A health care service intelligent scheduling method and system based on multi-factor fusion

Abstract

A health care service intelligent scheduling method and system based on multi-factor fusion, in the method, the health records and audio-visual data of users, the state data of nursing staff and the environmental data are collected, the audio-visual data is converted into vectors through a semantic model and is fused with other data. The user demand representation is obtained by analyzing the fused data by using a time sequence feature extraction model, a heterogeneous interaction graph containing multiple types of nodes is constructed, and the matching probability is calculated by a graph attention network. At the same time, a neural ordinary differential equation model is used to evaluate the service risk, and finally the target nursing staff is determined and the scheduling instruction is generated by combining the matching probability and the risk level. The present application is used to improve the accuracy and response timeliness of health care service resource allocation in complex dynamic scenarios, and reduce the potential risks in the service process.

Description

Technical Field

[0001] This application belongs to the field of task coordination and execution, and in particular relates to an intelligent scheduling method and system for health and wellness services based on multi-factor fusion. Background Technology

[0002] To cope with massive service requests and allocate nursing resources rationally, traditional scheduling models often rely on manual experience for order dispatch or simply assign resources based on a single rule such as geographical proximity. When faced with large-scale concurrent orders, it is often difficult to make optimal decisions in a short period of time, resulting in long service response cycles and uneven distribution of nursing resources.

[0003] Related technologies typically employ an automated scheduling method based on tag matching and historical statistical data. This method uses a pre-established service tag system to match users' service needs (such as daily living care and rehabilitation nursing) with the skills and expertise of caregivers. Combined with map positioning technology, it uses historical travel data to estimate the time it takes for caregivers to reach the service location, thereby achieving automatic matching of service supply and demand, shortening order response time, and improving the basic accuracy of service matching.

[0004] However, this scheduling method, which relies on static labels and statistical data, still has limitations when dealing with complex and dynamic scenarios. For example, when an elderly person with chronic cardiovascular disease initiates "routine care" in severe weather with a sudden drop in temperature, the aforementioned technologies often only match based on the static label "routine care" and physical distance. In this case, the system may assign a general caregiver who is nearby but lacks emergency first aid experience, and this caregiver may arrive later than expected due to inaccurate road condition estimations. This lack of consideration for real-time status and environmental variables reduces the accuracy of matching service personnel skills with the user's actual condition in high-risk scenarios, and also increases the potential health risks to the user while waiting for service. Summary of the Invention

[0005] This application provides a method and system for intelligent scheduling of health and wellness services based on multi-factor fusion, which can improve the accuracy and timeliness of health and wellness service resource allocation in complex and dynamic scenarios and reduce potential risks in the service process.

[0006] In the first aspect, this application provides a smart scheduling method for health and wellness services based on multi-factor fusion, which collects multimodal data of users who initiate service requests, service status data of candidate caregivers, and environmental data of the service area. The multimodal data includes structured health record data and unstructured audio and video communication data.

[0007] The audio and video communication data is transformed into semantic embedding vectors using a pre-defined semantic model. The semantic embedding vectors are then normalized and concatenated with health record data, service status data, and environmental data to generate a multi-factor fusion input vector.

[0008] The multi-factor fusion input vector is input into the time-series feature extraction model, and the output is a user demand representation vector.

[0009] Construct a heterogeneous interaction graph containing user nodes, caregiver nodes, and service station nodes, and map the user demand representation vector to the user node features in the heterogeneous interaction graph;

[0010] Calculate the attention coefficients between user nodes and caregiver nodes in a heterogeneous interaction graph using graph attention networks;

[0011] Output the matching probability matrix between users and each candidate caregiver based on the attention coefficient;

[0012] The multi-factor fusion input vector is input into the neural ordinary differential equation model to output the service risk level;

[0013] Candidate caregivers with a matching probability higher than a preset threshold are selected based on the matching probability matrix;

[0014] Target caregivers are identified from the pool of candidate caregivers based on service risk levels in order to generate dispatch instructions.

[0015] By employing the aforementioned technical solutions, multimodal data is collected and semantically transformed and normalized for concatenation, resulting in a comprehensive fusion vector encompassing user health status, caregiver status, and environmental factors. Analyzing this fusion vector using a temporal feature extraction model allows for more accurate capture of the dynamic changes in user needs. Based on this, a heterogeneous interaction graph is constructed, combined with a graph attention network to calculate attention coefficients, improving the matching accuracy between users and caregivers. A neural network-based differential equation model is used to predict and assess service risks, considering both risk levels and matching probabilities. This ensures that the final scheduling decision considers both service quality and safety, improving the intelligent scheduling efficiency and service safety of elderly care services.

[0016] In conjunction with some implementation methods of the first aspect, in some implementation methods, the multi-factor fusion input vector is input into the time-series feature extraction model to output a user demand representation vector, specifically including:

[0017] A temporal feature extraction model containing a bidirectional long short-term memory network layer and a self-attention mechanism layer was constructed. The multi-factor fusion input vector sequence within a preset historical time period was used as the training set input, and the corresponding historical manual evaluation demand category was used as the training set label. The temporal feature extraction model was trained under supervision until convergence.

[0018] The multi-factor fusion input vector generated at the current moment is input into the trained bidirectional long short-term memory network layer to extract the hidden state sequence containing the time dimension;

[0019] The weight coefficients in the hidden state sequence are calculated using a self-attention mechanism layer;

[0020] The hidden state sequence is weighted and summed based on the weight coefficients to output a user demand representation vector.

[0021] By adopting the above technical solution, a temporal feature extraction model is constructed using a bidirectional long short-term memory network layer and a self-attention mechanism layer. Through supervised training on historical data, the model can learn the temporal evolution patterns of user needs. The hidden state sequences extracted by the bidirectional long short-term memory network layer contain forward and backward temporal dependencies. The self-attention mechanism layer differentiates the weights of features at different time steps by calculating weight coefficients, thereby improving the model's ability to model long-term dependencies of user needs and the accuracy of identifying key temporal features, thus enhancing the temporal representation of user needs.

[0022] In conjunction with some implementation methods of the first aspect, in some implementation methods, the multi-factor fusion input vector is input into the neural ordinary differential equation model to output the service risk level, specifically including:

[0023] A neural ordinary differential equation model containing a parameterized derivative function network and an ordinary differential equation solver is constructed. Historical service data with sudden health event markers are selected as the training set, and historical service data not used in training are selected as the test set. The model parameters are optimized by minimizing the error loss function between the predicted risk and the actual event.

[0024] The multi-factor fusion input vector at the current moment is used as the initial state variable input to the ordinary differential equation solver;

[0025] The derivative relationship of state variables over time is defined using a parameterized derivative function network, and the ordinary differential equation solver is controlled to integrate from the current time to the preset service end time to obtain the state variables at the predicted time.

[0026] The state variables at the predicted time are input into a fully connected classification layer, and the service risk level is output by mapping through the Softmax function.

[0027] By employing the above technical solution, a neural ordinary differential equation model is constructed, incorporating a parameterized derivative function network and an ordinary differential equation solver. This model is trained using historical service data tagged with sudden health events, enabling it to learn the state evolution patterns during the service process. Using the current state as initial conditions, future state changes are predicted through continuous integration, and the risk level is obtained via a fully connected classification layer, thus improving the accuracy of predicting potential risks during the service process.

[0028] In conjunction with some implementations of the first aspect, in some implementations, target caregivers are identified from candidate caregivers based on service risk levels to generate dispatch instructions, specifically including:

[0029] If the service risk level is determined to exceed the preset high-risk level, obtain the emergency care skills certificate level for each candidate caregiver;

[0030] Calculate the weighted sum of the matching probability and the first aid skills certificate level;

[0031] The candidate caregiver with the largest weighted sum will be selected as the target caregiver.

[0032] Obtain the current geographic coordinates of the target caregiver and the service address coordinates of the user;

[0033] Determine the emergency access route based on the minimum time cost to generate dispatch instructions that include the target caregiver's identity ID, emergency access route, estimated arrival time, and emergency plan;

[0034] If the service risk level is determined to be no higher than the preset high risk level, the candidate caregiver with the highest probability value in the matching probability matrix is ​​selected as the target caregiver.

[0035] Obtain the current geographic coordinates of the target caregiver and the service address coordinates of the user;

[0036] The system determines the regular travel route based on the shortest spatial distance to generate a scheduling instruction that includes the target caregiver's ID, the regular travel route, and the standard service process.

[0037] By adopting the above technical solution, different caregiver selection strategies are used based on the level of service risk. In high-risk situations, the weighted average of emergency skills certificate level and matching probability is considered, while in normal situations, the caregiver with the highest matching probability is directly selected. For different risk levels, the system determines the travel route using the minimum time cost and shortest spatial distance, and generates dispatch instructions containing corresponding contingency plans or standard procedures. This improves the service response speed in emergency situations while ensuring the economy and convenience of routine services, enhancing the adaptability of the elderly care service dispatch system to different service scenarios.

[0038] In some embodiments of the first aspect, after identifying target caregivers from candidate caregivers based on service risk levels to generate scheduling instructions, the method further includes:

[0039] Obtain the current performance progress data of the target caregiver and calculate the expected arrival time based on the performance progress data;

[0040] If the estimated arrival time is later than the planned arrival time in the scheduling instruction, then the difference between the estimated arrival time and the planned arrival time is calculated.

[0041] The expected arrival time is input into the neural ordinary differential equation model as a new time observation variable. Taking advantage of the continuous evolution characteristics of the neural ordinary differential equation model, the user state corresponding to the planned arrival time is used as the initial value, and the solution is obtained by integration to the expected arrival time to obtain the delayed state vector.

[0042] Calculate the degree of state difference between the delayed state vector and the initial state vector on which the scheduling command is based;

[0043] If the state difference exceeds the preset service tolerance threshold, service correction parameters are generated.

[0044] Update the service operation sequence in the scheduling instruction using service correction parameters.

[0045] By adopting the above technical solution, the expected arrival time is calculated by acquiring the target caregiver's performance progress data and compared with the planned arrival time. In the event of a delay, the continuous evolution characteristics of the neural network differential equation model are used to predict the delay state. Combined with the state difference degree, service correction parameters are generated to update the scheduling instructions. This allows the system to dynamically perceive time deviations during service execution and quantify the risk of service quality impairment by modeling the impact of delays on user states. This dynamic adjustment mechanism based on state evolution improves the system's adaptability to delays, reduces the negative impact of delays by adjusting the service operation sequence in a timely manner, and considers the different sensitivities of different state dimensions to delays, making service corrections more targeted, thereby improving the system's service reliability in actual execution environments.

[0046] In conjunction with some implementation methods of the first aspect, in some implementation methods, the estimated arrival time is calculated based on the performance progress data, specifically including:

[0047] Extract the target caregiver's current geographical coordinates and the remaining service time of the currently performing task from the performance progress data;

[0048] Calculate the travel time for the target caregiver to reach the user service address based on the geographical coordinates and the user service address recorded in the dispatch instructions;

[0049] The current system time, remaining service time, and travel time are added together to calculate the estimated arrival time of the target caregiver at the user's service address.

[0050] By employing the aforementioned technical solution, and comprehensively considering three key factors—the target caregiver's current geographical location, remaining service time, and travel time—to calculate the estimated arrival time, the system can more accurately assess the caregiver's actual available time. This multi-dimensional time prediction method considers the actual constraints faced by caregivers in completing their current task and traveling to the next service location, thus improving the accuracy of arrival time prediction. By combining geospatial information with task time information, the system reduces errors caused by relying solely on distance or task duration predictions, enabling a more reasonable assessment of service time feasibility and thereby improving the overall time accuracy of the scheduling plan.

[0051] In conjunction with some implementations of the first aspect, in some implementations, the service operation sequence in the scheduling instruction is updated using service correction parameters, specifically including:

[0052] Extract the target state dimension identifier from the service correction parameters that causes the state difference to exceed the preset service tolerance threshold;

[0053] Traverse the service operation sequence in the scheduling instructions, and based on the preset state-service mapping relationship, identify the specific service sub-item that matches the target state dimension identifier;

[0054] The execution order of a specific service sub-item in the service operation sequence is adjusted to the first position to obtain the updated service operation sequence.

[0055] By adopting the above technical solution, and by identifying the target dimensions causing state differences and adjusting the service operation sequence using a preset state-service mapping relationship, dynamic adjustment of service priorities based on state sensitivity is achieved. This service sequence rearrangement method reduces the impact of state differences on service quality by executing service sub-items that are more affected by latency earlier. The automatic identification mechanism based on the mapping relationship improves the system's response efficiency to different types of state changes, enhances the system's service compensation effect when dealing with latency, and strengthens the overall service continuity and stability.

[0056] Secondly, embodiments of this application provide a smart scheduling system for health and wellness services based on multi-factor fusion. This smart scheduling system for health and wellness services based on multi-factor fusion includes: one or more processors and a memory; the memory is coupled to one or more processors, and the memory is used to store computer program code, which includes computer instructions. One or more processors call the computer instructions to cause the system to execute the method described in the first aspect and any possible implementation thereof.

[0057] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a system, cause the system to perform the method described in the first aspect and any possible implementation thereof.

[0058] Fourthly, embodiments of this application provide a computer program product that, when run on a system, causes the system to execute the method described in any possible implementation of the first aspect.

[0059] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0060] 1. This application provides an intelligent scheduling method for elderly care services based on multi-factor fusion. By collecting multimodal data and performing semantic transformation and normalization concatenation, a comprehensive fusion vector is formed, encompassing user health status, caregiver status, and environmental factors. Analyzing the fusion vector using a temporal feature extraction model can more accurately capture the dynamic changes in user needs. Based on this, a heterogeneous interaction graph is constructed, combined with a graph attention network to calculate the attention coefficient, improving the matching accuracy between users and caregivers. Service risks are predicted and assessed using a neural network-based differential equation model, and the risk level is considered in conjunction with the matching probability. This ensures that the final scheduling decision considers both service quality and safety, improving the intelligent scheduling efficiency and service safety of elderly care services.

[0061] 2. This application provides a multi-factor fusion-based intelligent scheduling method for elderly care services. It calculates the expected arrival time by acquiring the target caregiver's performance progress data and compares it with the planned arrival time. Upon detecting delays, it uses the continuous evolution characteristics of a neural network differential equation model to predict the delay state. Combined with state difference, it generates service correction parameters to update scheduling instructions. This allows the system to dynamically perceive time deviations during service execution and quantifies the risk of service quality impairment by modeling the impact of delays on user states. This dynamic adjustment mechanism based on state evolution improves the system's adaptability to delays, reduces the negative impact of delays by adjusting service operation sequences in a timely manner, and considers the different sensitivities of different state dimensions to delays, making service corrections more targeted, thereby improving the system's service reliability in actual execution environments. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating an intelligent scheduling method for health and wellness services based on multi-factor fusion, as described in an embodiment of this application.

[0063] Figure 2 This is another flowchart illustrating an intelligent scheduling method for health and wellness services based on multi-factor fusion, as described in this application.

[0064] Figure 3 This is a schematic diagram of the physical device structure of an intelligent scheduling system for health and wellness services based on multi-factor fusion, provided in an embodiment of this application. Detailed Implementation

[0065] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0066] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0067] The following example is used in conjunction with Figure 1This application describes an intelligent scheduling method for elderly care services based on multi-factor fusion in its embodiments:

[0068] Please see Figure 1 This is a flowchart illustrating an intelligent scheduling method for health and wellness services based on multi-factor fusion, as described in an embodiment of this application.

[0069] S101. Collect multimodal data of the user who initiated the service request, service status data of the candidate caregiver, and environmental data of the service area;

[0070] The system collects multimodal data from users initiating service requests, service status data from candidate caregivers, and environmental data from the service area. Multimodal data includes structured health record data and unstructured audio-visual communication data. Multimodal data refers to heterogeneous data sets originating from different sensors and data sources with different data formats, specifically including but not limited to structured health record data and unstructured audio-visual communication data. Structured data refers to data that can be logically expressed using a two-dimensional table structure, such as numerical data like blood pressure, heart rate history records, and past medical history codes; unstructured data refers to data that cannot be expressed using a unified database field format, such as voice streams during user calls and body movement images captured by video surveillance. Service status data refers to the current operational status of candidate caregivers, including whether they are available, their current location coordinates, and the type of task they are performing. Environmental data refers to the objective physical conditions within the service area, such as weather conditions, traffic congestion index, temperature, and humidity. The system obtains the above information in real-time or periodically from user terminals, medical databases, caregiver handheld terminals, and public environmental monitoring platforms through preset data interfaces and sensor networks. The system performs preliminary cleaning and verification of the collected data, removing invalid or erroneous data packets to ensure the accuracy of subsequent processing. The data collection trigger mechanism can be set to activate when a user initiates a request, or automatically when the system detects abnormal indicators. The collection process must strictly adhere to data security transmission protocols to ensure user privacy is not compromised. For the synchronization of multimodal data, the system employs timestamp alignment technology to ensure consistency between audio / video data and physiological parameters in the time dimension, thereby providing an accurate spatiotemporal reference for subsequent multi-factor fusion.

[0071] To implement this step, at least two methods can be used. The first method is a combined data acquisition technology based on IoT sensor networks and RESTful API interfaces. Wearable medical devices (such as smart bracelets and blood pressure monitors) are deployed at the user end, transmitting physiological parameters to the gateway via Bluetooth or ZigBee. The gateway then uploads the data to a cloud server via MQTT. Simultaneously, a mobile app uses the phone's microphone and camera to collect audio and video data, which is uploaded to an object storage server via HTTPS. For environmental data and caregiver status, the system performs real-time queries by calling third-party map service APIs and the database interface of the internal personnel management system. The second method is a distributed data acquisition technology based on edge computing nodes. Edge computing boxes are deployed at the gateway in the health and wellness community or home. These devices directly connect to various sensors and monitoring equipment, performing preliminary frame extraction and noise reduction on the raw audio and video streams locally, and formatting and encapsulating the health data. The edge nodes push the processed multimodal data packets to the central processing system via message queues (such as Kafka). This method effectively reduces network bandwidth pressure and improves the real-time performance of data acquisition, making it particularly suitable for high-concurrency service request scenarios.

[0072] S102. Use a preset semantic model to convert audio and video communication data into semantic embedding vectors, and normalize and splice the semantic embedding vectors with health record data, service status data and environmental data to generate a multi-factor fusion input vector.

[0073] A pre-trained semantic model refers to a deep learning model that has been pre-trained on a large-scale corpus and has the ability to map natural language or audio signals to a high-dimensional vector space. Semantic embedding vectors are the numerical representation of audio-visual data in a high-dimensional space, preserving the semantic information and emotional features of the original data. Normalization refers to the process of mapping data with different dimensions and value ranges (such as age, blood pressure, and distance) to the same numerical interval (such as [0,1] or [-1,1]) to eliminate the influence of numerical differences on model weights. Concatenation refers to connecting the processed vectors along the feature dimension to form a long vector containing all information. The system first extracts features from unstructured audio-visual data, transforming it into a dense vector of fixed length; simultaneously, it performs numerical encoding and normalization on structured health records, service status, and environmental data. Then, the system merges these feature vectors in a predetermined order to construct a multi-factor fusion input vector that comprehensively represents the current service scenario. This vector not only contains the user's explicit needs (expressed through voice), but also implies the user's physical condition and external environmental constraints, providing comprehensive data support for subsequent intelligent scheduling.

[0074] The specific technical approaches to implementing this step include at least the following two. The first approach combines a pre-trained model based on the Transformer architecture with Min-Max normalization. For text and audio data, pre-trained models such as BERT or Wav2Vec are used to extract context-dependent semantic feature vectors; for numerical data, the min-max normalization formula is used to scale it to the [0,1] interval; for categorical data (such as caregiver levels), One-Hot encoding is used. Finally, tensor concatenation operations in deep learning frameworks are used to merge all feature tensors along the channel dimension. The second approach combines convolutional neural network (CNN) feature extraction with Z-Score normalization. For video frame data, 3D-CNN is used to extract spatiotemporal feature vectors; for audio data, it is converted to a Mel spectrogram and then features are extracted using 2D-CNN; for continuous numerical variables, their mean and standard deviation are calculated and Z-Score normalization is performed to make them conform to a standard normal distribution. Fully connected layers are used to map features of different dimensions to the same dimensional space, and then feature fusion is performed. This approach can better handle dynamic features in video stream data.

[0075] S103. Input the multi-factor fusion input vector into the time series feature extraction model and output the user demand representation vector.

[0076] The system inputs a multi-factor fusion input vector into a temporal feature extraction model and outputs a user demand representation vector. Specifically, the system constructs a temporal feature extraction model containing a bidirectional long short-term memory network layer and a self-attention mechanism layer. It uses a sequence of multi-factor fusion input vectors within a preset historical time period as the training set input and the corresponding historical manual assessment demand categories as the training set labels. The temporal feature extraction model is then trained under supervision until convergence. The multi-factor fusion input vector generated at the current moment is input into the trained bidirectional long short-term memory network layer to extract a hidden state sequence containing the time dimension. The self-attention mechanism layer is used to calculate the weight coefficients in the hidden state sequence. Finally, the hidden state sequence is weighted and summed based on the weight coefficients to output the user demand representation vector.

[0077] The temporal feature extraction model is a deep neural network architecture specifically designed to process data sequences with temporal dependencies. The bidirectional Long Short-Term Memory (LSTM) network layer, composed of forward and backward LSTM networks, can simultaneously capture past and future information in the sequence data (in offline training or real-time processing with latency), thus providing a more comprehensive understanding of the temporal context. The self-attention mechanism layer is a mechanism that calculates the degree of correlation between different positions in the sequence; it is independent of temporal distance and can directly capture long-distance dependencies. The hidden state sequence refers to the set of internal states output by the LSTM layer after processing the input at each time step, containing a summary of information from that time and previous times. The weight coefficients are numerical values ​​calculated by the attention mechanism, representing the importance of the hidden state at each time step in the sequence to the final result. The user demand representation vector is a fixed-length vector obtained by weighted summation, which highly condenses the user's service demand features.

[0078] The specific implementation methods include at least the following two approaches. The first approach is to build a hierarchical deep learning network based on the TensorFlow or PyTorch framework. First, a Bi-LSTM layer is defined, setting the number of hidden units and receiving an input tensor of shape (time step, feature dimension). The output of the Bi-LSTM layer is a sequence containing the concatenation of the forward and backward hidden states at each time step. This is followed by a custom self-attention layer, which maps the Bi-LSTM output to three matrices (Query, Key, and Value) through a fully connected network. Attention scores (weight coefficients) are calculated using dot product operations and the Softmax function. These weight coefficients are then multiplied by the Value matrix and summed to obtain the context vector. Finally, a fully connected layer outputs the final user demand representation vector. The second approach utilizes a variant of a gated recurrent unit (GRU) and a multi-head attention mechanism. While the core remains a bidirectional recurrent network and attention, the implementation can use a GRU (Gated Recurrent Unit) instead of an LSTM to improve computational efficiency, and a scaled dot product attention algorithm is employed in the attention layer. During the training phase, the cross-entropy loss function is used to measure the difference between the predicted demand category and the true label, and the Adam optimizer is used for backpropagation to update the network parameters until the loss value converges. This approach can accelerate the model's inference speed while maintaining performance.

[0079] S104. Construct a heterogeneous interaction graph containing user nodes, caregiver nodes, and service station nodes, and map the user demand representation vector to the user node features in the heterogeneous interaction graph.

[0080] The system constructs a heterogeneous interaction graph containing user nodes, caregiver nodes, and service station nodes, and maps user demand representation vectors to user node features in the heterogeneous interaction graph. A heterogeneous interaction graph is a graph data structure characterized by containing multiple types of nodes (users, caregivers, and stations) and multiple types of edges (such as potential service relationships between users and caregivers, and membership relationships between caregivers and stations). Node features refer to the attribute vectors assigned to each node, describing its specific state and attributes. Mapping refers to the process of assigning the previously calculated vectors to the corresponding node attributes in the graph. The system first establishes the graph's topology based on entity information within the current service area. User nodes represent elderly people initiating requests, caregiver nodes represent all candidate service personnel, and service station nodes represent the management station or emergency center to which the caregiver belongs. Edges are connected based on geographical proximity, historical service records, or membership relationships. After the system is built, it directly assigns the user demand representation vector output in step S103 to the corresponding user node as its initial feature; for caregiver nodes, it converts their service status, skill level and other data into feature vectors and assigns them; for site nodes, it converts their resource reserves into feature vectors.

[0081] The specific technical approaches to implementing this step include at least the following two methods. The first method combines a graph database (such as Neo4j) with a graph computing framework (such as DGL or PyG). First, node labels and relationship types are defined in the graph database. Entity data is loaded into memory using query statements, and an adjacency matrix or edge list is constructed to represent the graph structure. Then, using the API provided by a graph neural network library, a heterogeneous graph object is created, and the user demand representation vector is used as tensor data, bound to the User node feature attributes of the graph object in the form of key-value pairs. The second method is an algebraic construction method based on sparse matrices. The system does not rely on an external graph database but directly constructs multiple sparse matrices in memory to represent the connection relationships between different types of nodes (such as user-caregiver adjacency matrices, caregiver-site adjacency matrices). Simultaneously, a feature matrix is ​​constructed, where each row corresponds to a node and each column corresponds to a feature dimension. The user demand representation vector is filled into the row corresponding to the user index in the feature matrix. This method avoids complex database operations and is suitable for quickly constructing large-scale heterogeneous graphs in memory-constrained computing environments, facilitating direct input into deep learning models for matrix operations.

[0082] S105. Calculate the attention coefficients between user nodes and caregiver nodes in a heterogeneous interaction graph using graph attention networks.

[0083] The system utilizes a graph attention network (GAT) to calculate the attention coefficients between user nodes and caregiver nodes in a heterogeneous interaction graph. A GAT is a graph neural network-based architecture that introduces an attention mechanism to process graph-structured data. The attention coefficient is a scalar value that quantifies the degree of mutual influence between two connected nodes in the graph; in this scenario, it specifically refers to the importance or matching degree of a candidate caregiver node in meeting the needs of a particular user node. The system inputs the constructed heterogeneous interaction graph into the GAT model. In each graph convolutional operation, the model does not simply average the features of neighboring nodes, but first calculates the correlation score between the features of the central node (user) and neighboring nodes (caregivers). This process typically involves concatenating or transforming the feature vectors of two nodes, followed by a non-linear transformation using a learnable weight vector and an activation function (such as LeakyReLU). The calculated raw scores are then normalized using Softmax to transform them into probabilistic attention coefficients. These coefficients reflect the caregiver's suitability relative to other candidates within the current specific user needs and service environment. By stacking multiple layers of GAT, node information can be propagated and aggregated in the graph, so that the final calculated attention coefficient incorporates higher-order neighborhood information.

[0084] The specific technical approaches to implementing this step include at least the following two methods. The first method employs a single-head graph attention mechanism. In each layer of the graph neural network, a shared parameterized linear transformation matrix is ​​defined to map user node features and caregiver node features to the same feature space. Then, for each pair of connected user and caregiver nodes, their transformed feature vectors are concatenated and input into a single-layer feedforward neural network (attention mechanism function), outputting a real-valued score. This score is processed using the LeakyReLU activation function and normalized using Softmax across all neighboring nodes to obtain the final attention coefficients. The second method employs a multi-head attention mechanism. The system runs K independent attention calculation processes in parallel, each with its own transformation matrix and attention parameters, thereby capturing the correlation features between nodes in different subspaces (such as skill matching, distance, personality compatibility, etc.). Each "head" independently calculates a set of attention coefficients, and finally, the K sets of coefficients are averaged or concatenated to obtain the final attention coefficient output.

[0085] S106. Output the matching probability matrix between the user and each candidate caregiver based on the attention coefficient;

[0086] The system outputs a matching probability matrix between users and each candidate caregiver based on attention coefficients. This matching probability matrix is ​​a two-dimensional or one-dimensional data structure containing the matching probability value between the user and each candidate caregiver. The probability values ​​are typically between 0 and 1, and the sum of the probability values ​​for all candidate caregivers is not necessarily 1 (depending on the specific normalization strategy; in single-user, multi-candidate scenarios, the focus is usually on relative ranking or independent matching). In this step, the system directly uses the final layer attention coefficients calculated in S105, or performs further calculations based on the node embedding vectors aggregated from the attention coefficients. If the last layer of the GAT model directly outputs normalized attention coefficients, these coefficients themselves can serve as approximations of the matching probability. If the GAT outputs updated node feature vectors, the system calculates the similarity (such as cosine similarity or dot product) between the user node vector and each caregiver node vector, and maps the similarity score to a probability value using a Sigmoid or Softmax function. This matrix visually demonstrates the likelihood that each caregiver is suitable for the current task; a higher value indicates a higher matching degree, providing a quantitative basis for subsequent screening and scheduling decisions.

[0087] The specific technical approaches to implementing this step include at least the following two methods. The first method is to directly extract the Softmax output of the GAT layer. When designing the graph attention network, the output dimension of the last layer is set to 1, and the Softmax function is forced to normalize the weights of all caregiver edges connected to a specific user. After the network's forward propagation, the weight values ​​on these edges are directly read and arranged in order of caregiver ID to construct a probability matrix. This method is computationally simple and directly reuses the calculation results of the attention mechanism. The second method is based on dot product scoring of node embeddings. After updating the feature vectors of all nodes in the GAT layer, the final feature vectors of the user node and all candidate caregiver nodes are extracted. The dot product of the user vector and each caregiver vector is calculated to obtain an unnormalized matching score. These scores are input into a sigmoid activation function, independently mapping each score to a probability value between 0 and 1, forming a matching probability matrix.

[0088] S107. Input the multi-factor fusion input vector into the neural ordinary differential equation model and output the service risk level.

[0089] The system inputs a multi-factor fusion input vector into a neural ordinary differential equation (NDE) model and outputs a service risk level. Specifically, the system constructs an NDE model containing a parameterized derivative function network and an NDE solver. Historical service data with markers of sudden health events is selected as the training set, and historical service data not used in training is used as the test set. The model parameters are optimized by minimizing the error loss function between the predicted risk and the actual event. The current multi-factor fusion input vector is used as the initial state variable input to the NDE solver. The parameterized derivative function network defines the derivative relationship of the state variable over time, and the NDE solver is controlled to integrate from the current time to the preset service end time to obtain the state variable at the predicted time. The predicted state variable is input into a fully connected classification layer, and the service risk level is output through a Softmax function mapping.

[0090] The Ordinary Differential Equation (ODE) model is an architecture that combines deep neural networks with continuous mathematical modeling. It does not rely on a fixed number of network layers but treats changes in data state as a continuous evolution over time. The parameterized derivative function network is a neural network submodule used to simulate the rate of change of state; it learns the trend of data change at any given time point rather than the direct output result. The ODE solver is a numerical computation component responsible for extrapolating the state of data at future times based on the rate of change. Service risk level is a classification prediction of the severity of potential future health or safety hazards (such as falls, sudden heart attacks, etc.). The system first uses historical full-process data containing event markers to supervise the model's training, enabling it to grasp the dynamic laws governing the evolution of health state over time. In practical applications, the system uses the multi-factor fusion input vector of the current time as the initial state input to the solver. The solver uses the derivative relationship defined by the parameterized derivative function network to progressively integrate along the time axis from the current time until the preset service end time, thereby calculating the final state variable at the predicted time. Finally, the system inputs the final state variable into a fully connected classification layer, maps it to the probability distribution of different risk categories, and selects the category with the highest probability as the final service risk level.

[0091] The specific implementation methods include at least the following two approaches. The first approach uses a basic step-size iterative accumulation method as the solver implementation. The system constructs a multilayer perceptron network to fit the instantaneous rate of change of the state. During inference, the solver does not perform complex intermediate sampling, but instead uses a fixed, small time step, using the current state plus the product of the current rate of change and the time step to estimate the state at the next time step. This process is repeated until the accumulation reaches the predicted target time. Finally, the accumulated final state vector is input into a linear classifier for risk rating. The second approach uses a multi-order slope estimation method combined with adjoint state sensitivity analysis. The system uses a high-precision numerical solution algorithm to sample the slope at different positions multiple times within each time step to calculate the weighted average rate of change, thereby obtaining a more accurate state update. During the model training phase, to address the issue of excessive memory consumption, the forward integration process is not directly backpropagated; instead, an auxiliary adjoint state is constructed, and the gradient is calculated by inversely solving another ordinary differential equation.

[0092] S108. Select candidate caregivers whose matching probability is higher than a preset threshold based on the matching probability matrix;

[0093] The system filters candidate caregivers whose matching probability is higher than a preset threshold based on the matching probability matrix. In this step, the preset threshold is a numerical standard used to determine suitability, typically derived from statistical analysis of historical big data. It represents the minimum matching degree required to guarantee basic service quality and can also be dynamically adjusted by the administrator based on current capacity constraints. The filtering process is a screening operation for all candidates. The system reads the matching probability matrix generated in the previous step, which contains the compatibility values ​​between the current user and each candidate caregiver. The system compares each probability value in the matrix with the preset threshold one by one. Only when a caregiver's matching probability value reaches or exceeds the threshold is the caregiver considered a valid candidate and retained in the subsequent scheduling list; conversely, if the probability value is lower than the threshold, the caregiver is considered unsuitable and removed from the current scheduling task. This process aims to quickly eliminate obviously unsuitable service personnel, ensuring that subsequent decisions are made from a high-quality candidate set, thereby improving the success rate and satisfaction of scheduling.

[0094] The specific implementation of this step includes at least the following two methods. The first method is a conditional loop-based filtering technique. The system writes an iterative program that sequentially accesses each element in the matching probability data structure. In each iteration, the program performs logical judgments, checking whether the value of the current element meets the condition of being greater than or equal to a threshold. If it does, the corresponding caregiver identifier and probability value are copied to a new result list; otherwise, the element is skipped. After the iteration is complete, the new list is the set of qualified candidates. The second method is a parallel filtering technique based on a Boolean mask matrix. The system utilizes the advanced indexing function of the numerical computation library to directly perform comparison operations on the entire probability matrix, generating a Boolean matrix (mask) with the same shape as the original matrix, where positions that meet the condition are marked as true, and positions that do not meet the condition are marked as false. Subsequently, the system uses this Boolean mask to extract all elements with true corresponding positions from the original caregiver index list in one go.

[0095] S109. Target caregivers are selected from the candidate caregivers based on the service risk level to generate dispatch instructions.

[0096] The system determines the target caregiver from candidate caregivers based on service risk levels to generate dispatch instructions. Specifically, if the service risk level exceeds a preset high-risk level, the system obtains the emergency medical skills certificate level of each candidate caregiver; calculates the weighted sum of the matching probability and the emergency medical skills certificate level; selects the candidate caregiver with the largest weighted sum as the target caregiver; obtains the current geographic coordinates of the target caregiver and the user's service address coordinates; determines the emergency travel route based on the minimum time cost to generate dispatch instructions containing the target caregiver's ID, emergency travel route, estimated arrival time, and emergency plan; if the service risk level does not exceed a preset high-risk level, the system selects the candidate caregiver with the highest probability value in the matching probability matrix as the target caregiver; obtains the current geographic coordinates of the target caregiver and the user's service address coordinates; determines the conventional travel route based on the shortest spatial distance to generate dispatch instructions containing the target caregiver's ID, conventional travel route, and standard service procedures.

[0097] Service risk level is the risk category predicted by previous steps, typically divided into high-risk and non-high-risk. The target caregiver is the person ultimately assigned to perform the task. The dispatch instruction is an execution data package containing task details, route planning, and operational procedures. The system adopts differentiated dispatch strategies based on risk level: When determined to be high-risk, the system prioritizes life-saving, focusing on assessing the caregiver's emergency medical qualifications and arrival speed. The system reads the candidate's emergency medical skills certificate level and performs a weighted summation calculation with the matching probability, assigning different weight coefficients to each and adding them together; the candidate with the highest score is selected as the target caregiver. Subsequently, based on the principle of minimum time cost, the system plans the fastest emergency route considering real-time traffic conditions. When determined to be non-high-risk, the system prioritizes service matching, directly selecting the candidate with the highest matching probability as the target caregiver. In this case, the system plans a conventional route based on the principle of shortest spatial distance to optimize service costs. The final generated dispatch instruction will include the target caregiver's identity information, the planned route coordinate sequence, and the corresponding service plan.

[0098] The specific implementation methods include at least the following two. The first method is a combination of a weighted scoring model and a breadth-first search variant. In high-risk scenarios, the system defines a linear weighted calculation logic, multiplying the matching probability and certificate level values ​​by preset weight factors and summing the results to obtain a comprehensive score, which is then used to select the best candidate. Route planning uses Dijkstra's algorithm to search for the optimal path from the starting point to the destination in a road network map weighted by travel time. In low-risk scenarios, the system directly sorts based on matching probabilities. Route planning also uses Dijkstra's algorithm, but the weights of the road network map are switched to geographical distance. The second method is a combination of decision tree logic and a heuristic search algorithm. The system internally runs a logical branch program that enters different processing flows based on the risk level variable. In the high-risk branch, the weighted calculation logic is executed to select the candidate, and the A* search algorithm is called, using Euclidean distance as a heuristic function and real-time traffic congestion coefficient as a cost function to quickly plan the shortest path. In the low-risk branch, the person with the highest probability value is selected directly using the A* search algorithm, but only geographical distance is used as the cost function for path planning. Finally, the calculation results are packaged into a standard format command and issued.

[0099] In the above embodiments, by collecting multimodal data and performing semantic transformation and normalization concatenation, a comprehensive fusion vector containing user health status, caregiver status, and environmental factors is formed. Analyzing the fusion vector using a temporal feature extraction model can more accurately capture the dynamic changes in user needs. Based on this, a heterogeneous interaction graph is constructed, combined with a graph attention network to calculate the attention coefficient, improving the matching accuracy between users and caregivers. Service risks are predicted and assessed using a neural network-based differential equation model, and the risk level is considered in conjunction with the matching probability. This ensures that the final scheduling decision considers both service quality and safety, improving the intelligent scheduling efficiency and service safety of elderly care services.

[0100] The above embodiments mainly describe the basic process of intelligent scheduling of elderly care services based on multi-factor fusion, realizing the complete process from data collection to final scheduling instruction generation. However, in practical application scenarios, various uncertainties may occur during the service execution process, causing deviations in the actual arrival time. In this case, it is necessary to dynamically correct the original scheduling instructions to ensure service quality. Therefore, this application also provides another intelligent scheduling method for elderly care services based on multi-factor fusion. By real-time monitoring of the performance status of the target caregiver and combining it with a neuronormal differential equation model to predict the impact of delays, intelligent correction of scheduling instructions is achieved. The following is a combination of... Figure 2 Another intelligent scheduling method for health and wellness services based on multi-factor fusion is described in the embodiments of this application:

[0101] Please see Figure 2 This is another flowchart illustrating an intelligent scheduling method for health and wellness services based on multi-factor fusion, as described in this application.

[0102] S201. Obtain the current performance progress data of the target caregiver and calculate the expected arrival time based on the performance progress data;

[0103] The system acquires the current performance progress data of the target caregiver and calculates the estimated arrival time based on this data. Specifically: it extracts the target caregiver's current geographic coordinates and remaining service time for the currently performed task from the performance progress data; it calculates the travel time from the geographic coordinates to the user service address recorded in the dispatch instructions; and it sums the current system time, remaining service time, and travel time to obtain the estimated arrival time of the target caregiver at the user service address. Performance progress data refers to a real-time data set reflecting the target caregiver's current task status and location information, including but not limited to the target caregiver's real-time geographic coordinates, remaining service time for the currently performed task, movement speed, and traffic conditions. The target caregiver refers to a caregiver assigned by the system to perform a specific health and wellness service task. The estimated arrival time refers to the specific time estimated based on the current performance progress data when the target caregiver will arrive at the next user service address. Geographic coordinates refer to latitude and longitude information obtained using the Global Positioning System (GPS) or other positioning technologies. Remaining service time refers to the remaining time required for the target caregiver to complete the currently ongoing service task. Travel time refers to the time required for the target caregiver to move from their current location to the next user service address. The system first collects real-time data on the target caregiver's progress via a pre-defined communication interface or mobile terminal device. This data includes not only static location information but also dynamic task execution status. The system parses this data to extract the target caregiver's precise current geographic coordinates and the remaining service time for the current task. Then, the system reads the next user service address pre-stored in the scheduling instruction. Based on the current geographic coordinates and the user service address, the system combines real-time traffic information to calculate the travel time from the current location to the target location. Finally, the system uses the current system time as a baseline, sequentially adding the remaining service time and travel time to estimate the target caregiver's expected arrival time at the user service address. This process is dynamically updated; as the caregiver's location changes and the task progresses, the estimated arrival time will be adjusted accordingly.

[0104] This can be achieved in at least two specific ways: The first method is a route planning algorithm based on a third-party map service API. The system uses the current latitude and longitude coordinates of the target caregiver as the starting point and the latitude and longitude coordinates of the user's service address as the ending point, calling the route planning interface provided by the map service provider. This interface comprehensively considers real-time traffic conditions, traffic congestion index, traffic light waiting time, and the caregiver's mode of travel (e.g., walking, cycling, driving), returning an accurate travel time. After obtaining this travel time, the system performs a simple addition operation with the current system time and remaining service time to obtain the estimated arrival time. The second method is a machine learning prediction model based on historical trajectory data. The system pre-collects a large amount of historical travel data of caregivers in the area, including movement speed and travel time under different time periods and weather conditions. A prediction model is trained using regression analysis algorithms or long short-term memory networks. During actual calculation, the current starting point, ending point, time period, and weather information are input, and the model directly outputs the predicted travel time. This method has high autonomy and adaptability in specific closed areas or when map data is not updated in a timely manner. The remaining service time can be obtained either by the caregiver manually entering the estimated value on the terminal, or by the system automatically calculating it by subtracting the service time already served from the standard service hours.

[0105] S202. If the estimated arrival time is later than the planned arrival time in the scheduling instruction, calculate the difference between the estimated arrival time and the planned arrival time.

[0106] The planned arrival time refers to the time when the target caregiver should arrive at the user's service address, as set by the system based on ideal conditions or preset rules during the initial scheduling instruction generation phase. The difference refers to the time interval between the expected arrival time and the planned arrival time, used to quantify the degree of delay. This step is a process combining logical judgment and numerical calculation. The system compares the expected arrival time calculated in step S201 with the planned arrival time stored in the scheduling instruction. If the timestamp of the expected arrival time is greater than the timestamp of the planned arrival time, it indicates a service delay. Under this premise, the system performs a subtraction operation, subtracting the planned arrival time from the expected arrival time to obtain a time difference. This difference represents the specific duration of the caregiver's lag and is an important input parameter for subsequent assessment of the impact of the delay on service quality. If the expected arrival time is earlier than or equal to the planned arrival time, it indicates that the performance is normal or ahead of schedule, requiring no further correction, and the process ends here or transitions to routine monitoring.

[0107] This step can be implemented in at least two specific ways: The first method is a direct comparison based on timestamps. The system converts the expected arrival time and the planned arrival time into milliseconds or seconds since the Unix epoch (January 1, 1970, 00:00:00 UTC). The numerical comparison operators in the programming language are used to directly determine the size of the two long integer values. If the expected arrival time in milliseconds is greater than the planned arrival time in milliseconds, the two are subtracted, and the result is converted to readable time units such as minutes or hours as the difference. This method is computationally efficient and suitable for systems with extremely high real-time requirements. The second method is a calendar comparison based on time objects. The system uses the date and time library provided by the programming language to construct two time object instances. The comparison methods provided by the library (such as isAfter, compareTo, etc.) are used to determine the order of the times. If a delay is determined, the time interval calculation methods provided by the library (such as Duration.between, etc.) are used to directly obtain the time span object between the two time points and extract the specific difference value.

[0108] S203. Input the expected arrival time as a new time observation variable into the neural constant differential equation model, and take advantage of the continuous evolution characteristics of the neural constant differential equation model. Use the user state corresponding to the planned arrival time as the initial value, integrate and solve to the expected arrival time to obtain the delayed state vector.

[0109] The neural network ordinary differential equation (NDE) model is a deep learning model that combines neural networks and NDEs to model continuous time series data and simulate the continuous evolution of system state over time. Time observation variables are parameters representing the time dimension in the model input. User state is a vector representation describing a user's physical function, health indicators, or service needs at a specific moment. Initial value refers to the initial state value in the differential equation solution process. Integration solution refers to calculating the solution of the differential equation within a specific time interval using numerical integration methods. Delayed state vector refers to the predicted user state vector at the expected arrival time. The system calls a pre-trained neural network NDE model, which has learned the inherent laws governing user state changes over time. The system extracts user state data (such as heart rate, blood pressure, mood index, hunger level, etc.) at the planned arrival time as the initial conditions for the differential equation. Then, the expected arrival time calculated in step S201 is set as the upper limit of integration, and the planned arrival time is set as the lower limit of integration. Using the derivative function defined by the constant differential equation model, the evolution trajectory of the user's state during this delay time is simulated through a numerical integration algorithm. Finally, the state that the user may be in at the expected arrival time is calculated, i.e., the delay state vector.

[0110] The process of solving for the delayed state vector using the NCD model in the above steps can be implemented in at least two specific ways: The first method is numerical integration based on the Euler method or the Runge-Kutta method. The core of the NCD model is a parameterized neural network used to approximate the derivative of the state with respect to time. The system uses the user state at the planned arrival time as the initial value z(t_0) and the expected arrival time as t_1. Using an adaptive step-size Runge-Kutta algorithm (such as Dopri5), it integrates step-by-step from t_0 to t_1. In each integration step, the neural network is called to calculate the derivative of the current state and update the state value. The final state obtained at time $t_1$ is the delayed state vector. This method has high accuracy and is suitable for scenarios sensitive to state changes. The second method is based on the adjoint sensitivity method. During the model inference stage, the system directly calls the API provided by a deep learning framework that supports NCD (such as torchdiffeq). The initial state and the time span sequence (including the planned arrival time and the expected arrival time) are input into the ODE solver. The solver automatically handles the integration process internally and efficiently calculates the gradient using the adjoint method. The solver outputs a state tensor corresponding to the expected arrival time, which the system resolves into a delayed state vector.

[0111] S204. Calculate the degree of state difference between the delayed state vector and the initial state vector on which the scheduling instruction is based;

[0112] The initial state vector refers to the user state data used to generate the original scheduling instruction, typically corresponding to the expected state at the planned arrival time. State difference is a numerical indicator that quantifies the deviation between the delayed state vector and the initial state vector, reflecting the actual impact of delay on the user's state. The system obtains the delayed state vector output in step S203 and simultaneously retrieves the initial state vector used to generate the original scheduling instruction from the database or cache. These two vectors have the same dimension, each representing a specific user state characteristic (such as physiological indicators, psychological state, environmental needs, etc.). The system uses a specific distance metric algorithm or similarity calculation formula to compare and analyze these two vectors. The calculation process comprehensively considers the magnitude of change in each state dimension, ultimately outputting a scalar value, namely the state difference. The larger the difference value, the more drastic the change in user state caused by delay, and the less applicable the original service plan may be; the smaller the value, the more stable the user state, and the smaller the impact of delay.

[0113] The process of calculating state difference in the above steps can be implemented in two specific ways: The first method is based on Euclidean distance. The system treats the delayed state vector and the initial state vector as two points in a multi-dimensional space. For each dimension of the vector (e.g., heart rate, blood pressure), the square of the difference between the corresponding values ​​is calculated. The squares of the differences in all dimensions are summed, and then the sum is squared to obtain the Euclidean distance between the two vectors. This distance is directly used as the state difference. To eliminate the influence of different units (e.g., a large heart rate value and a small body temperature value), the data for each dimension is usually normalized before calculation. The second method is based on weighted cosine similarity. Considering that different state dimensions have different weights in influencing service demand (e.g., the weight of sudden illness is higher than that of ordinary life needs), the system presets a weight coefficient for each state dimension. First, the cosine similarity between the two vectors is calculated, reflecting their isotropy in direction. Then, combined with the weighted Manhattan distance, the absolute numerical difference is calculated. Specifically, one approach could be to calculate the weighted sum of the absolute values ​​of the differences in each dimension, or to combine the weighted distance with the cosine similarity (e.g., subtract the similarity from 1 and then multiply by the weighted distance factor) to obtain a state difference that comprehensively reflects both directional and numerical deviations.

[0114] S205. If the state difference exceeds the preset service tolerance threshold, generate service correction parameters.

[0115] The preset service tolerance threshold refers to the maximum allowable change in user state by the system. Exceeding this threshold indicates that the original service plan can no longer meet the current user's needs. This threshold can be a globally fixed value or a dynamically set value based on different user levels or service types. Service correction parameters refer to a set of control parameters used to adjust the original service operation sequence, including but not limited to the target state dimension identifier requiring priority adjustment and the intensity level of adjustment. The target state dimension identifier is the unique identifier of the feature dimension with the largest or most critical change in the state vector. The system compares the state difference calculated in step S204 with the preset service tolerance threshold. If the state difference is greater than the threshold, a correction mechanism is triggered. The system further analyzes the specific differences between the delayed state vector and the initial state vector in each dimension, identifying the main factors causing the difference to exceed the limit (i.e., the state dimension with the most drastic change). The system records the identifier of this dimension as the target state dimension identifier and determines the urgency of correction based on the magnitude of the difference, combining them to generate service correction parameters. If the state difference does not exceed the threshold, it indicates that although delay exists, it is within an acceptable range, and the system maintains the original scheduling instructions unchanged.

[0116] The process of generating service correction parameters in the above steps can be implemented in two specific ways: The first method is parameter generation based on the dimension with the largest deviation. The system iterates through each dimension of the delayed state vector and the initial state vector, calculating the relative deviation rate of each dimension (i.e., |delay value - initial value| / initial value). The dimension with the largest relative deviation rate is identified and designated as the target state dimension identifier. Simultaneously, based on the range to which this deviation rate belongs (e.g., slight, moderate, severe), the corresponding correction level parameter is set. The dimension identifier and correction level are packaged into JSON format or a structure object and output as the service correction parameters. The second method is parameter generation based on a rule engine. The system has a pre-set business rule library, where each rule defines the mapping relationship between specific state dimension changes and service corrections. For example, "If the blood glucose value deviation exceeds 20%, then generate correction parameters for the 'blood glucose management' dimension." The system inputs the difference data for each dimension into the rule engine for matching. If multiple rules are triggered simultaneously, the highest priority rule (e.g., vital signs are prioritized over daily care) is selected based on the preset rule priority. The status dimension identifier and the corresponding service adjustment instruction code are extracted and combined to generate service correction parameters.

[0117] S206. Update the service operation sequence in the scheduling instruction using the service correction parameters.

[0118] The system updates the service operation sequence in the scheduling instruction using service correction parameters. Specifically, this includes: extracting the target state dimension identifier from the service correction parameters that causes the state difference to exceed a preset service tolerance threshold; traversing the service operation sequence in the scheduling instruction, and based on the preset state-service mapping relationship, identifying specific service sub-items that match the target state dimension identifier; and adjusting the execution order of the specific service sub-items in the service operation sequence to the first position to obtain the updated service operation sequence. A service operation sequence refers to a series of specific service actions or sub-tasks (such as measuring blood pressure, feeding, turning over, or chatting) arranged in a specific order within the scheduling instruction. A specific service sub-item refers to a specific service action in the service operation sequence that is directly associated with the target state dimension identifier. The preset state-service mapping relationship refers to a data table or knowledge graph stored in the system regarding the correspondence between state dimensions and service actions. Updating the scheduling instruction involves reordering, adding, or deleting existing service operation sequences to generate new execution plans. The system first parses the service correction parameters generated in step S205 and extracts the target state dimension identifier. Then, the system loads the service operation sequence from the original scheduling instruction. Using a pre-defined state-service mapping, the system searches the operation sequence for a specific service sub-item that matches the target state dimension identifier. For example, if the target state dimension is "hunger," the matching service sub-item might be "meal preparation" or "feeding." After finding the specific service sub-item, the system moves it to the first position in the sequence, ensuring that the most urgent problem is addressed first upon the caregiver's arrival. After the adjustment, the system saves the new service operation sequence, completing the update of the dispatch instructions.

[0119] The process of updating the service operation sequence in the above steps can be implemented in two specific ways: The first method is a reordering algorithm based on a priority queue. The system treats each sub-item in the original service operation sequence as an object, and each object contains a priority attribute. Before correction, these sub-items may have a default priority or be arranged in chronological order. Upon receiving the service correction parameters, the system identifies the specific service sub-item that matches the target state dimension identifier and sets the priority attribute of that sub-item to the highest level (e.g., MAX_VALUE). Subsequently, using a priority queue data structure or sorting algorithm (e.g., quicksort), all service sub-items are reordered according to their priority attributes. The sorted sequence naturally places the key service sub-item at the beginning, thus obtaining the updated service operation sequence. The second method is a top-positioning algorithm based on list operations. The system stores the service operation sequence as an ordered list. It iterates through the list, searching for a service sub-item corresponding to the target state dimension identifier. If found, the system first removes the sub-item from its current position in the list (Remove operation), and then inserts it at index 0 of the list (Insert / AddFirst operation). If the corresponding service sub-item does not exist in the original sequence (for example, the service was not originally planned, but a change in status has triggered a new requirement), the system will instantiate a new service sub-item from the service library according to the mapping relationship and insert it directly at the beginning of the list.

[0120] In the above embodiments, the expected arrival time is calculated by acquiring the target caregiver's performance progress data and compared with the planned arrival time. If a delay is detected, the continuous evolution characteristics of the neural network differential equation model are used to predict the delay state. Combined with the state difference degree, service correction parameters are generated to update the scheduling instructions. This allows the system to dynamically perceive time deviations during service execution and quantify the risk of service quality impairment by modeling the impact of delays on user states. This dynamic adjustment mechanism based on state evolution improves the system's adaptability to delays, reduces the negative impact of delays by adjusting the service operation sequence in a timely manner, and considers the different sensitivities of different state dimensions to delays, making service corrections more targeted, thereby improving the system's service reliability in actual execution environments.

[0121] The system in the embodiments of this invention is described below from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 3 This is a schematic diagram of the physical device structure of an intelligent scheduling system for health and wellness services based on multi-factor fusion, provided in an embodiment of this application.

[0122] It should be noted that, Figure 3 The structure of the system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0123] like Figure 3 As shown, the system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on a program stored in Read-Only Memory (ROM) 302 or a program loaded from storage portion 308 into Random Access Memory (RAM) 303, such as executing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0124] The following components are connected to I / O interface 305: input section 306 including a camera, infrared sensor, etc.; output section 307 including a liquid crystal display (LCD) and speakers, etc.; storage section 308 including a hard disk, etc.; and communication section 309 including a network interface card such as a LAN (Local Area Network) card and a modem, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0125] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the various functions defined in the present invention.

[0126] It should be noted that the computer-readable medium shown in the embodiments of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, wherein a computer-readable computer program is carried. The transmitted data signal can take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof.

[0127] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0128] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the system described in the above embodiments; or it may exist independently and not assembled into the system. The storage medium carries one or more computer programs that, when executed by a processor of a system, cause the system to implement the methods provided in the above embodiments.

[0129] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0130] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0131] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0132] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for intelligent scheduling of health and wellness services based on multi-factor fusion, characterized in that, include: Collect multimodal data of users who initiate service requests, service status data of candidate caregivers, and environmental data of the service area. The multimodal data includes structured health record data and unstructured audio and video communication data. The audio-visual communication data is converted into a semantic embedding vector using a preset semantic model. The semantic embedding vector is then normalized and concatenated with the health record data, the service status data, and the environmental data to generate a multi-factor fusion input vector. The multi-factor fusion input vector is input into the time-series feature extraction model, which outputs a user demand representation vector. Construct a heterogeneous interaction graph containing user nodes, caregiver nodes, and service station nodes, and map the user demand representation vector to the user node features in the heterogeneous interaction graph; The attention coefficients between the user nodes and each caregiver node in the heterogeneous interaction graph are calculated using a graph attention network. Based on the attention coefficient, output the matching probability matrix between the user and each of the candidate caregivers; The multi-factor fusion input vector is input into the neural ordinary differential equation model to output the service risk level; Candidate caregivers with a matching probability higher than a preset threshold are selected based on the matching probability matrix; Target caregivers are selected from the candidate caregivers based on the service risk level to generate dispatch instructions.

2. The method according to claim 1, characterized in that, The step of inputting the multi-factor fusion input vector into the time-series feature extraction model and outputting a user demand representation vector specifically includes: A temporal feature extraction model containing a bidirectional long short-term memory network layer and a self-attention mechanism layer is constructed. The multi-factor fusion input vector sequence within a preset historical time period is used as the training set input, and the corresponding historical manual evaluation demand category is used as the training set label. The temporal feature extraction model is trained under supervision until convergence. The multi-factor fusion input vector generated at the current moment is input into the trained bidirectional long short-term memory network layer to extract the hidden state sequence containing the time dimension; The weight coefficients in the hidden state sequence are calculated using the self-attention mechanism layer. The hidden state sequence is weighted and summed according to the weight coefficients to output the user demand representation vector.

3. The method according to claim 1, characterized in that, The process of inputting the multi-factor fusion input vector into the neural network constant differential equation model and outputting the service risk level specifically includes: A neural ordinary differential equation model containing a parameterized derivative function network and an ordinary differential equation solver is constructed. Historical service data with sudden health event markers are selected as the training set, and historical service data not used in training are selected as the test set. The model parameters are optimized by minimizing the error loss function between the predicted risk and the actual event. The multi-factor fusion input vector at the current moment is used as the initial state variable and input into the ordinary differential equation solver; The parameterized derivative function network is used to define the derivative relationship of the state variables over time, and the ordinary differential equation solver is controlled to integrate from the current time to the preset service end time to obtain the state variables at the predicted time. The state variables at the predicted time are input into a fully connected classification layer, and the service risk level is output through the Softmax function.

4. The method according to claim 1, characterized in that, The step of determining the target caregiver from the candidate caregivers based on the service risk level to generate a dispatch instruction specifically includes: If the service risk level is determined to exceed the preset high risk level, the emergency care skills certificate level of each candidate caregiver is obtained; Calculate the weighted sum of the matching probability and the first aid skill certificate level; The candidate caregiver with the largest weighted sum is selected as the target caregiver; Obtain the current geographic coordinates of the target caregiver and the service address coordinates of the user; Determine the emergency access route based on the minimum time cost to generate a dispatch instruction that includes the target caregiver's identity ID, the emergency access route, the estimated arrival time, and the emergency plan; If the service risk level is determined to be no higher than the preset high risk level, the candidate caregiver with the highest probability value in the matching probability matrix is ​​selected as the target caregiver. Obtain the current geographic coordinates of the target caregiver and the service address coordinates of the user; The normal travel path is determined based on the shortest spatial distance to generate a scheduling instruction that includes the target caregiver's identity ID, the normal travel path, and the standard service process.

5. The method according to claim 1, characterized in that, After determining the target caregiver from the candidate caregivers based on the service risk level to generate a scheduling instruction, the method further includes: Obtain the current performance progress data of the target caregiver, and calculate the expected arrival time based on the performance progress data; If it is determined that the expected arrival time is later than the planned arrival time in the scheduling instruction, then the difference between the expected arrival time and the planned arrival time is calculated. The expected arrival time is input into the neural network constant differential equation model as a new time observation variable. Taking advantage of the continuous evolution characteristics of the neural network constant differential equation model, the user state corresponding to the planned arrival time is used as the initial value, and the solution is obtained by integration to the expected arrival time to obtain the delayed state vector. Calculate the state difference degree between the delayed state vector and the initial state vector on which the scheduling instruction is based; If the state difference is determined to exceed a preset service tolerance threshold, service correction parameters are generated. The service operation sequence in the scheduling instruction is updated using the service correction parameters.

6. The method according to claim 5, characterized in that, The calculation of the estimated arrival time based on the performance progress data specifically includes: Extract the current geographical coordinates of the target caregiver and the remaining service time of the currently performing task from the performance progress data; Based on the geographical coordinates and the user service address recorded in the dispatch instruction, calculate the travel time for the target caregiver to travel to the user service address; The current system time, the remaining service time, and the travel time are added together to obtain the estimated arrival time of the target caregiver at the user's service address.

7. The method according to claim 5, characterized in that, The step of updating the service operation sequence in the scheduling instruction using the service correction parameters specifically includes: Extract the target state dimension identifier from the service correction parameters that causes the state difference to exceed the preset service tolerance threshold; Traverse the service operation sequence in the scheduling instruction, and based on the preset state-service mapping relationship, identify the specific service sub-item that matches the target state dimension identifier; The execution order of the specific service sub-item in the service operation sequence is adjusted to the first position to obtain the updated service operation sequence.

8. A smart scheduling system for health and wellness services based on multi-factor fusion, characterized in that, The system includes: One or more processors and a memory; the memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on the system, the system performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the system, the system performs the method as described in any one of claims 1-7.

Images