Online management method and system for hospital infection prevention and control related data

Abstract

The invention provides a hospital infection prevention and control related data online management method and system, and relates to the technical field of data processing, and the method comprises the steps: carrying out the dynamic graph calculation of an infection transmission path through a multi-modal deep learning model, and recognizing a high-risk department, equipment and personnel interaction mode, so as to obtain a real-time calculation result; generating a dynamic risk threshold based on the real-time calculation result, and generating a risk event vector; mapping the risk event vector into an executable supervision task list, and calculating a final task allocation path through a reinforcement learning algorithm; based on the final task distribution path, constructing a department-level risk portrait and a hospital-level risk topological graph; expected loss values of different prevention and control strategies are calculated through Monte Carlo simulation, and a department specific prevention and control scheme is generated; and when a multi-drug-resistant bacterium propagation risk is detected, automatically associating historical data of related departments, and calculating a final isolation region division scheme. The false alarm rate can be reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an online management method and system for hospital infection prevention and control related data. Background Art

[0002] Hospital infection (HIV) management is a core component of medical quality and safety, directly impacting patient treatment outcomes and public health safety. Traditional HAI management relies heavily on manual record-keeping, regular spot checks, and post-event tracing. Consequently, the following technical flaws may exist: For example, data such as infection cases, disinfection records, and medical waste disposal are scattered across different systems (such as HIS, LIS, and material management systems), lacking the ability to integrate data using a unified temporal and spatial benchmark, making it difficult to trace the source of infection. Existing systems often employ threshold alarm mechanisms, which are unable to dynamically model infection transmission paths, making it difficult to identify key risk sources such as high-risk departments and super-spreaders in a timely manner, and often miss the best time for intervention. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and system for online management of hospital infection prevention and control related data, which can reduce the false alarm rate.

[0004] In order to solve the above technical problems, the technical solutions of the present invention are as follows: In a first aspect, a method for online management of hospital infection prevention and control related data comprises: Step S1: Receive and integrate multi-source heterogeneous data, including patient infection records, disinfection operation logs, medical waste disposal data, and medical staff operation behavior data; perform spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized data set; use a streaming computing engine to verify the standardized data set in real time to form a verified structured data stream; Step S2: Based on the verified structured data stream, a trained multimodal deep learning model is used to extract multidimensional risk features. The multimodal deep learning model performs dynamic graph calculations on the infection transmission path, identifying high-risk departments, equipment, and personnel interaction patterns to obtain real-time calculation results. Based on the real-time calculation results, a dynamic risk threshold is generated, and a risk event vector is generated. Step S3: Map the risk event vector into an executable inspection task list and calculate the final task allocation path through the reinforcement learning algorithm; In step S4, based on the final task allocation path, a department-level risk profile and a hospital-level risk topology map are constructed; the expected loss values of different prevention and control strategies are calculated through Monte Carlo simulation to generate department-specific prevention and control plans; when the risk of multidrug-resistant bacteria transmission is detected, the historical data of the relevant departments are automatically associated to calculate the final isolation area division plan.

[0005] In the second aspect, an online management system for hospital infection prevention and control related data includes: The receiving module is used to receive and integrate multi-source heterogeneous data, including patient infection records, disinfection operation logs, medical waste disposal data, and medical staff operation behavior data; perform spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized data set; and use a streaming computing engine to verify the standardized data set in real time to form a verified structured data stream. A generation module is used to extract multidimensional risk features from the verified structured data stream using a trained multimodal deep learning model. The multimodal deep learning model performs dynamic graph calculations on infection transmission paths, identifying high-risk departments, equipment, and personnel interaction patterns to obtain real-time calculation results. Based on the real-time calculation results, dynamic risk thresholds are generated, and risk event vectors are generated. The calculation module is used to map the risk event vector into an executable inspection task list and calculate the final task allocation path through the reinforcement learning algorithm; The partitioning module is used to construct department-level risk profiles and hospital-level risk topology maps based on the final task allocation path; calculate the expected loss values of different prevention and control strategies through Monte Carlo simulation to generate department-specific prevention and control plans; and automatically associate historical data of related departments when the risk of multidrug-resistant bacteria transmission is detected to calculate the final isolation area partitioning plan.

[0006] The above solution of the present invention includes at least the following beneficial effects: Solving the heterogeneity problem of multi-system data through the spatiotemporal alignment algorithm can improve the correlation accuracy of data such as infection events and disinfection operations; the streaming computing engine implements millisecond-level data verification, improving the efficiency of detecting abnormal data. The dynamic graph computing model based on the spatiotemporal propagation graph can predict the risk of infection outbreaks in advance and improve the accuracy of identifying high-risk departments. The adaptive threshold mechanism dynamically adjusts the warning standards based on department characteristics, reducing the false alarm rate.

[0007] The task allocation strategy driven by reinforcement learning improves the response speed of cross-departmental collaborative processing and avoids the problem of task omission in traditional management. The results of risk event processing are fed back into the model training process, achieving automatic iterative updates every week and continuously improving the model prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 The present invention provides a flowchart of an online management method for hospital infection prevention and control related data.

[0009] Figure 2 The figure is a schematic diagram of an online management system for hospital infection prevention and control related data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0010] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0011] like Figure 1 As shown, an embodiment of the present invention provides a method for online management of hospital infection prevention and control related data, the method comprising the following steps: Step S1: Receive and integrate multi-source heterogeneous data, including patient infection records, disinfection operation logs, medical waste disposal data, and medical staff operation behavior data; perform spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized data set; use a streaming computing engine to verify the standardized data set in real time to form a verified structured data stream; Step S2: Based on the verified structured data stream, a trained multimodal deep learning model is used to extract multidimensional risk features. The multimodal deep learning model performs dynamic graph calculations on the infection transmission path, identifying high-risk departments, equipment, and personnel interaction patterns to obtain real-time calculation results. Based on the real-time calculation results, a dynamic risk threshold is generated, and a risk event vector is generated. Step S3: Map the risk event vector into an executable inspection task list and calculate the final task allocation path through the reinforcement learning algorithm; In step S4, based on the final task allocation path, a department-level risk profile and a hospital-level risk topology map are constructed; the expected loss values of different prevention and control strategies are calculated through Monte Carlo simulation to generate department-specific prevention and control plans; when the risk of multidrug-resistant bacteria transmission is detected, the historical data of the relevant departments are automatically associated to calculate the final isolation area division plan.

[0012] In this embodiment of the present invention, multi-dimensional data, including patient infection, disinfection procedures, medical waste, and medical and nursing behaviors, is integrated and validated in real time through spatiotemporal alignment and a streaming computing engine to form a structured data stream. This addresses the issues of traditional data fragmentation and inconsistent formats, improving data availability and accuracy. For example, time-matching disinfection logs with infection records can accurately pinpoint the correlation between disinfection omissions and infection occurrences. Streaming computing enables real-time data validation and processing, avoiding the lag inherent in traditional offline analysis and ensuring the timeliness of hospital infection risk monitoring. By extracting multidimensional risk features (such as department environment, equipment usage frequency, and personnel contact patterns) through a multimodal model and combining it with dynamic graph calculations to simulate infection transmission pathways, high-risk departments, equipment, and personnel interaction scenarios (such as the risk of cross-infection caused by medical staff operating across departments) can be accurately identified. Dynamic risk thresholds are automatically generated based on real-time calculation results, rather than fixed thresholds, to adapt to risk fluctuations across time periods and departments. Risk event vectors transform abstract risks into quantifiable and traceable digital signals, facilitating subsequent task scheduling and intervention. By generating the optimal task allocation path through reinforcement learning algorithms, intelligent scheduling of inspection tasks (such as disinfection of key departments and personnel training) can be achieved, reducing manual decision-making costs and improving the efficiency of the use of prevention and control resources (human and material resources).

[0013] The system constructs department-level risk profiles (e.g., equipment contamination rates and personnel compliance within a department) and hospital-level risk topology maps, enabling managers to intuitively understand risk distribution. Monte Carlo simulations quantify the expected losses of different prevention and control strategies and generate department-specific plans (e.g., enhanced air disinfection protocols for the ICU), achieving "precision prevention and control" rather than uniform management. When a multidrug-resistant bacteria risk is detected, the system automatically links historical data to calculate isolation zone demarcation plans, shortening emergency response times and reducing the risk of the spread of drug-resistant bacteria. This approach is particularly useful for rapidly handling outbreaks of infections.

[0014] In a preferred embodiment of the present invention, step S1 receives and integrates multi-source heterogeneous data, including patient infection records, disinfection operation logs, medical waste disposal data, and medical staff operation behavior data; performs spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized data set; and uses a streaming computing engine to perform real-time verification on the standardized data set to form a verified structured data stream, including: Using the infection timestamp in the patient infection record as the reference axis, a sliding window calibration is performed on the equipment disinfection time in the disinfection operation log to generate a time-aligned infection-disinfection association matrix. The positioning coordinates of medical staff's operation behavior data are mapped to the hospital's three-dimensional spatial model, and spatially superimposed with the medical waste disposal trajectory to generate a weighted personnel-object contact network; Perform compliance check on the infection-disinfection association matrix: calculate the standard deviation of the time intervals between adjacent disinfection operations, and mark it as an abnormal disinfection sequence when the standard deviation exceeds 1.5 times the benchmark value of the department type; Performing density cluster analysis on the human-object contact network to identify abnormal interaction events that exceed the 3σ range of normal contact frequency; Abnormal disinfection sequences and abnormal interaction events are encoded into standardized event tuples in a unified spatiotemporal coordinate system. Each tuple contains [timestamp, spatial grid code, event type, and associated object ID] to form a verified structured data stream.

[0015] In an embodiment of the present invention, through time correlation, it is possible to directly analyze whether the disinfection before the infection event is timely and compliant (for example, whether there is a disinfection record within 24 hours before the infection), calculate the time interval between disinfection and infection, and evaluate the impact of disinfection measures on infection control (for example, the shorter the interval, the better the disinfection effect may be). Medical staff who frequently come into contact with medical waste from patients with multidrug-resistant bacteria may become transmission media. By analyzing the hot spots of personnel-object contact, the equipment placement or work process can be adjusted (for example, reducing cross-regional movement).

[0016] When the above steps are specifically applied, they can be implemented by the following steps, for example: Disinfection records are grouped by equipment and department, and the time interval between two adjacent disinfection operations is calculated (for example, the disinfection interval for equipment A in the ICU is [8h, 10h, 6h, 12h]). The standard deviation of each group of time intervals is calculated (for example, the standard deviation in the above example is 2.5h). This is compared with the department type benchmark value (for example, the benchmark standard deviation for the ICU is 4.0h). If it exceeds 1.5 times (i.e., 4.0h × 1.5 = 6.0h), it is marked as an abnormality. Abnormal disinfection sequences are marked (for example, disinfection intervals that are long or short may indicate non-standard operation) and the abnormality type is classified (for example, disinfection intervals that are too long or the frequency is unstable). This invention can promptly discover disinfection loopholes: for example, a large standard deviation of the disinfection interval for a certain equipment may indicate that the disinfection process is not strictly followed. Different departments (such as ICU vs. general wards) use different benchmark values.

[0017] The method extracts the weights (i.e., contact frequency) of all edges from the person-object contact network, calculates the mean (μ) and standard deviation (σ) of the contact frequency, sets a threshold (e.g., μ + 3σ), and considers contacts outside this range as abnormal. A density clustering algorithm (e.g., DBSCAN) is then used to group contact events, identify areas of dense interaction, and analyze participants in abnormal interactions (e.g., frequent contact between specific medical staff and patients with multidrug-resistant bacteria). This method can identify potential transmission chains. For example, if a nurse's contact frequency with multiple infected patients far exceeds the normal range, she may be the source of transmission. This can reduce the cross-departmental flow of medical staff in high-risk areas and reduce the risk of cross-infection.

[0018] The timestamp, equipment ID, and department ID are extracted from abnormal disinfection sequences. The personnel ID, item ID, contact time, and contact location coordinates are extracted from abnormal interaction events. Three-dimensional coordinates (e.g., (10m, 5m, 0m)) are converted into a grid code (e.g., Grid-20-10-0). Event types are defined (e.g., "abnormal disinfection," "abnormal personnel-equipment contact," and "abnormal personnel-waste contact"). Each abnormal event is encoded in the format of [timestamp, spatial grid code, event type, associated object ID] (e.g., [2024-05-22 08:30:00, Grid_20_10_0, abnormal disinfection, equipment A]). Abnormal events from different sources are standardized into a unified format to facilitate subsequent analysis and system integration. Each event contains clear time and space information, supporting rapid location and tracing (e.g., frequent abnormal interactions in a certain area). This invention uses a streaming computing engine to verify data in real time, promptly detecting anomalies and avoiding the lag associated with traditional offline analysis.

[0019] In a preferred embodiment of the present invention, step S2, using the infection timestamp in the patient infection record as the reference axis, performs a window sliding calibration on the equipment disinfection time in the disinfection operation log to generate a time-aligned infection-disinfection association matrix, including: Centered on the timestamp of each infection event, we extend the set time window forward and backward to extract the disinfection records of all associated devices within the window. Calculate the time deviation coefficient between the disinfection operation and the infection event: if the disinfection time is later than the infection event, it is marked as a delayed disinfection event; Construct a two-dimensional matrix with the row dimension being the infection event ID and the column dimension being the medical device ID. The matrix element value is the weighted value of the time deviation coefficient and the disinfection compliance rate. The matrix was subjected to singular value decomposition to extract the main correlation patterns and generate a time-aligned infection-disinfection correlation matrix.

[0020] In the embodiment of the present invention, when the above steps are specifically applied, the above steps can be implemented by the following steps, for example: With the timestamp of each infection event (e.g., 2024-05-22 08:30:00) as the center, a fixed time period (e.g., ±48 hours) is extended forward / backward to form a time window (e.g., 2024-05-20 08:30:00 to 2024-05-24 08:30:00). The window size can be adjusted according to the infection type (e.g., the window for viral infections can be set to ±72 hours, and for bacterial infections to ±48 hours). Devices associated with the event (e.g., ventilators, surgical instruments) are extracted from the infection records. All disinfection records for these devices within the time window are filtered out from the disinfection log, and invalid records (e.g., incomplete disinfection, repeated operations) are eliminated. A unified time format (e.g., converting timestamps from different systems to UTC time) can avoid interference from irrelevant disinfection records and focus on analyzing disinfection operations that may be related to the infection event.

[0021] For each disinfection record, calculate the difference between it and the infection time (e.g., if the disinfection time is 2024-05-22 10:00:00 and the infection time is 2024-05-22 08:30:00, the deviation is +1.5 hours). The closer the time difference is to 0 (i.e., the closer it is to the infection time), the higher the coefficient (e.g., disinfection 1 hour before infection, the coefficient is 0.9); the larger the time difference is (e.g., disinfection 48 hours before infection), the lower the coefficient (e.g., 0.1); a negative coefficient, with a larger absolute value, indicates a more severe lag (e.g., disinfection 6 hours after infection, the coefficient is -0.5). If the disinfection time is later than the infection time, mark it as a delayed disinfection event, and record the lag duration (e.g., a lag of 1.5 hours). The coefficient intuitively reflects the temporal relationship between disinfection and infection. A positive number indicates disinfection before infection, and a negative number indicates delayed disinfection. Delayed disinfection may indicate that the infection has spread, requiring close attention and tracing the source.

[0022] Rows: infection event ID (such as infection event_001, infection event_002), columns: medical device ID (such as ventilator_01, surgical instrument_23), for each combination of infection event and device, calculate the element value: Time deviation coefficient (calculated in the above steps).

[0023] Disinfection compliance rate: This is assessed based on indicators such as the completeness of the disinfection operation and the concentration of the disinfectant (e.g., a compliance rate of 0.9 indicates that the disinfection operation meets 90% of the standards).

[0024] The matrix is usually multiplied by the device type and usage scenario (e.g., the weight of a ventilator is 1.5, and the weight of an ordinary bed unit is 0.8). If there is no disinfection record for a device within the time window of the infection event, the matrix element value is set to 0. Abnormal values (e.g., a disinfection compliance rate of 0 but a high time deviation coefficient) are manually reviewed or corrected.

[0025] It also takes into account the timeliness, quality, and equipment risks of disinfection, comprehensively reflects the disinfection effect, and areas where matrix element values deviate from normal values (such as close to 0 or negative values) can quickly locate high-risk disinfection omissions.

[0026] For zero values in the matrix (no disinfection record), use mean filling or KNN interpolation to supplement reasonable values, subtract the mean of the column from each element of the matrix to make the data distribution more concentrated and easier to decompose, and decompose the matrix into the product of three matrices: Original matrix = U × Σ × V T , where U represents the characteristic vector of the infection event (such as infection type clustering), Σ represents a diagonal matrix, and the diagonal elements are singular values reflecting the importance of each feature, V T Represent the eigenvectors of the devices (such as device function grouping), retain the first k largest singular values and their corresponding vectors (such as k=3), ignore minor information, achieve data dimensionality reduction, analyze the eigenvectors corresponding to the first k singular values, extract the main association patterns (such as the strong association between a certain type of infection event and the disinfection of specific equipment), reconstruct the matrix based on the retained singular values and eigenvectors, generate a simplified infection-disinfection association matrix, remove noise and redundant information in the matrix, focus on the core association patterns (such as the strong correlation between ICU infection events and ventilator disinfection), and use the decomposition results to intuitively show which infection events are most related to which equipment disinfection operations, assist in decision-making (such as increasing the disinfection frequency of specific equipment), and predict high-risk scenarios based on the extracted patterns (such as when the ventilator disinfection compliance rate drops, providing early warning of related infection risks).

[0027] Through window sliding calibration, the temporal relationship between infection events and disinfection operations is clarified, disinfection omissions are identified, and the weighted value of the time deviation coefficient and the disinfection compliance rate is used to provide a risk score for each infection-disinfection combination. The association patterns extracted by SVD decomposition can be used to predict high-risk scenarios (for example, when a certain type of infection event occurs frequently, it is necessary to strengthen the disinfection of related equipment in advance). The matrix results support hospitals in optimizing disinfection processes, such as adjusting the disinfection cycle of high-risk equipment or strengthening emergency disinfection after infection.

[0028] In a preferred embodiment of the present invention, the positioning coordinates of the medical staff's operation behavior data are mapped to the hospital's three-dimensional spatial model, and spatially superimposed with the medical waste disposal trajectory to generate a weighted personnel-object contact network, including: Discretize the three-dimensional spatial model of the hospital into 1m×1m×1m grids and assign a unique code to each grid; Map the location coordinates of medical staff and the spatiotemporal points of medical waste disposal trajectories to corresponding grids; Count the co-occurrence time between people and medical waste in each grid and calculate the time series of contact frequency; Perform Fourier transform on the time series to extract the characteristic frequencies of the periodic contact pattern; When the contact frequency peak of a single grid exceeds 3 times the standard deviation of the department type benchmark value, it is marked as a high-risk contact node; A weighted bipartite graph network was constructed: the left node was the person ID, the right node was the medical waste container ID, and the edge weight was calculated by the product of contact frequency and duration.

[0029] In the embodiment of the present invention, when the above steps are specifically applied, the above steps can be implemented by the following steps, for example: Determine the scope of the hospital's three-dimensional spatial model, for example, the length, width, and height are L meters, W meters, and H meters respectively. Divide the space into 1m×1m×1m dimensions to form a three-dimensional grid, and assign a unique code to each grid. The coding method can be determined according to the position of the grid in space. For example, it can be represented by three-dimensional coordinates (x, y, z), where x represents the grid number along the length direction from the starting point of the space, y represents the grid number along the width direction, and z represents the grid number along the height direction. The discretized grid provides a unified basic unit for subsequent spatial positioning and analysis, allowing complex three-dimensional space to be accurately divided and described. The unique code facilitates the identification and management of each grid, and facilitates the rapid positioning and query of relevant information in subsequent calculations.

[0030] For each location coordinate of a medical staff member, such as (x1, y1, z1), it is mapped to the corresponding grid by rounding. For each spatiotemporal point in the medical waste treatment trajectory, its spatial coordinates are mapped to the corresponding grid in the same way. The location information of medical staff and medical waste is uniformly mapped to the grid, so that data from different sources can be compared and analyzed within the same spatial framework. This mapping method helps to intuitively understand the distribution of personnel and medical waste in the hospital space.

[0031] For each grid, we record the time periods when medical staff and medical waste co-occur within that grid. We then count the co-occurrence time of medical staff and medical waste within each time period—that is, the duration of their simultaneous presence within that grid. We then calculate the contact frequency within each time interval (e.g., hourly, daily, etc.), that is, the number of times medical staff and medical waste co-occur within that grid within that time interval. This yields a time series of contact frequencies. By counting co-occurrence times and contact frequencies, we can quantify the degree and frequency of contact between medical staff and medical waste within different grids. This time series analysis helps analyze how contact behavior changes over time and uncover potential cyclical or trending patterns.

[0032] The Fourier transform algorithm is applied to the contact frequency time series of each grid. The Fourier transform converts the time series from the time domain to the frequency domain to obtain a series of frequency components and their corresponding amplitudes. The frequency components with larger amplitudes are extracted from the frequency domain results. These frequency components are the characteristic frequencies of the periodic contact pattern. The Fourier transform can decompose complex time series into a superposition of sine and cosine waves of different frequencies, thereby revealing the hidden periodic patterns. Extracting characteristic frequencies helps to discover the periodic laws in the contact behavior of people with medical waste, such as whether there are high-frequency contact periods on fixed periods such as daily and weekly.

[0033] For each grid, the mean and standard deviation of its contact frequency are calculated, and the corresponding benchmark value is set according to the type of department. When the peak contact frequency of a single grid exceeds 3 times the standard deviation of the benchmark value of the department type, the grid is marked as a high-risk contact node. By setting a threshold to mark high-risk contact nodes, areas with abnormally high contact frequency can be quickly screened out. Taking into account the differences in department types, risk assessment is made more personalized and accurate, and different departments can formulate corresponding prevention and control strategies based on their own characteristics.

[0034] A bipartite graph network is constructed, where the left-side nodes represent person IDs and the right-side nodes represent medical waste container IDs. For each contact event between a person and a medical waste container, an edge weight is calculated. The edge weight is the product of the contact frequency and duration. For example, if person A and medical waste container B come into contact n times within a certain time period, with each contact lasting t, the edge weight between them is n × t. This bipartite graph network intuitively illustrates the contact relationships between people and medical waste containers, and the edge weights reflect the intensity and importance of the contact. This network structure clearly identifies which people have frequent and prolonged contact with which medical waste containers, helping to track potential transmission pathways of infection. Spatial Localization and Quantitative Analysis: By discretizing the hospital space and mapping the locations of people and medical waste, precise spatial localization and quantitative analysis of contact behavior are achieved. By counting co-occurrence times, calculating contact frequency, extracting characteristic frequencies, and marking high-risk nodes, high-risk contact areas and patterns can be accurately identified, enabling timely warnings.

[0035] In a preferred embodiment of the present invention, step S2 extracts multidimensional risk features based on the verified structured data stream using a trained multimodal deep learning model; the multimodal deep learning model performs dynamic graph calculations on the infection propagation path, identifies high-risk departments, equipment, and personnel interaction patterns, and obtains real-time calculation results; and generates dynamic risk thresholds based on the real-time calculation results, and generates risk event vectors, including: Taking the verified structured data stream as input, a spatiotemporal propagation graph is constructed: the nodes are wards, equipment, or personnel IDs, and the edge weights are calculated by the spatiotemporal proximity of the interaction events in the tuples; The node infection risk value is iteratively updated through the graph neural network. When the risk spread degree of any ward node is detected to exceed the warning baseline for three consecutive hours, a high-risk department mark is generated. Based on the historical distribution of high-risk department identification, calculate the dynamic risk threshold related to department type and determine early warning events; The warning event is encoded as a risk event vector of [risk type, spatial grid code, urgency, and associated object chain].

[0036] In the embodiment of the present invention, when the above steps are specifically applied, the above steps can be implemented by the following steps, for example: Hospital infection data is converted into a visual graph structure to clearly display the potential transmission paths between wards, equipment, and personnel. The edge weights reflect the closeness of interaction between nodes, helping to quickly locate high-risk transmission links. Through iterative calculations of GNN, the dynamic spread of infection risks between different nodes is captured in real time, avoiding the limitations of static assessments, automatically identifying departments with abnormally elevated risks, reducing manual investigation costs, and improving the timeliness of early warnings.

[0037] Collect data such as the historical frequency and duration of high-risk department identification in each department (for example, the number of times the ICU has been marked as a high-risk department in the past 30 days). Based on historical data, set dynamic risk thresholds for different department types. For example, for high-risk departments (such as ICU), the threshold may be set to 0.8; the threshold for general departments (such as rehabilitation departments) is set to 0.5. The threshold can be dynamically adjusted according to factors such as season and the type of patients admitted to the hospital (for example, lowering the threshold during the flu season). Compare the current node risk value with the dynamic threshold of the corresponding department. If the threshold is exceeded, an early warning event is triggered (for example, the node risk value of a ward reaches 0.85, exceeding the ICU threshold of 0.8). Dynamic adjustment of the threshold can adapt to risk changes in the hospital at different times and improve the flexibility of prevention and control strategies.

[0038] Extract key information from the nodes and associated edges that trigger the warning, including: risk type: such as "cross-infection between departments" and "spread of equipment contamination"; Spatial grid code: The specific grid where the event occurred (e.g. Grid_15_20_0).

[0039] Urgency: The risk is graded based on the gap between the risk value and the threshold (such as high, medium, and low).

[0040] Chain of related objects: sequence of personnel, equipment, and ward IDs involved (e.g., Doctor_007 → Ventilator_03 → Ward_01).

[0041] This information is integrated into a fixed-format risk event vector (e.g., ["Department Cross-Infection", "Grid_15_20_0", "High", ["Doctor_007", "Ventilator_03", "Ward_01"]]). This unified vector format facilitates data sharing between systems and enables multiple departments to quickly obtain key risk information. The associated object chain and spatial information help managers quickly trace the source of infection and develop precise isolation or disinfection plans.

[0042] Combining spatiotemporal propagation graphs with deep learning, we comprehensively analyze the interactive data of personnel, equipment, and environment to achieve full-scenario risk assessment. Through dynamic graph calculation and threshold adjustment, we can timely discover and locate high-risk areas, shorten risk response time, and the risk event vector provides a structured basis for subsequent task allocation and prevention and control plan formulation, thereby improving the scientific nature and efficiency of hospital infection management.

[0043] In a preferred embodiment of the present invention, a spatiotemporal propagation graph is constructed using a verified structured data stream as input: nodes are wards, equipment, or personnel IDs, and edge weights are calculated based on the spatiotemporal proximity of interaction events in the tuples, including: Perform differential calculation on the timestamps in the normalized event tuple to generate a sequence of event time intervals; Convert the spatial grid code into three-dimensional coordinates and calculate the Euclidean distance between the event locations; Based on the Euclidean distance between the locations of events, a dynamic spatiotemporal propagation graph is generated for updating by the graph neural network.

[0044] In the embodiment of the present invention, when the above steps are specifically applied, the above steps can be implemented by the following steps, for example: Arrange standardized event tuples in ascending timestamp order (e.g., 2025-05-22 08:30:00, 2024-05-22 08:45:30), ensuring that events at the same node (ward / equipment / personnel) are arranged in chronological order. Subtract the timestamps of adjacent events to generate a time interval sequence. For example, the time interval between event A (08:30:00) and event B (08:45:30) is 15 minutes and 30 seconds. Set a valid time window (e.g., 24 hours) and remove intervals outside the window (e.g., event pairs with intervals exceeding 24 hours do not generate edges). Time intervals reflect the closeness of events. Short-interval event pairs (e.g., within 15 minutes) may indicate a high risk of transmission. Sequential changes can be used to discover temporal patterns in infection transmission (e.g., dense events during an outbreak).

[0045] Spatial grid codes (such as Grid_15_20_0) are parsed into three-dimensional coordinates (such as (15, 20, 0)), corresponding to locations in the hospital's three-dimensional spatial model. For the coordinates of any two event locations (such as (x1, y1, z1) and (x2, y2, z2)), the spatial straight-line distance is calculated, and a valid spatial range is defined (such as 10 meters). Edges are only generated for event pairs within the range (for example, event pairs with a distance exceeding 10 meters are not associated). The abstract grid code is converted into physical distance to intuitively reflect the distance relationship between event locations. Irrelevant events are filtered out using thresholds, focusing on short-distance transmission paths (such as interactions in the same ward or adjacent areas).

[0046] Ward, equipment, and personnel IDs (such as Ward_01 and Ventilator_03) are used as nodes of the graph; for edge existence, if the time interval between two events is ≤ the time window threshold and the spatial distance is ≤ the distance threshold, then an edge is added between the corresponding nodes.

[0047] Edge weight calculation: Time factor: the shorter the time interval, the higher the weight (for example, the weight of a 15-minute interval is 0.9, and the weight of a 2-hour interval is 0.3).

[0048] Spatial factor: The closer the distance, the higher the weight (for example, the weight of a distance of 2 meters is 0.8, and the weight of a distance of 8 meters is 0.2).

[0049] The comprehensive weight is usually the product of the time factor and the space factor (e.g., time weight 0.9 × space weight 0.8 = 0.72).

[0050] Dynamic update of the graph: As new events flow in, new nodes and edges are continuously added, or the weights of existing edges are updated (for example, if a device frequently interacts with multiple wards, the edge weight increases dynamically).

[0051] Taking both time and space into account, it comprehensively captures potential paths of infection transmission (such as the activities of personnel in adjacent wards within a short period of time), supports real-time reflection of the spatiotemporal evolution of hospital infection risks (for example, when a sudden infection occurs in a certain area, the graph structure is quickly adjusted). The generated graph structure can be directly used as input to the graph neural network, supporting subsequent iterative calculation of risk values, converting discrete hospital infection events into a continuous spatiotemporal network, revealing potential transmission patterns, and intuitively displaying high-risk nodes (such as frequently accessed devices) and transmission links (such as Ward_01 → Nurse_005 → Ventilator_03) through the graph structure. The spatiotemporal transmission graph provides the data structure for subsequent GNN iterative calculations, supporting the accurate identification of high-risk areas and transmission patterns.

[0052] In a preferred embodiment of the present invention, the node infection risk value is iteratively updated through a graph neural network. When the risk spread degree of any ward node is detected to exceed the warning baseline for three consecutive hours, a high-risk department identification is generated, including: Initialize the risk value of each node to the number of associated infection events, and aggregate the weighted risk values of adjacent nodes through the message passing mechanism of the graph neural network; Exponentially smooth the hourly risk values of ward nodes to generate a risk propagation time series curve. The slope of the curve within a continuous 3-hour window is calculated, and an alert is triggered when the slope exceeds the department type threshold. Bind the early warning ward node with its associated medical equipment and medical staff ID to generate a high-risk department identification containing [department code, risk level, and associated object chain].

[0053] In the embodiment of the present invention, when the above steps are specifically applied, the above steps can be implemented by the following steps, for example: The number of infection events associated with each node is counted (for example, if a ward has three infection events, the initial risk value is 3). Special nodes (such as high-risk equipment and key monitoring departments) can be given additional initial weights. Neighborhood information aggregation: Each node collects the risk values of its neighboring nodes and weights them together based on edge weights (spatiotemporal proximity). For example, ward A is connected to nurse B and equipment C, with edge weights of 0.8 and 0.6, respectively. If B has a risk value of 5 and C has a risk value of 2, then A's aggregated neighborhood risk value is 5 × 0.8 + 2 × 0.6 = 5.2. The aggregation result is then combined with the node's own risk value (e.g., taking the average or weighted sum) to update the node's risk value. This process is repeated to spread the risk value across the graph (e.g., a high risk in the ICU is transmitted to other departments through medical staff). Neighborhood aggregation simulates the transmission path of infection among personnel, equipment, and wards, prioritizing areas with high infection events to avoid risk concealment caused by averaging.

[0054] Exponential smoothing is performed on the hourly risk values of the ward nodes to eliminate random fluctuations and highlight trend changes. For example, if the current risk value is 8, the smoothed value of the previous hour is 6, and the smoothing coefficient is 0.3, then the current smoothed value is 0.3×8 + (1-0.3)×6 = 6.6. With time as the horizontal axis and the smoothed risk value as the vertical axis, a risk propagation time series curve is generated.

[0055] The slope of the curve (the rate of risk change) within a continuous three-hour window is calculated and compared to the department type threshold (e.g., 0.5 for the ICU and 0.3 for general wards). If the threshold is exceeded, an alert is triggered. For example, if the risk value of an ICU ward rises from 7 to 10 for three consecutive hours with a slope of 1, exceeding the threshold of 0.5, an alert is triggered. Through smoothing and slope analysis, departments with rapidly increasing risk are identified in advance, rather than focusing solely on high risk values. Different thresholds are set for different departments to avoid false alarms (e.g., if the ICU's daily risk value is high but the rate of change is low, no alert is triggered). Starting from the alert ward node, the graph structure is traced back to extract the medical equipment (e.g., ventilator, thermometer) and medical staff IDs directly connected to it. For example, alert ward A is associated with device C, nurse B, and doctor D. The risk level (e.g., high, medium, low) is comprehensively assessed based on the risk value, transmission rate, and risk weight of the associated objects. For example, a risk value of 10, a slope of 1, and associated high-risk equipment is assessed as "high risk."

[0056] Information is integrated into the format of [department code, risk level, and associated object chain], for example: ["ICU_01","High", ["Ventilator_03","Nurse_005","Doctor_007"]]. This associated object chain allows for rapid location of possible infection sources and transmission paths (e.g., a nurse frequently exposed to multiple high-risk areas). Unified identifiers facilitate sharing between systems and enable decision-makers to quickly understand the full risk picture. Through GNN iteration and time series analysis, the spatiotemporal evolution of infection risk is captured in real time, rather than static assessments. Warnings are based on the rate of risk change rather than absolute values, enabling the detection of potential outbreaks more than three hours in advance.

[0057] In a preferred embodiment of the present invention, step S3 maps the risk event vector into an executable inspection task list and calculates the final task allocation path through a reinforcement learning algorithm, including: Parse the associated object chain in the risk event vector and generate the inspection task parameter set: equipment-related risks are converted to [disinfection task, target equipment ID, standard operating procedure version]; personnel-related risks are converted to [operation assessment task, personnel ID, illegal operation fragment ID]; Construct task allocation state space: including real-time department load, personnel current location, and equipment availability; Define the reward function and use the deep deterministic policy gradient algorithm to iteratively optimize the task allocation strategy. When the cumulative reward value reaches the convergence condition, output the final task allocation path containing [executing department, responsible person, deadline, and related resources].

[0058] In the embodiment of the present invention, when the above steps are specifically applied, the above steps can be implemented by the following steps, for example: Extract the risk type and associated object chain from the risk event vector (e.g., ["Equipment Contamination Spread","Grid_15_20_0","High", ["Ventilator_03","Nurse_005","Ward_01"]]).

[0059] Task type conversion: Equipment risk: If the risk type is "Equipment Contamination Transmission," the associated device ID (such as Ventilator_03) is converted to a disinfection task and associated with the standard operating procedure version (such as the current latest version, "Ventilator Disinfection Specification V2.1").

[0060] Personnel risk: If the risk type is "personnel operation violation", the personnel ID (such as Nurse_005) is converted into an operation assessment task and associated with the illegal operation clip ID (such as the video number of a certain operation recorded by the system).

[0061] Parameter set integration: The task type, target object ID, standard / violation information, etc. are combined into a parameter set.

[0062] For example: the parameter set for the equipment disinfection task is [Disinfection Task, Ventilator_03, Ventilator Disinfection Specification V2.1]; the parameter set for the personnel assessment task is [Operation Assessment Task, Nurse_005, Operation Video_20250522_0830].

[0063] Convert abstract risk events into specific, executable tasks, clarify task objectives and execution standards, and unify parameter formats to facilitate the task scheduling system to quickly identify task types and reduce manual interpretation costs.

[0064] Statistics are collected on the number of currently assigned tasks, task priorities, and estimated completion time for each department. For example, the ICU currently has five high-priority tasks, which are expected to be completed in three hours and are marked as high-load. General wards have fewer tasks and are marked as low-load. The real-time location of medical staff (such as Grid_18_22_0) is obtained through the positioning system, and their current working status (such as "idle" or "performing disinfection tasks") is recorded.

[0065] Check whether the equipment is in use, under maintenance, or available. For example, Ventilator_03 is being used by a patient and marked as occupied; Sterilizer_01 is not in use and marked as available. The above information is integrated into a multi-dimensional state space as the input of the reinforcement learning algorithm.

[0066] For example, a state in the state space is {ICU: high load, Nurse_005: Grid_18_22_0 / idle, Ventilator_03: occupied}.

[0067] Understand the real-time status of departments, personnel, and equipment to avoid task allocation conflicts (such as assigning busy personnel to perform new tasks), update status over time, and support dynamic adjustment of task allocation strategies (such as giving priority to idle personnel or low-load departments).

[0068] Reward function design: Task completion efficiency: If the task is completed ahead of schedule, the reward value increases; if it is completed late, the reward value decreases.

[0069] Resource utilization: Priority will be given to allocating resources to low-load departments or idle personnel. Rewards will be given for improving resource utilization.

[0070] Risk reduction effect: If the risk level in the risk event vector decreases after the task is executed, the reward value will be greatly increased.

[0071] Deep Deterministic Policy Gradient (DDPG) algorithm iteration: Strategy generation: Based on the current state space, the algorithm generates an initial task allocation strategy (e.g., assigning Ventilator_03's disinfection task to Nurse_005, an available nurse in the ICU). After executing the strategy, the algorithm calculates a reward value based on the actual results (e.g., task completion time, risk change). The algorithm adjusts the strategy parameters through reward feedback, gradually optimizing the task allocation plan (e.g., prioritizing personnel closer to the equipment next time).

[0072] Convergence and result output: Continue iterating until the cumulative reward value no longer increases significantly (i.e., convergence conditions are reached), and output the final task assignment path in the format of [executing department, responsible person, deadline, associated resources].

[0073] For example: [ICU, Nurse_005, 2024-05-22 12:00, Ventilator_03].

[0074] Through reinforcement learning, task allocation is automatically optimized to replace manual experience-based decision-making, improve resource utilization efficiency, and dynamically adjust task allocation strategies based on real-time status and reward feedback to adapt to changes in the hospital environment (such as the emergence of new risk events).

[0075] A complete chain is formed from risk identification to task allocation to ensure that risk events are quickly transformed into actual prevention and control actions. Tasks are dynamically allocated based on real-time status and reinforcement learning to reduce resource waste (such as avoiding ineffective running around of personnel across departments). The effectiveness of task allocation is evaluated through a quantitative reward function.

[0076] In a preferred embodiment of the present invention, step S4 constructs a department-level risk profile and a hospital-level risk topology map based on the final task allocation path; calculates the expected loss values of different prevention and control strategies through Monte Carlo simulation, and generates department-specific prevention and control plans; when the risk of multidrug-resistant bacteria transmission is detected, automatically associates the historical data of related departments and calculates the final isolation area division plan, including: Based on the execution status data in the task allocation path, calculate the real-time risk index of the department: Map the risk index of all hospital departments to the hospital floor plan to generate a hospital-level risk topology map in the form of a heat map; Inject current risk topology data into the Monte Carlo simulation to simulate infection transmission paths under different strategies. When the spread of multidrug-resistant bacteria is detected, extract the pathogen distribution data of the associated department over the past 72 hours. According to the pathogen distribution data, the isolation area is divided using the Voronoi diagram algorithm to obtain the final isolation area division plan.

[0077] In the embodiment of the present invention, when the above steps are specifically applied, the above steps can be implemented by the following steps, for example: The execution status of each department's tasks is extracted from the final task allocation path, including task completion progress (e.g., completed, in progress, not started), task priority (high / medium / low), and execution results (e.g., whether disinfection meets standards, whether operational assessments are passed). Unstarted or delayed high-priority tasks significantly increase the department's risk index; completed tasks reduce the risk. If disinfection tasks fail to meet standards or operational assessments violate regulations, the risk weight is increased accordingly. The above factors are weighted and summed to obtain the department's real-time risk index (e.g., the ICU's risk index is 85 due to delays in multiple high-priority tasks; the general ward's risk index is 20 due to good task completion).

[0078] The task execution status can be used to reflect the department's risk level in real time, avoiding the lag of static assessment. The risk index can be linked to task execution, forcing departments to give priority to high-risk tasks and improve prevention and control efficiency.

[0079] Map the risk index of each department to a unified range of 0-100 for easy visual comparison (such as normalizing 85 to 90, and 20 to 25). Mark all the departments in the hospital on the hospital floor plan according to their actual locations. Fill each department area with different colors according to the normalized risk index (such as red for high risk and green for low risk), and generate a heat map. As the task execution status changes, the heat map color and value are refreshed in real time to reflect the dynamic evolution of the risk distribution of the entire hospital. The heat map can intuitively display the risk hotspots of the entire hospital (such as high-risk departments such as ICU and emergency department), assisting managers to quickly locate problems. The risk topology map provides an intuitive basis for the deployment of protective materials and personnel support (such as prioritizing the dispatch of disinfection personnel to red areas).

[0080] Simulation parameters are set, including basic assumptions such as the probability of infection transmission, personnel flow patterns, and the effectiveness of prevention and control measures. The current hospital-level risk topology map data (such as the risk index of each department, personnel and equipment distribution) is imported into the simulation model as the initial state. The infection transmission path under different prevention and control strategies (such as increasing the frequency of disinfection and restricting personnel flow) is simulated. When the spread of multidrug-resistant bacteria is detected, the simulation is paused and the pathogen distribution data of the related departments in the past 72 hours (such as the detection rate of drug-resistant bacteria in samples from a certain department and the type of strain) are extracted. The expected losses of different strategies (such as the number of infections and the cost of prevention and control) are quantified through simulation to provide data support for decision-making. When the risk of multidrug-resistant bacteria transmission is triggered, historical pathogen data is automatically retrieved to reduce manual retrieval time.

[0081] The extracted pathogen distribution data is cleaned, invalid samples (such as contaminated samples) are removed, and the density of drug-resistant bacteria in each area is counted. Departments or wards are used as nodes, and node weights are assigned according to the pathogen density. The Voronoi diagram algorithm is used to divide the hospital space into multiple areas to ensure that the distance from the points in each area to the central node is shortest (that is, the distribution of pathogens in the same area is similar). The division results are adjusted based on the actual layout of the hospital (such as corridors and ventilation systems) to avoid cross-area infection. The final isolation area division plan is output to clarify the departments and personnel that need to be isolated and the control measures (such as blocking specific channels and increasing the frequency of disinfection).

[0082] Dynamically divide isolation areas based on pathogen distribution to avoid resource waste or prevention and control loopholes caused by "one-size-fits-all" isolation. The automated division process significantly shortens isolation decision-making time and quickly blocks the transmission chain when multidrug-resistant bacteria break out. A complete chain is formed from real-time risk monitoring, strategy simulation to isolation execution. Through data-driven simulation and algorithms, empirical judgment is replaced to achieve department-specific prevention and control and dynamic isolation. In high-risk scenarios such as multidrug-resistant bacteria, automated solution generation significantly shortens response time and reduces the risk of epidemic spread.

[0083] When applied in practice, the specific case implementation steps of the above hospital infection management online manual can be achieved through the following content, for example: An online hospital infection registration platform is embedded in the hospital information system (HIS), allowing healthcare professionals to log in using their hospital accounts. This platform automatically synchronizes basic patient information (such as name, hospitalization number, and diagnosis) from the HIS, eliminating duplicate entry.

[0084] For example, when an infection department nurse sees a patient suspected of pneumonia, the system automatically retrieves the patient's medical record information from the HIS. The nurse only needs to add hospital infection-specific data such as infection symptoms and onset time.

[0085] Multi-mode input and real-time verification: Provides three methods: manual input, code scanning, and batch import: Disinfection records: Disinfection personnel use a PDA scanner to scan the RFID tag on the equipment to automatically record the disinfection time, personnel, and disinfectant concentration (such as "2024-06-01 14:00, Nurse Zhang San, 75% alcohol disinfection").

[0086] Medical waste treatment: By connecting weighing equipment to the system, waste weight, type, and treatment destination (such as "infectious waste, 5kg, has been transferred to a temporary storage point") are automatically uploaded. The system verifies data logic in real time (such as whether the disinfection frequency meets department standards and whether the waste treatment process is complete). Abnormal data (such as a disinfection interval exceeding 24 hours) will be marked in red and prohibited from submission.

[0087] The online problem record is combined with the AI intelligent risk warning module to identify potential risks through AI and achieve proactive warning. Case implementation steps: Establishment and classification of problem logs: When a nurse registers an infection case, the system automatically generates a problem log entry, including: Problem types: such as "Ventilator disinfection does not meet standards" and "Medical staff do not wear gloves properly."

[0088] Time and place: "2024-06-02 09:30, ICU ward bed 3".

[0089] Responsible person: The medical staff ID in the associated operation record (such as "Nurse Li Si").

[0090] The AI risk identification model was trained and applied based on three years of historical hospital data (including more than 100,000 disinfection records and more than 5,000 infection cases). The random forest algorithm model was trained to identify risk-related factors (such as "disinfection interval > 12 hours" and "the incidence of ventilator-associated pneumonia increased by 2.3 times").

[0091] Real-time data analysis: When a department has more than five records of "disinfection interval > 12 hours" for three consecutive days, the model triggers a yellow alert; if the number of infection cases in the department increases by 30% month-on-month, it is upgraded to a red alert.

[0092] Multi-channel early warning notification: When a red alert is triggered, the system automatically sends the following to the department director and the person in charge of hospital infection: Pop-up reminders (mandatory display when logging into the system), SMS notifications (such as "ICU has discovered a high-risk hospital infection event, please verify immediately"), and email attachments (including risk data reports and historical similar case references).

[0093] Second-level online quality supervision and improvement closed-loop module for the hospital department, case implementation steps: Inspection plan formulation: Hospital level: The hospital infection office formulates a hospital-wide inspection plan every month, such as "focusing on inspecting aseptic operation standards in operating rooms in June."

[0094] Department level: The emergency department conducts weekly self-inspections of the "hand hygiene implementation rate" and sends the results to the hospital infection office for filing through the system.

[0095] On-site inspection of mobile terminals: Inspectors (such as hospital infection specialists) use a tablet installed with the hospital infection management app to retrieve an on-site inspection checklist (such as "completeness of surgical instrument disinfection record") and check the inspection results item by item: Compliant items: directly mark "passed"; Non-conformities: Take photos of the scene (e.g. “Disinfection registration book not filled out according to specifications”), enter a description of the problem and associate the person responsible.

[0096] Corrective closed-loop management: The system automatically generates a rectification task order and pushes it to the responsible department (such as "the operating room needs to complete the disinfection registration process within 3 days"). The responsible person uploads the rectification measures (such as "all-staff training has been organized, and a new electronic registration template has been added") and supporting materials (training sign-in sheet, new registration book photo) through the APP. The hospital infection office reviews the rectification results in the system. If the standards are not met, a second rectification will be initiated, and a performance deduction warning will be triggered for the responsible person.

[0097] Secondary collaborative risk assessment module for the hospital and department, case implementation steps: Application of standardized risk assessment tools: The Hospital Infection Office and the Respiratory Department jointly conduct quarterly risk assessments using the risk matrix method: Risk factors: "Detection rate of multidrug-resistant bacteria" (incidence: high, weight 40%), "department staff turnover rate" (exposure frequency: medium, weight 30%), and "aging of disinfection equipment" (vulnerability: high, weight 30%).

[0098] The system automatically calculates the risk level: Risk value = Occurrence × Exposure frequency × Vulnerability = High × Medium × High = "High risk".

[0099] Risk control strategy formulation and tracking, formulate measures for "high risk" conclusions: Short-term: Increase the frequency of disinfection (from 2 times a day to 4 times a day), and the system will automatically update the disinfection task reminder; Long-term: Apply for the purchase of new disinfection equipment, create an "Equipment Update" project in the system, and track the purchase progress (such as "The bidding process has been completed and the goods are expected to arrive in July").

[0100] A risk trend report is automatically generated every two weeks, comparing data before and after the implementation of measures (for example, the detection rate of multidrug-resistant bacteria dropped from 15% to 8%) to evaluate the prevention and control effects.

[0101] Secondary training and assessment management system module of the institute: Hospital-level: Annual training plans are released in January each year, such as "Conduct hospital-wide hand hygiene standard training in Q2." Courses include: Video resources: "Hand Hygiene Standard Operation Demonstration"; Department level: ICU adds special training for "ventilator-related prevention and control" (such as "ventilator tube replacement operation simulation"), and medical staff learn courses independently through the system, and the system records the learning time (such as "Nurse Wang Wu has completed 2 hours of hand hygiene video learning").

[0102] After the training, the assessment is automatically triggered, and questions are randomly selected from the question bank (such as "What are the five moments of hand hygiene?" and "What is the qualified standard for ventilator disinfection?"). Those who fail the assessment need to retake the exam until they pass. The results are synchronized to the personnel system and linked to quarterly performance. The Hospital Infection Office uses the system to calculate the average assessment score of the entire hospital (such as "the average score of the hand hygiene assessment is 85 points") and weak knowledge points (such as "the accuracy rate of "multi-drug resistant bacteria isolation measures" is only 60%), and make targeted adjustments to the training content for the next quarter.

[0103] like Figure 2 As shown, an embodiment of the present invention further provides an online management system for hospital infection prevention and control related data, including: The receiving module is used to receive and integrate multi-source heterogeneous data, including patient infection records, disinfection operation logs, medical waste disposal data, and medical staff operation behavior data; perform spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized data set; and use a streaming computing engine to verify the standardized data set in real time to form a verified structured data stream. A generation module is used to extract multidimensional risk features from the verified structured data stream using a trained multimodal deep learning model. The multimodal deep learning model performs dynamic graph calculations on infection transmission paths, identifying high-risk departments, equipment, and personnel interaction patterns to obtain real-time calculation results. Based on the real-time calculation results, dynamic risk thresholds are generated, and risk event vectors are generated. The calculation module is used to map the risk event vector into an executable inspection task list and calculate the final task allocation path through the reinforcement learning algorithm; The partitioning module is used to construct department-level risk profiles and hospital-level risk topology maps based on the final task allocation path; calculate the expected loss values of different prevention and control strategies through Monte Carlo simulation to generate department-specific prevention and control plans; and automatically associate historical data of related departments when the risk of multidrug-resistant bacteria transmission is detected to calculate the final isolation area partitioning plan.

[0104] It should be noted that this system is a system corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0105] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for online management of hospital infection prevention and control related data, characterized in that: The method comprises: Step S1: Receive and integrate multi-source heterogeneous data, including patient infection records, disinfection operation logs, medical waste disposal data, and medical staff operation behavior data; perform spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized data set; use a streaming computing engine to verify the standardized data set in real time to form a verified structured data stream; Step S2: Based on the verified structured data stream, a trained multimodal deep learning model is used to extract multidimensional risk features. The multimodal deep learning model performs dynamic graph calculations on the infection transmission path, identifying high-risk departments, equipment, and personnel interaction patterns to obtain real-time calculation results. Based on the real-time calculation results, a dynamic risk threshold is generated, and a risk event vector is generated. Step S3: Map the risk event vector into an executable inspection task list and calculate the final task allocation path through the reinforcement learning algorithm; In step S4, based on the final task allocation path, a department-level risk profile and a hospital-level risk topology map are constructed; the expected loss values of different prevention and control strategies are calculated through Monte Carlo simulation to generate department-specific prevention and control plans; when the risk of multidrug-resistant bacteria transmission is detected, the historical data of the relevant departments are automatically associated to calculate the final isolation area division plan.

2. The online management method for hospital infection prevention and control related data according to claim 1 is characterized in that: Step S1: Receive and integrate multi-source heterogeneous data, including patient infection records, disinfection operation logs, medical waste disposal data, and medical staff operation behavior data; perform spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized data set; Use a streaming computing engine to perform real-time verification of standardized data sets to form a verified structured data stream, including: Using the infection timestamp in the patient infection record as the reference axis, a sliding window calibration is performed on the equipment disinfection time in the disinfection operation log to generate a time-aligned infection-disinfection association matrix. The positioning coordinates of medical staff's operation behavior data are mapped to the hospital's three-dimensional spatial model, and spatially superimposed with the medical waste disposal trajectory to generate a weighted personnel-object contact network; Perform compliance check on the infection-disinfection association matrix: calculate the standard deviation of the time intervals between adjacent disinfection operations, and mark it as an abnormal disinfection sequence when the standard deviation exceeds 1.5 times the benchmark value of the department type; Performing density cluster analysis on the person-object contact network to identify abnormal interaction events; Abnormal disinfection sequences and abnormal interaction events are encoded into standardized event tuples in a unified spatiotemporal coordinate system. Each tuple contains [timestamp, spatial grid code, event type, and associated object ID] to form a verified structured data stream.

3. The online management method for hospital infection prevention and control related data according to claim 2 is characterized in that: Step S2: Using the infection timestamp in the patient infection record as the reference axis, perform window sliding calibration on the equipment disinfection time in the disinfection operation log to generate a time-aligned infection-disinfection association matrix, including: Centered on the timestamp of each infection event, we extend the set time window forward and backward to extract the disinfection records of all associated devices within the window. Calculate the time deviation coefficient between the disinfection operation and the infection event: if the disinfection time is later than the infection event, it is marked as a delayed disinfection event; Construct a two-dimensional matrix with the row dimension being the infection event ID and the column dimension being the medical device ID. The matrix element value is the weighted value of the time deviation coefficient and the disinfection compliance rate. The matrix was subjected to singular value decomposition to extract the main correlation patterns and generate a time-aligned infection-disinfection correlation matrix.

4. The online management method for hospital infection prevention and control related data according to claim 3 is characterized in that: The positioning coordinates of medical staff's operation behavior data are mapped to the hospital's three-dimensional spatial model, and spatially superimposed with the medical waste disposal trajectory to generate a weighted personnel-object contact network, including: Discretize the three-dimensional spatial model of the hospital into 1m×1m×1m grids and assign a unique code to each grid; Map the location coordinates of medical staff and the spatiotemporal points of medical waste disposal trajectories to corresponding grids; Count the co-occurrence time between people and medical waste in each grid and calculate the time series of contact frequency; Perform Fourier transform on the time series to extract the characteristic frequencies of the periodic contact pattern; When the contact frequency peak of a single grid exceeds 3 times the standard deviation of the department type benchmark value, it is marked as a high-risk contact node; A weighted bipartite graph network was constructed: the left node was the person ID, the right node was the medical waste container ID, and the edge weight was calculated by the product of contact frequency and duration.

5. The online management method for hospital infection prevention and control related data according to claim 4 is characterized in that: Step S2: extracting multi-dimensional risk features based on the verified structured data stream through the trained multimodal deep learning model; The multimodal deep learning model dynamically calculates infection transmission paths, identifies high-risk departments, equipment, and personnel interaction patterns, and obtains real-time calculation results. Generate dynamic risk thresholds based on real-time calculation results and generate risk event vectors, including: Taking the verified structured data stream as input, a spatiotemporal propagation graph is constructed: the nodes are wards, equipment, or personnel IDs, and the edge weights are calculated by the spatiotemporal proximity of the interaction events in the tuples; The node infection risk value is iteratively updated through the graph neural network. When the risk spread degree of any ward node is detected to exceed the warning baseline for three consecutive hours, a high-risk department mark is generated. Based on the historical distribution of high-risk department identification, calculate the dynamic risk threshold related to department type and determine early warning events; The warning event is encoded as a risk event vector of [risk type, spatial grid code, urgency, and associated object chain].

6. The online management method for hospital infection prevention and control related data according to claim 5 is characterized in that: Using the validated structured data stream as input, we construct a spatiotemporal propagation graph: nodes are wards, equipment, or personnel IDs, and edge weights are calculated based on the spatiotemporal proximity of interaction events in the tuples, including: Perform differential calculation on the timestamps in the normalized event tuple to generate a sequence of event time intervals; Convert the spatial grid code into three-dimensional coordinates and calculate the Euclidean distance between the event locations; Based on the Euclidean distance between the locations of events, a dynamic spatiotemporal propagation graph is generated for updating by the graph neural network.

7. The online management method for hospital infection prevention and control related data according to claim 6 is characterized in that: The node infection risk value is iteratively updated through the graph neural network. When the risk spread of any ward node is detected to exceed the warning baseline for three consecutive hours, a high-risk department identification is generated, including: Initialize the risk value of each node to the number of associated infection events, and aggregate the weighted risk values of adjacent nodes through the message passing mechanism of the graph neural network; Exponentially smooth the hourly risk values of ward nodes to generate a risk propagation time series curve. The slope of the curve within a continuous 3-hour window is calculated, and an alert is triggered when the slope exceeds the department type threshold. Bind the early warning ward node with its associated medical equipment and medical staff ID to generate a high-risk department identification containing [department code, risk level, and associated object chain].

8. The online management method for hospital infection prevention and control related data according to claim 7 is characterized in that: Step S3: Map the risk event vector to an executable inspection task list and calculate the final task allocation path through the reinforcement learning algorithm, including: Parse the associated object chain in the risk event vector and generate the inspection task parameter set: equipment-related risks are converted to [disinfection task, target equipment ID, standard operating procedure version]; personnel-related risks are converted to [operation assessment task, personnel ID, illegal operation fragment ID]; Construct task allocation state space: including real-time department load, personnel current location, and equipment availability; Define the reward function and use the deep deterministic policy gradient algorithm to iteratively optimize the task allocation strategy. When the cumulative reward value reaches the convergence condition, output the final task allocation path containing [executing department, responsible person, deadline, and related resources].

9. The online management method for hospital infection prevention and control related data according to claim 8, characterized in that: Step S4: Based on the final task allocation path, a department-level risk profile and a hospital-level risk topology map are constructed; the expected loss values of different prevention and control strategies are calculated through Monte Carlo simulation to generate department-specific prevention and control plans; When the risk of multidrug-resistant bacteria transmission is detected, historical data from related departments is automatically linked to calculate the final isolation area division plan, including: Based on the execution status data in the task allocation path, calculate the real-time risk index of the department: Map the risk index of all hospital departments to the hospital floor plan to generate a hospital-level risk topology map in the form of a heat map; Inject current risk topology data into the Monte Carlo simulation to simulate infection transmission paths under different strategies. When the spread of multidrug-resistant bacteria is detected, extract the pathogen distribution data of the associated department over the past 72 hours. According to the pathogen distribution data, the isolation area is divided using the Voronoi diagram algorithm to obtain the final isolation area division plan.

10. An online management system for hospital infection prevention and control related data, characterized in that: Applied to the method according to any one of claims 1 to 9, comprising: The receiving module is used to receive and integrate multi-source heterogeneous data, including patient infection records, disinfection operation logs, medical waste disposal data, and medical staff operation behavior data; perform spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized data set; and use a streaming computing engine to verify the standardized data set in real time to form a verified structured data stream. A generation module is used to extract multidimensional risk features from the verified structured data stream using a trained multimodal deep learning model. The multimodal deep learning model performs dynamic graph calculations on infection transmission paths, identifying high-risk departments, equipment, and personnel interaction patterns to obtain real-time calculation results. Based on the real-time calculation results, dynamic risk thresholds are generated, and risk event vectors are generated. The calculation module is used to map the risk event vector into an executable inspection task list and calculate the final task allocation path through the reinforcement learning algorithm; The partitioning module is used to construct department-level risk profiles and hospital-level risk topology maps based on the final task allocation path; calculate the expected loss values of different prevention and control strategies through Monte Carlo simulation to generate department-specific prevention and control plans; and automatically associate historical data of related departments when the risk of multidrug-resistant bacteria transmission is detected to calculate the final isolation area partitioning plan.

Images