Pension community medical resource scheduling method and system based on graph embedding algorithm

Abstract

The invention discloses a method and a system for scheduling medical resources of a pension community based on a graph embedding algorithm. The method comprises the steps that S1, a heterogeneous medical resource relation graph is constructed, and the relation graph comprises nodes and edges; s2, mapping nodes in the heterogeneous graph into low-dimensional vector representation by adopting a multi-modal graph embedding algorithm, and retaining structural similarity and semantic relevance among the nodes; s3, calculating a resource matching degree based on the embedded vector, and searching an optimal medical care-equipment-medicine combination in the k-hop in real time for the first-aid request; s4, fusing the time sequence diagram convolutional network, and dynamically updating an embedded vector to respond to a resource state change and an emergency demand event; and S5, outputting a resource scheduling scheme which comprises a medical care dispatching path, an equipment start-stop instruction and a medicine allocation sequence. According to the invention, management of the aged population can be realized, reasonable use or distribution of medical resources is facilitated, health monitoring of the aged population is facilitated, the response speed of emergency rescue is increased, and the resource utilization rate is optimized.

Description

TECHNICAL FIELD

[0001] The application relates to a pension community medical resource scheduling method and system based on a graph embedding algorithm. BACKGROUND

[0002] With the acceleration of population aging, more problems of health and medical resources shortage of the aged population are brought; in order to facilitate the care and health management of the aged population, some areas have established pension institutions such as nursing homes and elderly activity centers; however, affected by many factors, many medical resources cannot be reasonably utilized, and many uneven phenomena occur, causing medical resource shortage in some areas, especially the aged population with relatively difficult action, which is difficult to care for or involve, which causes certain social problems; at the same time, the health status of the aged is also lack of care and supervision, so a pension community medical resource scheduling method and system based on a graph embedding algorithm need to be proposed. SUMMARY

[0003] The technical problem to be solved by the application is to provide a pension community medical resource scheduling method and system based on a graph embedding algorithm, which can realize the management of the aged population, is beneficial to the reasonable use or distribution of medical resources, and is beneficial to the health monitoring of the aged population, and has practicality and wide applicability.

[0004] To solve the above problems, the application adopts the following technical scheme: A pension community medical resource scheduling method based on a graph embedding algorithm, comprising the following steps: S1: constructing a heterogeneous medical resource relationship graph, the relationship graph comprising nodes and edges, wherein the nodes comprise: an old person node, a medical staff node, a medical device node, and a medicine node; The edges comprise: an old person-medical service relationship, a medical staff-device use relationship, and a device-medicine association relationship; S2: using a multi-modal graph embedding algorithm to map the nodes in the heterogeneous graph into low-dimensional vector representations, and retaining the structural similarity and semantic correlation between the nodes; S3: calculating the resource matching degree based on the embedding vectors, and for an emergency request, real-time searching for the optimal medical staff-device-medicine combination within k-hop; S4: fusing a time sequence graph convolution network to dynamically update the embedding vectors to respond to resource state changes and sudden demand events; S5: outputting a resource scheduling scheme, including a medical staff dispatch path, a device start-stop instruction, and a medicine allocation sequence.

[0005] Preferably, the multi-modal graph embedding algorithm of step S2 comprises: a. generating a heterogeneous walk sequence of old people-medical staff-devices through meta-path random walk; b. The attention mechanism is used to fuse the node attribute features and topological structure features; c. The loss function is: ; wherein, is the loss, used to measure the learning effect of the model, is the loss term based on the feature vector , difference of the node pair , is the weight coefficient, represents the reconstruction operation, is the adjacency matrix, is the node feature matrix, is the reconstruction loss term, used to constrain the error of feature reconstruction.

[0006] Preferably, the old person node in step S1 further comprises a social relationship subgraph, which is constructed by the following methods: Analyzing public activity area monitoring video to generate an old person social heat map; Constructing mutual assistance relationship edges based on smart bracelet close-range sensing data; When the old person has a sudden illness, the responsibility caregiver of the neighbor old person in the social relationship graph of the old person is preferentially dispatched.

[0007] Preferably, the resource matching degree calculation in step S3 adopts a bilinear function: ; wherein, is the patient node, is the device node, is the medicine node, is the learnable parameter matrix, represents vector splicing; is the Sigmoid activation function, , , is the corresponding input feature vector.

[0008] Preferably, the time sequence diagram convolution network in step S4 comprises: a spatial convolution layer: based on node embedding and adjacency matrix to aggregate neighbor information; a time convolution layer: using one-dimensional convolution to capture historical state change patterns; an emergency event injection module: when a fall alarm is detected, forcibly reset the related node embedding gradient.

[0009] An elderly care community medical resource scheduling system based on a graph embedding algorithm, comprising: IoT perception layer: deployed in the elderly community, including old people wearable devices, medical positioning badges, device status sensors, medicine RFID tags and environmental monitoring cameras; Graph computing engine layer: contains heterogeneous resource graph construction module, dynamic graph embedding training module, resource matching degree calculation module and time series graph convolution update module; Decision execution layer: including emergency dispatch terminal, medicine robot console, rehabilitation equipment management platform and nursing staff mobile APP; Feedback optimization layer: real-time collection of dispatch effect data and return to the streaming learning module of the graph computing engine layer.

[0010] Preferably, the hardware implementation of the graph computing engine layer includes: Deployed in each building access layer, the edge computing node performs local graph embedding calculation; The central graph database server stores the full community heterogeneous resource relationship graph; GPU accelerated cluster running time series graph convolution network training task.

[0011] Preferably, the heterogeneous resource graph construction module is configured to: Parse the perception layer data to generate four types of nodes, specifically: old people nodes, medical nodes, device nodes and medicine nodes; Establish three types of weighted edges, specifically: service relationship edges, operation relationship edges and dependency relationship edges.

[0012] Preferably, the resource matching degree calculation module performs the following operations: When receiving an emergency request, activate the subgraph with a radius of 50 meters with the patient node as the center; Call the embedding vector to calculate the adaptation score of the medical device medicine combination: filter the combination with a score greater than the threshold θ to generate a candidate dispatch scheme; , Score is the calculated score, Sigmoid activation function, 、 is a learnable weight matrix, 、 、 is a different feature vector, is a feature fusion operation.

[0013] Preferably, the time series graph convolution update module includes: a spatial convolution unit that aggregates neighbor node information based on an adjacency matrix; a time convolution unit that processes historical embedding sequences using causal convolution; an event response unit that injects a gradient reset signal when detecting an emergency call; Preferably, the interaction logic of the decision execution layer comprises: displaying a resource heat map based on embedding vectors, supporting drag adjustment of priority of an emergency dispatch terminal; receiving an optimal path embedding sequence, and automatically avoiding a first-aid channel to push a band social relationship prompt of a mutual aid task of a nursing staff APP of a medicine robot console.

[0014] Preferably, the feedback optimization layer realizes deviation analysis of an actual response time and a predicted value; dynamically adjusts graph embedding model parameters by comparing a loss function; and automatically triggers graph structure optimization when a device idle rate is greater than 15%.

[0015] The beneficial effects of the present application are: 1. Revolutionary speed-up of emergency response: through graph embedding space retrieval, resource matching time consumption is reduced from an average of 90 seconds to 0.3 seconds; emergency response enters the 3-minute era, breaking the traditional system response bottleneck of 8 minutes, and the median time is measured to be 2.8 minutes, so that the rescue success rate of cardiac arrest is increased to 42%.

[0016] 2. Optimization of resource utilization rate: the device idle rate is reduced to 11% (the industry is first below 15%), and the resource idle cost is reduced by more than 60%.

[0017] 3. Privacy compliance breakthrough: through edge computing + differential privacy + topology confusion, zero privacy leakage is realized in third-party penetration testing, and health data is "available but invisible" (third-party authentication complies with GDPR / CCPA).

[0018] 4. Personalized design for the elderly: the task misoperation rate of the nursing staff APP is reduced by 41%, the misoperation rate of the nursing staff APP is reduced to 9%, and the user satisfaction is 92%.

[0019] 5. Through the statistics of regional medical resources, the medical resources can be allocated on demand; at the same time, the medical resources can be adjusted in a timely and reasonable manner according to the actual situation, and full coverage of medical resources is realized.

[0020] 6. At the same time, the distance between part of the personnel and the medical system is shortened, and the operation of rapid medical treatment is realized.

[0021] 7. At the same time, through the collection of information of the elderly population, the health monitoring of the elderly population is facilitated, which is conducive to the health and medical management of the elderly population, and solves part of the problems brought by population aging.

[0022] 8. At the same time, through the use of smart bracelets, watches and the like, the safety of the elderly population is improved, and the elderly population can also understand their own conditions. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below, but are not a limitation on the present application.

[0024] Figure 1 The step flowchart of the present application. DETAILED DESCRIPTION

[0025] Referring to Figure 1 The method for scheduling medical resources in a pension community based on a graph embedding algorithm includes the following steps: S1: Constructing a heterogeneous medical resource relation graph, the relation graph including nodes and edges, The nodes include: old person nodes, medical staff nodes, medical equipment nodes, and medicine nodes. The edges include: old person-medical service relations, medical staff-equipment use relations, and equipment-medicine association relations. S2: Using a multi-modal graph embedding algorithm to map the nodes in the heterogeneous graph into low-dimensional vector representations, and preserving the structural similarity and semantic correlation between the nodes. S3: Calculating the resource matching degree based on the embedding vectors, and for an emergency request, real-time searching for the optimal medical staff-equipment-medicine combination within k-hop. S4: Fusing a time sequence graph convolution network to dynamically update the embedding vectors to respond to resource state changes and sudden demand events. S5: Outputting a resource scheduling scheme, including medical staff dispatch paths, equipment start-stop instructions, and medicine allocation sequences.

[0026] Further, the multi-modal graph embedding algorithm of step S2 includes: a. Generating a heterogeneous walk sequence of old person-medical staff-equipment through meta-path random walk; b. Using an attention mechanism to fuse node attribute features and topological structure features; c. The loss function is: ; Wherein, is the loss, used to measure the learning effect of the model, is the loss item based on the difference between the feature vectors of the node pairs, is the weight coefficient, represents the reconstruction operation, is the adjacency matrix, is the node feature matrix, is the reconstruction loss item, used to constrain the error of feature reconstruction.

[0027] ​Further, the old man node in step S1 further comprises a social relationship subgraph, which is constructed by the following way: Analyzing public activity area monitoring video to generate old man social heat map; Based on the close-range sensing data of the smart bracelet, the mutual assistance relationship edge is constructed; When the old man has an emergency, the neighbor's responsible caregiver in the social relationship graph is preferentially dispatched.

[0028] Further, the resource matching degree calculation of step S3 adopts a bilinear function: ; Wherein, is the patient node, is the device node, is the drug node, is the learnable parameter matrix, indicating the vector splicing; is the Sigmoid activation function, , , is the feature vector corresponding to the input.

[0029] Further, the time sequence diagram convolution network of step S4 comprises: Spatial convolution layer: aggregate neighbor information based on node embedding and adjacency matrix; Time convolution layer: use one-dimensional convolution to capture historical state change patterns; Emergency event injection module: when a fall alarm is detected, forcibly reset the related node embedding gradient.

[0030] Further, step S4 also includes a streaming graph learning engine, which continuously performs: update the device state node attribute every 5 minutes; Trigger local graph re-embedding when the medical position moves more than 20 meters; When the drug is consumed, automatically reduce the inventory edge weight and recalculate the embedding.

[0031] Further, the resource scheduling scheme generation comprises: Divide the old-age community into dynamic micro-service areas using embedding vector clustering; Based on the path embedding similarity, the drug delivery robot plans a route to avoid the emergency channel; Use graph matching algorithm to assign alternative device nodes for rehabilitation equipment appointment conflicts.

[0032] Further, the output interface of step S5 comprises: The medical terminal displays the resource heat map based on the embedding vector; The device console receives the start-stop probability instructions of the graph neural network; The family APP pushes the service progress graph with topological relationship visualization.

[0033] A kind of old-age community medical resource scheduling system based on graph embedding algorithm, comprising: Internet of Things perception layer, graph computing engine layer, decision-making execution layer and feedback optimization layer; Internet of Things perception layer: it is deployed in old-age community, including old person wearable device, medical positioning badge, equipment state sensor, drug RFID label and environmental monitoring camera; Graph computing engine layer: it includes heterogeneous resource graph construction module, dynamic graph embedding training module, resource matching degree calculation module and time series graph convolution update module; Decision-making execution layer: including first-aid dispatch terminal, drug robot console, rehabilitation equipment management platform and nursing staff mobile APP; Feedback optimization layer: real-time collection scheduling effect data and return to the flow learning module of graph computing engine layer.

[0034] Further, the hardware implementation of graph computing engine layer includes: Edge computing node: it is deployed in each building access layer, and performs local graph embedding calculation; Central graph database server: store the heterogeneous resource relationship graph of whole community; GPU acceleration cluster: run time series graph convolution network training task.

[0035] Further, the heterogeneous resource graph construction module is configured to: parse perception layer data to generate four types of nodes: Old person node: attribute = {location coordinates, health score, social relationship vector}; Medical node: attribute = {skill vector, real-time position, workload}; Device node: attribute = {type code, busy state, physical location}; Drug node: attribute = {drug ID, inventory, temperature control requirement}; Three types of weighted edges are established: Service relationship edge: connect old person and medical node, weight = service frequency; Operation relationship edge: connect medical and device nodes, weight = operation proficiency; Dependency relationship edge: connect device and drug node, weight = matching coefficient.

[0036] Further, the dynamic graph embedding training module includes: Meta-path random walker: traverse heterogeneous graph according to preset path mode; Attribute-structure fusion encoder: adopt multi-head attention mechanism to aggregate node features; Negative sampling optimizer: generate 128-dimensional embedding vector based on Skip-gram model.

[0037] Further, the resource matching degree calculation module performs the following operations: When receiving an emergency request, activate a subgraph with a radius of 50 meters centered on the patient node; Call the embedding vector to calculate the adaptation score of the medical-care-equipment-drug combination: , Score is the calculated score, is a Sigmoid activation function, 、 is a learnable weight matrix, 、 、 is a different feature vector, is a feature fusion operation, and combinations with a score greater than a threshold θ are generated as candidate scheduling schemes.

[0038] Further, the timing diagram convolution update module includes: Spatial convolution unit: aggregate neighbor node information based on adjacency matrix; Temporal convolution unit: process historical embedding sequences using causal convolution; Event response unit: inject gradient reset signal when detecting emergency call.

[0039] Further, the interaction logic of the decision execution layer includes: Emergency dispatch terminal: display resource heat map based on embedding vector, support drag adjustment priority; Drug robot console: receive optimal path embedding sequence, automatically avoid emergency passage; Nursing staff APP: push mutual aid tasks with social relationship prompts.

[0040] Further, the feedback optimization layer realizes deviation analysis of actual response time and predicted value; dynamically adjust graph embedding model parameters by comparing loss function; automatically trigger graph structure optimization when device idle rate > 15%.

[0041] Further, the interaction logic of the decision execution layer also includes an elastic scaling unit, which is used to: Turn off part of the edge computing nodes during the night low load period; When there are more than 3 consecutive emergency requests, automatically expand GPU computing resources; Dynamically expand graph database shards according to the growth of the number of old people.

[0042] Further, the security and privacy module integrates the following parts: Graph topology obfuscator: differential privacy processing on social relationship edges; Embedded trusted execution environment: encrypt health data on edge device side; Permission Isolation Gateway: Limit access roles of device control instructions.

[0043] Data desensitization for privacy security system: Add Laplace noise to old health score with differential privacy (ε=0.1) Instruction authentication: Device control instructions require medical biometrics + dynamic token two-factor authentication.

[0044] In the implementation process; Old people's wearable devices, such as old people's smart bracelets: sampling heart rate variability, UWB positioning accuracy 0.3 meters, at the same time, the old people's smart bracelet is equipped with a three-axis accelerometer (sampling rate ≥100Hz) and a PPG optical sensor, realizing fall detection (sensitivity 98.2%) and heart rate abnormality warning (specificity 93.5%).

[0045] Drug RFID tag: Temperature sensing error ±0.5℃, inventory automatic update delay <1s Spatial positioning network: Medical staff's chest card is equipped with a UWB module (accuracy 0.3 meters), which calculates the position of personnel in real time through TOF algorithm and transmits it to the edge node in encrypted form.

[0046] Device state perception chain: Medical devices are equipped with current transformers to identify working status through load current characteristics (such as "ECG machine running current >1.2A" to determine busy state).

[0047] The dynamic graph embedding training module: uses attribute-enhanced meta-path walk algorithm: generates walk sequence along the preset meta-path "old people → medical staff → device", and for each node in the sequence, fuses its attribute features and structure features, optimizes Skip-gram model through negative sampling, and outputs 128-dimensional embedding vector.

[0048] The decision execution layer executes the man-machine cooperative interface: specifically, First aid dispatch terminal: uses heat map visualization resource distribution, red highlight area represents first aid event location, green channel is the optimal navigation path. Supports nursing supervisor to drag and adjust priority, and the system checks in real time whether it meets the safety constraints.

[0049] Drug robot console: generates avoidance rules based on path embedding similarity: when detecting that the first aid medical staff's moving vector and the robot's path vector have an included angle <30°, triggers dynamic detour algorithm to ensure zero obstruction of first aid channel.

[0050] Fall event perception: bracelet accelerometer detects 3.5g impact force (threshold 2.7g), PPG monitors heart rate to 45bpm, triggers SOS signal.

[0051] First aid dispatch scenario example one: Event: Old man falls at a certain location, health score drops sharply; 1. Activate the 50-meter subgraph centered on the patient's coordinates; 2. Inject gradient reset signal into T-GCN; 3. Retrieve optimal combination: medical staff, defibrillator, coagulation factor; 4. Issue instructions: medical staff terminal receives navigation path: avoid robot path - defibrillator automatically unlocks, voice prompt "electrode pads in left compartment" - unmanned aerial vehicle takes off from pharmacy with cold chain box, path planning 128 meters; 5. On-site treatment: medical staff arrives at the scene, unmanned aerial vehicle arrives at the balcony at the same time, medicine retrieval time difference < 5 seconds; 6. Feedback optimization: analyze the deviation between actual response time and predicted value, increase the weight coefficient of "elevator waiting time" by 0.15.

[0052] The test effect of the scheme of the present application is as follows:

[0053] Example two of the emergency dispatching scenario: Event: Old man falls at a certain building, health score drops to 0.2; 1. System response: trigger T-GCN gradient reset, update corresponding building subgraph embedding; Then retrieve the optimal combination within a 50-meter radius, including: Medical staff: nurse; Equipment: defibrillator; Medicine: hemostatic gel; 2. And generate instructions: Nurse terminal pushes navigation path; Defibrillator automatically unlocks and preheats; Unmanned aerial vehicle delivers medicine to the balcony of a certain building.

[0054] The test effect of the scheme of the present application is as follows:

[0055] In the implementation process of the dispatching system, through the operation of the system, the medical resource dispatching system for the elderly community based on the graph embedding algorithm of the present application can also generate the following methods: Method one: For example, a medical resource optimization dispatching method includes the following steps: S1: Collect resource data of medical institutions in real time, including medical staff scheduling information, medical equipment status, bed occupancy rate, drug inventory, and emergency waiting queue; S2: Based on historical data and real-time demand prediction model, generate resource demand prediction results in future preset period; S3: Construct a multi-objective optimization model to minimize patient waiting time and maximize resource utilization, with constraints including resource availability limits and medical safety rules; S4: Solve the multi-objective optimization model using a heuristic algorithm to generate a dynamic resource scheduling plan; S5: Visualize the scheduling plan and output it to a terminal device, supporting manual adjustment and real-time updates; S6: After the scheduling plan is executed, collect actual execution deviation data through a feedback mechanism, and dynamically correct the prediction model parameters and optimization algorithm weights.

[0056] Further, the demand prediction model in step S2 uses a fusion model of time series analysis (ARIMA) and machine learning model (LSTM neural network), with input variables including seasonal factors, epidemic trend data, and sudden public health event warning data.

[0057] Further, the objective function of the multi-objective optimization model in step S3 is represented as: , where, is the average waiting time of patients, is the resource idle cost coefficient, is the weight parameter.

[0058] Further, the data collection in step S1 obtains medical device status data through Internet of Things sensors, and synchronizes the location and task status of medical staff through the hospital information system (HIS) interface.

[0059] Further, the heuristic algorithm in step S4 is an improved genetic algorithm, which includes: chromosome coding represents resource allocation combination; adaptive function is designed based on objective function value; simulated annealing mechanism is introduced to avoid local optimal solution.

[0060] Further, the resource scheduling plan includes: flexible scheduling table of emergency department medical staff; cross-department sharing plan of operating room and detection equipment; bed dynamic allocation strategy based on patient criticality level.

[0061] Further, in step S5, the visualization interface supports drag-and-drop manual adjustment, and the system automatically checks whether the adjusted plan meets the medical safety constraints.

[0062] Method two: An elderly community medical resource optimization scheduling method, comprising: S1: Real-time collection of multi-modal data in the elderly community, including: physiological indicators of wearable devices for the elderly, room emergency call status, location and skill level of nursing staff, inventory of medicine delivery robots, and usage status of rehabilitation equipment; S2: Based on the health portrait of the elderly, the health risk assessment model is used to classify the elderly into emergency, high-risk chronic disease, and routine care levels, and to predict the resource demand density of each category within the next 24 hours; S3: Construct a spatio-temporal constraint optimization model to minimize emergency response time and balance nursing workload as the core objective, with constraints including: maximum service radius of nursing staff, timeliness of drug distribution, and conflict rules for rehabilitation equipment sharing; S4: Use a dynamic task allocation algorithm to insert emergency requests into the scheduling queue in real time and re-optimize the non-urgent task sequence; S5: Generate visual scheduling instructions and push them to nursing staff mobile terminals, drug robot control systems, and rehabilitation equipment reservation management platforms; S6: Establish a resource scheduling effectiveness evaluation system, including: average emergency response time compression rate, chronic disease medication on-time delivery rate, rehabilitation equipment idle rate reduction, and scheduling scheme acceptance score based on nursing staff feedback.

[0063] Further, the health risk assessment model in step S2 integrates the following data: historical emergency trigger frequency, continuous monitoring of blood pressure / blood glucose abnormalities by wearable devices, fall risk level recorded in electronic medical records, and medication adherence score.

[0064] Further, the objective function of the optimization model in step S3 is: ; Where is the response delay time of the i-th emergency request, is the workload variance of the k-th nursing staff, 、 is the priority weight.

[0065] Further, the dynamic task allocation algorithm in step S4 uses a spatio-temporal conflict resolution mechanism, specifically: reserving a GPS track channel for emergency tasks for mobile nursing units; drug delivery robots avoid emergency mobile units in path planning; and the same floor's rehabilitation equipment usage time is forced to be separated by disinfection time.

[0066] Further, the data collection in step S1 includes: continuously monitoring heart rate variability (HRV) through smart wristbands for the elderly, tracking the real-time location of nursing staff using indoor UWB positioning systems, and automatically recording drug access through RFID medicine cabinets.

[0067] Further, the resource scheduling scheme specifically includes: dynamic formation of high-risk elderly exclusive nursing teams across buildings; unmanned aerial vehicle emergency supply paths for temperature-controlled drugs such as insulin; and anti-wandering monitoring schedules based on the activity patterns of cognitively impaired elderly.

[0068] Further, the visualization interface integration of step S5 is suitable for old design: specifically: The caregiver terminal displays a task urgency color block mark with large font; The medicine robot control interface supports voice instruction modification of destination; The family terminal APP synchronously pushes a delay warning of non-emergency service.

[0069] In the implementation process of the present application, part of the information needs to be collected and summarized, and it can also be executed in multiple schemes, as follows: a. Set the medical resource system with the county or district as the center, and establish the health archives of the old people in the region with the street or town as the sub-center, and count the old-age institutions in units of villages or communities, so as to realize that the medical system covers all old-age institutions; b. Take the street or town as a unit, count the resident old-age population status and the flow population status in the unit, and count and summarize according to gender and age group; count the number and structure of the old-age population of the old-age institutions in each unit; c. Count the medical resource status of each unit, including personnel, medicines, medical equipment, etc., calculate the medical resources and the old-age population in the unit area, and obtain the corresponding proportion value; d. Collect information such as diet, rest time, etc. of the old-age population, and optimize the best diet, rest, demand scheme; e. Establish an emergency medical channel and set up a medical and drug emergency center to realize fast commuting between two regions; f. Establish cooperation with medical institutions and introduce new treatment schemes in a timely manner; g. Through training and introduction, the service personnel such as doctors, nurses and social workers reach a certain proportion or quantity; h. Promote and popularize auxiliary medical or health and rehabilitation facilities; I. Optimize and promote the old-age insurance system.

[0070] In the implementation process, the collection of old-age information includes: I. Health status assessment: Health status assessment is the basis of old-age demand assessment. The assessment team needs to have a comprehensive understanding of the physical health status of the elderly, including but not limited to: 1. Chronic disease status: such as hypertension, diabetes, heart disease, etc.: 2. Cognitive status: such as Alzheimer's disease, vascular dementia, etc.: 3. Nutritional status: such as body weight, eating habits, etc.: 4. Motor ability: such as walking, going up and down stairs, etc.

[0071] II. Assessment of Daily Life Abilities: The assessment of daily life abilities focuses on whether the elderly can take care of themselves or need assistance in their daily lives. The assessment includes: 1. Ability to take care of daily life: such as bathing, dressing, eating, etc. 2. Ability to use tools: such as using a mobile phone, TV, washing machine, etc. 3. Ability to travel: such as taking transportation, identifying directions, etc.

[0072] III. Assessment of Mental State: The assessment of mental state focuses on the psychological condition of the elderly, including: 1. Emotional state: such as depression, anxiety, etc. 2. Social skills: such as communication and interaction with family and friends. 3. Interests and hobbies: such as reading, music, painting, etc.

[0073] IV. Assessment of Social Needs: The assessment of social needs aims to understand the social needs of the elderly, including: 1. Whether living with family members; 2. Participation in social activities; 3. Needs and preferences for social activities.

[0074] V. Assessment of Living Environment: The assessment of living environment focuses on the adaptation of the elderly to their living environment, including: 1. Whether the residential facilities meet the requirements of aging; 2. Whether the living environment is safe and comfortable; 3. Whether the living environment needs to be improved.

[0075] VI. Assessment of Technology Assistance Needs: The assessment of technology assistance needs mainly understands the needs of the elderly in technology products, including: 1. Whether to use smart phones, tablets and other technology products; 2. What technology products are needed to improve the quality of life; 3. Acceptance and use barriers of technology products.

[0076] VII. Assessment of Long-term Care Needs: The assessment of long-term care needs focuses on the care needs of the elderly, including: 1. Whether to need long-term care; 2. What types of care services are needed; 3. Quality and satisfaction of care services.

[0077] In the implementation process: for example, the configuration of the ambulance, take the county medical institutions as an example, suppose that the county medical institutions are configured with 10 ambulances, in addition to the part of the vehicle left for emergency, the rest of the vehicles are scattered in the region as the center, shorten the distance between the vehicle and the elderly, realize fast medical treatment.

[0078] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any change or replacement without creative labor should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be limited to the protection scope defined in the claims.

Claims

1. A method for scheduling medical resources in a retirement community based on a graph embedding algorithm, characterized in that: Comprising the steps of: S1: Constructing a heterogeneous medical resource relationship graph, the relationship graph containing nodes and edges, wherein the nodes include: elderly nodes, medical staff nodes, medical equipment nodes, and drug nodes; The edges include: elderly-medical service relationships, medical staff-equipment use relationships, and equipment-drug association relationships; S2: Using a multi-modal graph embedding algorithm to map the nodes in the heterogeneous graph to low-dimensional vector representations, preserving the structural similarity and semantic relevance between the nodes; S3: Calculating the resource matching degree based on the embedding vectors, for emergency requests, real-time retrieval of the optimal medical staff-equipment-drug combination within k-hop; S4: Fusing a time-series graph convolution network to dynamically update the embedding vectors in response to resource state changes and sudden demand events; S5: Outputting the resource scheduling scheme, including medical staff dispatch path, equipment start-stop instructions, and drug allocation sequence.

2. The method according to claim 1, characterized in that: The multi-modal graph embedding algorithm of step S2 includes: a. Generating a heterogeneous walk sequence of elderly-medical staff-equipment through meta-path random walk; b. Using an attention mechanism to fuse node attribute features and topological structure features; c. The loss function is: ； wherein, is a loss, used to measure the learning effect of the model, is a feature vector based on the node pair , , is a loss term based on the difference between the feature vectors, is a weight coefficient, represents a reconstruction operation, is an adjacency matrix, is a node feature matrix, is a reconstruction loss term, used to constrain the error of feature reconstruction.

3. The method of claim 1, wherein the method is based on a graph embedding algorithm. The elderly nodes in step S1 further include a social relationship subgraph, which is constructed by: Analyzing public activity area monitoring video to generate an elderly social heat map; Constructing mutual assistance relationship edges based on smart bracelet close-range sensing data; When an elderly person has a sudden illness, the responsible caregivers of their neighbors in the social relationship graph are preferentially dispatched.

4. The method of claim 1, wherein the method is based on a graph embedding algorithm. The resource matching degree calculation of step S3 adopts a bilinear function: ; wherein, is a patient node, is a device node, is a drug node, is a learnable parameter matrix, denotes vector concatenation; is a sigmoid activation function, , , is a feature vector corresponding to the input.

5. The method of claim 1, wherein the method is based on a graph embedding algorithm. The time-series graph convolution network of step S4 includes: Spatial convolution layer: aggregating neighbor information based on node embedding and adjacency matrix; Time convolution layer: using one-dimensional convolution to capture historical state change patterns; Emergency event injection module: when a fall alarm is detected, forcibly reset the related node embedding gradient.

6. A graph embedding algorithm-based system for performing the method of any one of claims 1-5, wherein, Comprising: Internet of Things perception layer: deployed in the elderly care community, including elderly wearable devices, medical staff positioning badges, equipment status sensors, drug RFID tags, and environmental monitoring cameras; Graph computing engine layer: including heterogeneous resource graph construction module, dynamic graph embedding training module, resource matching degree calculation module, and time-series graph convolution update module; Decision execution layer: including emergency dispatch terminal, drug robot console, rehabilitation equipment management platform, and nurse mobile APP; Feedback optimization layer: real-time collection of scheduling effect data and return to the streaming learning module of the graph computing engine layer.

7. The system according to claim 6, wherein the system is characterized in that: The hardware implementation of the graph computing engine layer includes: Edge computing nodes deployed in each building access layer to perform local graph embedding calculation; Central graph database server storing the full community heterogeneous resource relationship graph; GPU-accelerated cluster running time-series graph convolution network training tasks.

8. The system according to claim 6, characterized in that: The heterogeneous resource graph construction module is configured to: Parse the perception layer data to generate four types of nodes, specifically: elderly nodes, medical staff nodes, equipment nodes, and drug nodes; Establish three types of weighted edges, specifically: service relationship edges, operation relationship edges, and dependency relationship edges.

9. The system according to claim 6, wherein the system is characterized in that: The resource matching degree calculation module performs the following operations: When an emergency request is received, activate a subgraph with a radius of 50 meters centered on the patient node; Call the embedding vector to calculate the adaptation score of the medical staff-equipment-drug combination: select combinations with a score greater than the threshold θ to generate candidate scheduling schemes; ， Score is the computed score, is a Sigmoid activation function, , is a learnable weight matrix, , are different feature vectors, is a feature fusion operation.

10. The system according to claim 6, characterized in that: The timing diagram convolution update module comprises a spatial convolution unit for aggregating neighbor node information based on an adjacency matrix; a time convolution unit for processing a historical embedding sequence using a causal convolution; and an event response unit for injecting a gradient reset signal when an emergency call is detected; The interaction logic of the decision execution layer comprises: an emergency dispatch terminal for displaying a resource heat map based on embedding vectors and supporting drag adjustment of priorities; a nursing staff APP for receiving an optimal path embedding sequence and automatically avoiding an emergency channel to push a mutual aid task with a social relationship prompt; The feedback optimization layer realizes deviation analysis of an actual response time and a predicted value; dynamically adjusts graph embedding model parameters by comparing a loss function; and automatically triggers graph structure optimization when a device idle rate is greater than 15%.

Images