in the next iteration.

[0166] The learning rate is a hyperparameter that controls the step size of each parameter update, and the stochastic gradient descent algorithm is an optimization method that uses a batch of samples to calculate the gradient and update the parameters each time.

[0167] The above formula is based on the following: the backpropagation algorithm calculates the output of each layer and the total loss during the forward propagation process, and backpropagates the total loss from the output layer to the input layer along the network layers, calculating the gradient of the total loss with respect to each trainable parameter layer by layer; the stochastic gradient descent algorithm or adaptive learning rate optimization algorithm adjusts the trainable parameters according to the gradient direction and learning rate, so that the total loss is iteratively updated along the descent direction.

[0168] The pain assessment terminal determines whether the updated pain component decoupling network to be trained meets the preset convergence condition. When the total loss does not meet the preset convergence condition, the pain assessment terminal returns to the step of inputting the multimodal sample dataset into the pain component decoupling network to be trained, and continues the next round of forward propagation and parameter update. When the total loss meets the preset convergence condition, the pain assessment terminal determines the current parameters as the parameters of the trained pain component decoupling network.

[0169] The above processing is based on the fact that if convergence is judged solely on the total loss data of a single iteration, it is easy to make a misjudgment due to fluctuations during the training process. Therefore, the preset convergence conditions usually combine multiple indicators, such as the number of iterations, the magnitude of loss reduction, and the total loss data.

[0170] In one embodiment, a decoupling loss is calculated, which characterizes the correlation between somatic pain latent variables and visceral pain latent variables generated within the pain component decoupling network to be trained, including:

[0171] (1) Obtain the hidden variables of somatic pain and visceral pain from the pain component decoupling network to be trained.

[0172] For example, the pain assessment terminal locates the output buffer of the decoupled attention mechanism of the pain component decoupling network to be trained, and reads the somatic pain latent variables and visceral pain latent variables temporarily stored during the forward propagation from the output buffer.

[0173] The output buffer includes a high-speed buffer storage unit used to store intermediate tensors during the forward propagation of the pain component decoupling network to be trained.

[0174] (2) Perform nonlinear mapping on the latent variables of somatic pain to obtain the projection representation of somatic pain.

[0175] For example, the pain assessment terminal inputs the latent somatic pain variables into the linear transformation layer of the somatic pain projection sub-network. The linear transformation layer performs feature dimension alignment processing on the latent somatic pain variables. The dimension-aligned features are then input into the nonlinear activation layer of the somatic pain projection sub-network. The nonlinear activation layer performs nonlinear transformation processing on the dimension-aligned features. The nonlinearly transformed features are then input into the layer normalization layer of the somatic pain projection sub-network. The layer normalization layer performs distribution normalization processing on the nonlinearly transformed features to obtain the somatic pain projection representation.

[0176] The somatic pain projection subnetwork comprises a cascaded structure of linear transformation layers, nonlinear activation layers, and layer normalization layers. The linear transformation layer includes a weight matrix and bias vector for mapping somatic pain latent variables to the dimension of the common projection space. The nonlinear activation layer includes a nonlinear transformation function of a modified linear unit or a parameterized modified linear unit. The layer normalization layer includes a statistical normalization operation unit for eliminating feature distribution offsets. The somatic pain projection representation includes the standardized embedding vector of somatic pain latent variables in the common projection space. The common projection space includes a unified feature semantic space shared by the somatic pain projection representation and the visceral pain projection representation.

[0177] (3) Perform nonlinear mapping on the latent variables of visceral pain to obtain the projection representation of visceral pain.

[0178] For example, the pain assessment terminal inputs the latent visceral pain variables into the linear transformation layer of the visceral pain projection sub-network. The linear transformation layer performs feature dimension alignment processing on the latent visceral pain variables. The dimension-aligned features are then input into the nonlinear activation layer of the visceral pain projection sub-network. The nonlinear activation layer performs nonlinear transformation processing on the dimension-aligned features. The nonlinearly transformed features are then input into the layer normalization layer of the visceral pain projection sub-network. The layer normalization layer performs distribution standardization processing on the nonlinearly transformed features to obtain the visceral pain projection representation.

[0179] The visceral pain projection subnetwork comprises a cascaded structure of linear transformation layers, nonlinear activation layers, and layer normalization layers. The linear transformation layer includes a weight matrix and bias vector for mapping visceral pain latent variables to the common projection space dimension. The nonlinear activation layer includes a nonlinear transformation function with modified linear units or parameterized modified linear units. The layer normalization layer includes a statistical normalization operation unit for eliminating feature distribution offsets. The visceral pain projection representation includes the standardized embedding vector of visceral pain latent variables in the common projection space. The parameters of the visceral pain projection subnetwork and the somatic pain projection subnetwork are independent of each other to avoid mutual interference in the projection processes of the two types of pain features.

[0180] (4) Calculate the distribution difference measure between the somatic pain projection representation and the visceral pain projection representation. The distribution difference measure characterizes the distance between the somatic pain projection representation and the visceral pain projection representation in the feature space.

[0181] The distribution difference metric is used as the decoupling loss.

[0182] For example, the pain assessment terminal traverses all samples in the current batch, performs kernel mapping processing on the somatic pain projection representation of each sample, maps the somatic pain projection representation of each sample to the regenerating kernel Hilbert space, obtains the somatic pain kernel mapping vector corresponding to each sample, and performs arithmetic mean processing on the somatic pain kernel mapping vectors of all samples to obtain the mean embedding vector of the somatic pain projection representation.

[0183] The kernel mapping process includes a transformation operation that maps the projected representation to a high-dimensional feature space using a Gaussian kernel function or a Laplace kernel function. The regenerated kernel Hilbert space includes the complete inner product space induced by the kernel function.

[0184] The pain assessment terminal iterates through all samples in the current batch, performs kernel mapping processing on the visceral pain projection representation of each sample, maps the visceral pain projection representation of each sample to the regenerating kernel Hilbert space, and obtains the visceral pain kernel mapping vector corresponding to each sample. The arithmetic mean processing is performed on the visceral pain kernel mapping vectors of all samples to obtain the mean embedding vector of the visceral pain projection representation.

[0185] The kernel mapping processing for visceral pain projection representation and somatic pain projection representation uses the same kernel function to ensure that the two types of mean embedding vectors are in the same regenerating kernel Hilbert space. The mean embedding vector for visceral pain projection representation includes the expected representation of the probability distribution of visceral pain projection representation in the regenerating kernel Hilbert space.

[0186] The pain assessment terminal calculates the Euclidean distance between the mean embedding vector of the somatic pain projection representation and the mean embedding vector of the visceral pain projection representation in the regeneration kernel Hilbert space. The Euclidean distance is used as a measure of the distribution difference between the somatic pain projection representation and the visceral pain projection representation. The magnitude of the distribution difference measure is positively correlated with the distance between the two types of projection representations in the feature space.

[0187] The pain assessment terminal uses the distribution difference measure as the decoupling loss. The decoupling loss is used to adjust all trainable parameters of the pain component decoupling network to be trained in the subsequent backpropagation process through the gradient descent algorithm, so that the somatic pain latent variable and the visceral pain latent variable are independent of each other in the feature space.

[0188] like Figure 2 As shown, based on the predicted values ​​of somatic pain intensity and visceral pain intensity, a decoupled assessment of the components of post-gynecological postoperative pain is generated for the patient, including:

[0189] S201: Compare the predicted values ​​of somatic pain intensity and visceral pain intensity to determine the dominant pain component type, which includes somatic pain-dominant, visceral pain-dominant, and mixed types.

[0190] For example, the pain assessment terminal reads the generated predicted values ​​of somatic pain intensity and visceral pain intensity from the local cache, performs amplitude validity verification and alignment verification with the same monitoring period on the predicted values ​​of somatic pain intensity and visceral pain intensity in sequence, compares the amplitude of the two verified predicted values ​​bit by bit, matches the preset dominant type determination rules, and outputs the dominant pain component type.

[0191] Among them, amplitude validity verification includes removing outliers that exceed the pain rating scale range, and alignment verification for the same monitoring period includes ensuring that two predicted values ​​correspond to the same postoperative monitoring window. The preset dominant type determination rules include determining that a higher amplitude of somatic pain indicates a somatic pain-dominant type, a higher amplitude of visceral pain indicates a visceral pain-dominant type, and a difference in amplitude less than a preset threshold indicates a mixed type.

[0192] S202: Based on the type of dominant pain component and the numerical relationship between the predicted values ​​of somatic pain intensity and visceral pain intensity, pain component composition description information is generated. The pain component composition description information represents the composition ratio of somatic pain component and visceral pain component in postoperative gynecological pain.

[0193] For example, the pain assessment terminal uses the pain dominant component type as an index to perform a normalized summation operation on the verified predicted values ​​of somatic pain intensity and visceral pain intensity to obtain the somatic pain component proportion coefficient and the visceral pain component proportion coefficient. The two proportion coefficients are then input into a pre-trained natural language template mapping engine, which matches the description template corresponding to the pain dominant component type and fills in the proportion coefficients to generate a description of the pain component composition.

[0194] The natural language template mapping engine includes a pre-built corpus that stores descriptive sentence patterns corresponding to different dominant types.

[0195] S203: The results of the decoupling assessment of postoperative pain components in gynecological surgery include the type of dominant pain component and descriptive information on the composition of pain components.

[0196] For example, the pain assessment terminal creates a standardized assessment result structure, writes the type of the dominant pain component into the classification label field of the assessment result structure, writes the description information of the pain component into the description field of the assessment result structure, performs full-field compliance verification on the assessment result structure, and after the verification passes, encapsulates it into a decoupled assessment result of post-gynecological pain components.

[0197] The evaluation result structure includes a standardized data container adapted to the clinical electronic medical record system, and the full-field compliance verification includes two types of verification items: field integrity verification and data type matching verification.

[0198] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0199] In one embodiment, such as Figure 3 As shown, this application also provides a gynecological postoperative pain assessment system 300, which includes:

[0200] The multimodal data acquisition module 301 is used to simultaneously acquire multimodal physiological signals and facial video data of the patient to be evaluated during postoperative recovery. The multimodal physiological signals include electrocardiogram signals, skin conductance signals, abdominal surface electromyography signals, and respiratory signals.

[0201] The multimodal feature extraction module 302 is used to preprocess multimodal physiological signals and facial video data, and extract features from the preprocessed signals within a unified time window to generate multimodal feature vectors. The multimodal feature vectors include electrocardiogram features, skin conductance features, electromyography features, respiratory features, and facial expression dynamic features.

[0202] The pain component decoupling prediction module 303 is used to input multimodal feature vectors into a pre-constructed pain component decoupling network. The pain component decoupling network includes a somatic pain prediction head and a visceral pain prediction head. The pain component decoupling network performs feature fusion on the multimodal feature vectors to obtain a multimodal fused feature vector. The multimodal fused feature vector is dynamically allocated through a decoupling attention mechanism. The somatic pain intensity prediction value is output through the somatic pain prediction head, and the visceral pain intensity prediction value is output through the visceral pain prediction head.

[0203] The assessment result generation module 304 is used to generate decoupled assessment results of postoperative gynecological pain components for patients based on the predicted values ​​of somatic pain intensity and visceral pain intensity.

[0204] Specifically, the pain assessment terminal includes a multimodal data acquisition module 301, a multimodal feature extraction module 302, a pain component decoupling prediction module 303, and an assessment result generation module 304.

[0205] The multimodal data acquisition module 301 calls the hardware synchronization trigger interface to start the body surface sensor array to continuously sample the patient's electrocardiogram signal, skin conductance signal, abdominal surface electromyography signal, and respiratory signal during the postoperative recovery period. At the same time, the camera device is started to continuously acquire frames of the patient's facial area. The timestamps of the four types of physiological signals are aligned with the timestamps of the facial video data to obtain synchronized multimodal physiological signals and facial video data.

[0206] The multimodal feature extraction module 302 performs a preprocessing process of signal quality screening, noise filtering, and abnormal segment removal on the synchronized multimodal physiological signals and facial video data. It defines a unified time window covering the same postoperative monitoring period and extracts electrocardiogram, electrodermal, electromyographic, and respiratory features segment by segment from the preprocessed multimodal physiological signals. At the same time, it performs face detection, facial key point localization, and facial action unit intensity calculation frame by frame from the preprocessed facial video data. It performs temporal encoding on the facial action unit intensity sequence to obtain facial expression dynamic features. Then, it concatenates the five types of modal features into a multimodal feature vector according to a unified dimension.

[0207] The pain component decoupling prediction module 303 inputs the multimodal feature vector into the pre-constructed pain component decoupling network. The pain component decoupling network performs cross-modal gated weighted fusion processing on the multimodal feature vector to obtain a multimodal fused feature vector. It calls the decoupling attention mechanism built into the pain component decoupling network to perform dynamic allocation operations on the multimodal fused feature vector in sequence, including query vector generation, key-value vector generation, similarity calculation, weight distribution generation, and feature weighted splitting, to obtain somatic pain latent variables and visceral pain latent variables, respectively. Then, the somatic pain latent variables are input into the somatic pain prediction head built into the pain component decoupling network to perform nonlinear mapping and regression operations, and output the predicted value of somatic pain intensity. At the same time, the visceral pain latent variables are input into the visceral pain prediction head built into the pain component decoupling network to perform nonlinear mapping and regression operations, and output the predicted value of visceral pain intensity.

[0208] The assessment result generation module 304 retrieves the output predicted values ​​of somatic pain intensity and visceral pain intensity, performs a bitwise amplitude comparison operation on the two predicted values ​​to determine the type of dominant pain component, and performs a proportion quantification calculation based on the numerical relationship between the type of dominant pain component and the two predicted values ​​to generate pain component composition description information that characterizes the proportion of somatic pain component and visceral pain component in postoperative gynecological pain. Then, the pain dominant component type and pain component composition description information are encapsulated into a structured assessment result document that conforms to clinical data specifications, resulting in the decoupled assessment result of postoperative gynecological pain components.

[0209] The pain component decoupling prediction module 303 is also used for:

[0210] A nonlinear transformation is performed on the multimodal fusion feature vector to generate query vector and key value vector respectively;

[0211] Calculate the similarity between the query vector and the key value vector to obtain the weight distribution. The weight distribution represents the proportion of information corresponding to somatic pain mode and visceral pain mode in the multimodal fusion feature vector.

[0212] The multimodal fusion feature vector is weighted and split using weight distribution to obtain somatic pain latent variables and visceral pain latent variables;

[0213] Input the hidden variables of somatic pain into the somatic pain prediction head for pain intensity mapping processing, and output the predicted value of somatic pain intensity.

[0214] The latent variable of visceral pain is input into the visceral pain prediction head for pain intensity mapping processing, and the predicted value of visceral pain intensity is output.

[0215] The pain component decoupling prediction module 303 is also used for:

[0216] The multimodal fusion feature vector is subjected to a first linear projection to obtain a first linear projection vector, and a nonlinear activation is applied to the first linear projection vector to generate a query vector;

[0217] The multimodal fusion feature vector is subjected to a second linear projection to obtain a second linear projection vector, and a nonlinear activation is applied to the second linear projection vector to generate a key value vector.

[0218] The parameter matrices of the first linear projection and the second linear projection are learned independently during the training of the pain component decoupling network, enabling the query vector and key value vector distribution to capture discriminative information corresponding to different pain modes in the multimodal fusion feature vector.

[0219] The pain component decoupling prediction module 303 is also used for:

[0220] Obtain a multimodal sample dataset, which includes multimodal physiological signal samples and facial video data samples from multiple samples, as well as the corresponding somatic pain intensity annotation values ​​and visceral pain intensity annotation values ​​for each sample;

[0221] Each sample in the multimodal sample dataset is input into the pain component decoupling network to be trained, and the pain component decoupling network to be trained outputs the corresponding predicted values ​​of somatic pain intensity and visceral pain intensity.

[0222] Calculate the prediction loss, which represents the deviation between the predicted value of somatic pain intensity and the labeled value of somatic pain intensity, as well as the deviation between the predicted value of visceral pain intensity and the labeled value of visceral pain intensity.

[0223] Calculate the decoupling loss, which characterizes the correlation between the somatic pain latent variables and the visceral pain latent variables generated within the pain component decoupling network to be trained;

[0224] The parameters of the pain component decoupling network to be trained are updated based on the prediction loss and decoupling loss until the preset convergence condition is met, thus obtaining the trained pain component decoupling network.

[0225] The pain component decoupling prediction module 303 is also used for:

[0226] Obtain somatic pain latent variables and visceral pain latent variables from the pain component decoupling network to be trained;

[0227] Nonlinear mapping is applied to the latent variables of somatic pain to obtain the projected representation of somatic pain.

[0228] Nonlinear mapping is applied to the latent variables of visceral pain to obtain the projected representation of visceral pain;

[0229] Calculate the distribution difference measure between the somatic pain projection representation and the visceral pain projection representation. The distribution difference measure characterizes the distance between the somatic pain projection representation and the visceral pain projection representation in the feature space.

[0230] The distribution difference metric is used as the decoupling loss.

[0231] The evaluation result generation module 304 is also used for:

[0232] By comparing the predicted values ​​of somatic pain intensity and visceral pain intensity, the dominant pain component type was determined. The dominant pain component types include somatic pain dominant type, visceral pain dominant type, and mixed type.

[0233] Based on the dominant pain component type and the numerical relationship between the predicted values ​​of somatic pain intensity and visceral pain intensity, pain component composition description information is generated. This pain component composition description information represents the composition ratio of somatic pain component to visceral pain component in postoperative gynecological pain.

[0234] The results of the decoupling assessment of pain components after gynecological surgery include the type of dominant pain component and descriptive information on the composition of pain components.

[0235] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0236] In one embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0237] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0238] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for assessing postoperative pain in gynecological surgery, characterized in that, The method includes: Simultaneously acquire multimodal physiological signals and facial video data of the patient to be evaluated during postoperative recovery. The multimodal physiological signals include electrocardiogram signals, skin conductance signals, abdominal surface electromyography signals, and respiratory signals. The multimodal physiological signals and facial video data are preprocessed, and features are extracted from the preprocessed signals within a unified time window to generate multimodal feature vectors. The multimodal feature vectors include electrocardiogram features, skin conductance features, electromyography features, respiratory features, and facial expression dynamic features. The multimodal feature vector is input into a pre-constructed pain component decoupling network, which includes a somatic pain prediction head and a visceral pain prediction head. The pain component decoupling network performs feature fusion on the multimodal feature vector to obtain a multimodal fused feature vector. The multimodal fused feature vector is dynamically allocated through a decoupling attention mechanism. The somatic pain intensity prediction value is output through the somatic pain prediction head, and the visceral pain intensity prediction value is output through the visceral pain prediction head. Based on the predicted values ​​of somatic pain intensity and visceral pain intensity, a decoupled assessment result of the gynecological postoperative pain components for the patient is generated.

2. The method for assessing postoperative pain in gynecological surgery according to claim 1, characterized in that, The step of dynamically allocating the multimodal fusion feature vector through a decoupled attention mechanism, and outputting the predicted value of somatic pain intensity through the somatic pain prediction head and the predicted value of visceral pain intensity through the visceral pain prediction head, includes: The multimodal fusion feature vector is subjected to a nonlinear transformation to generate a query vector and a key-value vector, respectively. The similarity between the query vector and the key value vector is calculated to obtain the weight distribution, which represents the proportion of information corresponding to somatic pain mode and visceral pain mode in the multimodal fusion feature vector; The multimodal fusion feature vector is weighted and split using the weight distribution to obtain somatic pain latent variables and visceral pain latent variables; The hidden somatic pain variables are input into the somatic pain prediction head for pain intensity mapping processing, and the predicted somatic pain intensity value is output. The visceral pain latent variable is input into the visceral pain prediction head for pain intensity mapping processing, and the predicted value of visceral pain intensity is output.

3. The method for assessing postoperative pain in gynecological surgery according to claim 2, characterized in that, The step of performing a nonlinear transformation on the multimodal fused feature vector to generate a query vector and a key-value vector includes: The multimodal fusion feature vector is subjected to a first linear projection to obtain a first linear projection vector, and a nonlinear activation is applied to the first linear projection vector to generate the query vector; The multimodal fusion feature vector is subjected to a second linear projection to obtain a second linear projection vector, and a nonlinear activation is applied to the second linear projection vector to generate the key value vector; The parameter matrices of the first linear projection and the second linear projection are learned independently during the training of the pain component decoupling network, so that the query vector and the key value vector distribution capture the discriminative information corresponding to different pain modes in the multimodal fusion feature vector.

4. The method for assessing postoperative pain in gynecological surgery according to claim 1, characterized in that, The pre-training process of the pain component decoupling network includes: A multimodal sample dataset is obtained, which includes multimodal physiological signal samples and facial video data samples from multiple samples, as well as the corresponding somatic pain intensity labeling value and visceral pain intensity labeling value for each sample; Each sample in the multimodal sample dataset is input into the pain component decoupling network to be trained, and the pain component decoupling network to be trained outputs the corresponding predicted values ​​of somatic pain intensity and visceral pain intensity. Calculate the prediction loss, which represents the deviation between the predicted value of somatic pain intensity and the labeled value of somatic pain intensity, and the deviation between the predicted value of visceral pain intensity and the labeled value of visceral pain intensity; Calculate the decoupling loss, which characterizes the correlation between the somatic pain latent variables and the visceral pain latent variables generated within the pain component decoupling network to be trained; The parameters of the pain component decoupling network to be trained are updated based on the prediction loss and the decoupling loss until the preset convergence condition is met, thus obtaining the trained pain component decoupling network.

5. The method for assessing postoperative pain in gynecological surgery according to claim 4, characterized in that, The calculation of decoupling loss, which characterizes the correlation between somatic pain latent variables and visceral pain latent variables generated within the pain component decoupling network to be trained, includes: The somatic pain latent variable and the visceral pain latent variable are obtained from the pain component decoupling network to be trained; The somatic pain latent variables are subjected to nonlinear mapping processing to obtain the somatic pain projection representation; The latent variables of visceral pain are subjected to nonlinear mapping to obtain the projected representation of visceral pain. Calculate the distribution difference measure between the somatic pain projection representation and the visceral pain projection representation, wherein the distribution difference measure characterizes the distance between the somatic pain projection representation and the visceral pain projection representation in the feature space; The distribution difference metric is used as the decoupling loss.

6. The method for assessing postoperative pain in gynecological surgery according to claim 1, characterized in that, The decoupled assessment results of postoperative gynecological pain components for patients, generated based on the predicted values ​​of somatic pain intensity and visceral pain intensity, include: The predicted values ​​of somatic pain intensity and visceral pain intensity are compared to determine the dominant pain component type, which includes somatic pain dominant type, visceral pain dominant type, and mixed type. Based on the type of dominant pain component and the numerical relationship between the predicted values ​​of somatic pain intensity and visceral pain intensity, pain component composition description information is generated, which characterizes the composition ratio of somatic pain component and visceral pain component in postoperative gynecological pain. The decoupling assessment results of postoperative gynecological pain components include the type of dominant pain component and the descriptive information of the composition of the pain components.

7. A gynecological postoperative pain assessment system, characterized in that, The system includes: The multimodal data acquisition module is used to simultaneously acquire multimodal physiological signals and facial video data of the patient to be evaluated during postoperative recovery. The multimodal physiological signals include electrocardiogram signals, skin conductance signals, abdominal surface electromyography signals, and respiratory signals. The multimodal feature extraction module is used to preprocess the multimodal physiological signals and the facial video data, and extract features from the preprocessed signals within a unified time window to generate multimodal feature vectors. The multimodal feature vectors include electrocardiogram features, skin conductance features, electromyography features, respiratory features, and facial expression dynamic features. The pain component decoupling prediction module is used to input the multimodal feature vector into a pre-constructed pain component decoupling network. The pain component decoupling network includes a somatic pain prediction head and a visceral pain prediction head. The pain component decoupling network performs feature fusion on the multimodal feature vector to obtain a multimodal fused feature vector. The multimodal fused feature vector is dynamically allocated through a decoupling attention mechanism. The somatic pain intensity prediction value is output through the somatic pain prediction head, and the visceral pain intensity prediction value is output through the visceral pain prediction head. The assessment result generation module is used to generate a decoupled assessment result of the gynecological postoperative pain components for the patient based on the predicted values ​​of somatic pain intensity and visceral pain intensity.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the gynecological postoperative pain assessment method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for assessing postoperative gynecological pain according to any one of claims 1 to 6.

Images