Postoperative pain monitoring method based on wearable flexible sensor

Abstract

The invention discloses a postoperative pain monitoring method based on a wearable flexible sensor, and relates to the technical field of postoperative pain monitoring. According to the method, a postoperative pain monitoring model mapping relation is constructed based on patient basic feature data in a patient information database, and a corresponding postoperative pain monitoring model is determined based on the postoperative pain monitoring model mapping relation; patient pain feature data of a postoperative patient is input into a postoperative pain monitoring model, pain level monitoring is performed, a postoperative pain monitoring model mapping relation is constructed by integrating historical patient physiological data and individual features, and a dual mechanism of rapid screening of a first classification model and accurate verification of a secondary multimode monitoring model is adopted. According to the invention, the method achieves the grading monitoring of the pain grade of a postoperative patient, enhances the adaptability to different patient groups based on the model adaptation of individual features, and solves the problems that the existing postoperative pain monitoring method is too limited in single-mode evaluation, and is insufficient in individual difference adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of postoperative pain monitoring, and in particular to a postoperative pain monitoring method based on a wearable flexible sensor. Background Art

[0002] Currently, the most common method for assessing a patient's pain level is to use a self-rating pain scale filled out by the patient. However, due to differences in the conceptualization of pain among patients, it can be difficult for patients to accurately complete the self-rating scale, making it difficult for doctors to determine the patient's true pain level. Another method uses a neural network to determine whether the patient's facial expression is painful based on facial images to determine the patient's pain level.

[0003] A Chinese patent application with publication number CN119831983A discloses a method for auxiliary identification of post-anesthesia pain in patients based on near-infrared technology. The method includes: obtaining near-infrared images of the surgical site of a patient to be detected at historical moments and current moments after anesthesia; determining the temperature rise coefficient, temperature fluctuation factor and post-operative pain factor corresponding to each pixel point in the near-infrared image; determining the pain occurrence factor corresponding to each pixel point in the current near-infrared image based on the difference between the temperature rise coefficient, the post-operative pain factor and the temperature rise coefficient corresponding to the pixel points at the same position; determining the target pain level corresponding to each pixel point in the current near-infrared image based on the pain occurrence factor and the post-operative pain factor corresponding to each pixel point in the current near-infrared image.

[0004] However, existing postoperative pain monitoring methods have the problem that single-modality assessment is too limited and lacks adaptability to individual differences. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a postoperative pain monitoring method based on a wearable flexible sensor, which solves the problems of the existing postoperative pain monitoring method, such as the single modality evaluation is too limited and the adaptability to individual differences is insufficient.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a postoperative pain monitoring method based on a wearable flexible sensor, comprising the following steps: collecting historical patient physiological data based on a wearable flexible sensor, and simultaneously obtaining historical patient individual characteristic information, integrating it into historical patient basic characteristic data, and storing it in a patient information database; constructing a postoperative pain monitoring model mapping relationship based on the patient basic characteristic data in the patient information database, the postoperative pain monitoring model including a pain level classification model and a pain level multimodal monitoring model; obtaining the patient basic characteristic data of the postoperative patient, and determining the corresponding postoperative pain monitoring model based on the postoperative pain monitoring model mapping relationship, the patient basic characteristic data of the postoperative patient including patient pain characteristic data and patient individual characteristic information; inputting the patient pain characteristic data of the postoperative patient into the postoperative pain monitoring model to perform pain level monitoring, wherein: the pain level classification model is used to determine the pain level for the first time, and if the pain level cannot be determined, the pain level is determined for the second time through the pain level multimodal monitoring model.

[0007] Furthermore, the wearable flexible sensor includes a flexible pressure sensor, a brain wave collector, a heart rate collector, an electromyographic signal sensor, and a flexible body temperature sensor; the historical patient physiological data includes historical patient surface pressure data, historical patient brain wave data, historical patient heart rate data, historical patient electromyographic data, and historical patient temperature data; the historical patient individual characteristic information includes the historical patient's age, body mass index, and underlying disease history;

[0008] The integration into basic characteristic data of historical patients includes the following steps: extracting time domain features, frequency domain features and time-frequency domain features of historical patients' physiological data to obtain time domain features, frequency domain features and time-frequency domain features of historical patients; preprocessing individual characteristic information of historical patients: removing units from age, encoding basic disease history to obtain encoded basic disease history, and obtaining processed individual characteristic information of historical patients; splicing the time domain features, frequency domain features, time-frequency domain features and processed individual characteristic information of historical patients to obtain basic characteristic data of historical patients.

[0009] Furthermore, a postoperative pain monitoring model mapping relationship is constructed based on the patient basic characteristic data in the patient information database, including the following steps: obtaining a determined first neighborhood radius and a minimum number of samples of the first core point, performing cluster analysis on the patient basic characteristic data in the patient information database based on the DBSCAN algorithm, and determining several first clusters; determining the patient basic characteristic data corresponding to each first cluster and performing feature processing to obtain the processed patient basic characteristic data corresponding to each first cluster, including processed patient physiological data and processed patient individual characteristic information; constructing a model based on the historical patient time domain characteristics, historical patient frequency domain characteristics and historical patient time-frequency domain characteristics in the patient basic characteristic data corresponding to each first cluster, and obtaining a pain level classification model and a pain level multimodal monitoring model; making one-to-one correspondence between the processed patient basic characteristic data corresponding to each first cluster and the corresponding pain level classification model and the pain level multimodal monitoring model to obtain a postoperative pain monitoring model mapping relationship.

[0010] Furthermore, the processed patient basic characteristic data corresponding to each first cluster is obtained, including the following steps: performing mean processing on the historical patient time domain characteristics, historical patient frequency domain characteristics and historical patient time-frequency domain characteristics in the patient basic characteristic data corresponding to each first cluster, respectively, to obtain averaged historical patient time domain characteristics, averaged historical patient frequency domain characteristics and averaged historical patient time-frequency domain characteristics, and the combination is recorded as processed patient physiological data; processing the processed historical patient individual characteristic information in the patient basic characteristic data corresponding to each first cluster: performing mean processing on age and body mass index to obtain averaged age and averaged body mass index; counting the number of coded basic disease histories, and taking the top N coded basic disease histories as the coded basic disease history set; and recording the averaged age, averaged body mass index and the coded basic disease history set as the processed patient individual characteristic information.

[0011] Furthermore, a pain level classification model is obtained, comprising the following steps: recording the historical patient time domain features, historical patient frequency domain features, and historical patient time-frequency domain features in the patient basic feature data as historical patient pain feature data, and annotating the historical patient pain feature data with pain levels; obtaining a determined second neighborhood radius and a second core point minimum sample number, performing cluster analysis on the annotated historical patient pain feature data based on the DBSCAN algorithm, and obtaining a plurality of second clusters; counting the pain level distribution data corresponding to the historical patient pain feature data in each second cluster, that is, the proportion of each pain level, and determining the pain level with the largest proportion from the pain level distribution data, recording it as the pain level corresponding to the second cluster, and establishing an association;

[0012] If there is only one second cluster associated with a certain pain level, the historical patient pain feature data in the second cluster is averaged to obtain the averaged historical patient pain feature data, and the averaged historical patient pain feature data is associated with the corresponding pain level to form a pain feature-pain level branch;

[0013] If there is more than one second cluster associated with a certain pain level, the historical patient pain feature data in each second cluster is averaged to obtain averaged historical patient pain feature data. A group of averaged historical patient pain feature data with the smallest average distance from other averaged historical patient pain feature data is determined based on Euclidean distance, and the averaged historical patient pain feature data is associated with the corresponding pain level to form a pain feature-pain level branch.

[0014] If the second cluster associated with a certain pain level is zero, then the annotated historical patient pain feature data corresponding to the pain level is obtained from the adjacent second cluster associated with the pain level, the annotated historical patient pain feature data is averaged to obtain the averaged historical patient pain feature data, and the averaged historical patient pain feature data is associated with the corresponding pain level to form a pain feature-pain level branch;

[0015] All pain feature-pain level branches are arranged in order from small to large pain levels to obtain a pain level classification model.

[0016] Furthermore, a multimodal monitoring model for pain levels is obtained, comprising the following steps: training a single model layer based on historical patient pain feature data, including a support vector machine layer, a random forest layer and a multilayer perceptron layer, and the support vector machine layer, the random forest layer and the multilayer perceptron layer each have only one output, including the support vector machine pain level prediction result, the random forest pain level prediction result and the multilayer perceptron pain level prediction result; collecting statistics on the performance data of the single model layer, including the support vector machine layer performance data, the random forest layer performance data and the multilayer perceptron layer performance data; determining the credibility of the prediction results of the support vector machine layer, the random forest layer and the multilayer perceptron layer based on the single model layer performance data; and obtaining the output of the multimodal monitoring model for pain levels based on the support vector machine pain level prediction result, the random forest pain level prediction result, the multilayer perceptron pain level prediction result and the credibility of the prediction results.

[0017] Furthermore, the support vector machine layer performance data includes the SVM accuracy Zq SVM , SVM recall rate Zh SVM , SVMF1 value F1 SVM ; Random forest layer performance data includes RF accuracy Zq RF , RF recall rate Zh RF 、RFF1 value F1RF ; Multilayer perceptron layer performance data including ANN accuracy Zq ANN , ANN recall rate Zh ANN ,ANNF1 value F1 ANN ;

[0018] Determining the credibility of the prediction results of the support vector machine layer, the random forest layer, and the multi-layer perceptron layer includes the following steps: calculating the comprehensive scores of the support vector machine layer, the random forest layer, and the multi-layer perceptron layer; normalizing the comprehensive scores of each model to obtain the credibility of the prediction results of the support vector machine layer, the random forest layer, and the multi-layer perceptron layer.

[0019] Furthermore, the comprehensive scores of the support vector machine layer, random forest layer, and multilayer perceptron layer are calculated as follows:

[0020]

[0021] Among them, F SVM is the comprehensive score of the support vector machine layer, F RF is the comprehensive score of the random forest layer, F ANN is the composite score of the multilayer perceptron layer.

[0022] Furthermore, the calculation method for the credibility of the prediction results of the support vector machine layer, random forest layer and multi-layer perceptron layer is:

[0023]

[0024] Among them, K SVM is the reliability of the prediction results of the support vector machine layer, K RF is the credibility of the prediction results of the random forest layer, K ANN is the confidence level of the prediction result of the multi-layer perceptron layer.

[0025] Furthermore, the pain level classification model determines the pain level for the first time, including the following steps: comparing the basic feature data of the postoperative patient with the processed basic feature data of the patient in the mapping relationship of the postoperative pain monitoring model one by one to obtain a cosine similarity value comparison value; determining the maximum cosine similarity comparison value, and determining the corresponding pain level based on the maximum cosine similarity comparison value; obtaining the attribution threshold corresponding to the pain level, if the maximum cosine similarity comparison value is greater than the attribution threshold, the first determination of the pain level is successful, if the maximum cosine similarity comparison value is not greater than the attribution threshold, the first determination of the pain level is unsuccessful, that is, the pain level cannot be determined.

[0026] The present invention has the following beneficial effects:

[0027] This postoperative pain monitoring method based on wearable flexible sensors integrates historical patient physiological data and individual characteristics to construct a mapping relationship for the postoperative pain monitoring model. It uses a dual mechanism of rapid screening of the first classification model and precise verification of the second multimodal monitoring model to achieve graded monitoring of the postoperative patient's pain level. At the same time, model adaptation based on individual characteristics enhances the adaptability to different patient groups, solving the problems of existing postoperative pain monitoring methods, such as the single modality evaluation being too limited and the lack of adaptability to individual differences.

[0028] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of the postoperative pain monitoring method based on a wearable flexible sensor of the present invention.

[0030] Figure 2 The present invention is a flowchart for constructing a postoperative pain monitoring model mapping relationship for a postoperative pain monitoring method based on a wearable flexible sensor.

[0031] Figure 3 The present invention is a flow chart of a multimodal pain level monitoring model obtained by a postoperative pain monitoring method based on a wearable flexible sensor. DETAILED DESCRIPTION

[0032] See also Figure 1 The embodiment of the present invention provides a technical solution: a postoperative pain monitoring method based on a wearable flexible sensor, comprising the following steps: collecting historical patient physiological data based on the wearable flexible sensor, and simultaneously obtaining historical patient individual characteristic information, integrating it into historical patient basic characteristic data, and storing it in a patient information database.

[0033] Wearable flexible sensors include flexible pressure sensors (attached to the patient's back, buttocks and other postural support areas, collecting contact pressure values ​​at a frequency of 5 times / minute), EEG collectors (collecting EEG signals in α, β, γ and other frequency bands through forehead electrodes), heart rate collectors, myoelectric signal sensors and flexible body temperature sensors (attached to the skin surface of the armpit or clavicle, monitoring body temperature at a frequency of 1 time / minute to exclude physiological fluctuations caused by non-pain factors such as infection); historical patient physiological data include historical patient surface pressure data, historical patient EEG data, historical patient heart rate data, historical patient EMG data and historical patient temperature data; historical patient individual characteristic information includes historical patient age, body mass index and underlying disease history;

[0034] The integration into basic characteristic data of historical patients includes the following steps: extracting time domain features, frequency domain features and time-frequency domain features of historical patient physiological data to obtain time domain features, frequency domain features and time-frequency domain features of historical patients; preprocessing individual characteristic information of historical patients: removing units from age, encoding basic disease history (using One-Hot Encoding to convert disease type into binary feature vector, for example, "hypertension" is encoded as [1,0,0] and "diabetes" is encoded as [0,1,0]), obtaining encoded basic disease history, and obtaining processed individual characteristic information of historical patients; splicing the time domain features, frequency domain features, time-frequency domain features and processed individual characteristic information of historical patients to obtain basic characteristic data of historical patients.

[0035] By coupling analysis of characteristics such as age, BMI, and history of underlying diseases with physiological data, the pain response patterns of different groups can be identified (for example, the heart rate increase of elderly patients when in pain may be smaller than that of young patients, but the changes in EEG delta waves are more significant), providing a basis for subsequent clustering and personalized model training.

[0036] For example, the parameters included in the historical patient time domain features, historical patient frequency domain features, and historical patient time-frequency domain features are as follows:

[0037]

[0038]

[0039] Through multimodal sensor array acquisition, multidimensional feature extraction, and structured individual information integration, a standardized dataset covering physiological responses and individual characteristics was constructed. This process not only addresses the subjectivity and lag issues of traditional assessment methods, but also, through data-driven feature engineering, provides high-quality input for subsequent cluster analysis and model training. Ultimately, this enables real-time, objective, and personalized monitoring of postoperative pain, facilitating the precise formulation of clinical analgesia plans.

[0040] A postoperative pain monitoring model mapping relationship is constructed based on the patient's basic characteristic data in the patient information database. The postoperative pain monitoring model includes a pain level classification model and a pain level multimodal monitoring model. Traditional pain assessment relies on patient self-reports (such as VAS scores), which are subject to subjective bias and lag. This solution forms an objective mapping relationship of "physiological response-pain level" through cross-validation of multimodal data such as pressure, EEG, heart rate, electromyography, and body temperature to avoid misjudgment of a single signal.

[0041] like Figure 2As shown, the first neighborhood radius and the minimum number of samples of the first core point are obtained (which can be set based on the density of the data set, the formula is MinPts = floor (log2 (N)) (N is the total number of samples). If the data set contains 500 patients, then MinPts = 9, ensuring that each cluster contains at least enough samples for model training), and the basic characteristic data of patients in the patient information database are clustered and analyzed based on the DBSCAN algorithm to determine several first clusters.

[0042] The radius of the first neighborhood can be determined by the "k-distance graph" method, drawing a distance curve from all sample points to their kth nearest neighbor (k is MinPts-1), and the distance value corresponding to the curve mutation point is used as the initial ε. For example, when k = 7 (MinPts = 8), if the inflection point where the curve turns from steep to flat is 0.8, the initial value of ε is set to 0.8, and then fine-tuned by ±0.1 through cross-validation. The minimum number of samples for the first core point can be set based on the density of the data set, and the formula is MinPts = floor(log2(N)) (N is the total number of samples). If the data set contains 500 patients, then MinPts = 9 to ensure that each cluster contains at least enough samples for model training.

[0043] Determine the basic patient characteristic data corresponding to each first cluster and perform feature processing to obtain the processed patient basic characteristic data corresponding to each first cluster, including the processed patient physiological data and the processed patient individual characteristic information; construct a model based on the historical patient time domain characteristics, historical patient frequency domain characteristics and historical patient time-frequency domain characteristics in the patient basic characteristic data corresponding to each first cluster, and obtain a pain level classification model and a pain level multimodal monitoring model; make one-to-one correspondence between the processed patient basic characteristic data corresponding to each first cluster and the corresponding pain level classification model and the pain level multimodal monitoring model to obtain a mapping relationship of the postoperative pain monitoring model.

[0044] DBSCAN was used to divide patients into subgroups with similar physiological response patterns (such as the "elderly + hypertension + high myoelectric RMS" cluster and the "young + healthy + high EEG β wave" cluster) to avoid the mixing of characteristics of different groups leading to a decrease in the model's generalization ability.

[0045] Obtaining processed patient basic characteristic data corresponding to each first cluster includes the following steps: performing mean processing on the historical patient time domain characteristics, historical patient frequency domain characteristics, and historical patient time-frequency domain characteristics in the patient basic characteristic data corresponding to each first cluster, obtaining averaged historical patient time domain characteristics, averaged historical patient frequency domain characteristics, and averaged historical patient time-frequency domain characteristics, which are combined and recorded as processed patient physiological data; for example, the time domain mean (HR_mean) and frequency domain LF power mean (LF_power_mean) of the heart rate data in a cluster form a "standard physiological characteristic template" for the cluster. The mean processing eliminates individual data fluctuations and forms a "typical physiological response pattern" within the cluster.

[0046] Through mean processing, data fluctuations between individuals are compressed and common features within the cluster are retained. For example, thousands of original physiological data points can be simplified into dozens of mean features, reducing the model input dimension and improving training efficiency.

[0047] The processed historical patient individual characteristic information in the patient basic characteristic data corresponding to each first cluster is processed: the age and body mass index are averaged to obtain the averaged age and averaged body mass index; the number of coded basic disease histories is counted, and the top N coded basic disease histories are taken as the coded basic disease history set; the top N (such as N = 3) disease codes with the highest frequency of occurrence in the cluster are counted, such as "hypertension (45%), diabetes (30%), coronary heart disease (15%)" to form a disease set, reflecting the comorbidity characteristics of the patients in this cluster.

[0048] The averaged age, averaged body mass index and coded underlying disease history were recorded as the individual characteristic information of the patients after treatment.

[0049] The processed feature data of each cluster represents a class of patients with similar physiological basis and pain response patterns. The model trained based on this can capture the pain characteristics of this group in a targeted manner. Through the mean aggregation of physiological characteristics and high-frequency statistics of individual characteristics, the generated processed patient basic feature data realizes the transformation from "raw data" to "clinically interpretable feature templates." This process not only solves the problem of difficulty in quantifying individual differences in traditional assessments, but also improves the model's adaptability to different patient groups through structured feature representation, providing technical support that is both scientific and clinically practical for the rapid and accurate assessment of postoperative pain.

[0050] Obtaining a pain level classification model includes the following steps: recording the historical patient time domain features, historical patient frequency domain features, and historical patient time-frequency domain features in the patient basic feature data as historical patient pain feature data, and annotating the historical patient pain feature data with pain levels; pain level annotation can use clinical standard pain assessment methods (such as the numerical rating method NRS) to annotate the historical patient physiological feature data, and divide the pain level into mild (1-3 points), moderate (4-6 points), and severe (7-10 points). When annotating, the pain onset time window recorded by medical staff (such as assessment every 1 hour within 24 hours after surgery) is combined to ensure the time correspondence between the physiological feature data and the pain level.

[0051] Obtain the determined second neighborhood radius and the minimum number of samples of the second core point, perform cluster analysis on the annotated historical patient pain feature data based on the DBSCAN algorithm, and obtain multiple second clusters; calculate the pain level distribution data corresponding to the historical patient pain feature data in each second cluster, that is, the proportion of each pain level, and determine the pain level with the largest proportion from the pain level distribution data, record it as the pain level corresponding to the second cluster, and establish an association;

[0052] If there is only one second cluster associated with a certain pain level, the historical patient pain feature data in the second cluster is averaged to obtain the averaged historical patient pain feature data, and the averaged historical patient pain feature data is associated with the corresponding pain level to form a pain feature-pain level branch; the moderate pain cluster contains 300 samples, whose electromyography RMS mean is 45μV and the heart rate mean is 85 beats / minute. [45μV, 85 beats / minute,...] is used as the feature template of moderate pain, and a direct mapping of "moderate pain-feature mean" is established.

[0053] If there is more than one second cluster associated with a certain pain level, the historical patient pain feature data in each second cluster is averaged to obtain averaged historical patient pain feature data. A group of averaged historical patient pain feature data with the smallest average distance from other averaged historical patient pain feature data is determined based on Euclidean distance, and the averaged historical patient pain feature data is associated with the corresponding pain level to form a pain feature-pain level branch.

[0054] If the second cluster associated with a certain pain level is zero, then the annotated historical patient pain feature data corresponding to the pain level is obtained from the adjacent second cluster associated with the pain level, the annotated historical patient pain feature data is averaged to obtain the averaged historical patient pain feature data, and the averaged historical patient pain feature data is associated with the corresponding pain level to form a pain feature-pain level branch;

[0055] Arrange all pain feature-pain level branches in ascending order of pain level to obtain a pain level classification model. Build a spatial index of the feature template (e.g., a KD tree). After a new patient feature is input, use a nearest neighbor search (NNSearch) to quickly locate the branch with the smallest distance.

[0056] Clustering integrates multi-dimensional features such as pressure, EEG, and heart rate to avoid misjudgment of single signals. Abandoning traditional assessment methods that rely on patients' subjective descriptions, the automated process of "physiological feature clustering and hierarchical association" enables objective determination of pain levels. For example, a patient with an average RMS electromyography value of 55μV and a heart rate of 95 beats / minute can be directly diagnosed as experiencing severe pain through feature template matching, avoiding missed diagnoses due to patients being unconscious and unable to report the pain.

[0057] like Figure 3 As shown, a multimodal monitoring model for pain level is obtained, which includes the following steps: training a single model layer based on historical patient pain feature data, including a support vector machine layer, a random forest layer and a multilayer perceptron layer, and the support vector machine layer, the random forest layer and the multilayer perceptron layer each have only one output, including the support vector machine pain level prediction result, the random forest pain level prediction result and the multilayer perceptron pain level prediction result.

[0058] Support vector machine layer training:

[0059] Input the normalized feature data. Since the data is linearly inseparable, the radial basis function (RBF) is selected as the kernel function. Assuming that the penalty factor is 2 after multiple adjustments and the kernel function parameter gamma is 0.1, the model will output the probability value of each data sample belonging to different pain levels (mild pain, moderate pain, severe pain) after training. In order to obtain a unique output, the "maximum probability method" is used, that is, the pain level with the highest probability value is selected as the final prediction result of the sample. For example, the probability of mild pain in a sample is 0.3, the probability of moderate pain is 0.6, and the probability of severe pain is 0.1, then the sample is finally predicted by SVM as moderate pain. After predicting the entire test set data, the predicted results are compared one by one with the actual pain level label, and the number of samples predicted correctly is counted.

[0060] Random Forest layer training:

[0061] 500 subsets were extracted from the original training data with replacement, and a decision tree was trained on each subset. When each decision tree was constructed, half of the features were randomly selected for node splitting. After training, RF outputs the predicted pain level for each data sample. Since each decision tree has a prediction result, the "voting method" is used here to integrate the results of all decision trees. That is, the number of pain levels predicted by all decision trees is counted, and the pain level with the most votes is used as the final prediction result for the sample. For example, if 300 of the 500 decision trees predict moderate pain, 150 predict mild pain, and 50 predict severe pain, then the sample will ultimately be predicted by RF as moderate pain.

[0062] Multilayer Perceptron layer training:

[0063] A multilayer perceptron (MLP) architecture was used, with three hidden layers containing 10, 8, and 6 nodes, respectively. Reluctant Unit (ReLU) activation function was used. Training was performed using the backpropagation algorithm. After 100 iterations, the ANN outputted the probability of each data sample belonging to a different pain level. Similar to the SVM, the maximum probability method was used to determine a unique output, with the pain level with the highest probability selected as the predicted outcome. For example, if a sample had a probability of 0.2 for mild pain, 0.7 for moderate pain, and 0.1 for severe pain, the ANN would ultimately predict the sample as experiencing moderate pain.

[0064] The performance data of the single model layers were collected, including the support vector machine layer, random forest layer, and multi-layer perceptron layer. The random forest layer, random forest layer, and multi-layer perceptron layer were validated based on the validation datasets. Specifically, the validation datasets included validation datasets corresponding to mild, moderate, and severe pain, and moderate and severe pain, respectively. The single model layer performance data corresponding to the three single models at different pain levels were obtained.

[0065] Support vector machine layer performance data includes SVM accuracy Zq SVM , SVM recall rate Zh SVM , SVMF1 value F1 SVM ,Further, the support vector machine layer performance data includes the performance data of each pain level, that is, if the pain level includes mild, moderate and severe, then the support vector machine layer performance data includes the accuracy Zq corresponding to mild, moderate and severe pain levels respectively. SVM , SVM recall rate Zh SVM , SVMF1 value F1 SVM , the random forest layer performance data below is the same as the multilayer perceptron layer performance data.

[0066] Random Forest layer performance data including RF accuracy Zq RF , RF recall rate ZhRF 、RFF1 value F1 RF ; Multilayer perceptron layer performance data including ANN accuracy Zq ANN , ANN recall rate Zh ANN ,ANNF1 value F1 ANN ;

[0067] Determining the credibility of prediction results of the support vector machine layer, the random forest layer, and the multilayer perceptron layer includes the following steps: calculating a comprehensive score of the support vector machine layer, the random forest layer, and the multilayer perceptron layer;

[0068]

[0069] Among them, F SVM is the comprehensive score of the support vector machine layer, F RF is the comprehensive score of the random forest layer, F ANN is the composite score of the multilayer perceptron layer.

[0070] The comprehensive scores of each model were normalized to obtain the credibility of the prediction results of the support vector machine layer, random forest layer and multi-layer perceptron layer.

[0071]

[0072] Among them, K SVM is the reliability of the prediction results of the support vector machine layer, K RF is the credibility of the prediction results of the random forest layer, K ANN is the confidence level of the prediction result of the multi-layer perceptron layer.

[0073] The credibility of the prediction results of the support vector machine layer, random forest layer and multi-layer perceptron layer is determined based on the performance data of the single model layer; the output of the multimodal monitoring model of pain level is obtained based on the support vector machine pain level prediction results, random forest pain level prediction results, multi-layer perceptron pain level prediction results and the credibility of the prediction results.

[0074] For example, the prediction results of the support vector machine layer, random forest layer, and multilayer perceptron layer are mild pain, moderate pain, and moderate pain, respectively. The confidence levels of the prediction results of the support vector machine layer, random forest layer, and multilayer perceptron layer are 0.3, 0.3, and 0.4, respectively. According to the weighted calculation, the mild pain score is 0.3 × 1 = 0.3, and the moderate pain score is 0.3 × 1 + 0.4 × 1 = 0.7. The final pain level is determined to be moderate pain.

[0075] The multimodal pain level monitoring model addresses the limitations of single models in pain assessment by integrating multiple algorithms and dynamically assigning weights. This improves the accuracy, robustness, and clinical adaptability of postoperative pain assessment. This solution not only provides more reliable decision-making support for medical staff but also, through data-driven automated processes, advances the technical advancement of postoperative analgesia from "empirical adjustment" to "precise quantitative management."

[0076] Obtain the basic characteristic data of the postoperative patient, and determine the corresponding postoperative pain monitoring model based on the mapping relationship of the postoperative pain monitoring model. The basic characteristic data of the postoperative patient includes the patient pain characteristic data and the patient individual characteristic information.

[0077] The postoperative pain characteristic data of the patients were input into the postoperative pain monitoring model for pain level monitoring, wherein: the pain level classification model was used to determine the pain level for the first time. If the pain level could not be determined, the pain level was determined for the second time through the pain level multimodal monitoring model.

[0078] The pain level classification model determines the pain level for the first time, including the following steps: comparing the basic feature data of the postoperative patient with the processed basic feature data of the patient in the mapping relationship of the postoperative pain monitoring model one by one to obtain a cosine similarity value comparison value.

[0079] Example: A postoperative patient is 62 years old (62 without the unit), has a BMI of 23.5 (after averaging), has no underlying diseases (coded as [0,0,0]), has an EMG RMS of 48μV, a heart rate of 88 beats / minute, and an EEG alpha wave power of 12μV. 2 / Hz, forming the feature vector [62, 23.5, 0, 0, 0, 48, 88, 12, ...]. Traverse all clusters and record the maximum similarity value and its corresponding cluster label (e.g., "Cluster 3: Moderate Pain"). The similarity between the patient's characteristics and "Cluster 3" is 0.82, and the similarity with "Cluster 2: Mild Pain" is 0.65, with the maximum similarity corresponding to Cluster 3.

[0080] Determine the maximum cosine similarity comparison value, and determine the corresponding pain level based on the maximum cosine similarity comparison value; obtain the attribution threshold corresponding to the pain level, if the maximum cosine similarity comparison value is greater than the attribution threshold, the initial determination of the pain level is successful; if the maximum cosine similarity comparison value is not greater than the attribution threshold, the initial determination of the pain level is unsuccessful, that is, the pain level cannot be determined.

[0081] For example, if the maximum similarity is greater than the threshold (e.g., 0.82 > 0.80), the patient is determined to belong to that cluster, and the "pain feature-level branch" of the corresponding pain level classification model is directly called to output the pain level (e.g., moderate pain). If the maximum similarity is less than or equal to the threshold (e.g., 0.78 ≤ 0.80), the patient's characteristics are considered insufficiently matched with the existing clusters, triggering a secondary evaluation of the multimodal monitoring model.

[0082] Each cluster corresponds to the characteristic template of a specific patient group. After matching, the targeted training classification model is directly called to avoid errors caused by "one-size-fits-all" evaluation.

[0083] An electronic device includes: a processor; and a memory, wherein computer program instructions are stored in the memory, and when the computer program instructions are executed by the processor, the processor executes the postoperative pain monitoring method based on the wearable flexible sensor as described above.

[0084] A computer-readable storage medium is used to store a program, which, when executed by a processor, implements the postoperative pain monitoring method based on a wearable flexible sensor as described above.

[0085] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0086] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0087] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0088] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0089] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0090] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A postoperative pain monitoring method based on a wearable flexible sensor, characterized in that: The following steps are involved: Based on wearable flexible sensors, historical patient physiological data is collected, and at the same time, individual characteristic information of historical patients is obtained, integrated into historical patient basic characteristic data, and stored in the patient information database; Constructing a postoperative pain monitoring model mapping relationship based on the patient basic feature data in the patient information database, wherein the postoperative pain monitoring model includes a pain level classification model and a pain level multimodal monitoring model; Obtaining basic patient characteristic data of the postoperative patient, and determining a corresponding postoperative pain monitoring model based on a mapping relationship of the postoperative pain monitoring model, wherein the basic patient characteristic data of the postoperative patient includes patient pain characteristic data and patient individual characteristic information; The postoperative pain characteristic data of the patients were input into the postoperative pain monitoring model for pain level monitoring, wherein: the pain level classification model was used to determine the pain level for the first time. If the pain level could not be determined, the pain level was determined for the second time through the pain level multimodal monitoring model.

2. The postoperative pain monitoring method based on a wearable flexible sensor according to claim 1, characterized in that: Wearable flexible sensors include flexible pressure sensors, brain wave collectors, heart rate collectors, myoelectric signal sensors, and flexible body temperature sensors; Historical patient physiological data includes historical patient surface pressure data, historical patient brain wave data, historical patient heart rate data, historical patient electromyography data, and historical patient temperature data; historical patient individual characteristic information includes historical patient age, body mass index, and underlying disease history; Integration into historical patient basic characteristic data includes the following steps: Extracting time domain features, frequency domain features and time-frequency domain features of historical patient physiological data to obtain historical patient time domain features, historical patient frequency domain features and historical patient time-frequency domain features; Preprocessing of historical patient individual characteristic information: De-unitize the age and encode the basic disease history to obtain the coded basic disease history and the processed individual characteristic information of the historical patients; The historical patients' time domain features, historical patients' frequency domain features, historical patients' time-frequency domain features and processed historical patients' individual feature information are spliced ​​to obtain the historical patients' basic feature data.

3. The postoperative pain monitoring method based on a wearable flexible sensor according to claim 2, characterized in that: Constructing a postoperative pain monitoring model mapping relationship based on the patient's basic characteristic data in the patient information database includes the following steps: Obtaining the determined first neighborhood radius and the minimum number of samples of the first core point, performing cluster analysis on the basic characteristic data of the patient in the patient information database based on the DBSCAN algorithm, and determining a plurality of first clusters; Determine the patient basic characteristic data corresponding to each first cluster and perform characteristic processing to obtain the processed patient basic characteristic data corresponding to each first cluster, including the processed patient physiological data and the processed patient individual characteristic information; A model is constructed based on the historical patient time domain features, historical patient frequency domain features, and historical patient time-frequency domain features in the patient basic feature data corresponding to each first cluster, to obtain a pain level classification model and a pain level multimodal monitoring model; The basic characteristic data of the processed patients corresponding to each first cluster and the corresponding pain level classification model and pain level multimodal monitoring model are matched one-to-one to obtain the mapping relationship of the postoperative pain monitoring model.

4. The postoperative pain monitoring method based on a wearable flexible sensor according to claim 3, characterized in that: Obtaining the processed patient basic characteristic data corresponding to each first cluster includes the following steps: Performing mean processing on the historical patient time domain features, historical patient frequency domain features, and historical patient time-frequency domain features in the patient basic feature data corresponding to each first cluster, respectively, to obtain the averaged historical patient time domain features, averaged historical patient frequency domain features, and averaged historical patient time-frequency domain features, and the combination is recorded as the processed patient physiological data; The processed historical patient individual feature information in the patient basic feature data corresponding to each first cluster is processed: Perform mean processing on age and body mass index to obtain averaged age and averaged body mass index; count the number of coded basic disease histories, and take the top N coded basic disease histories as the set of coded basic disease histories; The averaged age, averaged body mass index and coded underlying disease history were recorded as the individual characteristic information of the patients after treatment.

5. The postoperative pain monitoring method based on a wearable flexible sensor according to claim 4, characterized in that: Obtaining a pain level classification model includes the following steps: Record the historical patient time domain features, historical patient frequency domain features, and historical patient time-frequency domain features in the patient basic feature data as historical patient pain feature data, and label the historical patient pain feature data with pain levels; Obtain the determined second neighborhood radius and the minimum number of samples of the second core point, perform cluster analysis on the annotated historical patient pain feature data based on the DBSCAN algorithm, and obtain multiple second clusters; Counting the pain level distribution data corresponding to the historical patient pain characteristic data in each second cluster, that is, the proportion of each pain level, and determining the pain level with the largest proportion from the pain level distribution data, recording it as the pain level corresponding to the second cluster and establishing an association; If there is only one second cluster associated with a certain pain level, the historical patient pain feature data in the second cluster is averaged to obtain the averaged historical patient pain feature data, and the averaged historical patient pain feature data is associated with the corresponding pain level to form a pain feature-pain level branch; If there is more than one second cluster associated with a certain pain level, the historical patient pain feature data in each second cluster is averaged to obtain averaged historical patient pain feature data. A group of averaged historical patient pain feature data with the smallest average distance from other averaged historical patient pain feature data is determined based on Euclidean distance, and the averaged historical patient pain feature data is associated with the corresponding pain level to form a pain feature-pain level branch. If the second cluster associated with a certain pain level is zero, then the annotated historical patient pain feature data corresponding to the pain level is obtained from the adjacent second cluster associated with the pain level, the annotated historical patient pain feature data is averaged to obtain the averaged historical patient pain feature data, and the averaged historical patient pain feature data is associated with the corresponding pain level to form a pain feature-pain level branch; All pain feature-pain level branches are arranged in order from small to large pain levels to obtain a pain level classification model.

6. The postoperative pain monitoring method based on a wearable flexible sensor according to claim 5, characterized in that: The multimodal pain level monitoring model is obtained, comprising the following steps: A single model layer is trained based on historical patient pain feature data, including a support vector machine layer, a random forest layer, and a multilayer perceptron layer. Each of the support vector machine layer, the random forest layer, and the multilayer perceptron layer has only one output, including the support vector machine pain level prediction result, the random forest pain level prediction result, and the multilayer perceptron pain level prediction result; Collect performance data of single model layers, including support vector machine layer performance data, random forest layer performance data, and multi-layer perceptron layer performance data; Determine the credibility of predictions from the support vector machine layer, random forest layer, and multilayer perceptron layer based on the performance data of the single model layer; The output of the pain level multimodal monitoring model is obtained based on the support vector machine pain level prediction results, random forest pain level prediction results, multilayer perceptron pain level prediction results and the credibility of the prediction results.

7. The postoperative pain monitoring method based on a wearable flexible sensor according to claim 6, characterized in that: Support vector machine layer performance data includes SVM accuracy Zq SVM , SVM recall rate Zh SVM SVMF1 value F1 SVM ; Random Forest layer performance data including RF accuracy Zq RF , RF recall rate Zh RF 、RFF1 value F1 RF ; Multilayer perceptron layer performance data including ANN accuracy Zq ANN , ANN recall rate Zh ANN ,ANNF1 value F1 ANN ; Determining the credibility of the predictions from the support vector machine layer, random forest layer, and multilayer perceptron layer involves the following steps: Calculate the composite score of the support vector machine layer, random forest layer, and multilayer perceptron layer; The comprehensive scores of each model were normalized to obtain the credibility of the prediction results of the support vector machine layer, random forest layer and multi-layer perceptron layer.

8. The postoperative pain monitoring method based on a wearable flexible sensor according to claim 7, characterized in that: The combined score of the support vector machine layer, random forest layer, and multilayer perceptron layer is calculated as: Among them, F SVM is the comprehensive score of the support vector machine layer, F RF is the comprehensive score of the random forest layer, F ANN is the composite score of the multilayer perceptron layer.

9. The postoperative pain monitoring method based on a wearable flexible sensor according to claim 8, characterized in that: The calculation method for the credibility of the prediction results of the support vector machine layer, random forest layer and multilayer perceptron layer is: Among them, K SVM is the reliability of the prediction results of the support vector machine layer, K RF is the credibility of the prediction results of the random forest layer, K ANN is the confidence level of the prediction result of the multi-layer perceptron layer.

10. The postoperative pain monitoring method based on a wearable flexible sensor according to claim 6, characterized in that: The pain level classification model is used to determine the pain level for the first time, and includes the following steps: Compare the basic feature data of the postoperative patients with the basic feature data of the processed patients in the mapping relationship of the postoperative pain monitoring model one by one to obtain a cosine similarity comparison value; Determining a maximum cosine similarity comparison value, and determining a corresponding pain level based on the maximum cosine similarity comparison value; The attribution threshold corresponding to the pain level is obtained. If the maximum cosine similarity comparison value is greater than the attribution threshold, the pain level is successfully determined for the first time. If the maximum cosine similarity comparison value is not greater than the attribution threshold, the pain level is not successfully determined for the first time, that is, the pain level cannot be determined.

Images