Deep learning-based intelligent control method for analgesia pump

Abstract

The invention discloses an intelligent control method for an analgesia pump based on deep learning. The intelligent control method comprises the following steps: S1, acquiring multi-modal original signals of a postoperative patient and synchronously normalizing the multi-modal original signals; s2, removing artifacts and noise in the synchronous data, and performing interpolation compensation on missing data; s3, carrying out modal reliability weight distribution on the sensing sequence based on a Hemma optimization algorithm; s4, inputting the node set into a PainNet relation statistical network to generate a dynamic graph topological structure; s5, predicting the pain score of the patient in the future, and generating a low-dose high-frequency micro-pulse administration instruction; s6, according to the circadian rhythm of the patient, dynamically adjusting the edge weight and the administration interval of the PainNet network; and S7, when the pain score fluctuation exceeds the limit, triggering an emergency mode, quickly converging to the most conservative parameter combination, and outputting a safety locking instruction. According to the invention, the control accuracy and safety of postoperative analgesia are improved.

Description

Technical Field

[0001] This invention relates to the field of smart medical technology, and in particular to a deep learning-based intelligent control method for analgesic pumps. Background Technology

[0002] With the rapid integration of artificial intelligence (AI) technology into the medical field, AI-based intelligent pain management methods have been gradually applied in clinical practice, particularly in postoperative pain management in orthopedics. Current postoperative pain management techniques primarily rely on healthcare professionals' subjective assessment of the patient's pain level and their experience in determining the dosage and administration method of analgesics. Healthcare professionals typically determine the medication regimen based on the patient's subjective pain expression, facial expressions, and physiological signals, through visual observation and manual judgment. This method is relatively simple in principle, utilizing subjective experience and routine clinical indicators as the main basis for pain management, and has been widely used in medical practice.

[0003] In recent years, some technical solutions have attempted to use physiological signal monitoring and machine learning algorithms for quantitative pain assessment. However, these methods are often limited to the collection and analysis of single-modal or limited-modal data, such as using only heart rate or blood pressure signals, and often employ fixed model parameters or simple linear regression models for pain assessment, making it difficult to accurately reflect the individual differences and dynamic changes in postoperative pain. Furthermore, existing technologies lack effective mechanism design for the dynamic closed-loop connection between pain assessment and drug dosage control, typically employing a dosing pattern with long intervals and coarse dose changes, making it difficult to achieve refined and real-time dynamic analgesic control.

[0004] The main shortcomings of existing technologies in practical applications are as follows:

[0005] (1) The assessment of pain level relies too much on the patient’s complaints and the experience of the nursing staff, resulting in highly subjective and unstable assessment results;

[0006] (2) There is a lack of effective fusion analysis mechanism for multimodal data, and the rich information contained in multimodal data has not been fully utilized to accurately reflect the dynamic changes in the patient's pain status.

[0007] (3) The method of controlling the dosage of analgesics is rough and the dosing interval is fixed, which can easily cause dosage fluctuations or delayed pain management.

[0008] Therefore, how to provide a deep learning-based intelligent control method for analgesia pumps is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] One objective of this invention is to propose a deep learning-based intelligent control method for analgesia pumps, which achieves real-time and precise analgesia control through multimodal perception, PainNet relational statistics network and Hippo optimization algorithm. This invention has the advantages of objective and accurate pain assessment and precise and real-time dosage control.

[0010] The intelligent control method for analgesia pump based on deep learning according to an embodiment of the present invention includes the following steps:

[0011] S1. Collect the original multimodal signals of the postoperative patient, and perform clock synchronization and amplitude normalization processing on the original multimodal signals to obtain multimodal synchronization data;

[0012] S2. Remove image artifacts and signal noise from the multimodal synchronization data respectively, and perform interpolation compensation for missing frames and short-term drift to obtain the multimodal sensing sequence.

[0013] S3. Based on the Hippo optimization algorithm, in the first hour after surgery, the initial node weights of each modality data in the multimodal sensing sequence are assigned according to the dynamic reliability index to generate a weighted multimodal initial node set.

[0014] S4. Input the weighted multimodal initialization node set into the PainNet relation statistics network to generate a dynamic graph topology.

[0015] S5. Perform forward inference in the dynamic graph topology to predict the postoperative pain score of the patient at several future time steps, and generate low-dose, high-frequency micropulse dosing instructions based on the prediction results.

[0016] S6. Based on the patient's postoperative diurnal physiological rhythm detection results, dynamically adjust the edge weight update rate of the PainNet relational statistical network and the micropulse dosing interval of the micropulse dosing command.

[0017] S7. When the fluctuation range of the pain score exceeds the safety threshold, the emergency mode is triggered. The Hippo optimization algorithm quickly converges to the most conservative parameter combination and outputs a safety lock control command.

[0018] Optionally, the multimodal raw signals specifically include heart rate signals, blood pressure signals, patient facial muscle activity image sequences, patient voice signals, and patient medical record text data;

[0019] Optionally, S2 specifically includes:

[0020] S21. Generate time synchronization markers based on the acquisition time of multimodal synchronous data, and mark the start and end times of heart rate signals, blood pressure signals, patient facial muscle activity image sequences, and patient voice signals.

[0021] S22. Periodically insert dual-frequency calibration sampling segments into the heart rate and blood pressure signals for instantaneous calibration of the sensing channel and elimination of measurement drift. The dual-frequency calibration sampling segments include low-frequency sampling segments and high-frequency sampling segments. The low-frequency sampling segments have a sampling frequency of 1Hz and are used to calibrate the signal baseline drift. The high-frequency sampling segments have a sampling frequency of 100Hz and are used to suppress transient noise.

[0022] S23. For the patient's facial muscle activity image sequence, calculate the affine transformation between sequences based on facial key points and perform image registration to eliminate global artifacts caused by head movement.

[0023] S24. Simultaneously acquire infrared reflected light signals of the corresponding patient's facial muscle activity image sequence, construct a frame-by-frame illumination compensation mapping, the frame-by-frame illumination compensation mapping calculates the illumination correction factor based on the ratio of the infrared reflected light intensity of each frame to the average brightness of the corresponding RGB image frame, and applies the illumination correction factor frame by frame to the RGB image of the patient's facial muscle activity image sequence to balance the changes in ambient light between different frames.

[0024] S25. For abnormal intervals of heart rate and blood pressure signal fluctuations, extract the infrared reflected light signal of the patient's facial muscle activity image sequence at the corresponding time of the abnormal interval, and perform cross-validation to determine whether to remove or retain the heart rate and blood pressure signals at this time.

[0025] S26. For continuous missing frames in the patient's facial muscle activity image sequence due to occlusion or frame loss, use the facial skeleton motion path and color gradient information of adjacent frames to generate intermediate compensation frames and insert them into the missing frame positions of the original image sequence.

[0026] S27. The processed heart rate and blood pressure signals are re-synchronized and fused with the patient's facial muscle activity image sequence to generate a continuous multimodal sensing sequence.

[0027] Optionally, S3 specifically includes:

[0028] S31. The multimodal sensing sequences are divided into heart rate signal sequences, blood pressure signal sequences, patient facial muscle activity image sequences, patient speech signal sequences, and patient medical record text data sequences according to modal type.

[0029] S32. Using a fixed-length sliding window as a unit, extract local statistical features of the heart rate signal sequence and blood pressure signal sequence window by window. The local statistical features include the mean amplitude, variance, and signal-to-noise ratio of the heart rate and blood pressure sequences in each window, forming a physiological signal reliability index.

[0030] S33. Extract the spatial location information of facial feature key points in the patient's facial muscle activity image sequence frame by frame, calculate the motion amplitude based on the spatial displacement of facial feature key points between adjacent frames, and calculate the image clarity and contrast by combining the pixel gray-level gradient distribution of the facial region to form an image signal reliability index.

[0031] S34. Extract the signal energy, spectral entropy and fundamental frequency stability of each speech window from the patient's speech signal sequence in a window-by-window manner to generate speech signal reliability index;

[0032] S35. Perform sentence-by-sentence natural language word segmentation and semantic embedding calculation on the patient medical record text data sequence. Calculate the semantic coherence, word frequency stability and sentiment fluctuation of the text sentences based on word embedding vectors to generate text data reliability indicators.

[0033] S36. Define a reliability index weight function and perform iterative optimization using the Hippo optimization algorithm. During the optimization process, substitute the physiological signal reliability index, image signal reliability index, speech signal reliability index, and text data reliability index into the weight function. With the goal of maximizing comprehensive reliability, obtain the globally optimal combination of data weights for each modality within the initial predetermined time period after surgery.

[0034] S36. Define the reliability index weight function and iteratively optimize it using the Hippo optimization algorithm to maximize the overall reliability as the optimization objective, and obtain the globally optimal combination of data weights for each modality within the initial predetermined time period after surgery.

[0035] S37. Using the globally optimal combination of modal data weights, the heart rate signal sequence, blood pressure signal sequence, patient facial muscle activity image sequence, patient speech signal sequence, and patient medical record text data sequence are respectively processed by feature vector weighting to obtain a weighted multimodal initialization node set representing the patient's postoperative pain state.

[0036] Optionally, S4 specifically includes:

[0037] S41. Input the weighted multimodal initial node set into the PainNet relational statistics network, use the feature vectors in the node set as the initial state of each node in the PainNet relational statistics network, and construct the initial node feature space with the initial state of the nodes in the network.

[0038] S42. Perform high-dimensional topological mapping on the initial node feature space in the PainNet relation statistics network, and reconstruct the initial connection relationship between nodes based on the high-order neighborhood similarity between node feature vectors to form the initial high-dimensional topological structure of the network.

[0039] S43. Based on the initial high-dimensional topology, a similarity gradient field is constructed. The similarity gradient field is defined by the rate of change of the feature similarity between nodes over time. The trend of node connection status is judged based on the migration direction and migration speed of node pairs in the similarity gradient field.

[0040] S44. Calculate the temporal dynamic stability index of the similarity between nodes based on the similarity gradient field. The temporal dynamic stability index represents the consistency of the change in the similarity between node pairs over multiple consecutive time steps. When the temporal dynamic stability index exceeds a preset threshold, a connection between nodes is established. If the threshold is not exceeded, the existing connection between nodes is disconnected.

[0041] S45. Update the topology of the PainNet relational statistics network based on the dynamic connection relationship, and use the topology change to perform adaptive spatial compression mapping on the node feature space, so that the node feature vector is compressed from high-dimensional space to low-dimensional space.

[0042] S46. Based on the spatial compression result of node feature vectors, recalculate the distance matrix between nodes at the current time step, and update the connection relationship between nodes according to the compressed node distance matrix to obtain the dynamic graph topology.

[0043] Optionally, the PainNet relationship statistics network specifically includes an initial node feature mapping layer, a high-dimensional topology mapping layer, a similarity gradient field construction layer, a dynamic stability determination layer, an adaptive spatial compression mapping layer, and a distance matrix update layer:

[0044] The initial node feature mapping layer receives a weighted multimodal initial node set, performs linear mapping with the feature vectors in the node set as input, generates the initial state of each node in the PainNet relation statistics network, and constructs the initial node feature space.

[0045] The high-dimensional topology mapping layer calculates the high-order neighborhood similarity between node feature vectors based on the initial node feature space, and determines the initial connection relationship between nodes based on the high-order neighborhood similarity, thus forming the initial high-dimensional topology structure of the network.

[0046] The similarity gradient field construction layer receives the initial high-dimensional topology of the network, calculates the rate of change of the similarity of feature vectors between adjacent time steps, and constructs the similarity gradient field of node feature vectors.

[0047] The dynamic stability determination layer calculates the temporal dynamic stability index of the similarity between nodes based on the similarity gradient field. When the temporal dynamic stability index exceeds a preset threshold, a connection between nodes is established; when it does not exceed the preset threshold, the existing node connection is disconnected, thus obtaining a dynamic connection relationship.

[0048] The adaptive spatial compression mapping layer updates the topology of the PainNet relation statistics network according to the dynamic connection relationship and performs adaptive compression mapping of the node feature space, so that the node feature vector is compressed from the high-dimensional space to the low-dimensional space.

[0049] The distance matrix update layer recalculates the distance matrix between nodes at the current time step based on the compressed node feature vectors, and updates the connection relationships between network nodes according to the distance matrix to obtain the dynamic graph topology.

[0050] Optionally, S5 specifically includes:

[0051] S51. Extract the sequence of low-dimensional node feature vectors from the dynamic graph topology as input for forward inference;

[0052] S52. Define the prediction window length, and calculate the node state for multiple consecutive time steps in the future through forward inference. The node state is the node feature vector of the prediction time step.

[0053] S53. Input the node feature vector of the predicted time step into the pain score prediction function for mapping, and generate the predicted pain score sequence of the postoperative patient at the corresponding time step.

[0054] S54. Based on the changes in scores at adjacent time steps in the predicted pain score sequence, determine the time interval in the predicted sequence where the pain score reaches a preset threshold. The preset threshold is set through preoperative calibration.

[0055] S55. Within the time interval when the pain score in the predicted sequence reaches a preset threshold, generate multiple consecutive low-dose micro-pulse drug administration instructions. The micro-pulse drug administration instructions include the micro-pulse drug administration time and the corresponding drug dosage. The drug dosage is determined based on the proportional relationship between the pain score at the current time step and the preset analgesic drug dosage benchmark.

[0056] S56. The generated micro-pulse drug delivery command is encapsulated according to the analgesia pump drug delivery control interface protocol and transmitted to the analgesia pump to execute micro-pulse drug infusion.

[0057] Optionally, S6 specifically includes:

[0058] S61. Receive the continuous heart rate signal sequence and blood pressure signal sequence of the postoperative patient, calculate the time domain average and standard deviation of the heart rate signal sequence and blood pressure signal sequence with a fixed-length sliding window, and generate the postoperative patient's diurnal physiological rhythm characteristic sequence.

[0059] S62. Perform periodic analysis on the diurnal physiological rhythm characteristic sequence to determine the starting position, duration, and rhythm phase of the patient's diurnal physiological rhythm cycle;

[0060] S63. Based on the starting position, duration and phase of the patient's diurnal physiological rhythm cycle, the postoperative time segment is divided into a high-sensitivity rhythm segment and a low-sensitivity rhythm segment.

[0061] S64. Set the edge weight update rate adjustment function of PainNet relational statistics network, using the difference in physiological rhythm characteristics between high-sensitive rhythm segments and low-sensitive rhythm segments as parameters, increase the edge weight update rate in high-sensitive rhythm segments, and decrease the edge weight update rate in low-sensitive rhythm segments.

[0062] S65. Define a micropulse dosing interval adjustment function, based on the change in pain score of adjacent micropulse dosing instructions in the patient's diurnal physiological rhythm cycle, shorten the micropulse dosing interval in the high-sensitivity rhythm segment, and prolong the micropulse dosing interval in the low-sensitivity rhythm segment.

[0063] S66. Based on the calculation results of the edge weight update rate adjustment function and the micro-pulse dosing interval adjustment function, update the edge weight update rate of the PainNet relational statistics network and the micro-pulse dosing interval of the micro-pulse dosing command in real time.

[0064] Optionally, S7 specifically includes:

[0065] S71. Calculate the magnitude of the change in scores between adjacent time steps in the predicted pain score sequence using a fixed-length sliding window;

[0066] S72. Set a safe threshold for the change in pain score, and use a sliding window to compare the change in score with the safe threshold step by step to determine whether the change in pain score exceeds the safe range.

[0067] S73. When the pain score change exceeds the safety threshold, the emergency mode is triggered, and the emergency mode iterative search process of the Hippo optimization algorithm is started.

[0068] S74. In the emergency mode iterative search process of the Hippo optimization algorithm, the minimum change amplitude of pain score is used as the only fitness function to quickly converge and determine the most conservative parameter combination of the PainNet relational statistics network. The most conservative parameter combination is the network layer depth parameter, edge weight update rate parameter, and node feature vector mapping parameter that minimizes the fluctuation of the network output pain score prediction.

[0069] S75. Based on the most conservative parameter combination, calculate the feature vector of the PainNet relational statistics network node and generate the corresponding safety lock control command. The safety lock control command includes a command to stop the current micropulse drug delivery and a command to restore to the preset safe dose.

[0070] S76. The safety lock control command is encapsulated in real time and transmitted to the analgesic pump through the analgesic pump drug delivery control interface protocol to perform drug infusion control under the safety lock state.

[0071] The beneficial effects of this invention are:

[0072] (1) By collecting multimodal raw signals and performing synchronous normalization, Kalman filtering and interpolation compensation processing, this invention effectively eliminates facial image artifacts and physiological signal noise, improves the integrity and stability of pain perception data, and enhances the reliability of postoperative pain assessment data.

[0073] (2) This invention uses the Hippo optimization algorithm to dynamically allocate reliability weights to multimodal sensing sequences and reconstruct the dynamic graph topology in the PainNet relational statistics network in real time, thereby achieving accurate prediction of postoperative patient pain status, significantly improving the accuracy of pain score prediction, and showing better adaptability in dynamically changing postoperative pain management scenarios.

[0074] (3) In terms of analgesic pump drug delivery control, this invention predicts pain scores and generates low-dose, high-frequency micro-pulse drug delivery instructions. At the same time, it dynamically adjusts the micro-pulse drug delivery interval according to the patient's diurnal physiological rhythm. This effectively solves the shortcomings of the existing technology in terms of coarse drug delivery dosage and poor timeliness. It breaks through the limitations of the traditional fixed-dose and fixed-time interval drug delivery mode, realizes precise dynamic infusion of analgesic drugs, and effectively improves the safety and real-time performance of postoperative analgesia management.

[0075] (4) This invention uses real-time monitoring of pain score fluctuation amplitude and the rapid convergence mechanism of the emergency mode of the hippo optimization algorithm to quickly determine and output the most conservative combination of network parameters and safety lock control instructions, which significantly reduces the risk of abnormal fluctuations in the postoperative analgesia process and improves the safety and reliability of postoperative analgesia control. Attached Figure Description

[0076] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0077] Figure 1 This is a flowchart illustrating the overall process of the deep learning-based intelligent control method for analgesia pumps proposed in this invention. Detailed Implementation

[0078] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0079] refer to Figure 1A deep learning-based intelligent control method for analgesia pumps includes the following steps:

[0080] S1. Collect the original multimodal signals of the postoperative patient, and perform clock synchronization and amplitude normalization processing on the original multimodal signals to obtain multimodal synchronization data;

[0081] S2. Remove image artifacts and signal noise from the multimodal synchronization data respectively, and perform interpolation compensation for missing frames and short-term drift to obtain the multimodal sensing sequence.

[0082] S3. Based on the Hippo optimization algorithm, in the first hour after surgery, the initial node weights of each modality data in the multimodal sensing sequence are assigned according to the dynamic reliability index to generate a weighted multimodal initial node set.

[0083] S4. Input the weighted multimodal initialization node set into the PainNet relation statistics network to generate a dynamic graph topology.

[0084] S5. Perform forward inference in the dynamic graph topology to predict the postoperative pain score of the patient at several future time steps, and generate low-dose, high-frequency micropulse dosing instructions based on the prediction results.

[0085] S6. Based on the patient's postoperative diurnal physiological rhythm detection results, dynamically adjust the edge weight update rate of the PainNet relational statistical network and the micropulse dosing interval of the micropulse dosing command.

[0086] S7. When the fluctuation range of the pain score exceeds the safety threshold, the emergency mode is triggered. The Hippo optimization algorithm quickly converges to the most conservative parameter combination and outputs a safety lock control command.

[0087] By synchronously normalizing and data augmenting the original multimodal signals, the accuracy and stability of pain perception data were improved. The Hippo optimization algorithm was used to achieve dynamic reliability weighting of data from each modality, effectively improving the accuracy of pain state assessment. A dynamic graph topology was constructed using the PainNet relational statistics network, enabling precise prediction of pain scores and accurate control of micropulse drug administration. At the same time, the network parameters and drug administration intervals were dynamically adjusted according to the patient's diurnal physiological rhythm, and a safe mode fast convergence mechanism was set up, which significantly improved the real-time performance, safety, and accuracy of postoperative analgesia control.

[0088] In this embodiment, the multimodal raw signals specifically include heart rate signals, blood pressure signals, patient facial muscle activity image sequences, patient voice signals, and patient medical record text data;

[0089] By comprehensively collecting multimodal raw signals such as heart rate, blood pressure, facial muscle activity image sequences, voice signals, and medical record texts, the data foundation for pain assessment is significantly expanded, the comprehensiveness and objectivity of pain state perception are enhanced, and the technical problems of insufficient information, strong subjectivity, and insufficient accuracy in existing single-modal or multimodal information assessment methods are effectively solved.

[0090] In this embodiment, S2 specifically includes:

[0091] S21. Generate time synchronization markers based on the acquisition time of multimodal synchronous data, and mark the start and end times of heart rate signals, blood pressure signals, patient facial muscle activity image sequences, and patient voice signals.

[0092] S22. Periodically insert dual-frequency calibration sampling segments into the heart rate and blood pressure signals for instantaneous calibration of the sensing channel and elimination of measurement drift. The dual-frequency calibration sampling segments include low-frequency sampling segments and high-frequency sampling segments. The low-frequency sampling segments have a sampling frequency of 1Hz and are used to calibrate the signal baseline drift. The high-frequency sampling segments have a sampling frequency of 100Hz and are used to suppress transient noise.

[0093] S23. For the patient's facial muscle activity image sequence, calculate the affine transformation between sequences based on facial key points and perform image registration to eliminate global artifacts caused by head movement.

[0094] S24. Simultaneously acquire infrared reflected light signals of the corresponding patient's facial muscle activity image sequence, construct a frame-by-frame illumination compensation mapping, the frame-by-frame illumination compensation mapping calculates the illumination correction factor based on the ratio of the infrared reflected light intensity of each frame to the average brightness of the corresponding RGB image frame, and applies the illumination correction factor frame by frame to the RGB image of the patient's facial muscle activity image sequence to balance the changes in ambient light between different frames.

[0095] S25. For abnormal intervals of heart rate and blood pressure signal fluctuations, extract the infrared reflected light signal of the patient's facial muscle activity image sequence at the corresponding time of the abnormal interval, and perform cross-validation to determine whether to remove or retain the heart rate and blood pressure signals at this time.

[0096] S26. For continuous missing frames in the patient's facial muscle activity image sequence due to occlusion or frame loss, use the facial skeleton motion path and color gradient information of adjacent frames to generate intermediate compensation frames and insert them into the missing frame positions of the original image sequence.

[0097] S27. The processed heart rate and blood pressure signals are re-synchronized and fused with the patient's facial muscle activity image sequence to generate a continuous multimodal sensing sequence.

[0098] By employing high-precision synchronization labeling of multimodal synchronized data, dual-frequency calibration sampling of heart rate and blood pressure signals, global artifact elimination and frame-by-frame illumination compensation of facial image sequences, cross-validation of abnormal intervals, and precise compensation processing for missing data, the synchronization quality, integrity, and stability of multimodal data are effectively improved, thus providing a reliable data foundation for subsequent accurate and real-time pain assessment and analgesia control.

[0099] In this embodiment, S3 specifically includes:

[0100] S31. The multimodal sensing sequences are divided into heart rate signal sequences, blood pressure signal sequences, patient facial muscle activity image sequences, patient speech signal sequences, and patient medical record text data sequences according to modal type.

[0101] S32. Using a fixed-length sliding window as a unit, extract local statistical features of the heart rate signal sequence and blood pressure signal sequence window by window. The local statistical features include the mean amplitude, variance, and signal-to-noise ratio of the heart rate and blood pressure sequences in each window, forming a physiological signal reliability index.

[0102] S33. Extract the spatial location information of facial feature key points in the patient's facial muscle activity image sequence frame by frame, calculate the motion amplitude based on the spatial displacement of facial feature key points between adjacent frames, and calculate the image clarity and contrast by combining the pixel gray-level gradient distribution of the facial region to form an image signal reliability index.

[0103] S34. Extract the signal energy, spectral entropy and fundamental frequency stability of each speech window from the patient's speech signal sequence in a window-by-window manner to generate speech signal reliability index;

[0104] S35. Perform sentence-by-sentence natural language word segmentation and semantic embedding calculation on the patient medical record text data sequence. Calculate the semantic coherence, word frequency stability and sentiment fluctuation of the text sentences based on word embedding vectors to generate text data reliability indicators.

[0105] S36. Define a reliability index weight function and perform iterative optimization using the Hippo optimization algorithm. During the optimization process, substitute the physiological signal reliability index, image signal reliability index, speech signal reliability index, and text data reliability index into the weight function. With the goal of maximizing comprehensive reliability, obtain the globally optimal combination of data weights for each modality within the initial predetermined time period after surgery.

[0106] S36. Define a reliability index weight function and iteratively optimize it using the Hippo Optimization Algorithm, taking maximizing overall reliability as the optimization objective, and obtain the globally optimal combination of data weights for each modality within the initial predetermined time period after surgery:

[0107] The reliability index weighting function W is:

[0108]

[0109] in, μ represents the reliability function of the heart rate signal sequence. HR σ is the mean amplitude of the heart rate signal within the window. HR λ is the standard deviation of the amplitude of the windowed heart rate signal. HR This represents the amount of noise interference in the window's internal rate signal. μ represents the reliability function of the blood pressure signal sequence. BP σ is the mean amplitude of the blood pressure signal within the window. BP λ represents the standard deviation of the blood pressure signal amplitude within the window. BP This represents the amount of noise interference in the blood pressure signal within the window. η represents the reliability function of a sequence of facial muscle activity images. F For image sharpness, κ F For inter-frame keypoint motion stability, ξ F τ represents the average pixel gradient fluctuation. F To determine the degree of artifact interference, E represents the reliability function of the patient's speech signal sequence. V For short-time speech energy, H V For the spectral entropy, C V This is due to fundamental frequency period instability. S represents the reliability function of a sequence of patient medical record text data. T For semantic coherence, Q T For semantic similarity between sentences, D T Let α, β, γ, δ, and ∈ represent the modal weight coefficients for the degree of emotional fluctuation, with values ​​ranging from [0,1], and satisfying α+β+γ+δ+∈=1;

[0110] The formula comprehensively considers the reliability indicators of heart rate, blood pressure, facial images, speech, and text data, and uses the Hippo Optimization Algorithm to dynamically optimize the weights of multimodal data to obtain a data fusion effect that accurately reflects the patient's pain status. The formula quantifies the reliability of each modality of data by using the mean, standard deviation, signal-to-noise ratio, image sharpness, artifact interference, speech spectral features, and text semantic features of the signal within the window. It uses modality weight coefficients for linear combination and normalization to ensure that different modalities of data can play an appropriate role in pain status assessment, ultimately achieving a significant improvement in the objectivity, stability, and accuracy of postoperative patient pain status assessment.

[0111] S37. Using the globally optimal combination of modal data weights, the heart rate signal sequence, blood pressure signal sequence, patient facial muscle activity image sequence, patient speech signal sequence, and patient medical record text data sequence are respectively processed by feature vector weighting to obtain a weighted multimodal initialization node set representing the patient's postoperative pain state.

[0112] By extracting features from physiological signals, image signals, speech signals, and text data in multimodal sensing sequences and constructing refined reliability indicators, the Hippo Optimization Algorithm is used to dynamically determine the weight combination of each modality, forming a weighted multimodal node set. This effectively solves the problem of insufficient evaluation accuracy caused by statically fixed or manually set weights in the multimodal data fusion process of existing technologies, and realizes more accurate and real-time dynamic perception of the postoperative patient's pain status.

[0113] In this embodiment, S4 specifically includes:

[0114] S41. Input the weighted multimodal initial node set into the PainNet relational statistics network, use the feature vectors in the node set as the initial state of each node in the PainNet relational statistics network, and construct the initial node feature space with the initial state of the nodes in the network.

[0115] S42. Perform high-dimensional topological mapping on the initial node feature space in the PainNet relation statistics network, and reconstruct the initial connection relationship between nodes based on the high-order neighborhood similarity between node feature vectors to form the initial high-dimensional topological structure of the network.

[0116] S43. Based on the initial high-dimensional topology, a similarity gradient field is constructed. The similarity gradient field is defined by the rate of change of the feature similarity between nodes over time. The trend of node connection status is judged based on the migration direction and migration speed of node pairs in the similarity gradient field.

[0117] S44. Calculate the temporal dynamic stability index of the similarity between nodes based on the similarity gradient field. The temporal dynamic stability index represents the consistency of the change in the similarity between node pairs over multiple consecutive time steps. When the temporal dynamic stability index exceeds a preset threshold, a connection between nodes is established. If the threshold is not exceeded, the existing connection between nodes is disconnected.

[0118] S45. Update the topology of the PainNet relational statistics network based on the dynamic connection relationship, and use the topology change to perform adaptive spatial compression mapping on the node feature space, so that the node feature vector is compressed from high-dimensional space to low-dimensional space.

[0119] S46. Based on the spatial compression result of node feature vectors, recalculate the distance matrix between nodes at the current time step, and update the connection relationship between nodes according to the compressed node distance matrix to obtain the dynamic graph topology.

[0120] By employing high-dimensional topological mapping of the node feature space within the PainNet relational statistics network, dynamic construction of the similarity gradient field, and adaptive updating of node connection states, dynamic optimization of the network topology is achieved. Simultaneously, through adaptive spatial compression of node feature vectors and real-time updating of the distance matrix, the real-time performance, accuracy, and sensitivity to dynamic changes in postoperative pain state prediction are effectively improved, addressing the problem of insufficient flexibility in pain assessment caused by static network topology and fixed node connection relationships in existing technologies.

[0121] In this embodiment, the PainNet relationship statistics network specifically includes an initial node feature mapping layer, a high-dimensional topology mapping layer, a similarity gradient field construction layer, a dynamic stability determination layer, an adaptive spatial compression mapping layer, and a distance matrix update layer:

[0122] The initial node feature mapping layer receives a weighted multimodal initial node set, performs linear mapping with the feature vectors in the node set as input, generates the initial state of each node in the PainNet relation statistics network, and constructs the initial node feature space.

[0123] The high-dimensional topology mapping layer calculates the high-order neighborhood similarity between node feature vectors based on the initial node feature space, and determines the initial connection relationship between nodes based on the high-order neighborhood similarity, thus forming the initial high-dimensional topology structure of the network.

[0124] The similarity gradient field construction layer receives the initial high-dimensional topology of the network, calculates the rate of change of the similarity of feature vectors between adjacent time steps, and constructs the similarity gradient field of node feature vectors.

[0125] The dynamic stability determination layer calculates the temporal dynamic stability index of the similarity between nodes based on the similarity gradient field. When the temporal dynamic stability index exceeds a preset threshold, a connection between nodes is established; when it does not exceed the preset threshold, the existing node connection is disconnected, thus obtaining a dynamic connection relationship.

[0126] The adaptive spatial compression mapping layer updates the topology of the PainNet relation statistics network according to the dynamic connection relationship and performs adaptive compression mapping of the node feature space, so that the node feature vector is compressed from the high-dimensional space to the low-dimensional space.

[0127] The distance matrix update layer recalculates the distance matrix between nodes at the current time step based on the compressed node feature vectors, and updates the connection relationships between network nodes according to the distance matrix to obtain the dynamic graph topology.

[0128] Through the refined design of the internal structure of the PainNet relational statistics network, including node feature mapping, high-dimensional topology mapping, similarity gradient field construction, dynamic stability determination, adaptive spatial compression mapping, and distance matrix update, dynamic compression of node feature space and real-time adjustment of topology structure are effectively realized. This improves the accuracy and timeliness of the network's prediction of postoperative patient pain status and overcomes the technical problem of insufficient pain status prediction accuracy caused by the lack of dynamic adaptability of network topology structure in existing technologies.

[0129] In this embodiment, S5 specifically includes:

[0130] S51. Extract the sequence of low-dimensional node feature vectors from the dynamic graph topology as input for forward inference;

[0131] S52. Define the prediction window length, and calculate the node state for multiple consecutive time steps in the future through forward inference. The node state is the node feature vector of the prediction time step.

[0132] S53. Input the node feature vector of the predicted time step into the pain score prediction function for mapping, and generate the predicted pain score sequence of the postoperative patient at the corresponding time step.

[0133] The pain score prediction function is:

[0134]

[0135] Among them, P t Let x be the predicted pain score of the patient at the t-th prediction time step. i,t Let W be the feature vector of node i within the t-th prediction time step. p Let b be the projection weight matrix of the node feature vectors. p w is the projection bias vector of the node feature vectors. i,t Let be the dynamic feature weight coefficient of node i in the t-th prediction time step, derived from the node importance coefficients in the dynamic graph topology updated in step S4, N represent the number of nodes participating in the prediction, and σ(·) be the Logistic function. Used to normalize the predicted score to the standard score range, tanh(·) is the hyperbolic tangent function, used to enhance the model's sensitivity to small changes in node feature vectors, and M is the maximum quantization level constant of the pain score, used to map the predicted score to the standard quantization range of the pain score.

[0136] The formula combines the feature vectors of each node in the dynamic graph topology with the corresponding dynamic weight coefficients. Through projection using a nonlinear tanh function and mapping with the weight matrix, it can capture subtle changes in node features and amplify feature differences, thereby characterizing the patient's pain state. The formula uses the Logistic function to normalize the prediction results, ensuring that the output value falls within the standard pain score range. It also uses the maximum quantization level constant for scaling, ultimately achieving accurate prediction of postoperative patient pain scores and improving the real-time performance and accuracy of analgesia control.

[0137] S54. Based on the changes in scores at adjacent time steps in the predicted pain score sequence, determine the time interval in the predicted sequence where the pain score reaches a preset threshold. The preset threshold is set through preoperative calibration.

[0138] S55. Within the time interval when the pain score in the predicted sequence reaches a preset threshold, generate multiple consecutive low-dose micro-pulse drug administration instructions. The micro-pulse drug administration instructions include the micro-pulse drug administration time and the corresponding drug dosage. The drug dosage is determined based on the proportional relationship between the pain score at the current time step and the preset analgesic drug dosage benchmark.

[0139] S56. The generated micro-pulse drug delivery command is encapsulated according to the analgesia pump drug delivery control interface protocol and transmitted to the analgesia pump to execute micro-pulse drug infusion.

[0140] By designing a pain score prediction function based on a dynamic graph topology, a prospective and accurate prediction of postoperative patient pain status can be achieved. Based on this, low-dose, high-frequency micro-pulse drug administration instructions can be generated, which effectively improves the real-time, accuracy and continuity of analgesic drug infusion. This overcomes the technical problems of insufficient pain score prediction accuracy and poor analgesic drug dosage control accuracy and timeliness in the existing technology.

[0141] In this embodiment, S6 specifically includes:

[0142] S61. Receive the continuous heart rate signal sequence and blood pressure signal sequence of the postoperative patient, calculate the time domain average and standard deviation of the heart rate signal sequence and blood pressure signal sequence with a fixed-length sliding window, and generate the postoperative patient's diurnal physiological rhythm characteristic sequence.

[0143] S62. Perform periodic analysis on the diurnal physiological rhythm characteristic sequence to determine the starting position, duration, and rhythm phase of the patient's diurnal physiological rhythm cycle;

[0144] S63. Based on the starting position, duration and phase of the patient's diurnal physiological rhythm cycle, the postoperative time segment is divided into a high-sensitivity rhythm segment and a low-sensitivity rhythm segment.

[0145] S64. Define the edge weight update rate adjustment function of the PainNet relational statistics network, using the difference in physiological rhythm characteristics between high-sensitivity rhythm segments and low-sensitivity rhythm segments as parameters. Increase the edge weight update rate in high-sensitivity rhythm segments and decrease the edge weight update rate in low-sensitivity rhythm segments:

[0146]

[0147] Where U(t) represents the edge weight update rate at the current time step, U base μ represents the preset edge weight baseline update rate. h (t) represents the average physiological rhythm characteristic value of the highly sensitive rhythm segment at the current time step, μ l (t) represents the average physiological rhythm characteristic value of the low-sensitivity rhythm segment;

[0148] The formula uses the differences in postoperative patients' diurnal physiological rhythms as the core parameter. By calculating the ratio of feature differences between high-sensitivity rhythm segments and low-sensitivity rhythm segments in real time, it dynamically adjusts the edge weight update rate of the PainNet relational statistical network. When the patient is in a high-sensitivity rhythm segment, the network update rate is increased, making the network more sensitive and responsive to changes in pain status. In low-sensitivity rhythm segments, the update rate is reduced to decrease unnecessary network update burden, ensuring the sensitivity and stability of analgesia control and effectively improving the network's adaptive ability.

[0149] S65. Define a micropulse dosing interval adjustment function, based on the change in pain score between adjacent micropulse dosing commands during the patient's diurnal physiological rhythm cycle, shortening the micropulse dosing interval in high-sensitivity rhythm segments and lengthening the micropulse dosing interval in low-sensitivity rhythm segments:

[0150]

[0151] Where T(t) represents the dosing interval of the current time step micropulse dosing command, T base To establish a preset micropulse dosing interval reference, ΔP h (t) represents the average change in predicted pain scores within the current highly sensitive rhythmic segment, ΔP l (t) represents the average change in predicted pain scores within the low-sensitivity rhythm segment;

[0152] The formula dynamically adjusts the dosing interval of the micropulse dosing command based on the difference in the average change of predicted pain scores between high-sensitivity and low-sensitivity segments within the postoperative circadian rhythm cycle. In the high-sensitivity rhythm segment, when the change in pain score is large, the micropulse dosing interval is adaptively shortened to improve the real-time responsiveness of analgesic drug administration. In the low-sensitivity rhythm segment, when the change in pain score is small, the dosing interval is correspondingly extended to reduce the frequency of administration and optimize drug use efficiency.

[0153] S66. Based on the calculation results of the edge weight update rate adjustment function and the micro-pulse dosing interval adjustment function, update the edge weight update rate of the PainNet relational statistics network and the micro-pulse dosing interval of the micro-pulse dosing command in real time.

[0154] By monitoring the patient's diurnal physiological rhythm characteristics in real time and dynamically adjusting the update rate of the PainNet relational statistical network edge weights and the micropulse dosing interval, more refined and personalized analgesia control is achieved for different physiological rhythm states after surgery. This effectively overcomes the problem in existing technologies where analgesia protocols cannot be adjusted in real time according to changes in physiological rhythm, resulting in unstable analgesia control effects, and significantly improves the accuracy and adaptability of analgesia treatment.

[0155] In this embodiment, S7 specifically includes:

[0156] S71. Calculate the magnitude of the change in scores between adjacent time steps in the predicted pain score sequence using a fixed-length sliding window;

[0157] S72. Set a safe threshold for the change in pain score, and use a sliding window to compare the change in score with the safe threshold step by step to determine whether the change in pain score exceeds the safe range.

[0158] S73. When the pain score change exceeds the safety threshold, the emergency mode is triggered, and the emergency mode iterative search process of the Hippo optimization algorithm is started.

[0159] S74. In the emergency mode iterative search process of the Hippo optimization algorithm, the minimum change amplitude of pain score is used as the only fitness function to quickly converge and determine the most conservative parameter combination of the PainNet relational statistics network. The most conservative parameter combination is the network layer depth parameter, edge weight update rate parameter, and node feature vector mapping parameter that minimizes the fluctuation of the network output pain score prediction.

[0160] S75. Based on the most conservative parameter combination, calculate the feature vector of the PainNet relational statistics network node and generate the corresponding safety lock control command. The safety lock control command includes a command to stop the current micropulse drug delivery and a command to restore to the preset safe dose.

[0161] S76. The safety lock control command is encapsulated in real time and transmitted to the analgesic pump through the analgesic pump drug delivery control interface protocol to perform drug infusion control under the safety lock state.

[0162] By monitoring the changes in pain scores in real time and setting a safety threshold, the Hippo Optimization Algorithm is used to quickly converge in emergency mode to determine the most conservative parameter combination of the PainNet relational statistical network and output safety lock control commands in a timely manner. This effectively solves the problem of lack of a rapid response mechanism in the face of abnormal fluctuations in the postoperative patient's pain status in existing technologies, and significantly improves the safety and emergency response efficiency in the analgesia control process.

[0163] Example 1:

[0164] To verify the feasibility of this invention in practice, it was applied to postoperative pain management in the orthopedic department of a large general hospital. Specifically, multiple patients underwent experimental testing of intelligent control of postoperative analgesia pumps for pain assessment and precise dynamic analgesic drug dosing. In traditional orthopedic postoperative pain management methods, the assessment of postoperative pain levels typically relies on patient complaints and manual observation of physiological responses by medical staff, resulting in significant subjectivity and assessment instability. Furthermore, traditional analgesic drug dosing regimens generally employ fixed doses and fixed time intervals, making it difficult to accurately capture individual patient differences and real-time changes in pain status, easily leading to inaccurate drug dosage, poor analgesic effect, or the risk of drug overdose. Therefore, how to accurately monitor postoperative pain levels and dynamically and precisely adjust analgesic drug dosage in real time is a pressing technical challenge that needs to be addressed.

[0165] In practice, heart rate monitoring equipment, continuous blood pressure monitoring equipment, high-definition video equipment, voice acquisition equipment, and electronic medical record systems are used to collect postoperative vital signs data, facial muscle activity images, patient complaints, and medical record text information in real time. After nonlinear clock synchronization and amplitude normalization processing, these collected raw multimodal data are transformed into unified multimodal synchronized data, eliminating time differences and amplitude fluctuations between different modalities and ensuring data consistency and integrity.

[0166] Subsequently, Kalman filtering and anomaly detection algorithms were used to further denoise and remove artifacts from the multimodal synchronization data, and interpolation compensation was performed for missing image frames and signal segments to obtain a complete and high-quality multimodal sensing sequence. Then, the Hippo Optimization Algorithm was used to perform dynamic reliability assessment on the data of each modality of the sensing sequence in the first hour after surgery, automatically determining the optimal weight allocation strategy for each modality, thereby generating a reliability-optimized multimodal initialization node set.

[0167] Next, the optimized node set is input into the PainNet relational statistics network. A stable initial dynamic graph structure is formed through initial node feature mapping, high-dimensional topology mapping, and similarity gradient field construction. Simultaneously, through temporal dynamic stability analysis in the similarity gradient field, the real-time connection state of the dynamic graph topology is automatically determined. An adaptive spatial compression mapping method is used to map high-dimensional node features to a low-dimensional space, thereby updating the dynamic graph structure. After multiple optimizations, the prediction error of the PainNet network for postoperative pain scores gradually decreases.

[0168] To verify the actual effect, a pain score of 0-10 was used as the quantification range. With a prediction window of 30 minutes, the feature vectors of the network output nodes were substituted into the pain score prediction function to predict the trend of the patient's pain score changes over multiple consecutive time steps, and to generate low-dose, high-frequency micro-pulse drug administration instructions. When the predicted score reaches a preset threshold, the corresponding drug administration operation is automatically initiated. In addition, the system automatically detects the patient's diurnal rhythm variation characteristics and dynamically adjusts the edge weight update rate of the PainNet network and the interval of the micro-pulse drug administration instructions according to the sensitivity of different rhythm periods, ensuring that the drug administration regimen is precisely synchronized with the patient's physiological rhythm, avoiding the problems of drug efficacy lag or dose fluctuation in traditional drug administration modes.

[0169] To ensure the safety of pain control, the system also sets a safety threshold for the range of pain score changes. When the system detects that the pain score changes exceed the safety threshold, it immediately triggers an emergency mode, quickly searches and determines the most conservative PainNet network parameter combination through the Hippo optimization algorithm, and generates a safety lockout control command to interrupt the current drug infusion and restore it to the preset safe dose.

[0170] Experimental data showed that multiple patients were tested in this embodiment, and the measured and predicted pain scores at different postoperative time points were recorded. The control effect of traditional methods was compared. The table below lists the comparison data of measured and predicted postoperative pain scores for some patients.

[0171] Table 1 Comparison of measured and predicted postoperative pain scores of patients.

[0172] Patient number Testing time Pain score measured value Pain score prediction Patient 001 1 hour after surgery 7.8 7.6 Patient 001 3 hours after surgery 5.5 5.7 Patient 001 6 hours after surgery 3.2 3.4 Patient 002 1 hour after surgery 8.1 7.9 Patient 002 3 hours after surgery 5.8 5.5 Patient 002 6 hours after surgery 3.5 3.3 Patient 003 1 hour after surgery 7.5 7.4 Patient 003 3 hours after surgery 5.2 5.3 Patient 003 6 hours after surgery 3.0 3.2

[0173] As clearly shown in Table 1, the pain score prediction values ​​of the method of this invention have a small difference from the measured values, with the overall prediction error remaining within ±0.3 points, significantly better than the ±1.5 point error range commonly achieved by traditional subjective assessment methods. Simultaneously, this method enables real-time and precise control of analgesic drugs, effectively avoiding delayed drug efficacy or overdose. Regarding safety, in this embodiment, two patients experienced rapid fluctuations in pain scores exceeding the safety threshold in the early postoperative period. The system immediately activated emergency mode, quickly converged to the most conservative parameter combination, and promptly switched to a safe dosage control scheme, rapidly stabilizing the patients' pain fluctuations and avoiding potential risks, demonstrating the system's outstanding advantages in analgesic safety management.

[0174] In summary, this embodiment fully demonstrates that the deep learning-based intelligent control method for analgesia pumps proposed in this invention has achieved significant practical effects and technical advantages in terms of the objectivity of postoperative pain assessment, precise control of analgesic drug dosage, and safety management, and has high clinical application value and promotion potential.

[0175] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent control method for analgesia pumps, characterized in that, Includes the following steps: S1. Collect the original multimodal signals of the postoperative patient, and perform clock synchronization and amplitude normalization processing on the original multimodal signals to obtain multimodal synchronization data; S2. Remove image artifacts and signal noise from the multimodal synchronization data respectively, and perform interpolation compensation for missing frames and short-term drift to obtain the multimodal sensing sequence. S3. Based on the Hippo optimization algorithm, in the first hour after surgery, the initial node weights of each modality data in the multimodal sensing sequence are assigned according to the dynamic reliability index to generate a weighted multimodal initial node set. S4. Input the weighted multimodal initialization node set into the PainNet relation statistics network to generate a dynamic graph topology. S5. Perform forward inference in the dynamic graph topology to predict the postoperative pain score of the patient at several future time steps, and generate low-dose, high-frequency micropulse dosing instructions based on the prediction results. S6. Based on the patient's postoperative diurnal physiological rhythm detection results, dynamically adjust the edge weight update rate of the PainNet relational statistical network and the micropulse dosing interval of the micropulse dosing command. S7. When the fluctuation range of the pain score exceeds the safety threshold, the emergency mode is triggered. The Hippo optimization algorithm quickly converges to the most conservative parameter combination and outputs a safety lock control command.

2. The intelligent control method for analgesia pump based on deep learning according to claim 1, characterized in that, The multimodal raw signals specifically include heart rate signals, blood pressure signals, patient facial muscle activity image sequences, patient voice signals, and patient medical record text data.

3. The intelligent control method for analgesia pump based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Generate time synchronization markers based on the acquisition time of multimodal synchronous data, and mark the start and end times of heart rate signals, blood pressure signals, patient facial muscle activity image sequences, and patient voice signals. S22. Periodically insert dual-frequency calibration sampling segments into the heart rate and blood pressure signals for instantaneous calibration of the sensing channel and elimination of measurement drift. The dual-frequency calibration sampling segments include low-frequency sampling segments and high-frequency sampling segments. The low-frequency sampling segments have a sampling frequency of 1Hz and are used to calibrate the signal baseline drift. The high-frequency sampling segments have a sampling frequency of 100Hz and are used to suppress transient noise. S23. For the patient's facial muscle activity image sequence, calculate the affine transformation between sequences based on facial key points and perform image registration to eliminate global artifacts caused by head movement. S24. Simultaneously acquire infrared reflected light signals of the corresponding patient's facial muscle activity image sequence, construct a frame-by-frame illumination compensation mapping, the frame-by-frame illumination compensation mapping calculates the illumination correction factor based on the ratio of the infrared reflected light intensity of each frame to the average brightness of the corresponding RGB image frame, and applies the illumination correction factor frame by frame to the RGB image of the patient's facial muscle activity image sequence to balance the changes in ambient light between different frames. S25. For abnormal intervals of heart rate and blood pressure signal fluctuations, extract the infrared reflected light signal of the patient's facial muscle activity image sequence at the corresponding time of the abnormal interval, and perform cross-validation to determine whether to remove or retain the heart rate and blood pressure signals at this time. S26. For continuous missing frames in the patient's facial muscle activity image sequence due to occlusion or frame loss, use the facial skeleton motion path and color gradient information of adjacent frames to generate intermediate compensation frames and insert them into the missing frame positions of the original image sequence. S27. The processed heart rate and blood pressure signals are re-synchronized and fused with the patient's facial muscle activity image sequence to generate a continuous multimodal sensing sequence.

4. The intelligent control method for analgesia pump based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. The multimodal sensing sequences are divided into heart rate signal sequences, blood pressure signal sequences, patient facial muscle activity image sequences, patient speech signal sequences, and patient medical record text data sequences according to modal type. S32. Using a fixed-length sliding window as a unit, extract local statistical features of the heart rate signal sequence and blood pressure signal sequence window by window. The local statistical features include the mean amplitude, variance, and signal-to-noise ratio of the heart rate and blood pressure sequences in each window, forming a physiological signal reliability index. S33. Extract the spatial location information of facial feature key points in the patient's facial muscle activity image sequence frame by frame, calculate the motion amplitude based on the spatial displacement of facial feature key points between adjacent frames, and calculate the image clarity and contrast by combining the pixel gray-level gradient distribution of the facial region to form an image signal reliability index. S34. Extract the signal energy, spectral entropy and fundamental frequency stability of each speech window from the patient's speech signal sequence in a window-by-window manner to generate speech signal reliability index; S35. Perform sentence-by-sentence natural language word segmentation and semantic embedding calculation on the patient medical record text data sequence. Calculate the semantic coherence, word frequency stability and sentiment fluctuation of the text sentences based on word embedding vectors to generate text data reliability indicators. S36. Define a reliability index weight function and perform iterative optimization using the Hippo optimization algorithm. During the optimization process, substitute the physiological signal reliability index, image signal reliability index, speech signal reliability index, and text data reliability index into the weight function. With the goal of maximizing comprehensive reliability, obtain the globally optimal combination of data weights for each modality within the initial predetermined time period after surgery. S36. Define the reliability index weight function and iteratively optimize it using the Hippo optimization algorithm to maximize the overall reliability as the optimization objective, and obtain the globally optimal combination of data weights for each modality within the initial predetermined time period after surgery. S37. Using the globally optimal combination of modal data weights, the heart rate signal sequence, blood pressure signal sequence, patient facial muscle activity image sequence, patient speech signal sequence, and patient medical record text data sequence are respectively processed by feature vector weighting to obtain a weighted multimodal initialization node set representing the patient's postoperative pain state.

5. The intelligent control method for analgesia pump based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Input the weighted multimodal initial node set into the PainNet relational statistics network, use the feature vectors in the node set as the initial state of each node in the PainNet relational statistics network, and construct the initial node feature space with the initial state of the nodes in the network. S42. Perform high-dimensional topological mapping on the initial node feature space in the PainNet relation statistics network, and reconstruct the initial connection relationship between nodes based on the high-order neighborhood similarity between node feature vectors to form the initial high-dimensional topological structure of the network. S43. Based on the initial high-dimensional topology, a similarity gradient field is constructed. The similarity gradient field is defined by the rate of change of the feature similarity between nodes over time. The trend of node connection status is judged based on the migration direction and migration speed of node pairs in the similarity gradient field. S44. Calculate the temporal dynamic stability index of the similarity between nodes based on the similarity gradient field. The temporal dynamic stability index represents the consistency of the change in the similarity between node pairs over multiple consecutive time steps. When the temporal dynamic stability index exceeds a preset threshold, a connection between nodes is established. If the threshold is not exceeded, the existing connection between nodes is disconnected. S45. Update the topology of the PainNet relational statistics network based on the dynamic connection relationship, and use the topology change to perform adaptive spatial compression mapping on the node feature space, so that the node feature vector is compressed from high-dimensional space to low-dimensional space. S46. Based on the spatial compression result of node feature vectors, recalculate the distance matrix between nodes at the current time step, and update the connection relationship between nodes according to the compressed node distance matrix to obtain the dynamic graph topology.

6. The intelligent control method for analgesia pump based on deep learning according to claim 5, characterized in that, The PainNet relation statistics network specifically includes an initial node feature mapping layer, a high-dimensional topology mapping layer, a similarity gradient field construction layer, a dynamic stability determination layer, an adaptive spatial compression mapping layer, and a distance matrix update layer. The initial node feature mapping layer receives a weighted multimodal initial node set, performs linear mapping with the feature vectors in the node set as input, generates the initial state of each node in the PainNet relation statistics network, and constructs the initial node feature space. The high-dimensional topology mapping layer calculates the high-order neighborhood similarity between node feature vectors based on the initial node feature space, and determines the initial connection relationship between nodes based on the high-order neighborhood similarity, thus forming the initial high-dimensional topology structure of the network. The similarity gradient field construction layer receives the initial high-dimensional topology of the network, calculates the rate of change of the similarity of feature vectors between adjacent time steps, and constructs the similarity gradient field of node feature vectors. The dynamic stability determination layer calculates the temporal dynamic stability index of the similarity between nodes based on the similarity gradient field. When the temporal dynamic stability index exceeds a preset threshold, a connection between nodes is established; when it does not exceed the preset threshold, the existing node connection is disconnected, thus obtaining a dynamic connection relationship. The adaptive spatial compression mapping layer updates the topology of the PainNet relation statistics network according to the dynamic connection relationship and performs adaptive compression mapping of the node feature space, so that the node feature vector is compressed from the high-dimensional space to the low-dimensional space. The distance matrix update layer recalculates the distance matrix between nodes at the current time step based on the compressed node feature vectors, and updates the connection relationships between network nodes according to the distance matrix to obtain the dynamic graph topology.

7. The intelligent control method for analgesia pump based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Extract the sequence of low-dimensional node feature vectors from the dynamic graph topology as input for forward inference; S52. Define the prediction window length, and calculate the node state for multiple consecutive time steps in the future through forward inference. The node state is the node feature vector of the prediction time step. S53. Input the node feature vector of the predicted time step into the pain score prediction function for mapping, and generate the predicted pain score sequence of the postoperative patient at the corresponding time step. S54. Based on the changes in scores at adjacent time steps in the predicted pain score sequence, determine the time interval in the predicted sequence where the pain score reaches a preset threshold. The preset threshold is set through preoperative calibration. S55. Within the time interval when the pain score in the predicted sequence reaches a preset threshold, generate multiple consecutive low-dose micro-pulse drug administration instructions. The micro-pulse drug administration instructions include the micro-pulse drug administration time and the corresponding drug dosage. The drug dosage is determined based on the proportional relationship between the pain score at the current time step and the preset analgesic drug dosage benchmark. S56. The generated micro-pulse drug delivery command is encapsulated according to the analgesia pump drug delivery control interface protocol and transmitted to the analgesia pump to execute micro-pulse drug infusion.

8. The intelligent control method for analgesia pump based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Receive the continuous heart rate signal sequence and blood pressure signal sequence of the postoperative patient, calculate the time domain average and standard deviation of the heart rate signal sequence and blood pressure signal sequence with a fixed-length sliding window, and generate the postoperative patient's diurnal physiological rhythm characteristic sequence. S62. Perform periodic analysis on the diurnal physiological rhythm characteristic sequence to determine the starting position, duration, and rhythm phase of the patient's diurnal physiological rhythm cycle; S63. Based on the starting position, duration and phase of the patient's diurnal physiological rhythm cycle, the postoperative time segment is divided into a high-sensitivity rhythm segment and a low-sensitivity rhythm segment. S64. Set the edge weight update rate adjustment function of PainNet relational statistics network, using the difference in physiological rhythm characteristics between high-sensitive rhythm segments and low-sensitive rhythm segments as parameters, increase the edge weight update rate in high-sensitive rhythm segments, and decrease the edge weight update rate in low-sensitive rhythm segments. S65. Define a micropulse dosing interval adjustment function, based on the change in pain score of adjacent micropulse dosing instructions in the patient's diurnal physiological rhythm cycle, shorten the micropulse dosing interval in the high-sensitivity rhythm segment, and prolong the micropulse dosing interval in the low-sensitivity rhythm segment. S66. Based on the calculation results of the edge weight update rate adjustment function and the micro-pulse dosing interval adjustment function, update the edge weight update rate of the PainNet relational statistics network and the micro-pulse dosing interval of the micro-pulse dosing command in real time.

9. The intelligent control method for analgesia pump based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Calculate the magnitude of the change in scores between adjacent time steps in the predicted pain score sequence using a fixed-length sliding window; S72. Set a safe threshold for the change in pain score, and use a sliding window to compare the change in score with the safe threshold step by step to determine whether the change in pain score exceeds the safe range. S73. When the pain score change exceeds the safety threshold, the emergency mode is triggered, and the emergency mode iterative search process of the Hippo optimization algorithm is started. S74. In the emergency mode iterative search process of the Hippo optimization algorithm, the minimum change amplitude of pain score is used as the only fitness function to quickly converge and determine the most conservative parameter combination of the PainNet relational statistics network. The most conservative parameter combination is the network layer depth parameter, edge weight update rate parameter, and node feature vector mapping parameter that minimizes the fluctuation of the network output pain score prediction. S75. Based on the most conservative parameter combination, calculate the feature vector of the PainNet relational statistics network node and generate the corresponding safety lock control command. The safety lock control command includes a command to stop the current micropulse drug delivery and a command to restore to the preset safe dose. S76. The safety lock control command is encapsulated in real time and transmitted to the analgesic pump through the analgesic pump drug delivery control interface protocol to perform drug infusion control under the safety lock state.

Images