Accurate temperature control negative pressure adsorption acupoint transdermal introduction method and device

Abstract

The invention provides a precise temperature control negative pressure adsorption acupoint transdermal introduction method and device, and relates to the technical field of transdermal administration, the method comprises the following steps: starting a negative pressure device, an infrared heating device and an electric heating device; first temperature data, second temperature data and pressure data are detected, a first state vector, a second state vector and a third state vector are formed, and then first predicted temperature data, second predicted temperature data and predicted pressure data at multiple moments in the future are predicted. Therefore, whether the on-off states of the negative pressure equipment, the infrared heating device and the electric heating device need to be changed or not is judged. According to the invention, dispersion and absorption of drugs can be promoted through the electric heating device, local skin blood circulation can be promoted through the infrared heating device, the absorption efficiency of the drugs is improved, and the skin temperature, the drug temperature and the pressure can be accurately controlled; and the probability that the transdermal rate of the medicine is reduced due to skin scald, medicine failure and inaccurate pressure control is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of transdermal drug delivery, and in particular to a method and device for precise temperature-controlled negative pressure adsorption through-acupoint transdermal drug delivery. Background Art

[0002] Negative pressure transdermal drug delivery has developed rapidly due to its minimal side effects on the liver and gastrointestinal tract. However, many factors need to be considered during the drug delivery process. For example, when heating the space within the negative pressure suction cup to promote drug absorption, it is difficult to accurately control the temperature and pressure within the space. This may cause the drug to become ineffective due to high temperature, the patient's skin to be burned due to high temperature, and inaccurate pressure control to cause a decrease in the drug's transdermal rate.

[0003] The information disclosed in the background technology section of this application is only intended to deepen the understanding of the general background technology of this application, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art. Summary of the Invention

[0004] The present invention provides a method and device for accurately controlling the temperature and pressure of acupuncture points by negative pressure adsorption, which can solve the technical problem that the related art is difficult to accurately control the temperature and pressure of the space inside the negative pressure suction cup.

[0005] According to a first aspect of the present invention, a method for precise temperature-controlled negative pressure adsorption transdermal acupuncture is provided, comprising:

[0006] The negative pressure device is activated to form a negative pressure in the space between the negative pressure suction cup and the skin, wherein medicine is placed in the space and an infrared heating device and an electric heating device are placed in the negative pressure suction cup;

[0007] The skin is heated by an infrared heating device, and the medicine is heated by an electric heating device;

[0008] detecting first temperature data of the skin, second temperature data of the medicine, and pressure data in the space;

[0009] The first temperature data and the first switching state of the infrared heating device form a first state vector, the second temperature data and the second switching state of the electric heating device form a second state vector, and the pressure data and the third switching state of the negative pressure device form a third state vector;

[0010] Obtaining first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple future moments based on the first state vector, the second state vector, and the third state vector at the i-th moment and multiple moments before the i-th moment;

[0011] Determine, based on the first predicted temperature data, the second predicted temperature data, and the predicted pressure data, a first predicted switch state of the infrared heating device, a second predicted switch state of the electric heating device, and a third predicted switch state of the negative pressure device at the (i+1)th moment;

[0012] At the (i+1)th moment, setting is performed according to the first predicted switch state, the second predicted switch state, and the third predicted switch state.

[0013] According to the present invention, first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple future moments are obtained based on the first state vector, the second state vector, and the third state vector at the i-th moment and multiple moments before the i-th moment, including:

[0014] Input the first state vector, the second state vector, and the third state vector at the same time into the first fully connected layer, the second fully connected layer, and the third fully connected layer, respectively, to obtain a first input vector corresponding to the first state vector, a second input vector corresponding to the second state vector, and a third input vector corresponding to the third state vector;

[0015] Obtaining a first graph structure according to the first input vector, the second input vector, and the third input vector;

[0016] The first graph structure is processed using the trained graph neural network model to obtain momentary state feature information;

[0017] The trained state prediction model is used to process the moment state feature information corresponding to the i-th moment and multiple moments before it to obtain the first predicted temperature data of the skin at multiple moments in the future, the second predicted temperature data of the medicine, and the predicted pressure data in the space.

[0018] According to the present invention, the first graph structure is processed by a trained graph neural network model to obtain momentary state feature information, including:

[0019] Obtaining an adjacency matrix of the first graph structure;

[0020] Processing the first input vector, the second input vector, and the third input vector through the attention mechanism layer of the trained graph neural network model to obtain a connection weight between the first input vector and the second input vector, a connection weight between the second input vector and the third input vector, and a connection weight between the first input vector and the third input vector;

[0021] Processing the connection weight between the first input vector and the second input vector, the connection weight between the second input vector and the third input vector, the connection weight between the first input vector and the third input vector, the adjacency matrix, the first input vector, the second input vector, and the third input vector through the trained graph neural network model to obtain first node feature information corresponding to the first input vector, second node feature information corresponding to the second input vector, and third node feature information corresponding to the third input vector;

[0022] The first node characteristic information, the second node characteristic information and the third node characteristic information are concatenated to obtain the momentary state characteristic information.

[0023] According to the present invention, the state feature information corresponding to the i-th moment and multiple moments before the i-th moment is processed by a trained state prediction model to obtain first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple future moments, including:

[0024] Input the moment state feature information at the i-n1+jth moment and the hidden state information at the i-n1+j-1th moment into the j-th input port of the trained state prediction model to obtain the hidden state information at the i-n1+jth moment, where n1 is the number of input ports of the state prediction model, 1≤j≤n1, and i, j, and n1 are all positive integers. When j=1, the hidden state information at the i-n1+j-1th moment is a zero vector;

[0025] Iterate the steps of obtaining hidden state information until the hidden state information at the i-th moment is obtained;

[0026] Input the hidden state information at the i-th moment into the fourth fully connected layer and the first activation layer to obtain the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the i+1-th moment;

[0027] Maintain the first switch state, the second switch state, and the third switch state at the i-th moment, and obtain the first predicted input vector, the second predicted input vector, and the third predicted input vector at the i+k-th moment based on the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the i+k-th moment, where 1≤k≤n2, and n2 is the number of moments in the future;

[0028] Obtaining a first prediction graph structure at the i+kth moment according to the first prediction input vector, the second prediction input vector, and the third prediction input vector at the i+kth moment;

[0029] The first prediction graph structure at the i+kth moment is processed through the trained graph neural network model to obtain the predicted moment state feature information at the i+kth moment;

[0030] The trained state prediction model is used to process the state feature information from the i-n1+k+1th moment to the i-th moment and the predicted state feature information from the i+1th moment to the i+kth moment to obtain the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the i+k+1th moment;

[0031] The process of acquiring the first predicted temperature data, the second predicted temperature data, and the predicted pressure data is iteratively performed until the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the (i+n2)th moment are acquired.

[0032] According to the present invention, the method further comprises:

[0033] Obtaining a first historical state vector, a second historical state vector, and a third historical state vector at multiple historical moments;

[0034] The first historical state vector, the second historical state vector, and the third historical state vector at the same historical moment are processed by a graph neural network model to obtain the historical feature information of the first node, the historical feature information of the second node, and the historical feature information of the third node;

[0035] Fully connect and activate the historical feature information of the first node, the historical feature information of the second node, and the historical feature information of the third node respectively to obtain skin temperature judgment results, drug temperature judgment results, and pressure judgment results;

[0036] Determine the state loss function based on the skin temperature judgment results, the drug temperature judgment results, the pressure judgment results, and the annotation information of the historical moments;

[0037] splicing the historical feature information of the first node, the historical feature information of the second node, and the historical feature information of the third node to obtain historical moment state feature information;

[0038] Processing historical moment state feature information corresponding to multiple historical moments through a state prediction model to obtain first training temperature data, second training temperature data, and training pressure data for subsequent historical moments of the multiple historical moments;

[0039] Determining a prediction loss function based on the first training temperature data, the second training temperature data, and the training pressure data, as well as the first historical temperature data of the skin, the second historical temperature data of the drug, and the historical pressure data in the space at subsequent historical moments of the plurality of historical moments;

[0040] According to the state loss function and the prediction loss function, the graph neural network model and the state prediction model are trained to obtain a trained graph neural network model and a trained state prediction model.

[0041] According to the present invention, a state loss function is determined based on the skin temperature judgment result, the drug temperature judgment result, the pressure judgment result, and the annotation information of the historical moment, including:

[0042] According to the formula

[0043]

[0044] Determine the state loss function L S , where p la,skin,x is the probability that the skin temperature is normal according to the labeled information at the xth historical moment, p skin,x is the skin temperature judgment result at the xth historical moment, p la,me,x is the probability that the drug temperature is normal according to the labeling information at the xth historical moment, p me,x is the drug temperature judgment result at the xth historical moment, p la,pr,x is the probability that the pressure is normal according to the annotation information at the xth historical moment, p pr,x is the pressure judgment result at the xth historical moment, w1, w2 and w3 are preset weights, N is the number of historical moments, x≤N, and both x and N are positive integers.

[0045] According to the present invention, a prediction loss function is determined based on the first training temperature data, the second training temperature data, and the training pressure data, as well as the first historical temperature data of the skin, the second historical temperature data of the drug, and the historical pressure data in the space at subsequent historical moments of a plurality of historical moments, including:

[0046] According to the formula

[0047]

[0048] Determine the prediction loss function L P , where T T,1,y is the first training temperature data at the yth historical moment, T H,2,y is the first historical temperature data at the yth historical moment, T T,2,y is the second training temperature data at the yth historical moment, T H,2,y is the second historical temperature data at the yth historical moment, P T,y is the training pressure data at the yth historical moment, P H,y is the historical pressure data at the yth historical moment, w4, w5 and w6 are preset weights, n1 is the number of input ports of the state prediction model, N is the number of historical moments, N-n1+1≤y≤N, and y, n1 and N are all positive integers.

[0049] According to a second aspect of the present invention, there is provided a precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction device, comprising:

[0050] a start module for starting the negative pressure device to form a negative pressure in the space between the negative pressure suction cup and the skin, wherein medicine is placed in the space and an infrared heating device and an electric heating device are placed in the negative pressure suction cup;

[0051] A heating module is used to heat the skin through an infrared heating device and heat the medicine through an electric heating device;

[0052] a detection module, configured to detect first temperature data of the skin, second temperature data of the medicine, and pressure data in the space;

[0053] a state vector module, configured to combine the first temperature data and the first switching state of the infrared heating device into a first state vector, combine the second temperature data and the second switching state of the electric heating device into a second state vector, and combine the pressure data and the third switching state of the negative pressure device into a third state vector;

[0054] a prediction module, configured to obtain first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple future moments based on the first state vector, the second state vector, and the third state vector at the i-th moment and multiple moments before the i-th moment;

[0055] a switch state module, configured to determine a first predicted switch state of the infrared heating device, a second predicted switch state of the electric heating device, and a third predicted switch state of the negative pressure device at the (i+1)th moment based on the first predicted temperature data, the second predicted temperature data, and the predicted pressure data;

[0056] The setting module is used to perform setting according to the first predicted switch state, the second predicted switch state and the third predicted switch state at the (i+1)th moment.

[0057] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0058] According to the present invention, not only can the dispersion and absorption of drugs be promoted by the electric heating device, but the local skin blood circulation can also be promoted by the infrared heating device, thereby improving the absorption efficiency of drugs. Moreover, by monitoring the skin temperature, drug temperature and pressure at multiple moments, the switch state of the infrared heating device, the electric heating device and the negative pressure device at the subsequent moment can be determined, thereby achieving accurate control of the skin temperature, drug temperature and pressure, reducing the probability of problems such as skin burns, drug failure and inaccurate pressure control causing a decrease in the transdermal rate of the drug. When obtaining the momentary state feature information, the features of each node can be aggregated through the graph neural network model. During the aggregation process, the connection weights between nodes are determined using the attention mechanism hierarchy, thereby accurately reflecting the mutual influence between each node, thereby obtaining the momentary state feature information. It can not only accurately describe the switch state of the infrared heating device, the switch state of the electric heating device, the switch state of the negative pressure device, the skin temperature, the drug temperature, and the pressure in space at that moment, but also describe the mutual influence of these data, thereby improving the accuracy of the momentary state feature information and the comprehensiveness of the feature description, and providing an accurate data basis for subsequent prediction processing. When predicting data at multiple future moments, a sliding time window model can be used to perform backward predictions based on data from multiple recent moments, discarding data from a longer period of time. This discards data that has little impact on current and subsequent predictions or may even have a negative impact on accuracy, thereby improving prediction accuracy. Furthermore, the switch state at the i-th moment can be kept unchanged during the prediction process to predict data at multiple future moments, thereby providing an accurate data basis for determining whether the switch state should be changed. When determining the state loss function, the training of the graph neural network model can be assisted by introducing fully connected layers and activation layers. This allows the graph neural network model to be trained using data from each historical moment without waiting for the state prediction model to complete processing the state feature information of multiple historical moments. This improves the training strength of the graph neural network model. While the state prediction model is processing the state feature information of multiple historical moments, the graph neural network model can be trained using data from multiple historical moments. This helps the state prediction model obtain more accurate state feature information of historical moments and make more accurate predictions, thereby improving the overall training efficiency of the graph neural network model and the state prediction model. When determining the prediction loss function, the prediction loss function can be obtained through the gap between the model output data and the actual historical data at each historical moment. Therefore, during the training process, the graph neural network model and the state prediction model can be trained as a whole through backpropagation of the prediction loss function. In the process of narrowing the gap between the model output data and the actual historical data, the overall error of the graph neural network model and the state prediction model is reduced, thereby improving the overall performance of the graph neural network model and the state prediction model and improving the prediction accuracy.

[0059] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and not limiting of the present invention. Other features and aspects of the present invention will become more apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can derive other embodiments based on these drawings without inventive efforts.

[0061] Figure 1 A schematic diagram exemplarily illustrates a process flow of a method for precise temperature-controlled negative pressure adsorption transdermal acupuncture according to an embodiment of the present invention;

[0062] Figure 2 A schematic diagram of obtaining first predicted temperature data, second predicted temperature data, and predicted pressure data according to an embodiment of the present invention is exemplarily shown;

[0063] Figure 3 A block diagram of a precise temperature-controlled negative pressure adsorption acupuncture point transdermal introduction device according to an embodiment of the present invention is exemplarily shown. DETAILED DESCRIPTION

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0065] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0066] Figure 1 The following is a flow chart showing a method for precise temperature-controlled negative pressure adsorption transdermal acupuncture according to an embodiment of the present invention, wherein the method comprises:

[0067] Step S1, starting the negative pressure device to form a negative pressure in the space between the negative pressure suction cup and the skin, wherein medicine is placed in the space, and an infrared heating device and an electric heating device are placed in the negative pressure suction cup;

[0068] Step S2, heating the skin by an infrared heating device and heating the medicine by an electric heating device;

[0069] Step S3, detecting first temperature data of the skin, second temperature data of the medicine, and pressure data in the space;

[0070] Step S4, combining the first temperature data and the first switching state of the infrared heating device into a first state vector, combining the second temperature data and the second switching state of the electric heating device into a second state vector, and combining the pressure data and the third switching state of the negative pressure device into a third state vector;

[0071] Step S5, obtaining first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple future moments based on the first state vector, the second state vector, and the third state vector at the i-th moment and multiple moments before the i-th moment;

[0072] Step S6, determining a first predicted switch state of the infrared heating device, a second predicted switch state of the electric heating device, and a third predicted switch state of the negative pressure device at the (i+1)th moment based on the first predicted temperature data, the second predicted temperature data, and the predicted pressure data;

[0073] Step S7: at the (i+1)th moment, setting is performed according to the first predicted switch state, the second predicted switch state, and the third predicted switch state.

[0074] According to the precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction method of the embodiment of the present invention, not only can the dispersion and absorption of drugs be promoted by the electric heating device, but also the local skin blood circulation can be promoted by the infrared heating device, thereby improving the absorption efficiency of drugs. Moreover, by monitoring the skin temperature, drug temperature and pressure at multiple moments, the switching states of the infrared heating device, the electric heating device and the negative pressure device at subsequent moments can be determined, thereby achieving accurate control of the skin temperature, drug temperature and pressure, and reducing the probability of problems such as skin burns, drug failure and decreased drug transdermal rate due to inaccurate pressure control.

[0075] According to one embodiment of the present invention, in step S1 and step S2, an infrared heating device and an electric heating device may be provided in the negative pressure suction cup. The infrared heating device can emit infrared light to irradiate the skin, causing the skin to heat up evenly after receiving the infrared radiation. The radiant energy can directly reach the subcutaneous tissue, accelerating microcirculation and increasing the absorption rate of the drug. The electric heating device can heat drugs in the form of gels or creams, making the molecular thermal motion of the drugs more intense, increasing the diffusion rate of the drugs, and helping to increase the absorption rate of the drugs. The negative pressure suction cup can be placed on the skin surface at a preset position, for example, the skin surface at the location of a meridian or acupoint, and the negative pressure device can be activated to form a negative pressure in the space between the negative pressure suction cup and the skin, for example, forming a negative pressure of -100 mmHg. Furthermore, multiple sensors can be used to monitor multiple data points in the space between the negative pressure suction cup and the skin. For example, a temperature sensor can be used to detect the temperature of the skin surface, thereby controlling the on / off state of the infrared heating device based on the skin surface temperature, keeping the skin surface temperature within a certain range (e.g., between 38°C and 40°C), improving microcirculation while preventing skin burns. For example, another temperature sensor can be used to detect the temperature of the drug. For example, the probe of this temperature sensor can be inserted into the drug gel to detect the drug temperature, and based on the drug temperature, the electric heating device can be controlled to keep the drug temperature within a certain range (e.g., between 38°C and 40°C), improving the diffusion rate of the drug while preventing the drug from becoming ineffective due to high temperature. For another example, a pressure sensor can be used to monitor the pressure in the space between the negative pressure suction cup and the skin, thereby controlling the on / off state of the negative pressure device to maintain the pressure within a certain range (e.g., between -80 mmHg and -200 mmHg), allowing the negative pressure suction cup to remain on the skin surface and causing the skin to deform and expand to facilitate drug penetration into the subcutaneous tissue.

[0076] According to one embodiment of the present invention, in step S3, the above-mentioned sensor can detect the first temperature data of the skin, the second temperature data of the medicine, and the pressure data in the space at multiple moments respectively. The time interval between adjacent moments can be 5 seconds or 10 seconds, etc., and the present invention does not impose any restrictions on this.

[0077] According to one embodiment of the present invention, in step S4, the first temperature data and the first switch state of the infrared heating device can be combined into a first state vector. For example, the first switch state can be represented by 0 or 1. When the first switch state is 0, it indicates that the infrared heating device is turned off. When the first switch state is 1, it indicates that the infrared heating device is turned on and heating is in progress. The first temperature data and the first switch state at a certain moment can form a first state vector. For example, if the first temperature data at this moment is 38.5°C and the first switch state is 1, then the first state vector at this moment is Similarly, the second state vector and the third state vector at multiple moments can be obtained.

[0078] According to one embodiment of the present invention, in step S5, after obtaining the first state vector, the second state vector and the third state vector at multiple moments, the temperature of the skin, the temperature of the medicine and the pressure in the space between the negative pressure suction cup and the skin at future moments can be predicted, so as to determine whether the temperature of the skin will soon exceed the allowable temperature upper limit, whether the temperature of the medicine will soon exceed the allowable temperature upper limit, and whether the pressure in the space will deviate from the allowable range if the current switch state is maintained, so as to judge whether the switch state needs to be changed.

[0079] Figure 2 A schematic diagram of obtaining first predicted temperature data, second predicted temperature data, and predicted pressure data according to an embodiment of the present invention is exemplarily shown.

[0080] According to one embodiment of the present invention, in step S5, based on the first state vector, second state vector and third state vector at the i-th moment and multiple moments before it, first predicted temperature data of the skin at multiple moments in the future, second predicted temperature data of the medicine and predicted pressure data in the space are obtained, including: inputting the first state vector, second state vector and third state vector at the same moment into the first fully connected layer, the second fully connected layer and the third fully connected layer respectively to obtain a first input vector corresponding to the first state vector, a second input vector corresponding to the second state vector and a third input vector corresponding to the third state vector; obtaining a first graph structure based on the first input vector, the second input vector and the third input vector; processing the first graph structure through a trained graph neural network model to obtain moment state feature information; processing the moment state feature information corresponding to the i-th moment and multiple moments before it through a trained state prediction model to obtain the first predicted temperature data of the skin at multiple moments in the future, the second predicted temperature data of the medicine and predicted pressure data in the space.

[0081] According to one embodiment of the present invention, the first state vector, the second state vector, and the third state vector are all two-dimensional vectors. The first state vector, the second state vector, and the third state vector can be processed by the first fully connected layer, the second fully connected layer, and the third fully connected layer, respectively, to obtain higher-dimensional first input vectors, second input vectors, and third input vectors, thereby describing richer feature information at each moment. For example, the first input vector, the second input vector, and the third input vector can be 16-dimensional vectors or 32-dimensional vectors, which is not limited to this aspect of the present invention.

[0082] According to one embodiment of the present invention, the first graph structure may include nodes and edges. The first state vector, the second state vector, and the third state vector may serve as nodes in the first graph structure. The edges may be used to connect the nodes. If there is an association between two nodes, there is a connecting edge between the two nodes. Conversely, if there is no association between the two nodes, there is no connecting edge between the two nodes.

[0083] According to one embodiment of the present invention, the first input vector can be used as the initial feature of the node corresponding to the first state vector, the second input vector can be used as the initial feature of the node corresponding to the second state vector, and the third input vector can be used as the initial feature of the node corresponding to the third state vector. In addition, when the infrared heating device heats the skin, it can cause the temperature in the space between the negative pressure suction cup and the skin to rise, thereby indirectly raising the temperature of the medicine. When the electric heating device heats the medicine, it can also cause the temperature in the space between the negative pressure suction cup and the skin to rise, thereby indirectly raising the temperature of the skin. The first state vector is related to the infrared heating device and the skin temperature, and the second state vector is related to the electric heating device and the medicine temperature. Based on the above mutual influence relationship, it can be seen that there is an association between the node corresponding to the first state vector and the node corresponding to the second state vector, and there is an edge connecting the two nodes. Furthermore, the temperature in the space may also affect the pressure in the space. For example, an increase in temperature causes the thermal motion of air molecules to intensify, which may cause an increase in air pressure. Both the infrared heating device and the electric heating device may cause temperature changes in the space, and the third state vector is related to the negative pressure device and the pressure in the space. Therefore, the nodes corresponding to the first state vector and the second state vector are associated with the nodes corresponding to the third state vector, that is, there is an edge connecting the node corresponding to the first state vector and the node corresponding to the third state vector, and there is an edge connecting the node corresponding to the second state vector and the node corresponding to the third state vector. Based on the above nodes and edges, the first graph structure can be obtained.

[0084] According to one embodiment of the present invention, a first graph structure is processed by a trained graph neural network model to obtain momentary state feature information, including: obtaining an adjacency matrix of the first graph structure; processing a first input vector, a second input vector, and a third input vector by an attention mechanism layer of a trained graph neural network model to obtain the connection weights between the first input vector and the second input vector, the connection weights between the second input vector and the third input vector, and the connection weights between the first input vector and the third input vector; processing the connection weights between the first input vector and the second input vector, the connection weights between the second input vector and the third input vector, the connection weights between the first input vector and the third input vector, the adjacency matrix, the first input vector, the second input vector, and the third input vector by a trained graph neural network model to obtain first node feature information corresponding to the first input vector, second node feature information corresponding to the second input vector, and third node feature information corresponding to the third input vector; and concatenating the first node feature information, the second node feature information, and the third node feature information to obtain momentary state feature information.

[0085] According to one embodiment of the present invention, based on the above description of edges, an adjacency matrix between nodes can be obtained. In the adjacency matrix, the data in the ath row and the bth column is used to indicate whether there is an edge connecting the ath node and the bth node. If so, the data in the ath row and the bth column is 1, otherwise it is 0. The elements on the diagonal of the adjacency matrix are 0. Both a and b are positive integers less than or equal to 3.

[0086] According to one embodiment of the present invention, in an adjacency matrix, it is possible to describe whether a connection relationship exists between nodes, but it is impossible to distinguish the strength of the connection relationship. For example, if two nodes have a high correlation and the other two nodes have a low correlation, it is difficult to reflect this in the adjacency matrix. Therefore, the first input vector, the second input vector, and the third input vector can be processed by the attention mechanism layer of the trained graph neural network model to obtain the connection weight between the first input vector and the second input vector, the connection weight between the second input vector and the third input vector, and the connection weight between the first input vector and the third input vector. The higher the connection weight, the higher the correlation between the two nodes. The first input vector and the second input vector can be processed by the attention mechanism layer to obtain the connection weight between the node corresponding to the first state vector and the node corresponding to the second state vector. The first input vector and the third input vector can be processed by the attention mechanism layer to obtain the connection weight between the node corresponding to the first state vector and the node corresponding to the third state vector. The second input vector and the third input vector can be processed by the attention mechanism layer to obtain the connection weight between the node corresponding to the second state vector and the node corresponding to the third state vector. Taking the connection weight between the node corresponding to the first state vector and the node corresponding to the second state vector as an example, the processing method of the attention mechanism layer may include using a weight matrix to multiply the first input vector and the second input vector respectively to obtain a first weighted input vector corresponding to the first input vector and a second weighted input vector corresponding to the second input vector. Further, the first weighted input vector and the second weighted input vector may be spliced ​​to obtain a weighted spliced ​​feature vector, and then a weight vector with the same dimension as the weighted spliced ​​feature vector is used to perform an inner product operation with the weighted spliced ​​feature vector to obtain a weight operation value. After the weight operation value is processed by an activation function (for example, a LeakyReLU function, a softmax function, etc.), the connection weight between the node corresponding to the first state vector and the node corresponding to the second state vector can be obtained. Similarly, the connection weight between the node corresponding to the first state vector and the node corresponding to the third state vector, as well as the connection weight between the node corresponding to the second state vector and the node corresponding to the third state vector can be obtained. Furthermore, the connection weights can be used to replace the data in the adjacency matrix to obtain an updated adjacency matrix. For example, in the original adjacency matrix, the data in row a, column b is 1. In the updated adjacency matrix, the data in row a, column b is the connection weight between node a and node b. The data in the above weight matrix and weight vector can be obtained through training.

[0087] According to one embodiment of the present invention, the updated adjacency matrix, as well as the first input vector, the second input vector, and the third input vector can be processed by a trained graph neural network model to obtain first node feature information corresponding to the first input vector, second node feature information corresponding to the second input vector, and third node feature information corresponding to the third input vector. Thus, each node feature information not only describes its own state, but also describes the influence of other nodes on itself. For example, the first node feature information can describe not only the first on / off state of the infrared heating device and the first temperature data of the skin, but also the influence of the on / off state of the electric heating device on the skin temperature, the influence of the on / off state of the negative pressure device on the skin temperature, etc.

[0088] According to one embodiment of the present invention, the first, second, and third node feature information can be concatenated to obtain moment-by-moment state feature information. This information describes the on / off status of the infrared heating device, the on / off status of the electric heating device, the on / off status of the negative pressure device, the skin temperature, the temperature of the medication, the pressure in space, and the interactions between these data at that moment. Similarly, moment-by-moment state feature information can be obtained for the i-th moment and multiple moments prior to that moment.

[0089] In this way, the features of each node can be aggregated through the graph neural network model. During the aggregation process, the attention mechanism hierarchy is used to determine the connection weights between nodes, so as to accurately reflect the mutual influence between each node and obtain the momentary state feature information. It can not only accurately describe the on-off state of the infrared heating device, the on-off state of the electric heating device, the on-off state of the negative pressure device, the temperature of the skin, the temperature of the medicine, and the pressure in space at that moment, but also describe the mutual influence of these data, thereby improving the accuracy of the momentary state feature information and the comprehensiveness of the description of the features, and providing an accurate data basis for subsequent prediction processing.

[0090] According to one embodiment of the present invention, the state feature information corresponding to the i-th moment and the multiple moments before it is processed by a trained state prediction model to obtain the first predicted temperature data of the skin at multiple moments in the future, the second predicted temperature data of the drug, and the predicted pressure data in the space, including: inputting the state feature information of the i-n1+j-th moment and the hidden state information of the i-n1+j-1-th moment into the j-th input port of the trained state prediction model to obtain the hidden state information of the i-n1+j-th moment, where n1 is the state The number of input ports of the prediction model is 1≤j≤n1, and i, j and n1 are all positive integers, and when j=1, the hidden state information at the i-n1+j-1th moment is a zero vector; iteratively execute the step of obtaining the hidden state information until the hidden state information at the i-th moment is obtained; the hidden state information at the i-th moment is input into the fourth fully connected layer and the first activation layer to obtain the first predicted temperature data, the second predicted temperature data and the predicted pressure data at the i+1th moment; the first switch state, the second switch state and the third switch state at the i-th moment are maintained, and according to the i+k-th moment, the hidden state information is obtained. The first predicted temperature data, second predicted temperature data and predicted pressure data at the moment are obtained to obtain the first predicted input vector, second predicted input vector and third predicted input vector of the i+kth moment, where 1≤k≤n2, and n2 is the number of moments in the future; the first predicted graph structure of the i+kth moment is obtained according to the first predicted input vector, second predicted input vector and third predicted input vector of the i+kth moment; the first predicted graph structure of the i+kth moment is processed by the trained graph neural network model to obtain the predicted moment state feature information of the i+kth moment; the moment state feature information from the i-n1+k+1th moment to the i-th moment and the predicted moment state feature information from the i+1th moment to the i+kth moment are processed by the trained state prediction model to obtain the first predicted temperature data, second predicted temperature data and predicted pressure data of the i+k+1th moment; the process of obtaining the first predicted temperature data, the second predicted temperature data and the predicted pressure data is iteratively executed until the first predicted temperature data, the second predicted temperature data and the predicted pressure data of the i+n2th moment are obtained.

[0091] According to one embodiment of the present invention, the state prediction model may be an LSTM model or an RNN model, and the state prediction model may have n1 input ports, and therefore, may receive the moment state feature information of n1 moments in chronological order. The input port backend may include a processing module, which may process the moment state feature information received by the input port and the hidden state information output by the processing module corresponding to the previous input port, and obtain the hidden state information output by the current processing module. In the example, at the beginning of the process of negative pressure adsorption transdermal drug delivery, the negative pressure device, the infrared heating device and the electric heating device may be turned on at the same time, and the moment state feature information of each moment after the start moment may be obtained. The time interval between adjacent moments may be 5 seconds or 10 seconds, etc. The present invention does not limit the specific value of the time interval between adjacent moments. At n1 moments after the start time, the state feature information from the 1st moment to the n1th moment can be input into n1 (e.g., 5) input ports of the state prediction model to predict the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the n1+1th moment and subsequent moments. Alternatively, it is not necessary to wait for n1 moments before inputting the state feature information at the same time. Instead, after the 1st moment, the state feature information of the 1st moment can be input into the first port, and after the 2nd moment, the hidden state information of the 1st moment and the state feature information of the 2nd moment can be input into the second port, and so on, thereby saving computation time. Similarly, if the current moment is the i-th moment (i>n1), the state feature information from the i-n1+1th moment to the i-th moment can be input into the n1 input ports of the state prediction model to predict the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the i+1th moment and subsequent moments. Specifically, the moment state feature information at the i-n1+jth moment can be input into the jth input port, and the hidden state information at the i-n1+j-1th moment can also be input into the jth input port. When j=1, the hidden state information at the i-n1+j-1th moment (i.e., the i-n1th moment) is a zero vector. In other words, each time the state prediction model predicts the first predicted temperature data of a subsequent moment, the second predicted temperature data of the drug, and the predicted pressure data in the space, the moment state feature information at the current moment and the n1-1 moments before it is used, and the hidden state information at the i-n1th moment and the moments before it is discarded. This means that backward prediction can be performed based on the data at the recent n1 moments using a sliding time window mode, while data from a longer period of time (i-n1th moment and the moments before it) is discarded. This allows discarding data that has little impact on the current and subsequent predictions, or may even have a negative impact on the accuracy, thereby improving prediction accuracy.At the i-th moment, the above method can be used to calculate the moment state feature information of multiple moments between the i-n1+1-th moment and the i-th moment to obtain the hidden state information of the i-th moment.

[0092] According to an embodiment of the present invention, each input port combines the current moment state feature information and the previous moment hidden state information to obtain the current moment hidden state information. In the operation, the hidden state information of the i-n1+1th moment and the momentary state feature information of the i-n1+2th moment are jointly calculated to obtain the hidden state information of the i-n1+2th moment. In other words, the hidden state information of the i-n1+2th moment is fused with the hidden state information of the i-n1+1th moment. Similarly, the hidden state information of the i-n1+3th moment is fused with the hidden state information of the i-n1+2th moment. Therefore, it is also fused with the hidden state information of the i-n1+1th moment... and so on. The hidden state information of the i-th moment is fused with the hidden state information from the i-n1+1th moment to the i-1th moment, and is also fused with the momentary state feature information of the i-th moment. In other words, the hidden state information of the i-th moment is fused with the features from the i-n1+1th moment to the i-th moment, and therefore, can be used to represent the overall features of the above n1 moments. The overall features of n1 moments can be used to predict the data at subsequent moments. For example, the hidden state information at the i-th moment is input into the fourth fully connected layer and the first activation layer to obtain the first predicted temperature data, the second predicted temperature data and the predicted pressure data at the i+1-th moment. The fourth fully connected layer can convert the high-dimensional hidden state information into a three-dimensional vector, and activate the vector through the first activation layer. For example, activation functions such as RELU are used for processing to obtain the first predicted temperature data, the second predicted temperature data and the predicted pressure data at the i+1-th moment. When obtaining the predicted pressure data, its absolute pressure data (i.e., pressure data based on vacuum) can be determined. For example, the pressure of -200 mmHg in the space (pressure data based on standard atmospheric pressure) is converted to absolute pressure data of 560 mmHg.

[0093] According to one embodiment of the present invention, in the subsequent prediction process, it can be assumed that the first switch state, the second switch state, and the third switch state at the i-th moment remain unchanged, so as to determine the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at multiple subsequent moments under the current switch state. For example, the first switch state at the i-th moment and the first predicted temperature data at the i+1-th moment are combined into the first predicted state vector at the i+1-th moment, the second switch state at the i-th moment and the second predicted temperature data at the i+1-th moment are combined into the second predicted state vector at the i+1-th moment, and the third switch state at the i-th moment and the predicted pressure data at the i+1-th moment are combined into the third predicted state vector at the i+1-th moment. Then, through the above-mentioned first fully connected layer, the second fully connected layer, and the third fully connected layer, the first predicted input vector, the second predicted input vector, and the third predicted input vector are obtained respectively, and then the first prediction graph structure at the i+1-th moment is obtained. Then, the trained graph neural network model is used for processing to obtain the predicted state feature information at the i+1-th moment. Since the state prediction model has n1 input ports, Therefore, the moment state feature information from the i-n1+2th moment to the i-th moment and the predicted moment state feature information at the i+1th moment can be input into the trained state prediction model to obtain the predicted hidden state information at the i+1th moment, and then processed through the fourth fully connected layer and the first activation layer to obtain the first predicted temperature data, the second predicted temperature data and the predicted pressure data at the i+2th moment, and the first switch state, the second switch state and the third switch state at the i-th moment are still maintained, and form the first predicted state vector, the second predicted state vector and the third predicted state vector at the i+2th moment together with the first predicted temperature data, the second predicted temperature data and the predicted pressure data at the i+2th moment, and then perform the above processing again to obtain the first predicted temperature data, the second predicted temperature data and the predicted pressure data at the i+3th moment. The above processing can be iteratively performed until the first predicted temperature data, the second predicted temperature data and the predicted pressure data at the i+n2th moment are obtained, where n2 is the number of moments in the future, for example, n2=5. The first predicted temperature data, the second predicted temperature data and the predicted pressure data at the future multiple moments can be referred to to determine whether the switch state at the i+1th moment should change.

[0094] In this way, backward prediction can be made based on data from multiple recent moments through a sliding time window mode, and data from a longer period of time can be discarded, thereby discarding data that has little impact on current and subsequent predictions or may even have a negative impact on accuracy, so as to improve prediction accuracy. In the prediction process, the switch state at the i-th moment can be kept unchanged to predict data from multiple future moments, thereby providing an accurate data basis for judging whether the switch state should be changed.

[0095] According to one embodiment of the present invention, in step S6, the average value of the first predicted temperature data at multiple future moments can be determined. If the average value exceeds the reasonable range of skin temperature (e.g., between 38°C and 40°C), the first switch state at the i-th moment can be changed, and the changed first switch state can be used as the first predicted switch state of the infrared heating device at the i+1-th moment. For example, if the first switch state at the i-th moment is 1, i.e., the on state, if the average value of the first predicted temperature data at multiple future moments exceeds 40°C, the first predicted switch state of the infrared heating device at the i+1-th moment is determined to be 0, i.e., the off state. Similarly, the second predicted switch state of the electric heating device and the third predicted switch state of the negative pressure device at the i+1-th moment can be determined. In step S7, the infrared heating device is set according to the first predicted switch state, the electric heating device is set according to the second predicted switch state, and the negative pressure device is set according to the third predicted switch state.

[0096] According to one embodiment of the present invention, the above graph neural network model and state prediction model can be trained before use. The method also includes: obtaining the first historical state vector, the second historical state vector and the third historical state vector of multiple historical moments; processing the first historical state vector, the second historical state vector and the third historical state vector of the same historical moment through the graph neural network model to obtain the first node historical feature information, the second node historical feature information and the third node historical feature information; performing full connection and activation processing on the first node historical feature information, the second node historical feature information and the third node historical feature information respectively to obtain the skin temperature judgment result, the drug temperature judgment result and the pressure judgment result; determining the state loss function according to the skin temperature judgment result, the drug temperature judgment result and the pressure judgment result, and the annotation information of the historical moment; performing full connection and activation processing on the first node historical feature information, the second node historical feature information and the third node historical feature information respectively to obtain the skin temperature judgment result, the drug temperature judgment result and the pressure judgment result; determining the state loss function according to the skin temperature judgment result, the drug temperature judgment result and the pressure judgment result, and the annotation information of the historical moment; performing full connection and activation processing on the first node historical feature information, the second node historical feature information and the third node historical feature information The node historical feature information and the third node historical feature information are spliced ​​to obtain the state feature information of the historical moment; the historical moment state feature information corresponding to multiple historical moments is processed by the state prediction model to obtain the first training temperature data, second training temperature data and training pressure data of the subsequent historical moments of the multiple historical moments; the prediction loss function is determined based on the first training temperature data, second training temperature data and training pressure data, as well as the first historical temperature data of the skin, the second historical temperature data of the medicine and the historical pressure data in the space at the subsequent historical moments of the multiple historical moments; the graph neural network model and the state prediction model are trained according to the state loss function and the prediction loss function to obtain the trained graph neural network model and the trained state prediction model.

[0097] According to one embodiment of the present invention, the first historical state vector, the second historical state vector and the third historical state vector of multiple historical moments can be obtained in a similar manner to the first state vector, the second state vector and the third state vector, which will not be repeated here.

[0098] According to one embodiment of the present invention, the first historical state vector, the second historical state vector and the third historical state vector at the same historical moment can be processed by a graph neural network model to obtain the first node historical feature information, the second node historical feature information and the third node historical feature information. The acquisition methods are similar to those of the first node feature information, the second node feature information and the third node feature information, and will not be repeated here.

[0099] According to one embodiment of the present invention, the graph neural network model can be trained separately by additionally setting up a fully connected layer and an activation layer. For example, the historical feature information of the first node can be input into the fully connected layer, and the output of the fully connected layer can be processed through an activation function (e.g., a sigmoid function) to obtain a skin temperature judgment result, which describes the probability that the skin temperature is within a reasonable range (e.g., between 38°C and 40°C). Similarly, drug temperature judgment results and pressure judgment results can be obtained.

[0100] According to one embodiment of the present invention, the graph neural network model can be trained based on the above skin temperature judgment results, drug temperature judgment results, and pressure judgment results, and the annotation information of the historical moments. According to the skin temperature judgment results, drug temperature judgment results, and pressure judgment results, and the annotation information of the historical moments, the state loss function is determined, including: determining the state loss function L according to formula (1) S ,

[0101]

[0102] Among them, p la,skin,x is the probability that the skin temperature is normal according to the labeled information at the xth historical moment, p skin,x is the skin temperature judgment result at the xth historical moment, p la,me,x is the probability that the drug temperature is normal according to the labeling information at the xth historical moment, p me,x is the drug temperature judgment result at the xth historical moment, p la,pr,x is the probability that the pressure is normal according to the annotation information at the xth historical moment, p pr,x is the pressure judgment result at the xth historical moment, w1, w2 and w3 are preset weights, N is the number of historical moments, x≤N, and both x and N are positive integers.

[0103] According to one embodiment of the present invention, the labeling information can be a 0 or 1 labeling. For example, if the skin temperature is normal (e.g., between 38°C and 40°C), the probability of normal skin temperature is 1, otherwise it is 0. The labeling method for the probability of normal drug temperature and normal pressure is similar to the probability of normal skin temperature, and will not be repeated here.

[0104] According to one embodiment of the present invention, -[p la,skin,x logp skin,x +(1-p la,skin,x )log(1-p skin,x )] is the cross entropy loss function composed of the probability of normal skin temperature determined by the annotation information of the xth historical moment and the skin temperature judgment result (the result in the form of probability, which is the probability of normal skin temperature output by the model). Similarly, -[p la,me,x logp me,x +(1-p la,me,x )log(1-p me,x )] is the cross entropy loss function composed of the probability of normal drug temperature determined by the annotation information at the xth historical moment and the drug temperature judgment result, -[p la,pr,x logp pr,x +(1-p la,pr,x )log(1-p pr,x )] is the cross entropy loss function composed of the probability of normal pressure determined by the annotation information of the x-th historical moment and the pressure judgment result. The loss function of the graph neural network model at the x-th historical moment can be obtained by weighted summation of the three cross entropy loss functions (for example, w1=0.3, w2=0.3, w3=0.4). Furthermore, the loss function of the graph neural network model at each moment can be summed to obtain the state loss function. During training, there is no need to wait for the state prediction model to complete the processing of the historical state feature information of multiple moments and obtain the prediction data before training. Instead, the data of each historical moment can be used to train the graph neural network model, thereby improving the training intensity and frequency of the graph neural network model, which helps to improve the accuracy of the graph neural network model.

[0105] In this way, the training of the graph neural network model can be assisted by introducing fully connected layers and activation layers, so that the data of each historical moment can be used to train the graph neural network model without waiting for the state prediction model to complete processing the historical moment state feature information of multiple moments, thereby improving the training intensity of the graph neural network model. During the period when the state prediction model processes the historical moment state feature information of multiple historical moments, the graph neural network model can be trained using data from multiple historical moments, which is conducive to the state prediction model obtaining more accurate historical moment state feature information and making more accurate predictions, thereby improving the overall training efficiency of the graph neural network model and the state prediction model.

[0106] According to one embodiment of the present invention, the historical feature information of the first node, the historical feature information of the second node, and the historical features of the third node are spliced ​​together to obtain the state feature information of the historical moment. The state feature information of the historical moment corresponding to multiple historical moments is processed by the state prediction model to obtain the first training temperature data, second training temperature data, and training pressure data of the subsequent historical moments of the multiple historical moments. For example, the historical moment state feature information of the yth historical moment and multiple historical moments before it is input into the state prediction model to obtain the training hidden state information of the yth historical moment, and then processed by the fourth fully connected layer and the first activation layer to obtain the first training temperature data, second training temperature data, and training pressure data of the y+1th historical moment. The first training temperature data, second training temperature data, and training pressure data of the y+1th historical moment can then be compared with the first historical temperature data, second historical temperature data, and historical pressure data of the y+1th historical moment. Based on the comparison results of multiple historical moments, a prediction loss function can be obtained.

[0107] According to one embodiment of the present invention, a prediction loss function is determined based on the first training temperature data, the second training temperature data, and the training pressure data, as well as the first historical temperature data of the skin, the second historical temperature data of the drug, and the historical pressure data in the space at subsequent historical moments of the plurality of historical moments, including: determining the prediction loss function L according to formula (2): P ,

[0108]

[0109] Among them, T T,1,y is the first training temperature data at the yth historical moment, T H,1,y is the first historical temperature data at the yth historical moment, T T,2,y is the second training temperature data at the yth historical moment, T H,2,y is the second historical temperature data at the yth historical moment, P T,y is the training pressure data at the yth historical moment, PH,y is the historical pressure data at the yth historical moment, w4, w5 and w6 are preset weights, n1 is the number of input ports of the state prediction model, N is the number of historical moments, N-n1+1≤y≤N, and y, n1 and N are all positive integers.

[0110] According to one embodiment of the present invention, |T T,1,y -T H,1,y | is the difference between the first training temperature data and the first historical temperature data at the yth historical moment, |T T,2,y -T H,2,y | is the difference between the second training temperature data and the second historical temperature data at the yth historical moment, |P T,y -P H,y | is the difference between the training pressure data and the historical pressure data at the yth historical moment. Taking the weighted sum of the three differences above yields the loss function for the yth historical moment (for example, w4 = 0.3, w5 = 0.3, and w6 = 0.4). Summing the loss functions across multiple historical moments yields the prediction loss function. Because the output data of the graph neural network model and the state prediction model are used to determine the prediction loss function, the graph neural network model and the state prediction model can be trained as a whole during backpropagation.

[0111] In this way, the prediction loss function can be obtained through the gap between the model output data and the actual historical data at each historical moment. Therefore, during the training process, the graph neural network model and the state prediction model can be trained as a whole through backpropagation of the prediction loss function. In the process of narrowing the gap between the model output data and the actual historical data, the overall error of the graph neural network model and the state prediction model is reduced, thereby improving the overall performance of the graph neural network model and the state prediction model and improving the prediction accuracy.

[0112] According to one embodiment of the present invention, a graph neural network model and a state prediction model are trained based on a state loss function and a prediction loss function. That is, the graph neural network model is trained separately by backpropagating the state loss function, and the graph neural network model and the state prediction model are trained as a whole by backpropagating the prediction loss function. After training using data from multiple historical moments and ensuring that the overall accuracy of the graph neural network model and the state prediction model meets the requirements, the training is completed, resulting in a trained graph neural network model and a trained state prediction model.

[0113] According to the embodiment of the present invention, the precise temperature-controlled negative pressure adsorption acupuncture transdermal delivery method not only promotes the dispersion and absorption of drugs through the electric heating device, but also promotes local skin blood circulation through the infrared heating device, thereby improving the absorption efficiency of drugs. Moreover, by monitoring the skin temperature, drug temperature, and pressure at multiple moments, the on / off state of the infrared heating device, the electric heating device, and the negative pressure device at a subsequent moment can be determined, thereby achieving accurate control of the skin temperature, drug temperature, and pressure, and reducing the probability of problems such as skin burns, drug failure, and inaccurate pressure control causing a decrease in drug transdermal rate. When obtaining the momentary state feature information, the features of each node can be aggregated through a graph neural network model. During the aggregation process, the connection weights between nodes are determined using a hierarchical attention mechanism to accurately reflect the mutual influence between each node, thereby obtaining the momentary state feature information. This not only accurately describes the on / off state of the infrared heating device, the on / off state of the electric heating device, the on / off state of the negative pressure device, the skin temperature, the drug temperature, and the pressure in the space at that moment, but also describes the mutual influence of these data, thereby improving the accuracy of the momentary state feature information and the comprehensiveness of the feature description, providing an accurate data foundation for subsequent prediction processing. When predicting data at multiple future moments, a sliding time window model can be used to perform backward predictions based on data from multiple recent moments, discarding data from a longer period of time. This discards data that has little impact on current and subsequent predictions or may even have a negative impact on accuracy, thereby improving prediction accuracy. Furthermore, the switch state at the i-th moment can be kept unchanged during the prediction process to predict data at multiple future moments, thereby providing an accurate data basis for determining whether the switch state should be changed. When determining the state loss function, the training of the graph neural network model can be assisted by introducing fully connected layers and activation layers. This allows the graph neural network model to be trained using data from each historical moment without waiting for the state prediction model to complete processing the state feature information of multiple historical moments. This improves the training strength of the graph neural network model. While the state prediction model is processing the state feature information of multiple historical moments, the graph neural network model can be trained using data from multiple historical moments. This helps the state prediction model obtain more accurate state feature information of historical moments and make more accurate predictions, thereby improving the overall training efficiency of the graph neural network model and the state prediction model. When determining the prediction loss function, the prediction loss function can be obtained through the gap between the model output data and the actual historical data at each historical moment. Therefore, during the training process, the graph neural network model and the state prediction model can be trained as a whole through backpropagation of the prediction loss function. In the process of narrowing the gap between the model output data and the actual historical data, the overall error of the graph neural network model and the state prediction model is reduced, thereby improving the overall performance of the graph neural network model and the state prediction model and improving the prediction accuracy.

[0114] Figure 3 A block diagram of a precise temperature-controlled negative pressure adsorption acupuncture point transdermal delivery device according to an embodiment of the present invention is exemplarily shown, wherein the device comprises:

[0115] a start module for starting the negative pressure device to form a negative pressure in the space between the negative pressure suction cup and the skin, wherein medicine is placed in the space and an infrared heating device and an electric heating device are placed in the negative pressure suction cup;

[0116] A heating module is used to heat the skin through an infrared heating device and heat the medicine through an electric heating device;

[0117] a detection module, configured to detect first temperature data of the skin, second temperature data of the medicine, and pressure data in the space;

[0118] a state vector module, configured to combine the first temperature data and the first switching state of the infrared heating device into a first state vector, combine the second temperature data and the second switching state of the electric heating device into a second state vector, and combine the pressure data and the third switching state of the negative pressure device into a third state vector;

[0119] a prediction module, configured to obtain first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple future moments based on the first state vector, the second state vector, and the third state vector at the i-th moment and multiple moments before the i-th moment;

[0120] a switch state module, configured to determine a first predicted switch state of the infrared heating device, a second predicted switch state of the electric heating device, and a third predicted switch state of the negative pressure device at the (i+1)th moment based on the first predicted temperature data, the second predicted temperature data, and the predicted pressure data;

[0121] The setting module is used to perform setting according to the first predicted switch state, the second predicted switch state and the third predicted switch state at the (i+1)th moment.

[0122] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0123] Those skilled in the art will appreciate that the embodiments of the present invention described above and shown in the accompanying drawings are intended to be illustrative only and are not intended to limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functional and structural principles of the present invention have been demonstrated and illustrated in the embodiments. Any variations or modifications may be made to the embodiments of the present invention without departing from the principles described.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction method, characterized in that: include: The negative pressure device is activated to form a negative pressure in the space between the negative pressure suction cup and the skin, wherein medicine is placed in the space and an infrared heating device and an electric heating device are placed in the negative pressure suction cup; The skin is heated by an infrared heating device, and the medicine is heated by an electric heating device; detecting first temperature data of the skin, second temperature data of the medicine, and pressure data in the space; The first temperature data and the first switching state of the infrared heating device form a first state vector, the second temperature data and the second switching state of the electric heating device form a second state vector, and the pressure data and the third switching state of the negative pressure device form a third state vector; Obtaining first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple future moments based on the first state vector, the second state vector, and the third state vector at the i-th moment and multiple moments before the i-th moment; Determine, based on the first predicted temperature data, the second predicted temperature data, and the predicted pressure data, a first predicted switch state of the infrared heating device, a second predicted switch state of the electric heating device, and a third predicted switch state of the negative pressure device at the (i+1)th moment; At the (i+1)th moment, setting is performed according to the first predicted switch state, the second predicted switch state, and the third predicted switch state.

2. The precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction method according to claim 1, characterized in that: Obtaining first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple future moments based on the first state vector, the second state vector, and the third state vector at the i-th moment and multiple moments before the i-th moment, including: Input the first state vector, the second state vector, and the third state vector at the same time into the first fully connected layer, the second fully connected layer, and the third fully connected layer, respectively, to obtain a first input vector corresponding to the first state vector, a second input vector corresponding to the second state vector, and a third input vector corresponding to the third state vector; Obtaining a first graph structure according to the first input vector, the second input vector, and the third input vector; The first graph structure is processed using the trained graph neural network model to obtain momentary state feature information; The trained state prediction model is used to process the moment state feature information corresponding to the i-th moment and multiple moments before it to obtain the first predicted temperature data of the skin at multiple moments in the future, the second predicted temperature data of the medicine, and the predicted pressure data in the space.

3. The precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction method according to claim 2, characterized in that: The first graph structure is processed using the trained graph neural network model to obtain momentary state feature information, including: Obtaining an adjacency matrix of the first graph structure; Processing the first input vector, the second input vector, and the third input vector through the attention mechanism layer of the trained graph neural network model to obtain a connection weight between the first input vector and the second input vector, a connection weight between the second input vector and the third input vector, and a connection weight between the first input vector and the third input vector; Processing the connection weight between the first input vector and the second input vector, the connection weight between the second input vector and the third input vector, the connection weight between the first input vector and the third input vector, the adjacency matrix, the first input vector, the second input vector, and the third input vector through the trained graph neural network model to obtain first node feature information corresponding to the first input vector, second node feature information corresponding to the second input vector, and third node feature information corresponding to the third input vector; The first node characteristic information, the second node characteristic information and the third node characteristic information are concatenated to obtain the momentary state characteristic information.

4. The precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction method according to claim 2, characterized in that: The trained state prediction model is used to process the state feature information corresponding to the i-th moment and multiple moments before the i-th moment to obtain first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple moments in the future, including: Input the moment state feature information at the i-n1+jth moment and the hidden state information at the i-n1+j-1th moment into the j-th input port of the trained state prediction model to obtain the hidden state information at the i-n1+jth moment, where n1 is the number of input ports of the state prediction model, 1≤j≤n1, and i, j, and n1 are all positive integers. When j=1, the hidden state information at the i-n1+j-1th moment is a zero vector; Iterate the steps of obtaining hidden state information until the hidden state information at the i-th moment is obtained; Input the hidden state information at the i-th moment into the fourth fully connected layer and the first activation layer to obtain the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the i+1-th moment; Maintain the first switch state, the second switch state, and the third switch state at the i-th moment, and obtain the first predicted input vector, the second predicted input vector, and the third predicted input vector at the i+k-th moment based on the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the i+k-th moment, where 1≤k≤n2, and n2 is the number of moments in the future; Obtaining a first prediction graph structure at the i+kth moment according to the first prediction input vector, the second prediction input vector, and the third prediction input vector at the i+kth moment; The first prediction graph structure at the i+kth moment is processed through the trained graph neural network model to obtain the predicted moment state feature information at the i+kth moment; The trained state prediction model is used to process the state feature information from the i-n1+k+1th moment to the i-th moment and the predicted state feature information from the i+1th moment to the i+kth moment to obtain the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the i+k+1th moment; The process of acquiring the first predicted temperature data, the second predicted temperature data, and the predicted pressure data is iteratively performed until the first predicted temperature data, the second predicted temperature data, and the predicted pressure data at the (i+n2)th moment are acquired.

5. The precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction method according to claim 2, characterized in that: The method further comprises: Obtaining a first historical state vector, a second historical state vector, and a third historical state vector at multiple historical moments; The first historical state vector, the second historical state vector, and the third historical state vector at the same historical moment are processed by a graph neural network model to obtain the historical feature information of the first node, the historical feature information of the second node, and the historical feature information of the third node; Fully connect and activate the historical feature information of the first node, the historical feature information of the second node, and the historical feature information of the third node respectively to obtain skin temperature judgment results, drug temperature judgment results, and pressure judgment results; Determine the state loss function based on the skin temperature judgment results, the drug temperature judgment results, the pressure judgment results, and the annotation information of the historical moments; splicing the historical feature information of the first node, the historical feature information of the second node, and the historical feature information of the third node to obtain historical moment state feature information; Processing historical moment state feature information corresponding to multiple historical moments through a state prediction model to obtain first training temperature data, second training temperature data, and training pressure data for subsequent historical moments of the multiple historical moments; Determining a prediction loss function based on the first training temperature data, the second training temperature data, and the training pressure data, as well as the first historical temperature data of the skin, the second historical temperature data of the drug, and the historical pressure data in the space at subsequent historical moments of the plurality of historical moments; According to the state loss function and the prediction loss function, the graph neural network model and the state prediction model are trained to obtain a trained graph neural network model and a trained state prediction model.

6. The precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction method according to claim 5, characterized in that: Based on the skin temperature judgment results, drug temperature judgment results, and pressure judgment results, as well as the annotation information of historical moments, the state loss function is determined, including: According to the formula Determine the state loss function L S , where p la,skin,x is the probability that the skin temperature is normal according to the labeled information at the xth historical moment, p skin,x is the skin temperature judgment result at the xth historical moment, p la,me,x is the probability that the drug temperature is normal according to the labeling information at the xth historical moment, p me,x is the drug temperature judgment result at the xth historical moment, p la,pr,x is the probability that the pressure is normal according to the annotation information at the xth historical moment, p pr,x is the pressure judgment result at the xth historical moment, w1, w2 and w3 are preset weights, N is the number of historical moments, x≤N, and both x and N are positive integers.

7. The precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction method according to claim 5, characterized in that: Determining a prediction loss function based on the first training temperature data, the second training temperature data, and the training pressure data, as well as the first historical temperature data of the skin, the second historical temperature data of the drug, and the historical pressure data in the space at subsequent historical moments of the plurality of historical moments, includes: According to the formula Determine the prediction loss function L P , where T T,1,y is the first training temperature data at the yth historical moment, T H,1,y is the first historical temperature data at the yth historical moment, T T,2,y is the second training temperature data at the yth historical moment, T H,2,y is the second historical temperature data at the yth historical moment, P T,y is the training pressure data at the yth historical moment, P H,y is the historical pressure data at the yth historical moment, w4, w5 and w6 are preset weights, n1 is the number of input ports of the state prediction model, N is the number of historical moments, N-n1+1≤y≤N, and y, n1 and N are all positive integers.

8. A precise temperature-controlled negative pressure adsorption acupuncture transdermal introduction device, characterized in that: include: a start module for starting the negative pressure device to form a negative pressure in the space between the negative pressure suction cup and the skin, wherein medicine is placed in the space and an infrared heating device and an electric heating device are placed in the negative pressure suction cup; A heating module is used to heat the skin through an infrared heating device and heat the medicine through an electric heating device; a detection module, configured to detect first temperature data of the skin, second temperature data of the medicine, and pressure data in the space; a state vector module, configured to combine the first temperature data and the first switching state of the infrared heating device into a first state vector, combine the second temperature data and the second switching state of the electric heating device into a second state vector, and combine the pressure data and the third switching state of the negative pressure device into a third state vector; a prediction module, configured to obtain first predicted skin temperature data, second predicted drug temperature data, and predicted pressure data in the space at multiple future moments based on the first state vector, the second state vector, and the third state vector at the i-th moment and multiple moments before the i-th moment; a switch state module, configured to determine a first predicted switch state of the infrared heating device, a second predicted switch state of the electric heating device, and a third predicted switch state of the negative pressure device at the (i+1)th moment based on the first predicted temperature data, the second predicted temperature data, and the predicted pressure data; The setting module is used to perform setting according to the first predicted switch state, the second predicted switch state and the third predicted switch state at the (i+1)th moment.

Images