Online assessment, early warning and empowerment system for professional health of nursing care workers

Abstract

This invention discloses an online occupational health assessment, early warning, and empowerment system for palliative care workers. The system includes a multimodal data acquisition module, a feature cross-modal imputation module, an early warning and assessment module, and a dynamic empowerment recommendation module. The system collects physiological time-series data and subjective psychological self-assessment data. When psychological data is missing, it performs feature imputation calculations based on a cross-modal attention mechanism triggered by physiological feature mutations. The system employs ensemble learning combined with a three-branch decision model to classify and rank occupational health risks into three categories. Simultaneously, the system calculates environmental stress gradient values ​​by combining clinical event timestamps and uses these as weights to dynamically extract subgraphs from the knowledge graph, generating empowerment strategies. This invention maintains the consistency of the input data to the prediction model and provides intervention strategies that fit the physical work environment.

Description

Technical Field

[0001] This invention relates to the fields of medical informatics and digital occupational health monitoring technology, and in particular to an online occupational health assessment, early warning and empowerment system for palliative care workers. Background Technology

[0002] Palliative care workers face continuous workload and psychological stress in their daily work environment, making the assessment and management of their occupational health a focus of medical management. Currently, occupational health assessments for this group often rely on periodically completed questionnaires or routine physical examinations, with relatively fixed data collection cycles. Regarding intervention and empowerment, existing systems typically depend on pre-built health guidance resource libraries. Upon receiving abnormal assessment results, they retrieve and push corresponding palliative advice or training content according to predetermined rules. In actual clinical scenarios, palliative care workers are often in situations involving high-frequency death stress or high-intensity emotional labor, making it easy for subjective reporting to be missed or delayed, leading to data sequence breaks in the assessment model's input. Furthermore, because sudden events and stress levels in the clinical physical work environment change dynamically over time, fixed-rule intervention strategy libraries struggle to balance the worker's current work schedule with the real-time clinical context when matching strategies, and the computational resources required for traversing and retrieving the entire data structure are substantial. Summary of the Invention

[0003] One of the objectives of this invention is to provide an online occupational health assessment, early warning, and empowerment system for palliative care workers, in order to address the problems mentioned in the background art.

[0004] This invention provides an online occupational health assessment, early warning, and empowerment system for palliative care workers, comprising: The multimodal data acquisition module is used to acquire real-time physiological time-series data of palliative care workers and receive subjective psychological self-report data actively reported by the palliative care workers. A feature cross-modal interpolation module, connected to the multimodal data acquisition module, is used to dynamically generate missing psychological feature vectors by triggering a cross-modal attention interpolation mechanism based on physiological feature mutations in the physiological time-series data when the subjective psychological self-evaluation data is missing. The early warning and assessment module, connected to the feature cross-modal interpolation module, is used to output the three-class classification results of occupational health early warning for palliative care workers based on the physiological time series data and the interpolated or original subjective psychological self-assessment data, using Stacking ensemble learning classification combined with a three-branch decision model. The dynamic empowerment recommendation module, connected to the early warning and assessment module, is used to construct an empowerment system based on knowledge graph technology and obtain event timestamp data in the clinical physical context to calculate the environmental pressure gradient value. When the three-category result is medium-risk or high-risk, the environmental pressure gradient value is used to dynamically extract subgraphs from the empowerment system. Based on the extracted emergency empowerment subgraphs, the optimal empowerment recommendation strategy is generated and output.

[0005] Optionally, the feature cross-modal interpolation module further includes a microsecond-level interaction and anomaly interception unit; The microsecond-level interaction and anomaly interception unit uses a sliding window to extract the physiological time series data in real time and calculates the first-order difference gradient of the physiological characteristics. When the first-order differential gradient exceeds the preset dynamic adaptive threshold, a command is sent to the front end to trigger the microsecond-level light interaction component. If no interactive feedback packet is received from the front end within the set time threshold, the exception is automatically intercepted, and the underlying cross-modal temporal convolutional attention network is activated to generate psychological feature vectors.

[0006] Optionally, after being activated, the cross-modal temporal convolutional attention network uses the aforementioned extracted abnormal physiological data fragments as the Query parameter, and the historical physiological-psychological multimodal alignment matrix of the palliative care worker as the Key and Value parameters to perform attention calculation, thereby dynamically generating the missing psychological feature vector.

[0007] Optionally, the three decision-making models in the early warning and evaluation module are used to determine the cost parameters and thresholds in the cost matrix based on the Stacking ensemble learning classification results. ; The early warning and assessment module is based on the threshold. With probability value , The size relationship between them is used to determine the three-classification result; the three-classification result is specifically divided into three levels: low risk, medium risk, and high risk.

[0008] Optionally, the early warning and assessment module is also configured with a model evaluation and interpretation unit; For the Stacking binary classification model, precision, recall, F1 score, and AUC are selected for model evaluation based on the binary classification confusion matrix. For the three-classification model guided by the three-branch decision model, accuracy and AUC are selected as evaluation indicators based on the three-classification confusion matrix; at the same time, the SHAP package is called to quantify and interpret the importance of key factors affecting the occupational health of palliative care workers.

[0009] Optionally, the empowerment system includes empowerment nodes in four aspects: physical self-protection, psychological self-protection, social self-protection, and emotional self-protection.

[0010] Optionally, the dynamic empowerment recommendation module obtains event timestamp data from the hospital's scheduling system and electronic medical record system in real time through the environment perception interface as heterogeneous event data in the clinical physical context; the event timestamp data includes at least the duration of consecutive night shifts, the frequency of critical illness alarms for patients in the ward under its responsibility in the past 12 hours, and death events.

[0011] Optionally, the dynamic empowerment recommendation module introduces a time decay function to fuse the acquired heterogeneous event data into a real-time continuous variable, which serves as the environmental pressure gradient value.

[0012] Optionally, when the three-classification result is medium-risk or high-risk, the dynamic empowerment recommendation module maps the environmental pressure gradient value at the current moment to the pruning weight of the knowledge graph; based on the pruning weight, it filters out long-term strategy nodes with long execution time and node depth, and dynamically extracts and caches the emergency empowerment subgraph containing the immediate adjustment strategy with short execution time.

[0013] Optionally, the evaluation indicators of the system include the occupational burnout index and the empathy fatigue index of the palliative care workers.

[0014] The present invention has achieved the following beneficial effects: This invention provides an online occupational health assessment, early warning, and empowerment system for palliative care workers. The system, through a cross-modal feature interpolation module, activates a cross-modal temporal convolutional attention network based on differential gradients of objective physiological features when subjective psychological self-assessment data is missing. It then generates psychological feature vectors using a historical multimodal alignment matrix, maintaining the consistency of the model's input data. The early warning and assessment module employs ensemble learning combined with a three-branch decision model, setting three-dimensional early warning intervals to provide objective risk grading criteria. The dynamic empowerment recommendation module calculates environmental stress gradient values ​​by collecting event timestamp data from electronic medical records and scheduling systems. These values ​​are used as pruning weights for knowledge graph traversal, dynamically extracting emergency empowerment subgraphs whose time consumption meets the current threshold. This saves computational resources during graph retrieval, ensuring that the output empowerment recommendation strategy is adapted to the worker's current clinical physical environment, increasing the operability of occupational health interventions in practice.

[0015] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart illustrating the construction and interpretation of the classification prediction model provided in this embodiment of the invention; Figure 2 This is an overall architecture diagram of the digital hierarchical management model for empathic fatigue among palliative care workers provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the health intervention interface structure provided in an embodiment of the present invention. Detailed Implementation

[0018] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] This application discloses an online occupational health assessment, early warning, and empowerment system for palliative care workers. Specifically, the system includes a multimodal data acquisition module, a feature cross-modal interpolation module, an early warning and assessment module, and a dynamic empowerment recommendation module. The modules are connected and interact with each other via a communication network to achieve the collection of occupational health status data, feature vector processing, classification and early warning, and recommendation and scheduling of intervention strategy map nodes. The early warning and assessment module's preset assessment indicators include occupational burnout and empathic fatigue indicators for palliative care workers.

[0021] Specifically, the multimodal data acquisition module is used to acquire real-time physiological time-series data of palliative care workers and receive subjective psychological self-assessment data actively reported by the workers. It is understood that the physiological time-series data is collected by wearable device sensor nodes worn by the palliative care workers. These wearable device sensor nodes are equipped with photoplethysmography (PPG) sensors and electrodermal activity (EDA) sensors. The multimodal data acquisition module receives the underlying waveform data from the sensors according to a preset sampling frequency and extracts and generates the corresponding physiological time-series data through filtering and denoising instructions executed by an edge computing node. The physiological time-series data includes heart rate, heart rate variability characteristics, and EDA levels. Further, the multimodal data acquisition module is equipped with a data receiving interface for acquiring subjective psychological self-assessment data input by the palliative care workers through a mobile terminal interface. This subjective psychological self-assessment data includes scale rating items used to characterize occupational health dimensions. The multimodal data acquisition module performs timestamp alignment processing on the received physiological time-series data and subjective psychological self-assessment data and outputs it to the system data bus.

[0022] Furthermore, the feature cross-modal interpolation module is connected to the multimodal data acquisition module. When the subjective psychological self-assessment data is missing, it triggers a cross-modal attention interpolation mechanism based on physiological feature mutations in the physiological time-series data to dynamically generate the missing psychological feature vector. Specifically, the feature cross-modal interpolation module is internally configured with a microsecond-level interaction and anomaly interception unit. This unit has a sliding data window for real-time reading of the physiological time-series data output by the multimodal data acquisition module and calculating the first-order difference gradient of key physiological features. The first-order difference gradient serves as a numerical parameter characterizing the rate of change of physiological feature data within adjacent time windows. The microsecond-level interaction and anomaly interception unit is configured with a dynamic adaptive threshold. Based on the variance of the stored historical physiological baseline data fluctuations, the system periodically updates this dynamic adaptive threshold using an exponentially weighted moving average algorithm.

[0023] Specifically, to ensure the rigorous consistency of the judgment dimensions, the calculation formula for periodically updating the dynamically adaptive threshold using the exponentially weighted moving average algorithm is as follows: ; in, For the current number Dynamic adaptive threshold for each update cycle The dynamic adaptive threshold of the previous update cycle; The system calculates the standard deviation by performing a square root reduction operation on the first-order difference variance of the historical physiological baseline data acquired in the current cycle. This ensures that the dimensions of the update factor are strictly consistent with the dimensions of the first-order difference gradient. The set smooth decay factor and satisfy ; The baseline mean constant is set. This is the abnormal sensitivity adjustment coefficient.

[0024] Understandably, the microsecond-level interaction and anomaly interception unit compares the real-time calculated first-order differential gradient with the current dynamic adaptive threshold. When the first-order differential gradient exceeds the preset dynamic adaptive threshold, the unit sends a trigger command to the terminal device of the corresponding palliative care worker, activating the microsecond-level lightweight interaction component configured on the terminal device. Simultaneously, the unit initiates a timeout listener at the underlying level. The microsecond-level lightweight interaction component is configured as a graphical confirmation component on the terminal interface or a vibration response module of a wearable device.

[0025] Specifically, if the timeout listener does not receive the interactive feedback data packet returned by the front end within the set time threshold, the microsecond-level interaction and anomaly interception unit determines that the subjective psychological self-assessment data has an abnormal missing component, performs an anomaly interception operation, and sends an activation command to the cross-modal temporal convolutional attention network. Further, after receiving the activation command, the cross-modal temporal convolutional attention network extracts the abnormal physiological data fragment that triggers the aforementioned first-order differential gradient threshold comparison condition, performs dimensionality reduction processing on it through a one-dimensional causal convolutional layer and a dilated convolutional layer to generate a high-dimensional temporal feature vector, and sets it as the query parameter of the attention mechanism. Simultaneously, the cross-modal temporal convolutional attention network retrieves the historical physiological-psychological multimodal alignment matrix corresponding to palliative care workers from the database. This matrix structure contains physiological feature data and psychological self-assessment scale data with complete records at historical time points.

[0026] Furthermore, the cross-modal temporal convolutional attention network maps and transforms the aforementioned historical multimodal alignment matrix, converting historical physiological feature data into Key parameters and historical psychological self-assessment data into Value parameters. The cross-modal temporal convolutional attention network performs dot product operations to calculate the similarity score between the Query parameter and each Key parameter, and then normalizes it using the Softmax function to generate an attention weight distribution vector. Subsequently, this attention weight distribution vector is used to perform a weighted summation calculation on the corresponding Value parameters, aggregating and generating the missing psychological feature vector at the current moment, which is then output to the warning and evaluation module.

[0027] Specifically, it should be noted that the aforementioned dimensionality reduction refers to compression and dimensionality reduction along the time series length dimension, while the generated high-dimensional temporal feature vector refers to expansion along the feature network channel dimension. A cross-modal temporal convolutional attention network performs feature mapping and aggregation to generate missing psychological feature vectors. The complete computational model is as follows: Setting Query Parameters ,in This is a high-dimensional temporal feature vector derived from abnormal physiological data fragments; the Key parameter is set. Value parameter ,in and These are the historical physiological feature data matrix and the historical psychological self-evaluation data matrix extracted from the historical multimodal alignment matrix, respectively; , , These are the learnable weight matrices in the network; Aggregation generates missing psychological feature vectors The formula is: ; in, This is the transpose of the Key parameter; For feature dimension parameters, As a scaling factor, it is used to prevent the gradient from vanishing due to excessively large multimodal feature dot product values; This is a normalization function for generating the attention weight distribution vector.

[0028] Specifically, the early warning and assessment module is connected to the feature cross-modal interpolation module. Based on the physiological time-series data output by the multimodal data acquisition module and the psychological feature vector output by the feature cross-modal interpolation module, it uses a Stacking ensemble learning classification model combined with a three-branch decision model to output a three-class classification result for occupational health early warning of palliative care workers. It can be understood that the early warning and assessment module concatenates the aligned physiological and psychological features into a joint feature matrix, which is then input into the Stacking ensemble learning classification model. The first layer of the Stacking ensemble learning classification model contains multiple heterogeneous base learners, each of which outputs an initial predicted probability, which is then concatenated into a secondary feature matrix. The second-layer meta-learner performs linear combination calculations on the secondary feature matrix, outputting a continuous comprehensive probability value. This comprehensive probability value represents the probability data of occupational burnout or empathic fatigue occurring in palliative care workers.

[0029] Furthermore, the three-branch decision-making model within the early warning and assessment module is configured with a cost matrix. This cost matrix includes cost parameters for actual risk samples being misclassified as safe samples, and cost parameters for actual safe samples being misclassified as risk samples. The three-branch decision-making model calculates and obtains the upper limit threshold based on Bayesian decision principles. and lower threshold Among them, those that meet the conditions .

[0030] Specifically, the parameter definitions and threshold Bayesian derivation principles in the cost matrix are as follows: Assume the set of true occupational health states of palliative care workers includes risk states. With safety status .set up , , These respectively indicate when the worker is actually in a state of risk. At that time, the system classifies and warns of the risk level as high-risk, medium-risk, or low-risk (i.e., incorrectly classified as safe) based on cost parameters; let... , , These respectively represent the actual state of safety. At that time, the system will issue a warning categorized the cost parameters as high-risk (i.e., error-prone), medium-risk, or low-risk. Based on the actual error-tolerance cost patterns of palliative care, the system will configure it to meet the following constraints: and Based on the Bayesian principle of minimum expected risk, an equivalent derivation is performed to accurately obtain the upper limit threshold. and lower threshold The calculation formula is: ; .

[0031] The early warning and evaluation module compares the comprehensive probability value output by the Stacking ensemble learning classification model with the aforementioned threshold parameters to make a judgment. When the comprehensive probability value is greater than or equal to the upper threshold... When the judgment logic falls into the positive domain, the early warning and assessment module outputs a three-category result of "high risk" level; when the comprehensive probability value is less than or equal to the lower limit threshold... When the decision logic falls into the negative domain, a three-class classification result of "low risk" is output; when the comprehensive probability value is between and When the condition falls within the boundary domain, the three-class classification result of "medium risk" level is output.

[0032] It is understood that the early warning and assessment module is also equipped with a model evaluation and interpretation unit. For the aforementioned Stacking binary classification model, the model evaluation and interpretation unit calculates and generates a binary classification confusion matrix based on the test sample set, and extracts the true positive, false positive, true negative, and false negative values. Based on this, it calculates precision, recall, F1 score, and area under the ROC curve (AUC) as evaluation metrics for the Stacking binary classification model. For the three-class classification model guided by the three-branch decision model, this unit extracts elements based on the three-class confusion matrix, calculates and outputs precision and multi-class AUC values. Furthermore, the model evaluation and interpretation unit calls the SHAP package to calculate the expected marginal contribution of each input feature variable, generating and outputting quantitative explanatory data. This explanatory data is used to characterize the proportion of importance of key factors affecting the occupational health of palliative care workers.

[0033] like Figure 1 The diagram illustrates the architecture of the classification prediction model construction and interpretation process in the early warning and assessment module of this system. The joint feature data set output by the multimodal data acquisition module is first divided into training and test sets according to a predetermined ratio, and then input to the base learner layer via a ten-fold cross-validation mechanism. The base learner layer is configured with Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) algorithm nodes, each performing independent feature learning and prediction operations. The prediction matrices output by the base learners are aggregated to the meta-learner layer, which uses the Logistic Regression algorithm to perform dimensionality reduction and combination calculations on the feature matrices. Subsequently, a three-branch decision model performs threshold determination, outputting three-class classification results: low-risk, medium-risk, and high-risk. Simultaneously, the early warning and assessment module's backend connects to the SHAP value calculation component, performing global and local interpretation operations based on the feature marginal contribution, outputting a quantitatively interpreted feature map of the predictor variables.

[0034] Specifically, the dynamic empowerment recommendation module is connected to the early warning and evaluation module, and is used to construct an empowerment system based on knowledge graph technology and output empowerment recommendation strategies. It can be understood that the empowerment system is divided into four categories of empowerment nodes in the graph database structure: physical self-protection, psychological self-protection, social self-protection, and emotional self-protection. These empowerment nodes are connected through relational edges.

[0035] like Figure 2As shown, the overall architecture of the digital hierarchical management model for empathic fatigue among palliative care workers is presented. The classification prediction model in the system's early warning and assessment module integrates multi-dimensional basic features such as organizational factors, work factors, personal factors, prescription factors, family factors, and social factors. It is constructed using a Stacking ensemble model and the SHAP feature interpretation algorithm, and outputs three classification results—high-risk, medium-risk, and low-risk—combined with a three-branch decision-making approach. Simultaneously, the system combines structured text data from qualitative interviews, literature reviews, and expert consultations to establish a hierarchical management model in a graph knowledge base, consisting of physical self-care, psychological self-care, social self-care, and emotional self-care. Specifically, the physical self-care node includes skills training and self-care units; the psychological self-care node includes role adaptation, stress management, and psychological crisis intervention units; the social self-care node includes communication skills, peer support, and teamwork units; and the emotional self-care node includes emotion monitoring, emotion management, and self-regulation units. The aforementioned classification and prediction models and hierarchical management intervention mechanisms are all embedded in the AI ​​mobile customized program deployed on terminal smart devices in the form of code modules, forming a closed-loop data flow path from implementation assessment and accurate prediction to intelligent decision-making and precise management.

[0036] Furthermore, the dynamic empowerment recommendation module is equipped with an environment-aware interface, which communicates with the hospital's scheduling system server and electronic medical record system server. This interface is used to request event timestamp data via the network as heterogeneous event data within the clinical physical context. The event timestamp data includes continuous night shift duration data for specific palliative care workers, the frequency of critical care alarms for patients in the ward under their care within the past 12 hours, and timestamp records of patient deaths.

[0037] Furthermore, upon receiving the aforementioned heterogeneous event data, the dynamic empowerment recommendation module invokes a preset time decay function module for numerical calculation. The time decay function module substitutes the occurrence time of each event data point into the exponential decay formula and performs integral fusion calculation based on preset weight coefficients for each event category, converting the discrete heterogeneous event data into a real-time continuous scalar data point, which is then defined as the environmental pressure gradient value. Specifically, the time decay function module performs the integral fusion calculation and outputs the environmental pressure gradient value. The mathematical model of calculus is: ; in, This represents the current environmental perception and recognition moment of the system. This represents the start time of the sliding time window; This represents the total number of heterogeneous event categories. Assigned to the first by default Weighting coefficients for event types (such as night shift duration, critical care alarms, and death events). For the first The exponential decay constant of stress relief caused by similar events; For the first Similar events at historical moments The occurrence intensity function. For sudden events such as patient death or critical illness alarms that occur only at discrete time points, the occurrence intensity function is a superposition of a series of impulse functions, i.e. ,in For Dirac function, This corresponds to the timestamp of a discrete event. This is achieved by introducing a Dirac delta-based time stamp. The system perfectly realizes the transformation of heterogeneous discrete pulse events and continuous shift duration data into a real scalar of the residual psychological stress at the current moment through the integral equation of the function.

[0038] Specifically, when the three-class classification result output by the early warning and assessment module is medium-risk or high-risk, the dynamic empowerment recommendation module maps the environmental pressure gradient value calculated at the current moment to pruning weight parameters of the empowerment system knowledge graph. In the empowerment system knowledge graph, each empowerment node is pre-configured with an execution time attribute value and a node network depth parameter. The dynamic empowerment recommendation module initiates a graph traversal retrieval algorithm, converting the pruning weight parameters into threshold judgment conditions during the traversal calculation process. When a long-term strategy node with an execution time attribute value greater than a preset time parameter and a node depth parameter greater than the corresponding threshold is found, the system interrupts the subsequent traversal calculation of that branch path based on the pruning weight parameters and performs a filtering operation. Specifically, the dynamic empowerment recommendation module maps the calculated environmental pressure gradient value... Mapped to have The pruning weight parameter in the range The nonlinear transformation logic formula is as follows: ,in This is an adjustment factor to control the sensitivity of the mapping curve.

[0039] Furthermore, the shrinkage operation model in which the system converts the knowledge graph traversal into corresponding threshold judgment conditions and performs filtering is as follows: Real-time calculation of the current dynamic node depth threshold. ,in A maximum allowed hierarchy depth constant is preset for the knowledge graph. This is the floor function. The system will... Set to the corresponding threshold mentioned above; due to the environmental pressure gradient value The higher the value, the greater the pruning weight. The closer the value is to 1, the higher the calculated depth threshold. The smaller the value, the more likely the system will directly trigger the condition at a very shallow graph level and completely interrupt the traversal and retrieval of long-term strategies, ensuring that only immediate empowerment emergency subgraphs that meet the time requirements are recommended in high-pressure scenarios such as emergency rescue. The system dynamically extracts from the knowledge graph, caches the remaining nodes containing immediate adjustment strategies with a time consumption shorter than the preset time parameter, and generates an emergency empowerment subgraph. The dynamic empowerment recommendation module performs node matching calculations based on this emergency empowerment subgraph, generates the best empowerment recommendation strategy, and sends the corresponding strategy content instructions to the terminal devices of palliative care workers.

[0040] like Figure 3 The diagram illustrates the structure of the health intervention interface in an application on a smart terminal device. The underlying functional modules of the terminal program include an AI recommendation component, which is presented as four specific types of strategy data package tabs in the front-end interface, corresponding to the strategy categories generated by the aforementioned emergency empowerment sub-diagram. Specifically, the strategy data packages include a "Boost" package, a "Reassurance" package, a "Support" package, and a "Relax" package. The "Boost" package contains hyperlinks or content files related to empathy training, self-care skills, and cutting-edge professional concepts and technologies; the "Reassurance" package contains resources related to music therapy, stress-relieving painting, and expressive arts interventions; the "Support" package contains functional guidance links related to peer support, family collaboration, and life narratives; and the "Relax" package contains guidance steps for laughter therapy, mindfulness meditation, and acceptance and commitment therapy. Upon receiving the optimal empowerment recommendation strategy instruction, the terminal device dynamically loads and highlights the strategy data package that matches the worker's current risk level and environmental stress, allowing the terminal user to click and execute it.

[0041] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An online occupational health assessment, early warning, and empowerment system for palliative care workers, characterized in that, include: The multimodal data acquisition module is used to acquire real-time physiological time-series data of palliative care workers and receive subjective psychological self-evaluation data actively reported by the palliative care workers. A feature cross-modal interpolation module, connected to the multimodal data acquisition module, is used to dynamically generate missing psychological feature vectors by triggering a cross-modal attention interpolation mechanism based on physiological feature mutations in the physiological time-series data when the subjective psychological self-evaluation data is missing. The early warning and assessment module, connected to the feature cross-modal interpolation module, is used to output the three-class classification results of occupational health early warning for palliative care workers based on the physiological time series data and the interpolated or original subjective psychological self-assessment data, using Stacking ensemble learning classification combined with a three-branch decision model. The dynamic empowerment recommendation module, connected to the early warning and assessment module, is used to construct an empowerment system based on knowledge graph technology and obtain event timestamp data in the clinical physical context to calculate the environmental pressure gradient value. When the three-category result is medium-risk or high-risk, the environmental pressure gradient value is used to dynamically extract subgraphs from the empowerment system. Based on the extracted emergency empowerment subgraphs, the optimal empowerment recommendation strategy is generated and output.

2. The system according to claim 1, characterized in that, The feature cross-modal interpolation module also includes a microsecond-level interaction and anomaly interception unit; The microsecond-level interaction and anomaly interception unit uses a sliding window to extract the physiological time series data in real time and calculates the first-order difference gradient of the physiological characteristics. When the first-order differential gradient exceeds the preset dynamic adaptive threshold, a command is sent to the front end to trigger the microsecond-level light interaction component. If no interactive feedback packet is received from the front end within the set time threshold, the exception is automatically intercepted, and the underlying cross-modal temporal convolutional attention network is activated to generate psychological feature vectors.

3. The system according to claim 2, characterized in that, After being activated, the cross-modal temporal convolutional attention network uses the aforementioned extracted abnormal physiological data fragments as the Query parameter, and the historical physiological-psychological multimodal alignment matrix of the palliative care worker as the Key and Value parameters to perform attention calculation, thereby dynamically generating the missing psychological feature vector.

4. The system according to claim 1, characterized in that, The three decision-making models in the early warning and assessment module are used to determine the cost parameters and thresholds in the cost matrix based on the Stacking ensemble learning classification results. The early warning and assessment module determines the three-classification result based on the relationship between the threshold and the probability value; the three-classification result is specifically divided into three levels: low risk, medium risk, and high risk.

5. The system according to claim 4, characterized in that, The early warning and assessment module is also equipped with a model evaluation and interpretation unit; For the Stacking binary classification model, precision, recall, F1 score, and AUC are selected for model evaluation based on the binary classification confusion matrix. For the three-classification model guided by the three-branch decision model, accuracy and AUC are selected as evaluation indicators based on the three-classification confusion matrix; at the same time, the SHAP package is called to quantify and interpret the importance of key factors affecting the occupational health of palliative care workers.

6. The system according to claim 1, characterized in that, The empowerment system includes four empowerment nodes: physical self-protection, psychological self-protection, social self-protection, and emotional self-protection.

7. The system according to claim 1, characterized in that, The dynamic empowerment recommendation module obtains event timestamp data from the hospital's scheduling system and electronic medical record system in real time through the environmental perception interface as heterogeneous event data in the clinical physical context; the event timestamp data includes at least the duration of consecutive night shifts, the frequency of critical illness alarms for patients in the ward under its responsibility in the past 12 hours, and death events.

8. The system according to claim 7, characterized in that, The dynamic empowerment recommendation module introduces a time decay function to fuse the acquired heterogeneous event data into a real-time continuous variable, which serves as the environmental pressure gradient value.

9. The system according to claim 8, characterized in that, When the three-classification result is medium-risk or high-risk, the dynamic empowerment recommendation module maps the environmental pressure gradient value at the current moment to the pruning weight of the knowledge graph; based on the pruning weight, it filters out long-term strategy nodes with long execution time and node depth, and dynamically extracts and caches the emergency empowerment subgraph containing the immediate adjustment strategy with short execution time.

10. The system according to any one of claims 1 to 9, characterized in that, The evaluation indicators of the system include occupational burnout and empathic fatigue indicators for palliative care workers.

Images