Pain discrimination method, system and device based on time sequence and multiple modes and storage medium

Abstract

The invention provides a pain discrimination method based on a time sequence and multiple modes. The pain discrimination method comprises the steps that conventional physiological indexes, behavior characteristic indexes and potential auxiliary indexes are collected; preprocessing the multi-modal data, including abnormal value processing, individual normalization and time alignment, and constructing a time sequence data block; establishing an individual baseline, dynamically calibrating a baseline threshold value every 7 days, and calibrating a blood glucose baseline once every 3 days for a diabetic patient; physiological features, behavior features and time sequence features of the time sequence data blocks are extracted through an AI model, multi-modal features are fused through an attention mechanism, and full-level pain is recognized through multi-modal index synchronous verification, multi-classification probability calculation and grading threshold judgment; and outputting a clinical intervention strategy corresponding to the pain level, and dynamically updating the model parameters and the baseline threshold based on clinical intervention feedback and new crowd data, so that various crowds can be covered, the hierarchical intervention strategy is output, the intervention delay is shortened, and the clinical practicability and accuracy of pain management are improved.

Description

Technical Field

[0001] This application relates to the field of medical monitoring technology, specifically to a time-series and multimodal pain discrimination method, system, device, and storage medium. Background Technology

[0002] The Numerical Rating Scale (NRS) is a quantitative method for comprehensively assessing the severity of nursing care. It effectively helps healthcare institutions evaluate the severity of patient care and provides valuable decision support for nursing teams. The NRS scoring system is one of the most commonly used assessment tools in the field of nursing. Pain, as an important stress response in the human body, especially severe pain with an NRS score of 8-10, can easily lead to complications such as a sudden increase in blood pressure, nausea and vomiting, worsening tissue damage, and anxiety and depression if not identified and intervened in a timely manner, seriously affecting patient prognosis and quality of life.

[0003] Traditional monitoring methods rely heavily on patient descriptions or visual assessments by healthcare professionals, failing to cover vulnerable populations such as those in a coma, infants, and those with cognitive impairments who cannot express pain. These individuals have low rates of identifying hidden pain, making their conditions highly susceptible to worsening due to undetected pain. Traditional methods also overly depend on subjective reporting, limiting their coverage. Furthermore, current technologies largely focus on modeling single physiological indicators (such as heart rate and skin conductance), making them susceptible to interference from irrelevant factors like exercise, diet, and ambient temperature. The accuracy rate for identifying severe pain is only 62%-75%, making it difficult to distinguish between pain stress and other physiological stress states. Single-indicator monitoring has weak anti-interference capabilities and low accuracy. Additionally, physiological differences exist among monitored individuals. Traditional methods do not consider individual baseline physiological variations (such as resting heart rate and basal blood glucose levels), using fixed thresholds for judgment, which easily leads to misdiagnosis, especially unsuitable for special populations with fluctuating physiological indicators, such as diabetic patients and the elderly.

[0004] Therefore, there is an urgent need to develop a method, system, device, and storage medium for temporal and multimodal pain discrimination to solve the above-mentioned technical problems.

[0005] It should be noted that the above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The aim is to address the aforementioned technical challenges by providing a time- and multimodal pain identification method, system, device, and storage medium that can cover various population groups, output tiered intervention strategies, shorten intervention delays, and improve the clinical applicability and accuracy of pain management.

[0007] In a first aspect, this application relates to a pain discrimination method based on temporal and multimodal data, comprising the following steps: S1: Collect multimodal data, including conventional physiological indicators, behavioral characteristic indicators, and potential auxiliary indicators; S2: Preprocess the multimodal data, including outlier handling, individual normalization and time alignment, to construct time-series data blocks; S3: Establish an individual baseline based on the patient’s resting state data from the previous 3 days, and dynamically calibrate the baseline threshold every 7 days. For diabetic patients, the blood glucose baseline is calibrated once every 3 days. S4: Extract physiological, behavioral, and temporal features from the time-series data blocks using an AI model, and fuse multimodal features using an attention mechanism. The fusion formula is as follows: +β +γ +δ ; in As a feature of fusion, These are normal physiological characteristics. For behavioral characteristics, As a potential auxiliary feature, For time-series features, α is the weight of conventional physiological features, β is the weight of behavioral features, γ is the weight of potential auxiliary features, and δ is the weight of time-series features, α+β+γ+δ=1; S5: Through simultaneous verification of multimodal indicators, multi-class probability calculation and grading threshold judgment, it identifies pain of all levels from NRS 0 to 10. S6: Output clinical intervention strategies corresponding to the pain level, and dynamically update model parameters and baseline thresholds based on clinical intervention feedback and data from newly added populations.

[0008] In one specific implementation scheme, the routine physiological indicators include cardiovascular system indicators, neuromuscular system indicators, nervous system indicators, respiratory system indicators, and endocrine system indicators; The cardiovascular system indicators include heart rate, systolic blood pressure, and heart rate variability; The neuromuscular system indicators include skin conductance and electromyography; The neurological indicators include electroencephalography (EEG) and pupil diameter; The respiratory system indicators include respiratory rate, and the endocrine system indicators include cortisol.

[0009] In one specific implementation scheme, the behavioral characteristic indicators include facial expressions, body movements, and voice characteristics; The facial expressions are captured by a camera and recognized using a CNN convolutional neural network. The limb movements are collected by a combination of a posture sensor and a camera, and identified using a skeletal key point detection algorithm. The sound features are collected through a microphone, and the fundamental frequency F0 and spectral entropy features are extracted.

[0010] In one specific feasible implementation, the potential auxiliary indicators include blood glucose, skin microcirculation blood flow velocity, and α-amylase activity; The blood glucose level was collected using a non-invasive near-infrared spectroscopy sensor. The skin microcirculation blood flow velocity was acquired using a wrist-mounted laser Doppler sensor; The α-amylase activity was collected using saliva test strips.

[0011] In one specific implementation scheme, step S2 uses the 3σ principle to remove extreme values ​​of physiological indicators and uses the inter-frame difference method to filter environmental interference to achieve abnormal value processing of behavioral characteristics. The individual normalization is achieved through the formula... Calculate the rate of change for each indicator, where R is the rate of change. These are real-time collected values ​​for the indicator. The baseline value for the individual indicator; The time alignment binds multimodal data through timestamps and constructs time-series data blocks in 5-minute units.

[0012] In one specific implementation scheme, step S4 extracts the physiological characteristics, behavioral characteristics, and time series characteristics of the time series data block using an AI model; The multimodal feature fusion using the attention mechanism includes physiological feature extraction, behavioral feature extraction, temporal feature extraction, and attention fusion. The physiological feature extraction is based on CNN convolutional neural network to extract time-domain and frequency-domain features; The behavioral feature extraction involves extracting graded features of facial expressions using a CNN convolutional neural network, extracting graded features of limb movement amplitude using a pose estimation algorithm, and extracting graded differences of fundamental frequency changes using a speech algorithm. The temporal feature extraction is achieved by using an LSTM (Long Short-Term Memory) network to capture the trend and peak synchronization of indicator changes over a continuous 25-minute period. The attention fusion uses the Transformer attention mechanism to assign modal weights, classifying pain levels into mild, moderate, and severe pain. The weights of the multimodal structures are dynamically adjusted according to different pain levels, with the weight allocation for different pain levels satisfying the following: for mild pain, the weight of behavioral characteristics is greater than or equal to the weight of conventional physiological characteristics; for severe pain, the weight of conventional physiological characteristics is greater than or equal to 0.5; and the weight of behavioral characteristics for special populations is 0.

[0013] In one specific implementation, the full-level pain recognition logic of step S5 includes: Simultaneous verification of multimodal indicators to determine whether changes in core indicators conform to the unidirectional trend and amplitude range of the corresponding level of pain; Multi-class probability calculation, outputting the probability distribution of each level of NRS 0-10 through a combination model of LSTM, Transformer and Softmax; A classification threshold is set, and a secondary verification is triggered when the probability of the classification falls below the threshold. Anti-interference mechanism: When a single modality index is abnormal, the weight of that modality is automatically reduced.

[0014] Secondly, this application provides a pain discrimination system based on temporal and multimodal modes, including: The multimodal data acquisition module integrates conventional physiological indicator acquisition equipment, behavioral feature acquisition equipment, and potential auxiliary indicator acquisition equipment to collect multimodal data. The data preprocessing module is used to perform outlier removal, individual normalization, and time alignment operations. The baseline calibration module is used to establish and dynamically calibrate individual baseline thresholds; The AI ​​feature fusion module is used to extract physiological features, behavioral features, and temporal features from time-series data blocks, and to fuse multimodal features through an attention mechanism. The full-level pain recognition module is used to perform NRS 0-10 full-level pain recognition through multimodal index synchronous verification, multi-class probability calculation and grading threshold judgment; The graded intervention output module is used to output clinical intervention strategies; The model update module is used to dynamically update model parameters based on clinical intervention feedback and data from newly added populations.

[0015] Thirdly, this application provides a computer device, characterized in that it includes: a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method.

[0016] Fourthly, this application provides a computer-readable storage medium, characterized in that a program or instructions are stored on the storage medium, and when the program or instructions are executed by a processor, the steps of the method are implemented.

[0017] The pain discrimination method, system, device, and storage medium based on time series and multimodal analysis provided in this application have at least the following beneficial effects: 1. By constructing multimodal data from conventional physiological indicators, behavioral characteristic indicators, and potential auxiliary indicators, and preprocessing it with individual normalization and time alignment, and extracting and fusing multidimensional features through an AI model, we have achieved accurate identification of pain levels from 0 to 10 on the NRS scale. The accuracy of full-level identification is high and the misjudgment rate is low, breaking through the limitations of single-dimensional monitoring and achieving accurate identification of pain levels. 2. Based on the patient's resting data from the previous 3 days, an individual baseline is established and dynamically calibrated. Combined with multimodal feature fusion and hierarchical threshold judgment logic, the influence of individual physiological baseline differences and clinical environment interference can be effectively eliminated. It maintains stable recognition performance in different populations and different clinical scenarios, which can meet diverse clinical monitoring needs. At the same time, it adapts to individual and scenario differences and improves the stability of recognition. 3. By integrating multimodal data collection with features, physiological and auxiliary indicators can be identified without relying on the patient's subjective expression. This makes it suitable for special populations who "cannot express pain," such as those in a coma, infants, and those with cognitive impairment. This improves the recognition rate of hidden pain in special populations, broadens the application scope of pain monitoring, and has a wide coverage. 4. This method covers the entire process from data collection to dynamic model updates. While outputting pain recognition results at all levels, it simultaneously matches graded intervention strategies, which significantly shortens the delay in pain intervention and improves the timeliness of clinical intervention. At the same time, it updates the model through clinical intervention feedback to continuously optimize the accuracy of recognition, thereby improving the clinical applicability of pain management. Attached Figure Description

[0018] Figure 1 This application provides a schematic flowchart of a pain discrimination method based on temporal and multimodal characteristics, according to an embodiment of the present application. Figure 2 : A schematic diagram of the structure of a time-series and multimodal pain discrimination system according to an embodiment of this application; Figure 3 : A schematic diagram of the structure of a computer device according to an embodiment of this application. Detailed Implementation

[0019] Preferred embodiments of this application are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of this application and are not intended to limit the scope of protection of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0020] It should be noted that, in the description of this application, the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, features defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and should not be construed as indicating or implying relative importance. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0021] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "setting" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0022] The following is in conjunction with the appendix Figures 1 to 3 This application will be described in further detail.

[0023] This application provides a pain discrimination method based on temporal and multimodal data, the method flow of which is as follows: Figure 1 As shown, it includes the following steps: S1: Collect multimodal data, including routine physiological indicators, behavioral characteristic indicators, and potential auxiliary indicators; S2: Preprocess the multimodal data, including outlier handling, individual normalization and time alignment, and construct time series data blocks; S3: Establish an individual baseline based on the patient’s resting state data from the previous 3 days, and dynamically calibrate the baseline threshold every 7 days. For diabetic patients, the blood glucose baseline is calibrated once every 3 days. S4: Extract physiological, behavioral, and temporal features from time-series data blocks using an AI model, and fuse multimodal features using a Transformer attention mechanism. The fusion formula is as follows: +β +γ +δ ; in As a feature of fusion, These are normal physiological characteristics. For behavioral characteristics, As a potential auxiliary feature, For time-series features, α is the weight of conventional physiological features, β is the weight of behavioral features, γ is the weight of potential auxiliary features, and δ is the weight of time-series features, α+β+γ+δ=1; S5: Through simultaneous verification of multimodal indicators, multi-class probability calculation and grading threshold judgment, it identifies pain of all levels from NRS 0 to 10. S6: Output clinical intervention strategies corresponding to the pain level, and dynamically update model parameters and baseline thresholds based on clinical intervention feedback and data from newly added populations.

[0024] The pain discrimination method based on temporal and multimodal characteristics proposed in this application has at least the following beneficial effects: (1) By constructing multimodal data from conventional physiological indicators, behavioral characteristic indicators and potential auxiliary indicators, and preprocessing them with individual normalization and time alignment, and extracting multidimensional features through AI model and fusing them, the accurate determination of pain at all levels of NRS 0-10 is achieved through the deep integration of the four-element weighted fusion formula and time series data blocks. The accuracy rate of full-level recognition is ≥88% (mild ≥85%, moderate ≥90%, severe ≥93%), which is 20%-30% higher than the single indicator scheme. The accuracy rate of full-level recognition is high and the level misjudgment rate is low (level misjudgment rate ≤5%), breaking through the limitations of single-dimensional monitoring and achieving accurate recognition of pain at all levels. (2) The pain recognition delay is ≤5 minutes for all levels, which is 25-55 minutes shorter than the traditional monitoring mode, providing a critical time window for timely intervention of different levels of pain and a fast response speed; (3) Based on the patient's resting data for the first 3 days, an individual baseline is established and dynamically calibrated. The combination of individual baseline calibration, multimodal redundancy verification and graded weight adjustment, through multimodal feature fusion and graded threshold judgment logic, effectively eliminates the influence of individual physiological baseline differences such as exercise, diet, environment, and clinical environment interference. It maintains stable recognition performance in different populations and different clinical scenarios, which can meet the diverse clinical monitoring needs, while adapting to individual and scenario differences and improving the stability of recognition. (4) By combining multimodal data collection and feature fusion, physiological indicators and auxiliary indicators can be identified without relying on the patient's subjective expression. The rate of hidden pain recognition in special populations (comatose, infants, and cognitively impaired individuals) is increased to 70%-85%. It can be adapted to special populations such as comatose, infants, and cognitively impaired individuals who "cannot express pain", thus improving the rate of hidden pain recognition in special populations, expanding the application scope of pain monitoring, and covering a wide range. (5) This method covers the entire process from data collection to dynamic model updates. While outputting pain recognition results of all levels, it simultaneously matches graded intervention strategies, which greatly shortens the delay of pain intervention and significantly improves the timeliness of clinical intervention. At the same time, it updates the model through clinical intervention feedback and continuously optimizes the accuracy of recognition, which can improve the clinical applicability of pain management.

[0025] In the embodiments of this application, in step S1, the conventional physiological indicators include cardiovascular system indicators, neuromuscular system indicators, nervous system indicators, respiratory system indicators, and endocrine system indicators. The cardiovascular system indicators include heart rate (HR), systolic blood pressure (SBP), and heart rate variability (HRV). Heart rate (HR) is acquired by a photoplethysmography (PPG) sensor or an electrocardiogram (ECG) module. The PPG sensor and ECG module acquire cardiovascular data such as heart rate (HR) and heart rate variability (HRV) in real time, achieving non-invasive monitoring through skin contact sensing. The heart rate measurement accuracy of the PPG sensor is ±1 beat / min; the input noise of the ECG module is ≤5μV. Systolic blood pressure (SBP) is automatically measured every 10 minutes by a non-invasive blood pressure sensor, with an accuracy of ±3mmHg. Heart rate variability (HRV) is acquired through ECG or PPG signal analysis, with a heart rate (HR) sampling rate ≥10Hz and systolic blood pressure (SBP) collected once every 10 minutes. Neuromuscular system indicators include electrical conductance analysis (EDA) and electromyography (EMG), acquired via skin electrodes and surface EMG sensors, respectively, capturing skin conductance and skeletal muscle electromyography signals to reflect neuromuscular stress responses. The EDA sampling rate is ≥5Hz, and the EMG sampling rate is ≥100Hz. Neurological system indicators include electroencephalography (EEG) and pupil diameter, acquired via a portable dry electrode headband and an infrared pupil monitoring module, respectively. The EEG sampling rate is ≥250Hz, and pupil diameter is measured once per minute. The infrared pupil monitoring module has an accuracy of ±0.1mm, capturing real-time changes in pupil diameter and simultaneously acquiring neurological responses. Respiratory system indicators include respiratory rate (RR), acquired via a chest patch respiratory sensor or a photoplethysmography (PPG) derived algorithm, with a sampling rate ≥2Hz. Endocrine system indicators include cortisol in saliva or blood, collected using saliva test strips or minimally invasive electrochemical sensors, once every 30 minutes. Saliva samples are collected in a minimally invasive manner to detect endocrine system indicators.

[0026] Behavioral feature indicators include facial expressions, body movements, and vocal features. Facial expressions are captured using a ≥1080P camera at a frame rate of 30 frames per second. A CNN convolutional neural network model is used, based on the revised Wong-Baker Faces Pain Scale (FPS-R) for recognition. Body movements are recognized using a combination of a pose sensor and a camera, employing a skeletal keypoint detection algorithm at a frame rate of ≥15 frames per second. In this embodiment, the skeletal keypoint detection algorithm uses the OpenPose open-source algorithm (or the MediaPipe pose estimation algorithm), extracting 17 core skeletal keypoints (such as head, shoulder, elbow, wrist, etc.), and calculating the displacement deviation of these keypoints as body movement features. Sound features are collected via a microphone, extracting the fundamental frequency F0 and spectral entropy features. The sampling rate is ≥16kHz. In this embodiment, the fundamental frequency F0 is extracted using the autocorrelation method (sampling window 20ms, step size 10ms). The spectral entropy is calculated based on a 1024-point FFT with a 50% overlap window, with a frequency range of 20Hz-8kHz and an entropy range of [0,1]. A posture sensor is integrated into the patient's clothing or wearable terminal. The posture sensor has a sampling rate of 100Hz, an accelerometer measurement range of ±2g (accuracy ±0.01g), and a gyroscope measurement range of ±500° / s (accuracy ±0.5° / s), monitoring limb movement displacement in real time. The microphone collects sound signals, filters environmental noise, and extracts pain-related acoustic features such as groans and rapid breathing.

[0027] Potential auxiliary indicators include blood glucose, skin microcirculation blood flow velocity, and α-amylase activity. Blood glucose was collected every 15 minutes using a non-invasive near-infrared spectroscopy sensor to monitor changes in blood glucose levels. The accuracy of the non-invasive near-infrared spectroscopy sensor is ±1.0 mmol / L. α-amylase activity was measured in U / mL, with a normal range of 100-300 U / mL. Changes during pain were calculated as the percentage increase relative to the individual's baseline. Skin microcirculation blood flow velocity was collected every 20 minutes using a wrist-type laser Doppler sensor to measure skin microcirculation blood flow velocity. α-amylase activity was collected using saliva test strips. The α-amylase activity detection procedure was as follows: after collecting the sample with the saliva test strip and allowing it to stand for 5 minutes, the activity value was read using an electrochemical sensor. The rate of change during pain was calculated as R = (detected value - individual baseline value) / individual baseline value. This, along with blood glucose and skin microcirculation blood flow velocity, constitutes a potential auxiliary characteristic and is included in the fusion calculation. Potential ancillary indicators, along with routine physiological and behavioral characteristics, constitute three types of input data. This approach fully considers the individual physiological baseline differences (resting heart rate and baseline blood glucose levels) of special populations such as diabetic patients and the elderly, effectively eliminating individual physiological baseline differences and demonstrating high stability and accuracy. Including potential ancillary indicators as a type of input data expands the dimensions of multimodal data, which is beneficial for improving the accuracy and applicability of data collection.

[0028] The computing and storage devices in this application perform real-time data preprocessing and model inference on edge computing terminals, model training, data storage and system updates on cloud servers (such as Alibaba Cloud), and a dedicated monitoring APP is installed on medical devices (such as tablets), while patient devices (such as wearable devices) are equipped with graded warning indicator lights.

[0029] Furthermore, pain levels were categorized into mild pain (levels 1-3), moderate pain (levels 4-6), and severe pain (levels 7-10). Eight hundred subjects from Wuxi People's Hospital were selected, distributed by pain level as follows: no pain (level 0) 100 cases, mild pain (levels 1-3) 200 cases, moderate pain (levels 4-6) 250 cases, and severe pain (levels 7-10) 150 cases. 100 special populations (ICU comatose / infants / cognitive impairment) were also included. The age range was 0-85 years, including 500 adults, 150 children, and 150 elderly individuals. 120 patients had diabetes. Data were collected from all 800 subjects. Subjects with severe arrhythmias or end-stage renal disease were excluded for this clinical validation. This study has been reviewed and approved by the Ethics Committee of Wuxi People's Hospital. All subjects or their guardians have signed informed consent forms. Data collection and use comply with the Personal Information Protection Law and the Medical Data Security Guidelines.

[0030] The data collection process included baseline data collection, pain state data collection, special population data collection, and data synchronization. Baseline data collection involved collecting 30-minute resting state data at 8:00, 12:00, 16:00, and 20:00 daily for the first 3 days after admission. Healthy volunteers (level 0 pain) also had their data collected simultaneously. Pain state data collection involved continuously collecting data on patients with different levels of pain during pain attacks until 1 hour after intervention. Healthy volunteers simulated mild, moderate, and severe pain through gradient cold stimulation (0℃, 5℃, 10℃), and data was collected during stimulation and for 30 minutes after stimulation ended. Cold stimulation used a high-precision temperature-controlled patch (temperature error ±0.5℃), and the stimulation site was the inner forearm. Each level of stimulation lasted 30 seconds, with a 1-minute interval between levels. After stimulation, there was a 5-minute rest period. The temperature gradient (10℃→5℃→0℃) was gradually increased. Before each level of stimulation, the subject's tolerance was assessed, and if significant discomfort (such as chills or skin redness and swelling) occurred, the stimulation was immediately terminated. Special population data collection involves continuous data collection from ICU comatose patients, infants, and patients with cognitive impairment. Pain levels are labeled daily by medical staff based on clinical symptoms (e.g., for postoperative comatose patients, labels are based on changes in indicators and wound condition). All devices connect to edge terminals via Bluetooth 5.0 or Wi-Fi to achieve real-time data synchronization with a timestamp error ≤10ms.

[0031] Data preprocessing employs the 3σ principle to remove extreme data such as HR > 200 Hz, SBP > 200 mmHg, and EDA > 100 μS. Behavioral features are filtered for environmental interference (such as background movement and irrelevant sounds) using inter-frame differencing to eliminate outliers. The 3σ principle is calculated based on individual baselines, removing extreme data exceeding "individual baseline mean ± 3 × individual baseline standard deviation" to ensure outlier removal is adapted to individual differences. In this embodiment, inter-frame differencing uses a 3-frame sliding window, with a frame difference pixel threshold set at 5% of the total image pixels. Values ​​exceeding this threshold are considered environmental interference and the behavioral feature data for that frame is removed. (The formula is used to...) Calculate the rate of change, example: a patient's baseline heart rate is 70 beats / min ( =70), 84 beats / min for mild pain ( =84), =0.2, meaning a 20% increase in heart rate, falls within the range of mild pain variation. This eliminates individual baseline differences and highlights the relative magnitude of change corresponding to different pain levels, achieving individual normalization. Data collected from multiple devices is bound using timestamps, and time-series data blocks are constructed in 5-minute units. These blocks are divided into non-overlapping, continuous blocks, with one block generated every 5 minutes. The first block covers 0-5 minutes, the second covers 5-10 minutes, and so on, generating a total of 20,000 valid data blocks. Each block contains the absolute values, rates of change, and trend characteristics of multimodal indicators. Labels are added according to pain level, matching the peak characteristics of indicators for different pain levels (mild pain peaks at 5-10 minutes, moderate at 10-15 minutes, and severe at 15-20 minutes), achieving time alignment.

[0032] The outlier handling, individual normalization, and time alignment in this application are combined to construct a time-series data block, providing a foundation for subsequent time-series feature extraction. Individual normalization is a solution to the problem of large individual differences in pain discrimination. By combining outlier handling, individual normalization, and time alignment, individual differences can be reduced while retaining time-series information, which can effectively solve the problem of large individual differences without losing time-series features and improve the accuracy of discrimination.

[0033] Individual baseline calibration is achieved through baseline establishment, dynamic calibration, and graded baseline adjustment. Baseline establishment involves collecting resting state data from patients for the first 3 days (30 minutes each at 8:00, 12:00, 16:00, and 20:00 daily), and calculating the mean ± standard deviation of each indicator as the individual baseline. Dynamic calibration involves re-collecting baseline data every 7 days and updating the baseline in real time based on changes in the patient's health status (such as postoperative recovery and medication adjustments). For diabetic patients, the blood glucose baseline calibration frequency is increased (once every 3 days) to adapt to blood glucose fluctuations. Graded baseline adjustment establishes "grade-specific baseline deviation ranges" for different pain levels: EDA (mild 10%-20%, moderate 20%-40%, severe 40%-60%); respiratory rate (mild 10%-25%, moderate 25%-50%, severe 50%-100%); blood glucose (mild 5%-15%, moderate 15%-30%, severe 30%-50%), providing a basis for grade identification. Individual baselines are established based on "resting state data from the previous 3 days". For general scenarios, dynamic calibration is performed every 7 days, while for diabetic patients, the blood glucose baseline is calibrated separately every 3 days. For special populations in pain assessment, such as diabetic patients who often have neuropathic pain and whose blood glucose fluctuations affect pain assessment, targeted baseline calibration rules are designed to achieve hierarchical calibration of personalized dynamic benchmarks. This can meet the needs of diverse clinical monitoring and adapt to individual and scenario differences, which is conducive to improving the stability of recognition.

[0034] The physiological feature extraction in this application is based on CNN convolutional neural networks to extract temporal features (mean, variance, rate of change) and frequency domain features (power spectral density), focusing on capturing the feature differences of different levels of pain (e.g., the rate of change of EDA is significantly higher in severe pain than in mild pain). In this embodiment, the CNN convolutional neural network is a ResNet18 architecture with pre-trained weight initialization and fine-tuning of the last 3 layers. The activation function of the convolutional layers is ReLU, the pooling layers are max pooling (2×2 window, stride 2), and the dropout layer is placed before the fully connected layers. The power spectral density is calculated by FFT, with 1024 sampling points and a frequency range of 0-50Hz. Behavioral features are extracted by CNN convolutional neural networks to extract hierarchical features of facial expressions (e.g., deep features corresponding to the revised Wong-Baker Facial Expression Pain Assessment Method FPS-R score), the pose estimation algorithm to extract the amplitude hierarchical features of limb movements, and the speech algorithm to extract the hierarchical differences in fundamental frequency changes.

[0035] The temporal features are captured using an LSTM (Long Short-Term Memory) network (256 hidden nodes, dropout=0.3) to capture the trend and peak synchronicity of indicator changes within 25 consecutive minutes, distinguishing between "continuous pain" and "transient discomfort". In this embodiment, the LSTM network has 2 hidden layers, the input dimension is the dimension after concatenation of multimodal features (512 dimensions), the output dimension is 256 dimensions, the activation function is tanh, the recursion step size is set to 5 (corresponding to 5 minutes of temporal data blocks, with 1 minute of data per step), and the dropout layer is placed between the hidden layers.

[0036] Feature fusion employs a Transformer attention mechanism to dynamically allocate modality weights. The Transformer encoder uses pre-normalization, the Feed-Forward network's intermediate layers have a dimension of 2048, the activation function is GELU, and both the encoder and attention layers have a dropout rate of 0.1. Weights are dynamically adjusted for different pain levels (behavioral feature weights are increased to 0.4 for mild pain and physiological feature weights are increased to 0.5 for severe pain), as shown in the following formula: +β +γ +δ ; Constraints: α+β+γ+δ=1, α,β,γ,δ∈[0,1] Among them, mild pain (levels 1-3): α=0.3, β=0.4, γ=0.2, δ=0.1; moderate pain (levels 4-6): α=0.4, β=0.3, γ=0.2, δ=0.1; severe pain (levels 7-10): α=0.5, β=0.2, γ=0.2, δ=0.1; special populations (comatose / infants): α=0.5, β=0, γ=0.3, δ=0.2.

[0037] (1) Mild pain (grades 1-3): +0.4 +0.2 +0.1 ; (2) Moderate pain (4-6): +0.3 +0.2 +0.1 ; (3) Severe pain (7-10): +0.5 +0.2 +0.1 ; (4) Special populations (comatose / infants and young children): +0 +0.3 +0.2 ; The fusion feature vectors for different scenarios are determined by optimizing each modality through 5-fold cross-validation and grid search.

[0038] The weighting of different pain levels in this application follows the logic that the higher the pain level, the greater the weight of physiological characteristics. The weighting of different pain levels satisfies the following conditions: for mild pain, the weight of behavioral characteristics is greater than or equal to the weight of conventional physiological characteristics; for severe pain, the weight of conventional physiological characteristics is greater than or equal to 0.5; and for special populations, the weight of behavioral characteristics is 0. The multimodal feature weighting strategy aligns with the physiological mechanisms and clinical characteristics of pain. This weighting is based on the significant differences in conventional physiological features, behavioral features, potential auxiliary features, and temporal features across different pain levels. For mild pain (levels 1-3), the human body's stress response is weak, and the fluctuation range of conventional physiological indicators (such as heart rate and blood pressure) is small. The human body will express pain through facial expressions, slight limb movements and other behaviors, which have higher recognition. Therefore, the weight of behavioral characteristics β=0.4 is set higher than the weight of conventional physiological characteristics α=0.3. For moderate pain (level 4-6), the systemic stress response triggered by pain is manifested, and significant detectable changes occur in physiological indicators such as heart rate variability and skin conductance. At the same time, the incremental value of behavioral characteristics decreases, and the distinguishing power of movement amplitude is not as good as that of physiological indicators. Therefore, the weight of conventional physiological characteristics is increased to α=0.4, which is slightly higher than the weight of behavioral characteristics β=0.3, to balance the contribution of physiological characteristics and behavioral characteristics. For severe pain (7-10), intense pain can cause significant fluctuations in physiological indicators such as heart rate and cortisol, which become the core basis for pain assessment. At the same time, patients are unlikely to generate effective behavioral signals due to severe pain. Therefore, the weight of conventional physiological characteristics is increased to α=0.5 to ensure the dominance of core characteristics, while the weight of behavioral characteristics is reduced to reduce interference from invalid signals. For special populations such as those in a coma or infants, who cannot generate identifiable behavioral characteristics, the weight of behavioral characteristics β is set to 0. At the same time, the weight of potential auxiliary characteristics γ is increased to 0.3 to supplement the physiological baseline differences of special populations. The weight of temporal characteristics δ is increased to 0.2 to capture the continuous trend of physiological indicators and effectively distinguish between continuous fluctuations caused by pain and occasional abnormal indicators.

[0039] The constraint α+β+γ+δ=1 in this scheme ensures that the contribution of the four features is a relatively quantified relationship, avoiding double calculation or weight imbalance, and guaranteeing the repeatability of the technical solution.

[0040] The full-level pain recognition model employs an LSTM (Long Short-Term Memory) network for temporal feature capture, a Transformer attention mechanism for multimodal fusion, and a Softmax multi-classification combination model. It takes the fused features and the individual baseline deviation as input and outputs the probability distribution (P0, P1, ..., P10) of each level of the NRS score from 0 to 10. The recognition logic is as follows: (1) Multimodal index synchronous verification: Determine whether the changes in core indicators conform to the "unidirectional trend and amplitude range" of the corresponding level of pain. Core indicators are defined as cardiovascular system indicators (HR, SBP, HRV) and neuromuscular system indicators (EDA, EMG) among routine physiological indicators. Mild pain requires that the change rate of ≥2 types of core indicators be in the range of 10%-20%. (2) Multi-class probability calculation: The probability of each pain level is output by the Softmax function, and the level corresponding to the maximum probability is taken as the recognition result; (3) Classification threshold judgment: preset classification probability threshold (level 0: P0≥0.8; levels 1-3: P1 3 ≥ 0.75; Levels 4-6: P4 6 ≥ 0.8; Levels 7-10: P7 When the value is ≥0.85, a "secondary verification" is triggered if it falls below the corresponding threshold. During the "secondary verification," the temporal feature weight δ is doubled (e.g., the original δ=0.1 is adjusted to 0.2, and other weights are normalized proportionally to maintain α+β+γ+δ=1), and the multi-class probability is recalculated; level 0 represents no pain, level 1 represents no pain. Level 3 is mild pain, levels 4-6 are moderate pain, and levels 7-10 are severe pain. The secondary verification in this application achieves more accurate calculation results by selectively increasing the weights of time-series features (doubling the weights) while normalizing the proportions of other weights and through refined identification. Formula: δ′=2 , α′= ,β′= , γ′= , For example, when δ0=0.1, δ'=0.2, α'=0.3× =0.267 Constraints: α′+β′+γ′+δ′=1, / δ′ represents the temporal feature weights before / after the second verification. / α′、 / β′、 / γ′ represents the weights before / after other modal validations. These are the weight normalization coefficients; Anti-interference mechanism: When a single modality indicator is abnormal but other modalities do not change synchronously, the weight of that modality is automatically reduced (e.g., when exercise only causes an increase in heart rate, the HR weight is reduced to 0.1) to avoid misjudgment of the level. The criterion for judging the abnormality of a single modality indicator is "the real-time value of the indicator exceeds the individual baseline mean ± 3 × individual baseline standard deviation", and the abnormality is not synchronously verified by other core modality indicators (e.g., only HR is abnormal, and EDA and EMG do not change synchronously).

[0041] Existing technologies for preventing interference mostly involve directly removing anomalous modal data. The interference prevention mechanism of this application, however, retains anomalous modal data but dynamically reduces its weight to avoid data loss. ′= K, where K , Derivative fusion formula: = + +…+ , Constraints: =1 (total weight remains unchanged) / ′ represents the weight before / after the i-th modality adjustment, and K is the anti-interference weight coefficient (0.1 in this invention, which can be fine-tuned according to the scenario). These are the features of each modality.

[0042] Existing technologies often employ global retraining when updating models based on clinical feedback, which is time-consuming and lacks specificity. In contrast, this application performs small-scale iterative updates on the weights of "pain level after intervention index decline," achieving precise fine-tuning and quantification.

[0043] = +η ( ) , Constraints: ≤0.1, =1, —Modal weight of the nth pain level in round t; —The adjusted weights in round t+1 after the intervention feedback; η—fine-tuning the learning rate (in this application, η=1e) 6); —Comprehensive score of actual pain level after intervention; —Comprehensive score of predicted pain level before intervention; Different intervention strategies and output formats are matched according to different pain levels, as detailed in the table below:

[0044] Note: The use of prescription drugs (such as codeine and morphine) involved in the above intervention strategies must strictly comply with the relevant national regulations on the management of narcotic drugs, and must be prescribed by a licensed physician based on the patient's condition. This application does not involve the production, sale or specific usage methods of the above drugs.

[0045] This application continuously optimizes the performance of graded pain recognition through dynamic updates. Intervention feedback updates involve acquiring multimodal data changes within one hour after a graded intervention (e.g., the index drops back to the mild range after a moderate pain intervention), and then fine-tuning the feature weights and probability thresholds for the corresponding grade. This fine-tuning uses a learning rate of 1e-6 and 20 iterations, adjusting only the modal weights for the corresponding pain level (with an adjustment range not exceeding ±0.1), and the probability threshold is fine-tuned by ±0.02. Population data updates incorporate ≥1000 new clinical cases (covering all pain levels) every quarter. Pain level labeling is independently completed by two attending physicians, with a Kappa value of ≥0.85 for labeling consistency; inconsistent cases are adjudicated by third-party experts. Transfer learning optimizes the model's generalization ability, ensuring an accuracy rate of ≥85% for all pain levels across different populations, including adults, children, and the elderly.

[0046] The dataset was divided into three independent parts: training set, validation set, and test set. The training set was allocated 14,000 data blocks (70%), covering all pain levels and population types; the validation set was allocated 4,000 data blocks (20%); and the test set was allocated 2,000 data blocks (10%), containing data from 150 new subjects (covering all pain levels and special populations) independent of the training set.

[0047] In the embodiments of this application, the AI ​​model is trained by first extracting features and then training and transferring them using a Transformer model. The CNN convolutional neural network adopts the ResNet18 architecture, with an output feature dimension of 512. Secondly, based on 5-fold cross-validation, the number of hidden layer nodes in the LSTM is determined to be 256, with dropout=0.3, to achieve feature extraction. Finally, the Transformer attention mechanism model is trained with an encoder layer of 6, an attention head of 8, and a hidden layer dimension of 512. The training configuration is AdamW optimizer β1=0.9, β2=0.999 (initial learning rate 1e-4, weight decay 1e-5), batch size 32, and a weighted cross-entropy loss function. The training data is augmented by random flipping and time-axis shifting. To address the imbalance problem of samples at different pain levels, in addition to the weighted cross-entropy loss function, a combination of oversampling (mild pain samples) and undersampling (moderate pain samples) is used to balance the sample distribution. Weights were inversely proportional to the sample size of pain level (0 level 0 weight 0.8, mild 1.0, moderate 0.9, severe 1.2, special populations 1.5), with 100 training epochs and an early stopping strategy (if the accuracy of full-level recognition on the validation set does not improve for 5 consecutive epochs). A model trained on adult data was used as the pre-trained model, and fine-tuning was performed for transfer learning in children, the elderly, and diabetic patients. Except for the learning rate of 1e-5 and 50 iterations, the batch size remained constant at 32, and the loss function weights were adjusted according to population characteristics (children: severe pain weight 1.3; elderly: moderate pain weight 1.1; diabetic patients: blood glucose-related feature weight increased by 0.1). Specifically, during transfer learning for diabetic patients, the weight of the blood glucose sub-feature in the potential auxiliary features was increased by 0.1, while the overall weight γ of the potential auxiliary features remained unchanged at 0.2, and the weights of other modalities remained unchanged.

[0048] The model performance metrics on the test set are shown in the table below:

[0049] The edge computing terminal receives data from various sensors in real time with a transmission latency of ≤100ms. It performs data preprocessing, feature extraction, and model inference with a single data block inference time of ≤200ms. Based on the recognition results, it triggers corresponding level warnings (indicator lights + text prompts) and simultaneously pushes intervention strategies to the patient and medical staff ends to achieve hierarchical warning output.

[0050] The medical staff app displays the patient's pain level, multimodal data change curves, baseline values, and intervention suggestions, enabling real-time monitoring of patient information. Medical staff can record analgesic interventions (dosage, method, and time) and upload them synchronously to the cloud. It also supports querying the patient's pain level change trend, intervention records, and data statistical reports over the past year. In addition, the medical staff app supports simultaneous monitoring of multiple patients, with priority alerts for patients with severe pain, enabling batch management.

[0051] The cloud server uses MySQL to store structured data (baseline, monitoring results, intervention records) and MongoDB to store unstructured data (images, audio), with AES-256 encryption protection, in compliance with the "Medical Data Security Guidelines" for data storage; the model is retrained quarterly based on new clinical data, and update packages are pushed to edge terminals via OTA for model management; hierarchical access permissions are set for doctors, nurses, and administrators, and data access logs are retained for one year and are traceable for access control.

[0052] It should be noted that clinical validation was conducted on the experimental and control groups. The experimental group consisted of 400 patients monitored using the methods and systems described in this application, while the control group consisted of 400 patients monitored using the traditional monitoring model, namely, ward rounds by medical staff and subjective reports from patients. The validation period was 7 days per case. Indicators such as the accuracy rate of pain identification at all levels, the misjudgment rate of pain levels, the intervention delay time, the incidence of complications, and patient satisfaction were recorded. The validation results are shown in the table below:

[0053] The experimental group showed significantly better performance than the control group in terms of accuracy in pain recognition across all levels, misjudgment rate of pain level, and intervention delay, and no device-related adverse events occurred, demonstrating the clinical effectiveness and safety of the proposed solution.

[0054] This application addresses existing pain discrimination methods by employing individual normalization, targeted baseline calibration for diabetic patients, and a three-tiered progressive discrimination to achieve accurate identification of NRS scores across the entire range of 0-10. In particular, it designs a separate calibration cycle for blood glucose baselines for diabetic patients, a special population experiencing pain, thus resolving the assessment bias problem in diabetic patients with pain. This approach enables accurate identification of diabetic patients with pain, improves identification accuracy, and offers broad coverage and high identification stability.

[0055] like Figure 2 As shown in the embodiments of this application, a pain discrimination system based on temporal and multimodal modes is also proposed, including: The multimodal data acquisition module integrates conventional physiological indicator acquisition equipment, behavioral feature acquisition equipment, and potential auxiliary indicator acquisition equipment to collect multimodal data. The data preprocessing module is used to perform outlier removal, individual normalization, and time alignment operations. The baseline calibration module is used to establish and dynamically calibrate individual baseline thresholds; The AI ​​feature fusion module is used to extract physiological features, behavioral features, and temporal features from time-series data blocks, and to fuse multimodal features through an attention mechanism. The full-level pain recognition module is used to perform NRS 0-10 full-level pain recognition through multimodal index synchronous verification, multi-class probability calculation and grading threshold judgment; The graded intervention output module is used to output clinical intervention strategies; The model update module is used to dynamically update model parameters based on clinical intervention feedback and data from newly added populations.

[0056] Furthermore, embodiments of this application also propose a computer device, including: a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the above-described method.

[0057] Furthermore, embodiments of this application also propose a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement a time-series and multimodal pain discrimination method.

[0058] It should be noted that the time-series and multimodal pain discrimination system provided in this embodiment is only an example of the above-described division of functional modules / units. In practical applications, the above functions can be assigned to different functional modules / units as needed, that is, the internal structure of the time-series and multimodal pain discrimination system can be divided into different functional modules / units to complete all or part of the functions described above. Furthermore, the implementation method of the time-series and multimodal pain discrimination method provided in the above-described method embodiment and the implementation method of the time-series and multimodal pain discrimination system provided in this embodiment belong to the same concept. For details of the specific implementation process of the time-series and multimodal pain discrimination system provided in this embodiment, please refer to the above-described method embodiment, which will not be repeated here.

[0059] Figure 3The diagram illustrates the structure of a computer device according to an embodiment of this application. The computer device can be a desktop computer, laptop computer, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. The processor and memory can be connected via a bus or other means. The processor can be a Central Processing Unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, graphics processing units (GPUs), embedded neural network processing units (NPUs) or other dedicated deep learning coprocessors, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0060] The processor may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array). The processor may also include a main processor and coprocessors. The main processor, also known as the CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content required to be displayed on the screen. In some embodiments, the processor may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0061] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the above embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the methods in the above embodiments. The memory may include a program storage area and a data storage area, wherein the program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0062] In some embodiments, the computer device may also optionally include: a peripheral device interface and at least one peripheral device. The processor, memory, and peripheral device interface can be connected via a bus or signal lines. Each peripheral device can be connected to the peripheral device interface via a bus, signal lines, or a circuit board. Specifically, the peripheral device includes at least one of: a radio frequency circuit, a display screen, and a keyboard.

[0063] Peripheral device interfaces can be used to connect at least one I / O (Input / Output) related peripheral device to the processor and memory. In some embodiments, the processor, memory, and peripheral device interface are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor, memory, and peripheral device interface can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0064] The display screen is used to display the UI (User Interface). This UI can include graphics, text, icons, videos, and any combination thereof. When the display screen is a touch screen, it also has the ability to collect touch signals on or above the surface of the display. These touch signals can be input as control signals to a processor for processing. In this case, the display screen can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there may be one display screen, located on the front panel of the computer device; in other embodiments, there may be at least two display screens, respectively located on different surfaces of the computer device or in a folded design; in still other embodiments, the display screen may be a flexible display screen, located on a curved or folded surface of the computer device. Furthermore, the display screen can be configured as a non-rectangular, irregular shape, i.e., a non-rectangular screen. The display screen can be made of materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

[0065] A power supply is used to power the various components in a computer device. The power supply can be alternating current (AC), direct current (DC), a disposable battery, or a rechargeable battery. When the power supply includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is charged via a wired connection, while a wireless rechargeable battery is charged via a wireless coil. The rechargeable battery can also be used to support fast charging technology.

[0066] Those skilled in the art will understand that the structure shown in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0067] This application also discloses a computer-readable storage medium. Specifically, the computer-readable storage medium is used to store a computer program, which, when executed by a processor, implements the methods described in the above-described method embodiments. Those skilled in the art will understand that implementing all or part of the processes in the methods described above can be accomplished by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0068] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

Claims

1. A pain discrimination method based on temporal and multimodal data, characterized in that, Includes the following steps: S1: Collect multimodal data, including conventional physiological indicators, behavioral characteristic indicators, and potential auxiliary indicators; S2: Preprocess the multimodal data, including outlier handling, individual normalization and time alignment, to construct time-series data blocks; S3: Establish an individual baseline based on the patient’s resting state data from the previous 3 days, and dynamically calibrate the baseline threshold every 7 days. For diabetic patients, the blood glucose baseline is calibrated once every 3 days. S4: Extract physiological, behavioral, and temporal features from the time-series data blocks using an AI model, and fuse multimodal features using a Transformer attention mechanism. The fusion formula is as follows: +b +g +d ; in As a feature of fusion, These are normal physiological characteristics. For behavioral characteristics, As a potential auxiliary feature, For time-series features, α is the weight of conventional physiological features, β is the weight of behavioral features, γ is the weight of potential auxiliary features, and δ is the weight of time-series features, α+β+γ+δ=1; S5: Through simultaneous verification of multimodal indicators, multi-class probability calculation and grading threshold judgment, it identifies pain of all levels from NRS 0 to 10. S6: Output clinical intervention strategies corresponding to the pain level, and dynamically update model parameters and baseline thresholds based on clinical intervention feedback and data from newly added populations.

2. The pain discrimination method based on temporal and multimodal data according to claim 1, characterized in that, The routine physiological indicators include cardiovascular system indicators, neuromuscular system indicators, nervous system indicators, respiratory system indicators, and endocrine system indicators; The cardiovascular system indicators include heart rate, systolic blood pressure, and heart rate variability; The neuromuscular system indicators include skin conductance and electromyography; The neurological indicators include electroencephalography (EEG) and pupil diameter; The respiratory system indicators include respiratory rate, and the endocrine system indicators include cortisol.

3. The pain discrimination method based on temporal and multimodal data according to claim 1, characterized in that, The behavioral characteristic indicators include facial expressions, body movements, and voice characteristics; The facial expressions are captured by a camera and recognized using a CNN convolutional neural network. The limb movements are collected by a combination of a posture sensor and a camera, and identified using a skeletal key point detection algorithm. The sound features are collected through a microphone, and the fundamental frequency F0 and spectral entropy features are extracted.

4. The pain discrimination method based on time series and multimodal as described in claim 1, characterized in that, The potential auxiliary indicators include blood glucose, skin microcirculation blood flow velocity, and α-amylase activity; The blood glucose level was collected using a non-invasive near-infrared spectroscopy sensor. The skin microcirculation blood flow velocity was acquired using a wrist-mounted laser Doppler sensor; The α-amylase activity was collected using saliva test strips.

5. The pain discrimination method based on temporal and multimodal data according to claim 1, characterized in that, In step S2, the 3σ principle is used to remove extreme values ​​of physiological indicators, and environmental interference is filtered out through the inter-frame difference method to achieve abnormal value processing of behavioral features. The individual normalization is achieved through the formula... Calculate the rate of change for each indicator, where R is the rate of change. These are real-time collected values ​​for the indicator. The baseline value for the individual indicator; The time alignment binds multimodal data through timestamps and constructs time-series data blocks in 5-minute units.

6. The pain discrimination method based on temporal and multimodal data according to claim 1, characterized in that, Step S4 uses an AI model to extract the physiological, behavioral, and temporal features of the time-series data block. The use of the Transformer attention mechanism to fuse multimodal features includes physiological feature extraction, behavioral feature extraction, temporal feature extraction, and attention fusion. The physiological feature extraction is based on CNN convolutional neural network to extract time-domain and frequency-domain features; The behavioral feature extraction involves extracting graded features of facial expressions using a CNN convolutional neural network, extracting graded features of limb movement amplitude using a pose estimation algorithm, and extracting graded differences of fundamental frequency changes using a speech algorithm. The temporal feature extraction is achieved by using an LSTM (Long Short-Term Memory) network to capture the trend and peak synchronization of indicator changes over a continuous 25-minute period. The attention fusion uses the Transformer attention mechanism to assign modal weights, classifying pain levels into mild, moderate, and severe pain. The weights of the multimodal structures are dynamically adjusted according to different pain levels, with the weight allocation for different pain levels satisfying the following: for mild pain, the weight of behavioral characteristics is greater than or equal to the weight of conventional physiological characteristics; for severe pain, the weight of conventional physiological characteristics is greater than or equal to 0.5; and the weight of behavioral characteristics for special populations is 0.

7. The pain discrimination method based on temporal and multimodal data according to claim 1, characterized in that, The full-level pain recognition logic in step S5 includes: Simultaneous verification of multimodal indicators to determine whether changes in core indicators conform to the unidirectional trend and amplitude range of the corresponding level of pain; Multi-class probability calculation, outputting the probability distribution of each level of NRS 0-10 through a combination model of LSTM, Transformer and Softmax; A classification threshold is set, and a secondary verification is triggered when the probability of the classification falls below the threshold. Anti-interference mechanism: When a single modality index is abnormal, the weight of that modality is automatically reduced.

8. A pain discrimination system based on temporal and multimodal modes, characterized in that, include: The multimodal data acquisition module integrates conventional physiological indicator acquisition equipment, behavioral feature acquisition equipment, and potential auxiliary indicator acquisition equipment to collect multimodal data. The data preprocessing module is used to perform outlier removal, individual normalization, and time alignment operations. The baseline calibration module is used to establish and dynamically calibrate individual baseline thresholds; The AI ​​feature fusion module is used to extract physiological features, behavioral features, and temporal features from time-series data blocks, and to fuse multimodal features through an attention mechanism. The full-level pain recognition module is used to perform NRS 0-10 full-level pain recognition through multimodal index synchronous verification, multi-class probability calculation and grading threshold judgment; The graded intervention output module is used to output clinical intervention strategies; The model update module is used to dynamically update model parameters based on clinical intervention feedback and data from newly added populations.

9. A computer device, characterized in that, include: A processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 7.

Images