Intelligent postoperative pain management system based on multi-modal data

Abstract

This invention discloses a postoperative pain intelligent management system based on multimodal data. The system includes: a data acquisition device for collecting sensor data from multiple modalities of a target patient after gynecological surgery; a data analysis device for analyzing the sensor data according to a preset neural network algorithm to determine the target patient's pain information, pelvic floor exercise information, and current rehabilitation stage, and outputting a corresponding postoperative rehabilitation guidance plan; and a data interaction device connected to the data analysis device and the target patient's mobile device for receiving input data from the target patient or sending the postoperative guidance information to the mobile device for display. Therefore, this invention achieves more automated and efficient postoperative pain monitoring and rehabilitation management, effectively improving the efficiency and effectiveness of postoperative rehabilitation and providing users with more intelligent and comprehensive postoperative rehabilitation services.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to an intelligent postoperative pain management system based on multimodal data. Background Technology

[0002] Postoperative rehabilitation is a crucial component of gynecological surgical treatment, and its quality directly impacts patient outcomes and quality of life. However, most existing postoperative rehabilitation techniques rely on patient-reported visual analog scales (VAS) or numerical rating scales (NRS). These assessment methods are heavily influenced by patient subjectivity and do not consider using sensor data and intelligent algorithms to identify and manage pain, pelvic floor exercises, and rehabilitation stages in a targeted manner. Therefore, existing technologies have shortcomings that urgently need to be addressed. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a postoperative pain intelligent management system based on multimodal data, which can realize more automated and efficient postoperative pain monitoring and rehabilitation management, effectively improve the efficiency and effect of postoperative rehabilitation, and provide users with more intelligent and comprehensive postoperative rehabilitation services.

[0004] To address the aforementioned technical problems, the first aspect of this invention discloses a postoperative pain intelligent management system based on multimodal data, the system comprising: A data acquisition device is used to collect sensor data from multiple modalities of the target patient after gynecological surgery. The data analysis device is used to analyze the sensor data according to a preset neural network algorithm to determine the target patient's pain information, pelvic floor exercise information and current rehabilitation stage, and output a corresponding postoperative rehabilitation guidance plan. A data interaction device, connected to the data analysis device and the target patient's mobile device, is used to receive input data from the target patient or to send postoperative guidance information to the mobile device for display.

[0005] As an optional implementation, the data acquisition device includes: Wearable physiological sensors for collecting heart rate variability signals and skin conductance signals; Camera device used to capture facial images and body posture images; The intelligent pelvic floor sensor is used to collect pressure signals and electromyographic signals of pelvic floor muscle contraction.

[0006] As an optional implementation, the data analysis device includes: The feature extraction module is used to extract data features from the sensing data based on a neural network algorithm; The pain analysis module is used to analyze the pain information of the target patient based on the data characteristics. The exercise analysis module is used to analyze the exercise information of the target patient when performing pelvic floor rehabilitation exercises based on the data characteristics; the exercise information includes exercise posture, exercise compensation information, and movement quality information.

[0007] As an optional implementation, the feature extraction module includes a multi-branch feature extraction network and a cross-modal fusion network, wherein the multi-branch feature extraction network includes multiple feature extraction branches; and the cross-modal fusion network uses an attention mechanism to adaptively weight and fuse the features extracted by each feature extraction branch to obtain fused feature data.

[0008] As an optional implementation, the multi-branch feature extraction network includes: The physiological signal feature extraction branch adopts a temporal convolutional network structure, including multiple one-dimensional causal convolutional layers and residual connections. The convolutional kernel size is 3 to 7, the dilation factor increases exponentially, the number of network layers is 4 to 8, and the output feature dimension is 64 to 256, which is used to extract temporal features from heart rate variability signals and skin conductance signals. The facial expression feature extraction branch employs a lightweight convolutional neural network structure, including depthwise separable convolutional layers and global average pooling layers. The number of network parameters does not exceed 5M, and the output feature dimension is 64 to 128. The facial expression feature extraction branch also includes a facial action unit detection subnetwork for detecting pain-related facial action units. The facial expression feature extraction branch also includes an occlusion perception subnetwork for detecting facial occlusion areas and dynamically adjusting feature extraction weights. The body pose feature extraction branch adopts a graph convolutional network structure, with human skeleton key points as graph nodes and joint connections as graph edges. It includes multiple spatiotemporal graph convolutional layers, with 4 to 10 layers in the network and output feature dimensions of 64 to 256, which are used to extract pose temporal features from the skeleton key point sequence. The pelvic floor signal feature extraction branch adopts a combination structure of a one-dimensional convolutional neural network and a long short-term memory network. The one-dimensional convolutional neural network is used to extract local features of pressure signals and electromyographic signals. The hidden layer dimension of the long short-term memory network is 32 to 128, which is used to model the temporal dependency of the contraction process, and the output feature dimension is 32 to 128.

[0009] As an optional implementation, the pain analysis module includes: The real-time pain assessment unit adopts a multi-layer fully connected neural network structure. The input is the fusion feature output by the cross-modal fusion network. The number of hidden layers is 2 to 4, and the number of neurons in each layer is 64 to 256. The output is a continuous pain level score from 0 to 10. The assessment cycle of the real-time pain assessment unit is 1 to 5 minutes, and it supports continuous real-time monitoring. The breakthrough pain prediction unit employs a temporal fusion Transformer network structure. The input is fused temporal data from the past 1 to 4 hours, including a multi-head self-attention layer, gated residual connections, and a temporal coding layer. The number of attention heads is 4 to 8, and the network has 2 to 6 layers. The prediction time windows of the breakthrough pain prediction unit include 15 minutes, 30 minutes, and 60 minutes, and the output is the probability value of breakthrough pain occurring within each time window. The breakthrough pain prediction unit also includes an uncertainty quantification subunit, which uses the Monte Carlo Dropout method to calculate the confidence interval of the prediction results. The analgesia regimen recommendation unit employs a reinforcement learning network structure based on a proximal policy optimization algorithm, comprising a policy network and a value network. The policy network takes into account the current pain assessment result, pain prediction result, cumulative medication dosage, and the patient's physiological state as input, and outputs recommended analgesia regimen actions, including suggestions for medication dosage adjustment, timing of medication administration, and non-pharmacological interventions. The value network is used to evaluate the expected rehabilitation benefits of the current state. The analgesia regimen recommendation unit also includes a safety constraint module to check whether the recommended regimen complies with maximum daily dose limits, minimum dosing intervals, and contraindication rules. When the breakthrough pain prediction unit outputs a breakthrough probability exceeding a preset threshold, the early warning generation unit generates graded early warning information and pushes it to the mobile application interaction module. The early warning levels include observation level, warning level, and emergency level.

[0010] As an optional implementation, the exercise analysis module includes: The posture recognition unit employs a human posture estimation network, including a skeletal keypoint detection subnetwork and a posture classification subnetwork. The skeletal keypoint detection subnetwork detects at least 17 human keypoints, including the shoulder, hip, knee, and ankle, and outputs the two-dimensional coordinates and confidence scores of each keypoint. The posture classification subnetwork calculates posture parameters such as trunk tilt angle, hip flexion angle, and knee flexion angle based on the keypoint coordinates, and identifies the current training posture type, including supine, sitting, and standing positions. The compensatory action detection unit employs a multi-input classification network structure. The inputs include the temporal sequence of posture parameters output by the posture recognition unit and the pelvic floor features output by the pelvic floor signal feature extraction branch. The compensatory action detection unit detects abdominal compensation, hip compensation, thigh adduction compensation, and breath-holding compensation. The unit outputs the probability of occurrence and severity level for each type of compensatory action. The unit also includes a real-time feedback trigger subunit, which triggers an immediate correction prompt when a compensatory action is detected. The training quality assessment unit employs a multi-dimensional scoring network. Inputs include pelvic floor contraction parameters and compensation detection results. The pelvic floor contraction parameters include maximum contraction pressure, contraction rise time, contraction maintenance time, relaxation time, and contraction symmetry. The multi-dimensional movement quality score output by the training quality assessment unit includes strength score, endurance score, coordination score, relaxation score, and comprehensive score, with each dimension ranging from 0 to 100 points. The training plan generation unit uses a multi-armed slot machine algorithm based on Thompson sampling to dynamically adjust the training plan according to historical training data and training quality scores. The training plan includes training mode, contraction duration, relaxation duration, number of sets, and daily training frequency.

[0011] As an optional implementation, the data analysis device further includes: A rehabilitation stage identification module is used to determine the current rehabilitation stage of the target patient based on the pain information and the exercise information. The postoperative guidance module is used to output a corresponding postoperative rehabilitation guidance plan based on the current rehabilitation stage. The coordination module is used to control the working status of the pain analysis module and the exercise analysis module according to the current rehabilitation stage.

[0012] As an optional implementation, the rehabilitation stage identification module performs the following steps: Based on the surgical records of the target patients, the corresponding surgical sites were determined; In a pre-defined rehabilitation stage database, historical site records corresponding to multiple rehabilitation stages for the surgically affected area are determined. Based on the exercise information, multiple movement data of the target patient whose movement quality information was lower than a preset quality score threshold were identified; Calculate the weighted summation average of the first site distances between the exercised area and the surgically affected area corresponding to each of the aforementioned action data to obtain action-related data; wherein, the calculation weight corresponding to each first site distance is inversely proportional to the action quality information corresponding to the corresponding action data; Based on the pain information, the pain locations of the target patient where multiple pain signals exceed a preset pain threshold are identified; The weighted summation average of the second site distances between each pain site and the surgically affected site is calculated to obtain pain correlation data; wherein, the calculation weight corresponding to each second site distance is proportional to the pain information corresponding to the corresponding pain site; The motion association data and the pain association data are vectorized to obtain site recording data; Calculate the similarity between the site-recorded data and the historical site-recorded data corresponding to each rehabilitation stage, and determine the rehabilitation stage with the highest similarity as the current rehabilitation stage of the target patient.

[0013] As an optional implementation, the coordination module performs the following steps: Identify multiple historical data records corresponding to the current rehabilitation stage in the historical database; Calculate the proportion of a first data record in each of the historical data records where the pain information is greater than the pain threshold, and the proportion of a second data record where the action quality information is lower than the quality score threshold; The power consumption parameters of the pain module are obtained by multiplying the average proportion of the first data record in all the historical data records with the preset first power consumption weight. The power consumption parameters of the training module are obtained by multiplying the average proportion of the second data record in all the historical data records with the preset second power consumption weight. The power consumption parameters of the pain module and the exercise module are sent to the pain analysis module and the exercise analysis module, respectively, to control the operating power consumption.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a postoperative pain intelligent management system based on multimodal data. The system includes: a data acquisition device for collecting sensor data of multiple modalities from a target patient after gynecological surgery; a data analysis device for analyzing the sensor data according to a preset neural network algorithm to determine the target patient's pain information, pelvic floor exercise information, and current rehabilitation stage, and outputting a corresponding postoperative rehabilitation guidance plan; and a data interaction device connected to the data analysis device and the target patient's mobile device for receiving input data from the target patient or sending the postoperative guidance information to the mobile device for display. Therefore, this invention can collect multimodal data from a patient after gynecological surgery based on the data acquisition device, and perform data analysis to determine the patient's pain information, exercise information, and rehabilitation stage for postoperative guidance and data interaction. This achieves more automated and efficient postoperative pain monitoring and rehabilitation management, effectively improving the efficiency and effectiveness of postoperative rehabilitation and providing users with more intelligent and comprehensive postoperative rehabilitation services. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the functional modules of a postoperative pain intelligent management system based on multimodal data disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall architecture of a postoperative full-cycle intelligent rehabilitation management system disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of a multimodal data acquisition module structure disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of a deep learning analysis engine network structure disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of a pain management module structure disclosed in an embodiment of the present invention; Figure 6 This is a schematic diagram of a pelvic floor rehabilitation module structure disclosed in an embodiment of the present invention; Figure 7 This is a schematic diagram of a mobile application interaction module interface disclosed in an embodiment of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] Before further elaborating on the embodiments of the present invention, the prior art should be explained: taking gynecological surgery as an example, postoperative rehabilitation usually includes two stages: acute pain management and recovery functional rehabilitation. These two stages overlap in time and are related in terms of management objectives.

[0021] Currently, postoperative rehabilitation management faces the following technical challenges: First, there is the issue of subjectivity in pain assessment. Current pain assessment methods mainly rely on patient-reported visual analog scales (VAS) or numeric rating scales (NRS). These methods are greatly influenced by patient subjectivity, making it difficult to accurately reflect the patient's true pain status. Furthermore, the assessment frequency is limited, making continuous monitoring impossible.

[0022] Second, there is the difficulty in predicting breakthrough pain. Breakthrough pain refers to sudden, severe pain that occurs even after basic analgesia has been effective; its occurrence is sudden and unpredictable. Current technologies lack the ability to predict breakthrough pain, meaning that patients often receive intervention only after the pain has already occurred, impacting their recovery experience.

[0023] Third, there is a lack of personalization in analgesia regimens. Existing analgesia regimens mostly use standardized dosing methods, making it difficult to dynamically adjust according to individual patient differences and changes in pain, which can easily lead to insufficient analgesia or overdose.

[0024] Fourth, there is a lack of guidance on pelvic floor rehabilitation training. Pelvic floor function rehabilitation is an important part of post-gynecological surgery rehabilitation, but current pelvic floor rehabilitation training mostly relies on patients' self-practice, lacking professional guidance and real-time feedback. Patients are prone to compensatory movements during training, such as abdominal compensation and gluteal compensation. These incorrect training methods not only fail to achieve the desired rehabilitation effect but may also cause new injuries.

[0025] Fifth, the problem of fragmented management between the two stages. Current technology manages pain management and functional rehabilitation as two separate processes, lacking a unified coordination mechanism. In reality, pain status affects the intensity and effectiveness of rehabilitation training, and the progress of rehabilitation training also provides feedback for adjustments to pain management strategies; both need to be managed collaboratively.

[0026] Sixth, the lack of visual monitoring technology. Existing pelvic floor rehabilitation products such as Elvie and Perifit mainly rely on vaginal sensors to collect pelvic floor muscle contraction signals, but this method cannot detect compensatory movements of other parts of the body during training. Computer vision technology has been applied in the field of sports rehabilitation, but it has not yet been applied to posture monitoring and compensation detection in pelvic floor rehabilitation training.

[0027] To address the aforementioned technical problems, this invention discloses a postoperative pain intelligent management system based on multimodal data. This system collects multimodal data from patients after gynecological surgery using a data acquisition device, and analyzes the data to determine the patient's pain information, exercise information, and rehabilitation stage. This facilitates postoperative guidance and data interaction, achieving more automated and efficient postoperative pain monitoring and rehabilitation management. It effectively improves the efficiency and effectiveness of postoperative rehabilitation, providing users with more intelligent and comprehensive postoperative rehabilitation services. The following is a detailed description: Please see Figure 1 , Figure 1 This is a schematic diagram of the functional modules of a postoperative pain intelligent management system based on multimodal data disclosed in an embodiment of the present invention. Figure 1 As shown, the postoperative pain intelligent management system based on multimodal data includes: A data acquisition device is used to collect sensor data from multiple modalities of the target patient after gynecological surgery. The data analysis device is used to analyze the sensor data according to a preset neural network algorithm to determine the target patient's pain information, pelvic floor exercise information and current rehabilitation stage, and output the corresponding postoperative rehabilitation guidance plan. A data interaction device, connected to the data analysis device and the target patient's mobile device, is used to receive input data from the target patient or to send postoperative guidance information to the mobile device for display.

[0028] More specifically, in one concrete implementation plan, a postoperative full-cycle intelligent rehabilitation management system was implemented, the overall architecture of which can be found in [reference needed]. Figure 2 It includes a multimodal data acquisition module 1, a deep learning analysis engine 2, a pain management module 3, a pelvic floor rehabilitation module 4, a stage collaboration module 5, and a mobile application interaction module 6. Specifically, in this solution: The multimodal data acquisition module 1 includes a wearable sensor 1-1, a camera device 1-2, and a smart pelvic floor sensor 1-3, which are used to acquire the patient's physiological signals, facial expression images, body posture images, and pelvic floor signals.

[0029] The deep learning analytics engine 2 includes a multi-branch feature extraction network 2-1 and a cross-modal fusion network 2-2. The multi-branch feature extraction network 2-1 extracts features from each modality of data separately, while the cross-modal fusion network 2-2 adaptively weights and fuses the features from each branch through an attention mechanism, outputting a unified fused feature representation.

[0030] The pain management module 3 is based on fusion features to realize pain assessment, prediction and treatment plan recommendation functions, and is mainly used for pain management in the acute postoperative period (1 to 14 days after surgery).

[0031] The pelvic floor rehabilitation module 4 is based on fusion features to realize posture recognition, compensation detection and training quality assessment functions, and is mainly used for functional rehabilitation during the postoperative recovery period (2 weeks to 6 months after surgery).

[0032] In this phase, the collaborative module 5 is responsible for determining the current rehabilitation stage and coordinating the working status of the pain management module 3 and the pelvic floor rehabilitation module 4, so as to achieve a smooth transition and data sharing between the two stages.

[0033] The mobile application interaction module 6 includes a patient-side application and a medical care-side application, providing a user interface, displaying assessment results, providing rehabilitation guidance, and receiving user feedback.

[0034] Specifically, the modules transmit and interact with each other via wireless communication protocols (Bluetooth 5.0, WiFi) and a cloud server. The system supports offline mode, allowing basic functions to be performed locally when the network is unavailable, with data synchronized once the network is restored.

[0035] As an optional embodiment, the data acquisition device includes: Wearable physiological sensors for collecting heart rate variability signals and skin conductance signals; Camera device used to capture facial images and body posture images; The intelligent pelvic floor sensor is used to collect pressure signals and electromyographic signals of pelvic floor muscle contraction.

[0036] For details, see Figure 3 In the above implementation scheme, the structure of the multimodal data acquisition module 1 is realized based on this embodiment, wherein: The wearable sensor 1-1 adopts a wristband structure, with overall dimensions of 45mm × 35mm × 12mm and a weight of no more than 50 grams. The wristband of the wearable sensor 1-1 is made of medical-grade silicone material, and its length is adjustable from 140 to 220 mm to fit different wrist sizes.

[0037] The wearable sensor 1-1 incorporates a photoplethysmography (PPG) sensor, employing a dual-wavelength design with a green LED (525nm wavelength) and an infrared LED (880nm wavelength) at a sampling frequency of 100Hz to acquire heart rate variability (HRV) signals. HRV analysis extracts time-domain metrics (SDNN, RMSSD) and frequency-domain metrics (LF / HF ratio) to reflect the patient's autonomic nervous system activity.

[0038] The wearable sensor 1-1 incorporates skin conductance electrodes made of silver / silver chloride, with an electrode area of ​​1 square centimeter, an electrode spacing of 15 millimeters, and a sampling frequency of 20 Hz. Skin conductance signals can reflect the patient's sympathetic nervous system activity and are associated with pain and emotional state.

[0039] The wearable sensor 1-1 also has a built-in triaxial accelerometer and a temperature sensor. The accelerometer is used to detect motion status and sleep quality, and the temperature sensor is used to monitor changes in skin temperature.

[0040] The wearable sensor 1-1 has a built-in Bluetooth 5.0 communication module with an effective transmission distance of 10 meters. It has a built-in 500mAh lithium battery, uses a Type-C interface for charging, and can work continuously for 30 hours on a single charge.

[0041] The camera devices 1-2 utilize the built-in camera of a smartphone or tablet, or a standalone USB camera. The RGB camera has a 1080P resolution, a frame rate of 30 frames per second, and a 75-degree field of view. Some high-end devices also feature a depth camera, employing structured light principles, with a resolution of 640×480 and a frame rate of 30 frames per second.

[0042] Camera devices 1-2 have built-in edge computing chips (such as NPU or DSP) that can perform computationally intensive tasks such as face detection, key point detection, and pose estimation locally. The processed feature data is encrypted and then uploaded to the cloud server; the original image data is not uploaded to protect user privacy.

[0043] In the pain assessment scenario, camera devices 1-2 acquire facial images and identify pain-related facial expression features, such as frowning (AU4), closing eyes (AU6+AU7), and deepening of nasolabial folds (AU9+AU10), through facial action unit (AU) detection algorithms.

[0044] In the pelvic floor rehabilitation scenario, camera devices 1-2 acquire images of body posture, extract the coordinates of 17 skeletal key points through a human posture estimation algorithm, and calculate posture parameters such as trunk angle and hip joint angle for compensatory movement detection.

[0045] The intelligent pelvic floor sensors 1-3 feature an insertable design, with an ellipsoidal or cylindrical shape, a length of 50 to 80 mm, and a maximum diameter of 25 to 35 mm. The outer shell is made of medical-grade silicone with a Shore A30 hardness, conforming to the ISO 10993 biocompatibility standard.

[0046] The intelligent pelvic floor sensors 1-3 incorporate a built-in pressure sensor array, including six independent MEMS piezoresistive pressure sensors distributed at different locations on the sensor surface. Each sensor has a measurement range of 0 to 200 mmHg, a measurement accuracy of ±2 mmHg, and a sampling frequency of 100 Hz. Through multi-point pressure acquisition, the pressure distribution and symmetry of pelvic floor muscle contractions can be detected.

[0047] The intelligent pelvic floor sensor 1-3 incorporates four flexible surface electromyography (EMG) electrodes made of silver nanowire conductive film, arranged in a ring. The EMG signal sampling frequency is 1000 Hz, with a bandwidth of 20 to 500 Hz. The EMG signal directly reflects the electrical activity of the pelvic floor muscles, complementing the pressure signal and improving the accuracy of contraction detection.

[0048] The intelligent pelvic floor sensors 1-3 feature a built-in Bluetooth 5.0 communication module and a 200mAh lithium battery, supporting wireless charging. A single full charge provides up to 8 hours of continuous operation. The sensor housing has an IPX7 waterproof rating and can be rinsed and disinfected with water after use.

[0049] As an optional embodiment, the data analysis device includes: The feature extraction module is used to extract data features from the sensor data based on a neural network algorithm; The pain analysis module is used to analyze the pain information of the target patient based on the data characteristics. The exercise analysis module is used to analyze the exercise information of the target patient when performing pelvic floor rehabilitation exercises based on the data characteristics; the exercise information includes exercise posture, exercise compensation information, and movement quality information.

[0050] For details, see Figure 4In the above implementation scheme, the network structure of the deep learning analysis engine 2 is implemented based on this embodiment, wherein: The deep learning analysis engine 2 includes a multi-branch feature extraction network 2-1 and a cross-modal fusion network 2-2. It adopts a hierarchical processing architecture. The data of each modality is first processed by a dedicated feature extraction branch, and then cross-modal information is integrated through the fusion network.

[0051] The physiological signal feature extraction branch 2-1a adopts a temporal convolutional network (TCN) structure, which is specifically designed to process heart rate variability signals and skin conductance signals collected by wearable sensor 1-1.

[0052] The input to the physiological signal feature extraction branch 2-1a is a 300-second (5-minute) time-series signal with 6000 sampling points (heart rate variability) and 1000 sampling points (skin conductance). The input signal is first standardized and then fed into the TCN network.

[0053] The TCN network consists of six one-dimensional causal convolutional layers with a kernel size of 5 and dilation factors of 1, 2, 4, 8, 16, and 32 respectively, ensuring that the network's effective receptive field covers the entire input sequence. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and residual connections are established between adjacent layers. The network outputs 128-dimensional features.

[0054] The facial expression feature extraction branch 2-1b adopts a lightweight convolutional neural network structure, which is optimized based on the MobileNetV3 architecture and is specifically designed to process facial images captured by camera device 1-2.

[0055] The input to the facial expression feature extraction branch 2-1b is a 112×112 pixel facial image that has been detected and aligned. The network backbone uses depthwise separable convolutional layers, including an initial convolutional layer, 15 inverse residual modules, and a global average pooling layer. The network has 2.5M parameters, and the single-frame inference time is less than 10 milliseconds (on a mobile NPU).

[0056] The facial expression feature extraction branch 2-1b also includes a facial action unit (AU) detection subnetwork, which uses a multi-task learning framework to predict the activation intensity of 12 pain-related facial action units in parallel on the features output by the backbone network.

[0057] The facial expression feature extraction branch 2-1b also includes an occlusion perception subnetwork to detect whether there are occlusions on the face (such as oxygen masks, hand occlusions, etc.). When occlusion is detected, the fusion weights of facial features are dynamically reduced to avoid noise in the occluded area affecting the final pain assessment result. The network output features are 96-dimensional.

[0058] The body pose feature extraction branch 2-1c adopts a spatiotemporal graph convolutional network (ST-GCN) structure, which is specifically designed to process the skeletal keypoint sequence extracted from body pose images.

[0059] The input to the body pose feature extraction branch 2-1c is a sequence of 30 consecutive frames (1 second) of skeletal keypoints. Each frame contains the two-dimensional coordinates and confidence of 17 keypoints, and the input tensor dimension is 30×17×3.

[0060] The graph convolutional network constructs a graph structure based on the natural connections of the human skeleton, with 17 keypoints as graph nodes and joint connections as graph edges. The network consists of 8 spatiotemporal graph convolutional layers, each containing two sub-layers: spatial graph convolution and temporal convolution. Spatial graph convolution learns the spatial relationships between joints, while temporal convolution learns the temporal evolution of actions. The network outputs 128-dimensional features.

[0061] The pelvic floor signal feature extraction branch 2-1d adopts a combination structure of one-dimensional convolutional neural network (1D-CNN) and long short-term memory network (LSTM), which is specifically designed to process the pressure signals and electromyographic signals collected by the intelligent pelvic floor sensors 1-3.

[0062] The input to the pelvic floor signal feature extraction branch 2-1d is a 10-second pelvic floor signal time series, including a 6-channel pressure signal (sampling frequency 100Hz, 1000 sampling points) and a 4-channel electromyography signal (sampling frequency 1000Hz, 10000 sampling points). The two types of signals are preprocessed separately and then fed into their respective feature extraction subnetworks.

[0063] The pressure signal subnetwork employs three 1D convolutional layers with a kernel size of 7 and a stride of 2, outputting a 128-dimensional feature vector. The electromyography (EMG) signal subnetwork first undergoes bandpass filtering and rectification smoothing preprocessing before being fed into three 1D convolutional layers, outputting a 128-dimensional feature vector. The output features of the two subnetworks are concatenated and then fed into a two-layer LSTM network with 64 hidden layers to model the temporal dependencies of pelvic floor muscle contraction. The final network output has a 64-dimensional feature vector.

[0064] The cross-modal fusion network 2-2 adopts a Transformer structure based on a cross-attention mechanism to adaptively weight and fuse the features extracted from each branch.

[0065] The input to the cross-modal fusion network 2-2 consists of feature vectors from four branches. These vectors are first unified to a 256-dimensional feature space through a linear projection layer. Then, the four feature vectors are used as input tokens for a Transformer, which are processed through four Transformer encoder layers.

[0066] Each layer of the Transformer encoder includes a multi-head self-attention module and a feedforward neural network module, with 8 attention heads and a 1024-dimensional hidden layer in the feedforward network. Through the self-attention mechanism, features from different modalities can pay attention to each other and learn cross-modal semantic associations.

[0067] The cross-modal fusion network 2-2 also includes a dynamic reliability weight calculation unit, which calculates the reliability weight based on the signal quality of each modality's data (such as signal-to-noise ratio, confidence level, etc.). When the data quality of a certain modality is poor (such as when the face is obscured or the sensor is detached), the fusion weight of that modality is automatically reduced to improve the robustness of the system.

[0068] The output of the cross-modal fusion network 2-2 is a 256-dimensional fusion feature vector, which integrates information from each modality and is used for subsequent pain assessment, prediction, and rehabilitation training assessment tasks.

[0069] The Deep Learning Analytics Engine 2 can be deployed on edge computing devices (such as smartphones and edge servers) or cloud servers. When deployed at the edge, model quantization and pruning techniques are used to compress the model size, enabling real-time inference. The latency of a single complete inference (from raw data to fused features) is no more than 200 milliseconds.

[0070] As an optional embodiment, the pain analysis module includes: The real-time pain assessment unit employs a multi-layer fully connected neural network structure. The input is the fusion feature output by this cross-modal fusion network, with 2 to 4 hidden layers and 64 to 256 neurons per layer. The output is a continuous pain level score from 0 to 10. The assessment cycle of this real-time pain assessment unit is 1 to 5 minutes, and it supports continuous real-time monitoring. The breakthrough pain prediction unit employs a temporal fusion Transformer network structure. The input is fused temporal data from the past 1 to 4 hours, including multi-head self-attention layers, gated residual connections, and temporal coding layers. The number of attention heads is 4 to 8, and the number of network layers is 2 to 6. The prediction time windows of this breakthrough pain prediction unit include 15 minutes, 30 minutes, and 60 minutes, and the output is the probability value of breakthrough pain occurring within each time window. This breakthrough pain prediction unit also includes an uncertainty quantification subunit, which uses the Monte Carlo Dropout method to calculate the confidence interval of the prediction results. The analgesia regimen recommendation unit employs a reinforcement learning network structure based on a proximal policy optimization algorithm, comprising a policy network and a value network. The policy network takes into account the current pain assessment result, pain prediction result, cumulative medication dosage, and the patient's physiological state as input, and outputs recommended analgesia regimen actions, including suggestions for medication dosage adjustment, timing of medication administration, and non-pharmacological interventions. The value network is used to evaluate the expected rehabilitation benefits of the current state. The analgesia regimen recommendation unit also includes a safety constraint module to check whether the recommended regimen complies with maximum daily dose limits, minimum dosing intervals, and contraindication rules. When the breakthrough pain prediction unit outputs a breakthrough probability exceeding a preset threshold, the early warning generation unit generates tiered early warning information and pushes it to the mobile application interaction module. The early warning levels include observation level, warning level, and emergency level.

[0071] For details, see Figure 5 In the above implementation scheme, the structure of pain management module 3 is implemented based on this embodiment, wherein: The pain management module 3 includes a real-time pain assessment unit 3-1, a breakthrough pain prediction unit 3-2, an analgesia plan recommendation unit 3-3, and an early warning generation unit 3-4, and is mainly used for pain management in the acute postoperative period.

[0072] The real-time pain assessment unit 3-1 adopts a 3-layer fully connected neural network structure. The input is the 256-dimensional fusion feature output by the cross-modal fusion network 2-2. The dimensions of the hidden layers are 128, 64, and 32 respectively. The output is a continuous pain level score from 0 to 10.

[0073] The assessment results of the real-time pain assessment unit 3-1 correspond to the clinically commonly used Visual Analogue Scale (VAS), with 0 points indicating no pain and 10 points indicating severe pain. The system divides the scores into four levels: 0 to 2 points for mild pain, 3 to 4 points for moderate pain, 5 to 6 points for severe pain, and 7 to 10 points for excruciating pain.

[0074] The real-time pain assessment unit 3-1 has an assessment cycle of 2 minutes, meaning it outputs a pain score every 2 minutes. The assessment results are displayed in real time on both the patient and healthcare application platforms and stored in a database for trend analysis.

[0075] To improve the accuracy of the assessment, the real-time pain assessment unit 3-1 also incorporates the patient's subjective feedback. Patients can record their perceived pain level at any time via a mobile application. The system compares and calibrates the subjective scores with the objective assessment results, gradually optimizing the personalized parameters of the assessment model.

[0076] The breakthrough pain prediction unit 3-2 adopts a temporal fusion Transformer network structure. The input is the fusion feature time series of the past 2 hours (feature vectors of a total of 60 time points), and the output is the probability of breakthrough pain occurring in the next 15 minutes, 30 minutes and 60 minutes.

[0077] The network structure of the breakthrough pain prediction unit 3-2 includes a temporal location encoding layer, a four-layer Transformer encoder, and a prediction head. The temporal location encoding layer adds location information to the features at each time point; the Transformer encoder learns the long-range dependencies between temporal features; the prediction head consists of three parallel fully connected networks that predict the breakthrough probability for three time windows.

[0078] The breakthrough pain prediction unit 3-2 also includes an uncertainty quantification subunit, employing the Monte Carlo Dropout method: Dropout is kept active during inference, with 20 forward propagations performed to calculate the mean and standard deviation of the prediction results. The standard deviation reflects the uncertainty of the prediction. When the uncertainty is high, the system considers breakthrough pain to be imminent and triggers an early warning process when the probability of breakthrough pain prediction exceeds a preset threshold (default 0.7).

[0079] The analgesia recommendation unit 3-3 adopts a reinforcement learning network structure based on the proximal policy optimization (PPO) algorithm to recommend personalized analgesia plans based on the current pain status and the patient's historical data.

[0080] The state space of the analgesia regimen recommendation unit 3-3 includes: current pain score, pain prediction result, cumulative opioid dosage, time since last administration, patient physiological status (heart rate, blood pressure, etc.), and patient demographic characteristics (age, weight, etc.).

[0081] The action space of the analgesia regimen recommendation unit 3-3 includes: drug dosage adjustment suggestions (increase, maintain, decrease, discontinue), drug administration timing suggestions (immediate administration, timed administration, as-needed administration), and non-drug intervention suggestions (deep breathing training, cold compress, hot compress, distraction, etc.).

[0082] The reward function in Unit 3-3 of the analgesia regimen recommendation takes into account both the pain control effect (reduction in pain score) and the dosage (cumulative dosage of opioids). The goal is to reduce the dosage and lower the risk of addiction while effectively controlling pain.

[0083] The analgesia regimen recommendation unit 3-3 also includes a safety constraint module, which performs safety checks before outputting the recommended regimen. These checks include: maximum daily dose limit (calculated based on drug type and patient weight), minimum dosing interval (set based on drug half-life), and contraindication checks (based on patient medical history and allergy history). Regimens that do not meet the safety constraints will be rejected, and the system will recommend alternative regimens.

[0084] The recommendations in Analgesia Recommendation Unit 3-3 must be reviewed and confirmed by medical staff before implementation. The system will record the review decisions of medical staff for continuous optimization of the recommendation model.

[0085] The warning generation unit 3-4 is responsible for generating and pushing various warning information, and the warning levels are divided into three levels: Observational warning: The probability of breakthrough pain prediction is between 0.5 and 0.7, or the pain score rises continuously but does not reach severe levels. The system only displays the alert on the patient's end, advising the patient to pay attention to their condition.

[0086] Warning level alert: Breakthrough pain prediction probability is between 0.7 and 0.9, or pain score reaches severe (5 to 6). The system simultaneously pushes the alert to both the patient and healthcare provider, recommending preventative intervention.

[0087] Emergency alert: If the probability of breakthrough pain prediction exceeds 0.9, or the pain score reaches severe (7 points or above), or abnormal physiological indicators appear, the system will immediately push an emergency alert to the healthcare provider and trigger a phone or SMS notification.

[0088] As an optional embodiment, the exercise analysis module includes: The pose recognition unit employs a human pose estimation network, including a skeletal keypoint detection subnetwork and a pose classification subnetwork. The skeletal keypoint detection subnetwork detects at least 17 human keypoints, including the shoulder, hip, knee, and ankle, and outputs the two-dimensional coordinates and confidence scores of each keypoint. The pose classification subnetwork calculates pose parameters such as trunk tilt angle, hip flexion angle, and knee flexion angle based on the keypoint coordinates, and identifies the current training pose type, including supine, sitting, and standing positions. The compensatory action detection unit employs a multi-input classification network structure. Inputs include the temporal sequence of posture parameters output by the posture recognition unit and the pelvic floor features output by the pelvic floor signal feature extraction branch. The compensatory action detection unit detects abdominal compensation, hip compensation, thigh adduction compensation, and breath-holding compensation. It outputs the probability of occurrence and severity level for each type of compensatory action. The unit also includes a real-time feedback trigger subunit, which triggers immediate correction prompts when a compensatory action is detected. The training quality assessment unit employs a multi-dimensional scoring network. Inputs include pelvic floor contraction parameters and compensation detection results. These pelvic floor contraction parameters include maximum contraction pressure, contraction rise time, contraction maintenance time, relaxation time, and contraction symmetry. The multi-dimensional movement quality score output by the training quality assessment unit includes strength score, endurance score, coordination score, relaxation score, and comprehensive score, with each dimension ranging from 0 to 100 points. The training plan generation unit uses a multi-armed slot machine algorithm based on Thompson sampling to dynamically adjust the training plan according to historical training data and training quality scores. The training plan includes training mode, contraction duration, relaxation duration, number of sets, and daily training frequency.

[0089] For details, see Figure 6 In the above implementation scheme, the structure of the pelvic floor rehabilitation module 4 is realized based on this embodiment, wherein: The pelvic floor rehabilitation module 4 includes a posture recognition unit 4-1, a compensatory movement detection unit 4-2, a training quality assessment unit 4-3, and a training plan generation unit 4-4, and is mainly used for pelvic floor function rehabilitation training during the postoperative recovery period.

[0090] The pose recognition unit 4-1 uses a human pose estimation network, which is optimized based on the OpenPose or MediaPipe Pose architecture, to extract skeletal key points from the body pose images acquired by the camera device 1-2.

[0091] The skeletal keypoint detection subnetwork of pose recognition unit 4-1 detects 17 human keypoints, including: head (nose tip, left and right eyes, left and right ears), upper limbs (left and right shoulders, left and right elbows, left and right wrists), trunk (chest center, hip center), and lower limbs (left and right hips, left and right knees, left and right ankles). Each keypoint outputs two-dimensional coordinates (x, y) and a confidence score.

[0092] The posture classification sub-network of posture recognition unit 4-1 calculates multiple posture parameters based on key point coordinates: trunk tilt angle (the angle between the trunk and the vertical direction), hip flexion angle (the angle between the thigh and the trunk), knee flexion angle (the angle between the lower leg and the thigh), and pelvic tilt angle (the angle between the pelvis and the horizontal plane).

[0093] The posture recognition unit 4-1 can also identify the current training posture type, including supine position (knee flexed at 90 degrees), sitting position (hip flexed at 90 degrees), and standing position (hip and knee joints nearly extended). Different postures correspond to different training difficulties and target muscle groups.

[0094] The compensatory action detection unit 4-2 adopts a multi-input classification network structure, which integrates the temporal features of posture parameters and pelvic floor signal features to detect compensatory actions during the training process.

[0095] Compensatory Action Detection Unit 4-2 detects four main types of compensatory actions: Abdominal compensation: The patient mistakenly tightens their abdominal muscles when performing pelvic floor muscle contractions. The detection method involves monitoring changes in trunk posture and abdominal muscle activity; abdominal compensation is defined as the detection of abdominal adduction or forward tilting of the trunk.

[0096] Hip compensation: Patients mistakenly tighten their gluteus maximus muscles when performing pelvic floor muscle contractions. The detection method involves monitoring changes in hip joint angle and hip contour; hip hyperextension or significant hip tightening is identified as hip compensation.

[0097] Compensatory adduction of the thighs: This occurs when a patient incorrectly tightens their thigh adductor muscles during pelvic floor muscle contraction. The detection method involves monitoring changes in the distance between the knees; when the knees are noticeably brought closer together, it is considered compensatory adduction of the thighs.

[0098] Breath-holding compensation: The patient holds their breath while contracting the pelvic floor muscles. The detection method is to monitor the amplitude of chest wall movement; when a significant decrease in respiratory rate or cessation of chest wall movement is detected, it is determined to be breath-holding compensation.

[0099] The compensatory action detection unit 4-2 outputs the probability of occurrence (0 to 1) and severity level (mild, moderate, severe) for each type of compensatory action. When a compensatory action is detected, the real-time feedback trigger subunit immediately notifies the patient to correct it via voice or visual cues.

[0100] The training quality assessment unit 4-3 uses a multi-dimensional scoring network to output a training quality score by integrating pelvic floor contraction parameters and compensation detection results.

[0101] Training quality assessment unit 4-3 extracts five contraction parameters from pelvic floor signals: maximum contraction pressure (reflecting pelvic floor muscle strength), contraction rise time (reflecting muscle activation speed), contraction maintenance time (reflecting muscle endurance), relaxation time (reflecting muscle relaxation ability), and contraction symmetry (reflecting coordination between the left and right sides of the muscles).

[0102] Training quality assessment unit 4-3 outputs scores across five dimensions: strength score (based on maximum contraction pressure), endurance score (based on contraction duration), coordination score (based on contraction symmetry and compensation detection results), relaxation score (based on relaxation time and baseline pressure), and overall score (a weighted average of the five dimensions). Each dimension's score ranges from 0 to 100.

[0103] Training quality assessment unit 4-3 also generates a detailed training report, including: various indicators of this training, comparison with the previous training, comparison with peers, and suggestions for improvement.

[0104] The training plan generation unit 4-4 uses a multi-armed slot machine algorithm based on Thompson sampling to dynamically optimize the training plan according to the patient's historical training data.

[0105] Training plan generation unit 4-4 parameterizes the training plan into multiple dimensions: training mode (rapid contraction, continuous contraction, step contraction), contraction duration (3 to 10 seconds), relaxation duration (3 to 10 seconds), number of repetitions per set (5 to 15 times), number of sets (3 to 5 sets), and daily training frequency (1 to 3 times).

[0106] Training plan generation unit 4-4 maintains a Beta distribution for each combination of training parameters, representing the probability that the combination will improve training quality. After each training session, the Beta distribution for the corresponding combination is updated based on the actual training quality score. When formulating the next training plan, samples are taken from each distribution to select the parameter combination with the highest expected return.

[0107] Through the Thompson sampling algorithm, the training plan generation unit 4-4 can strike a balance between exploring new parameter combinations and utilizing known effective combinations, gradually finding the most suitable personalized training plan for each patient.

[0108] As an optional embodiment, the data analysis apparatus further includes: The rehabilitation stage identification module is used to determine the current rehabilitation stage of the target patient based on the pain information and the exercise information. The postoperative guidance module is used to output the corresponding postoperative rehabilitation guidance plan based on the current rehabilitation stage; The coordination module is used to control the working status of the pain analysis module and the exercise analysis module according to the current rehabilitation stage.

[0109] Optionally, the current rehabilitation stage can be divided into the acute phase (0-2 weeks post-surgery), the subacute phase (2-6 weeks post-surgery), the early recovery phase (6-12 weeks post-surgery), the mid-to-late recovery phase (12 weeks-6 months post-surgery), or the functional reconstruction phase (more than 6 months post-surgery), without limitation by this invention.

[0110] Optionally, the rehabilitation guidance program may include a daily exercise plan, demonstration videos, pain management suggestions, dietary guidance, rest arrangements, or follow-up appointment reminders; this invention does not impose any limitations.

[0111] Optionally, the working states of the pain analysis module and the exercise analysis module may include high-frequency monitoring, low-power standby, data upload frequency, or sensor sampling rate, which are not limited in this invention.

[0112] As an optional embodiment, the rehabilitation stage identification module performs the following steps: Based on the target patient's surgical records, determine the corresponding surgical site affected; In the pre-defined rehabilitation stage database, historical site records corresponding to multiple rehabilitation stages for the site affected by the surgery are identified. Based on this exercise information, several movement data of the target patient were identified as having quality information below a preset quality score threshold. Calculate the weighted average of the distances between the exercised body part and the surgically affected body part corresponding to each movement data point to obtain the movement association data; wherein, the calculation weight corresponding to each distance of the first body part is inversely proportional to the movement quality information corresponding to the movement data. Based on this pain information, the pain locations of multiple patients whose pain information exceeds the preset pain threshold are identified. The weighted summation of the distances between each pain location and the surgically affected area to the second location is calculated to obtain pain correlation data; wherein, the calculation weight corresponding to each distance to the second location is proportional to the pain information corresponding to the pain location. Vectorize the action-related data and the pain-related data to obtain the site recording data; Calculate the similarity between the recorded data of this site and the historical recorded data of each site corresponding to this rehabilitation stage, and determine the rehabilitation stage with the highest similarity as the current rehabilitation stage of the target patient.

[0113] Optionally, the surgical procedure can affect other human tissues in the knee joint, hip joint, spine, or pelvic floor region; this invention is not limited to these areas. Optionally, the distance between the first or second part can be an anatomical distance or a functional distance, and the present invention does not limit it.

[0114] Optionally, the vectorized site recording data can be a multi-dimensional feature vector, including site distance, correlation strength, pain / action weight, etc., which is not limited in this invention.

[0115] Optionally, the similarity can be cosine similarity, the reciprocal of Euclidean distance, or Mahalanobis distance; this invention does not limit the specific similarity.

[0116] As an optional embodiment, the coordination module performs the following steps: Identify multiple historical data records corresponding to the current rehabilitation stage in the historical database; Calculate the proportion of a first data record in each historical data record where the pain information is greater than the pain threshold, and the proportion of a second data record where the action quality information is lower than the quality score threshold; The power consumption parameters of the pain module are obtained by multiplying the average proportion of the first data record in all the historical data records with the preset first power consumption weight. The power consumption parameters of the training module are obtained by multiplying the average proportion of the second data record in all the historical data records with the preset second power consumption weight. The power consumption parameters of the pain module and the exercise module are sent to the pain analysis module and the exercise analysis module respectively to control the operating power consumption.

[0117] Optionally, the power consumption can be expressed as sampling frequency, data upload cycle, calculation accuracy, or sensor on-time, and this invention does not limit it.

[0118] Optionally, the coordination module can further integrate a low-power strategy, which forces the system to reduce to the lowest monitoring mode when the battery level is below 20%, retaining only the key abnormal alarm function. This invention does not limit this.

[0119] Optionally, the coordination module can support user-defined power consumption preferences, such as "priority battery life mode" or "priority monitoring accuracy mode", and dynamically adjust the first power consumption weight and the second power consumption weight according to the preference. This invention does not limit this.

[0120] Specifically, the above implementation scheme also includes a stage coordination module 5 based on this embodiment, wherein: The collaborative module 5 for this stage includes a stage judgment unit 5-1, a module coordination unit 5-2, and a cross-stage data sharing unit 5-3, which is responsible for achieving unified management throughout the entire postoperative period.

[0121] This stage assessment unit 5-1 divides the postoperative rehabilitation process into three stages: Acute phase: Days 1 to 14 post-surgery, focusing on pain management. Patients primarily experience acute post-operative pain, requiring aggressive analgesia. During this phase, Pain Management Module 3 is fully activated, while Pelvic Floor Rehabilitation Module 4 is either deactivated (first 7 days) or under basic monitoring (days 8 to 14).

[0122] Transition period: Weeks 2 to 4 post-surgery, pain management and functional rehabilitation proceed in parallel. The patient's acute pain gradually subsides, and low-intensity pelvic floor rehabilitation training begins. During this stage, pain management module 3 and pelvic floor rehabilitation module 4 are activated simultaneously, with the training intensity of pelvic floor rehabilitation module 4 constrained by pain assessment results.

[0123] Recovery period: From the 4th week to the 6th month post-surgery, the focus is on functional rehabilitation. The patient's pain has largely subsided, and systematic pelvic floor rehabilitation training begins. During this stage, pelvic floor rehabilitation module 4 is fully activated, while pain management module 3 switches to low-power monitoring mode.

[0124] Stage Judgment Unit 5-1 determines the current stage based on multi-dimensional indicators.

[0125] Phase switching requires confirmation from healthcare professionals. When the system detects that the switching conditions are met, it sends a notification to the healthcare professional's application, allowing them to decide whether to proceed with the phase switch.

[0126] The module coordination unit 5-2 controls the working status of each module based on the current stage output by the stage judgment unit 5-1, for example: During the acute phase, module coordination unit 5-2 sets the assessment cycle of pain management module 3 to 2 minutes and the prediction update cycle to 10 minutes; it sets pelvic floor rehabilitation module 4 to either a closed state (first 7 days) or a basic monitoring state (days 8 to 14). In the basic monitoring state, only pelvic floor signals are collected for baseline assessment, and no active training guidance is provided.

[0127] During the transition period, module coordination unit 5-2 activates two modules simultaneously, but establishes a constraint relationship: when the pain score exceeds 4 points, the intensity of pelvic floor rehabilitation training is automatically reduced or training is paused; when a significant increase in pain score is detected during training, training is immediately stopped and an alert is sent.

[0128] During the recovery period, the module coordination unit 5-2 sets the pelvic floor rehabilitation module 4 to a fully active state, and the training intensity is no longer constrained by the pain score; the pain management module 3 is set to a low power monitoring state, the assessment cycle is extended to 10 minutes, and an alarm is triggered only when the pain score exceeds 3 points.

[0129] This cross-phase data sharing unit 5-3 enables data collaboration between the acute and recovery phases: Forward data sharing: Pain sensitivity characteristics accumulated during the acute phase (such as pain threshold, pain tolerance, and response to analgesics) are transferred to pelvic floor rehabilitation module 4 to optimize training intensity settings. For example, patients with high pain sensitivity will have lower initial training intensity settings and slower progression.

[0130] Reverse data sharing: Training feedback data during the recovery period (such as changes in post-training pain, exercise-induced discomfort, etc.) is fed back to the pain management module 3 for continuous optimization of the pain prediction model. Through this feedback, the system can learn the relationship between exercise and pain, improving prediction accuracy.

[0131] For details, see Figure 7 The above implementation scheme also includes the interface design of the mobile application interaction module 6, wherein: The mobile application interaction module 6 includes a patient-side application 6-1 and a medical staff-side application 6-2, which provide interactive interfaces for patients and medical staff, respectively.

[0132] The patient-side application 6-1 includes the following functional interfaces: Homepage dashboard: Displays overview information such as current pain score, rehabilitation stage, completion rate of today's training tasks, and cumulative rehabilitation days. Visual charts are used to display pain trends and rehabilitation progress curves.

[0133] Pain recording interface: Patients can record their subjective pain score, pain location, pain nature (dull, sharp, throbbing, etc.), and triggering factors at any time. The system will compare and analyze the subjective records with the objective assessments.

[0134] Training guidance interface: Adopting a gamified design, the interface guides patients through pelvic floor rehabilitation training using animated characters. The interface displays the contraction pressure curve, target area, and compensation test results in real time. Upon completion of the training, the score and evaluation are displayed.

[0135] The rehabilitation diary interface allows patients to record daily changes in symptoms, quality of life assessments, medication usage, etc. The system periodically sends standardized questionnaires (such as the Pain Catastrophism Scale and the Pelvic Floor Function Assessment Questionnaire) for evaluation.

[0136] Educational Resources Interface: Provides educational content such as postoperative rehabilitation knowledge, pain management techniques, and pelvic floor rehabilitation training tutorials, including text, images, videos, and interactive Q&A formats.

[0137] The healthcare application 6-2 includes the following functional interfaces: Patient Management Dashboard: Displays overview information of managed patients in list or card format, including current pain score, rehabilitation stage, today's training completion rate, and alert status. Supports filtering and sorting by alert level, rehabilitation stage, pain score, and other dimensions.

[0138] Patient Details Interface: Displays detailed data for each patient, including pain score trend charts, medication records, training history, and various assessment indicators. Healthcare staff can add medical orders and notes on this interface.

[0139] Early Warning Management Interface: Displays all unprocessed early warning information, sorted by time and level. Healthcare staff can view warning details, perform interventions, and record the results.

[0140] The protocol adjustment interface displays the system's recommended analgesia protocols and training plans, which medical staff can review, modify, or confirm. All adjustments are recorded for subsequent analysis and model optimization.

[0141] Data analysis interface: Provides multi-dimensional data statistics and analysis functions, including pain control effect analysis, medication use analysis, rehabilitation progress analysis, and patient compliance analysis. Supports exporting reports for clinical research or quality management.

[0142] The postoperative rehabilitation management method based on the system of the present invention includes the following steps: Step 1: System Initialization: Upon postoperative admission, medical staff create a patient file in the medical staff application, entering basic information, surgical information, and medical history. Wearable sensors and smart pelvic floor sensors are provided to the patient, and the patient is guided to install the patient application and complete device pairing.

[0143] Step Two: Acute Pain Management: The system automatically enters acute phase mode. Wearable sensors continuously collect physiological signals, and cameras periodically capture facial expression images. The system assesses pain levels in real time and predicts the risk of breakthrough pain. Healthcare professionals adjust the analgesia regimen based on system recommendations and the actual situation. Patients record their subjective pain perception and medication usage through a patient-facing application.

[0144] Step 3: Joint Management During the Transition Period: When the system detects that the phase switching conditions are met, it alerts medical staff to confirm the entry into the transition period. The intelligent pelvic floor sensor begins collecting pelvic floor signals, and the system assesses basic pelvic floor function. Under the system's guidance, the patient performs low-intensity pelvic floor rehabilitation training, with the training intensity constrained by the pain assessment results.

[0145] Step Four: Recovery Period Rehabilitation Training: When the system detects that the stage switching conditions are met, it reminds medical staff to confirm the start of the recovery period. The patient performs pelvic floor rehabilitation training according to the personalized training plan generated by the system. The system monitors compensatory movements in real time and provides corrective guidance, assesses training quality, and dynamically optimizes the training plan. The pain management module switches to monitoring mode to continuously monitor changes in pain during the rehabilitation process.

[0146] Step 5: Rehabilitation Assessment and Follow-up: After the patient completes the planned rehabilitation cycle, the system generates a comprehensive rehabilitation assessment report, including pain management effectiveness, pelvic floor function recovery status, medication statistics, and training adherence. Medical staff will decide whether to continue training or transfer the patient to follow-up management based on the assessment results.

[0147] This system can be applied to multiple clinical scenarios such as postoperative rehabilitation in gynecology (e.g., hysterectomy, pelvic organ prolapse repair), postoperative rehabilitation in urology (e.g., prostatectomy), and postpartum rehabilitation, and has broad application prospects.

[0148] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0149] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0150] Finally, it should be noted that the postoperative pain intelligent management system based on multimodal data disclosed in the embodiments of the present invention is only a preferred embodiment of the present invention and is only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A postoperative pain intelligent management system based on multimodal data, characterized in that, The system includes: A data acquisition device is used to collect sensor data from multiple modalities of the target patient after gynecological surgery. The data analysis device is used to analyze the sensor data according to a preset neural network algorithm to determine the target patient's pain information, pelvic floor exercise information and current rehabilitation stage, and output a corresponding postoperative rehabilitation guidance plan. A data interaction device, connected to the data analysis device and the target patient's mobile device, is used to receive input data from the target patient or to send postoperative guidance information to the mobile device for display.

2. The postoperative pain intelligent management system based on multimodal data according to claim 1, characterized in that, The data acquisition device includes: Wearable physiological sensors for collecting heart rate variability signals and skin conductance signals; Camera device used to capture facial images and body posture images; The intelligent pelvic floor sensor is used to collect pressure signals and electromyographic signals of pelvic floor muscle contraction.

3. The postoperative pain intelligent management system based on multimodal data according to claim 1, characterized in that, The data analysis device includes: The feature extraction module is used to extract data features from the sensing data based on a neural network algorithm; The pain analysis module is used to analyze the pain information of the target patient based on the data characteristics. The exercise analysis module is used to analyze the exercise information of the target patient when performing pelvic floor rehabilitation exercises based on the data characteristics; the exercise information includes exercise posture, exercise compensation information, and movement quality information.

4. The postoperative pain intelligent management system based on multimodal data according to claim 3, characterized in that, The feature extraction module includes a multi-branch feature extraction network and a cross-modal fusion network. The multi-branch feature extraction network includes multiple feature extraction branches. The cross-modal fusion network uses an attention mechanism to adaptively weight and fuse the features extracted by each feature extraction branch to obtain fused feature data.

5. The postoperative pain intelligent management system based on multimodal data according to claim 4, characterized in that, The multi-branch feature extraction network includes: The physiological signal feature extraction branch adopts a temporal convolutional network structure, including multiple one-dimensional causal convolutional layers and residual connections. The convolutional kernel size is 3 to 7, the dilation factor increases exponentially, the number of network layers is 4 to 8, and the output feature dimension is 64 to 256, which is used to extract temporal features from heart rate variability signals and skin conductance signals. The facial expression feature extraction branch employs a lightweight convolutional neural network structure, including depthwise separable convolutional layers and global average pooling layers. The number of network parameters does not exceed 5M, and the output feature dimension is 64 to 128. The facial expression feature extraction branch also includes a facial action unit detection subnetwork for detecting pain-related facial action units. The facial expression feature extraction branch also includes an occlusion perception subnetwork for detecting facial occlusion areas and dynamically adjusting feature extraction weights. The body pose feature extraction branch adopts a graph convolutional network structure, with human skeleton key points as graph nodes and joint connections as graph edges. It includes multiple spatiotemporal graph convolutional layers, with 4 to 10 layers in the network and output feature dimensions of 64 to 256, which are used to extract pose temporal features from the skeleton key point sequence. The pelvic floor signal feature extraction branch adopts a combination structure of a one-dimensional convolutional neural network and a long short-term memory network. The one-dimensional convolutional neural network is used to extract local features of pressure signals and electromyographic signals. The hidden layer dimension of the long short-term memory network is 32 to 128, which is used to model the temporal dependency of the contraction process, and the output feature dimension is 32 to 128.

6. The postoperative pain intelligent management system based on multimodal data according to claim 3, characterized in that, The pain analysis module includes: The real-time pain assessment unit adopts a multi-layer fully connected neural network structure. The input is the fusion feature output by the cross-modal fusion network. The number of hidden layers is 2 to 4, and the number of neurons in each layer is 64 to 256. The output is a continuous pain level score from 0 to 10. The assessment cycle of the real-time pain assessment unit is 1 to 5 minutes, and it supports continuous real-time monitoring. The breakthrough pain prediction unit employs a temporal fusion Transformer network structure. The input is fused temporal data from the past 1 to 4 hours, including a multi-head self-attention layer, gated residual connections, and a temporal coding layer. The number of attention heads is 4 to 8, and the network has 2 to 6 layers. The prediction time windows of the breakthrough pain prediction unit include 15 minutes, 30 minutes, and 60 minutes, and the output is the probability value of breakthrough pain occurring within each time window. The breakthrough pain prediction unit also includes an uncertainty quantification subunit, which uses the Monte Carlo Dropout method to calculate the confidence interval of the prediction results. The analgesia regimen recommendation unit employs a reinforcement learning network structure based on a proximal policy optimization algorithm, comprising a policy network and a value network. The policy network takes into account the current pain assessment result, pain prediction result, cumulative medication dosage, and the patient's physiological state as input, and outputs recommended analgesia regimen actions, including suggestions for medication dosage adjustment, timing of medication administration, and non-pharmacological interventions. The value network is used to evaluate the expected rehabilitation benefits of the current state. The analgesia regimen recommendation unit also includes a safety constraint module to check whether the recommended regimen complies with maximum daily dose limits, minimum dosing intervals, and contraindication rules. When the breakthrough pain prediction unit outputs a breakthrough probability exceeding a preset threshold, the early warning generation unit generates graded early warning information and pushes it to the mobile application interaction module. The early warning levels include observation level, warning level, and emergency level.

7. The postoperative pain intelligent management system based on multimodal data according to claim 4, characterized in that, The exercise analysis module includes: The posture recognition unit employs a human posture estimation network, including a skeletal keypoint detection subnetwork and a posture classification subnetwork. The skeletal keypoint detection subnetwork detects at least 17 human keypoints, including the shoulder, hip, knee, and ankle, and outputs the two-dimensional coordinates and confidence scores of each keypoint. The posture classification subnetwork calculates posture parameters such as trunk tilt angle, hip flexion angle, and knee flexion angle based on the keypoint coordinates, and identifies the current training posture type, including supine, sitting, and standing positions. The compensatory action detection unit employs a multi-input classification network structure. The inputs include the temporal sequence of posture parameters output by the posture recognition unit and the pelvic floor features output by the pelvic floor signal feature extraction branch. The compensatory action detection unit detects abdominal compensation, hip compensation, thigh adduction compensation, and breath-holding compensation. The unit outputs the probability of occurrence and severity level for each type of compensatory action. The unit also includes a real-time feedback trigger subunit, which triggers an immediate correction prompt when a compensatory action is detected. The training quality assessment unit employs a multi-dimensional scoring network. Inputs include pelvic floor contraction parameters and compensation detection results. The pelvic floor contraction parameters include maximum contraction pressure, contraction rise time, contraction maintenance time, relaxation time, and contraction symmetry. The multi-dimensional movement quality score output by the training quality assessment unit includes strength score, endurance score, coordination score, relaxation score, and comprehensive score, with each dimension ranging from 0 to 100 points. The training plan generation unit uses a multi-armed slot machine algorithm based on Thompson sampling to dynamically adjust the training plan according to historical training data and training quality scores. The training plan includes training mode, contraction duration, relaxation duration, number of sets, and daily training frequency.

8. The postoperative pain intelligent management system based on multimodal data according to claim 3, characterized in that, The data analysis device also includes: A rehabilitation stage identification module is used to determine the current rehabilitation stage of the target patient based on the pain information and the exercise information. The postoperative guidance module is used to output a corresponding postoperative rehabilitation guidance plan based on the current rehabilitation stage. The coordination module is used to control the working status of the pain analysis module and the exercise analysis module according to the current rehabilitation stage.

9. The postoperative pain intelligent management system based on multimodal data according to claim 8, characterized in that, The rehabilitation stage identification module performs the following steps: Based on the target patient's surgical records, the corresponding surgical sites were determined; In a pre-defined rehabilitation stage database, historical site records corresponding to multiple rehabilitation stages for the surgically affected area are determined. Based on the exercise information, multiple movement data of the target patient whose movement quality information was lower than a preset quality score threshold were identified; Calculate the weighted summation average of the first site distances between the exercised area and the surgically affected area corresponding to each of the aforementioned action data to obtain action-related data; wherein, the calculation weight corresponding to each first site distance is inversely proportional to the action quality information corresponding to the corresponding action data; Based on the pain information, the pain locations of the target patient where multiple pain signals exceed a preset pain threshold are identified; The weighted summation average of the second site distances between each pain site and the surgically affected site is calculated to obtain pain correlation data; wherein, the calculation weight corresponding to each second site distance is proportional to the pain information corresponding to the corresponding pain site; The motion association data and the pain association data are vectorized to obtain site recording data; Calculate the similarity between the site-recorded data and the historical site-recorded data corresponding to each rehabilitation stage, and determine the rehabilitation stage with the highest similarity as the current rehabilitation stage of the target patient.

10. The postoperative pain intelligent management system based on multimodal data according to claim 9, characterized in that, The coordination module performs the following steps: Identify multiple historical data records corresponding to the current rehabilitation stage in the historical database; Calculate the proportion of a first data record in each of the historical data records where the pain information is greater than the pain threshold, and the proportion of a second data record where the action quality information is lower than the quality score threshold; The power consumption parameters of the pain module are obtained by multiplying the average proportion of the first data record in all the historical data records with the preset first power consumption weight. The power consumption parameters of the training module are obtained by multiplying the average proportion of the second data record in all the historical data records with the preset second power consumption weight. The power consumption parameters of the pain module and the exercise module are sent to the pain analysis module and the exercise analysis module, respectively, to control the operating power consumption.

Images