A hospital ward visual management method and system based on multi-source data fusion

Abstract

This application relates to a hospital ward visualization management method and system based on multi-source data fusion. The method includes: collecting data from IoT devices, clinical systems, and nurse terminals in the hospital ward; generating a four-dimensional label for each data point; and outputting tagged data packets. Based on a data type priority matrix, hierarchical compression and transmission scheduling are performed on the data packets to generate hierarchical data streams. The data streams are spatiotemporally aligned to extract a standardized feature vector set. A dynamic weighting strategy based on individual medical attributes is invoked to perform weighted fusion of the feature vectors, generating a comprehensive risk score table. This table is then input into a context-aware rule engine to match alarm thresholds, generating a hierarchical alarm instruction set. This set drives the visualization system to generate a dynamic focused view of the target bed and a nursing task queue. Based on this queue, navigation paths and operation instructions are generated for nurses, and the system receives completion signals to update patient status. This application improves ward monitoring efficiency and nursing safety.

Description

Technical Field

[0001] This application relates to the field of medical information technology, and in particular to a method and system for visual management of hospital wards based on multi-source data fusion. Background Technology

[0002] With the deepening of medical informatization and the widespread application of IoT technology, the management and nursing models of modern hospital wards are undergoing a profound digital transformation. Traditional operations, heavily reliant on paper records and manual inspections, are gradually being supplemented or replaced by various independent digital systems. For example, the Hospital Information System (HIS) manages basic patient information and medical orders, the Electronic Medical Record (EMR) stores medical records, and an increasing number of IoT devices, such as vital sign monitors, smart mattresses, infusion pumps, and environmental sensors, are being deployed in wards, enabling automated and continuous data collection of patient physiological parameters, behavioral states, treatment processes, and the ward environment. This marks a new stage in ward management driven by data, offering potential for more refined patient care and more efficient ward operations.

[0003] However, because various clinical information systems and IoT devices are typically provided by different vendors and follow different data standards and communication protocols, critical information such as patient information, real-time monitoring data, and nursing records are stored in isolated data silos. Obtaining a complete profile of a patient often requires frequent switching between multiple independent software interfaces and comparing information from different sources. This information acquisition method is not only inefficient but also increases cognitive load and the risk of errors. Furthermore, although IoT devices can achieve continuous monitoring, the massive amount of alarm information they generate often leads to alarm fatigue, causing truly critical signals to be drowned out by numerous ordinary prompts. This results in insufficient overall situational awareness in ward management, and nursing workflows remain in a relatively passive response mode. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a method and system for visual management of hospital wards based on multi-source data fusion.

[0005] Firstly, this application provides a hospital ward visualization management method based on multi-source data fusion, employing the following technical solution: Collect monitoring data from IoT devices in hospital wards, data from clinical information systems, and data entered by nurses' terminals, and generate a four-dimensional tag for each data entry containing patient identification, bed identification, timestamp, and data type, and output a tagged data package; Based on a preset data type priority matrix, hierarchical compression and transmission scheduling are performed on the tagged data packets to generate hierarchical data streams; Based on the patient identification and bed identification, the hierarchical data stream is spatiotemporally aligned, and vital sign statistical features, behavioral frequency features, treatment progress features and environmental assessment features are extracted to output a standardized feature vector set. The system calls upon individual medical attributes from a pre-configured patient profile database, dynamically configures feature weight allocation strategies, and performs weighted fusion operations on the standardized feature vector set to generate a comprehensive risk score table with weight labels. The comprehensive risk scoring table is input into a preset context-aware rule engine, matched with preset multi-condition alarm thresholds, and a hierarchical alarm instruction set is generated. The visualization system is driven by the hierarchical alarm instruction set to generate a dynamic focused view of the target bed and a nursing task queue. Based on the nursing task queue, navigation paths and operation instructions are generated through augmented reality devices, and execution completion signals are received from nurse terminals to update patient status data to the clinical information system.

[0006] By adopting the above technical solutions, data is given a unified spatiotemporal identifier and clinical semantics, and hierarchical transmission is implemented according to data type, laying an efficient and reliable data foundation. Then, through spatiotemporal alignment, multi-dimensional feature extraction, and especially dynamic weighted fusion based on individual patient profiles, the raw data is transformed into a comprehensive risk profile with high individualization and clinical interpretability. Then, using a context-aware rule engine, the quantified risks are transformed into clearly graded and operable nursing instructions. Finally, through intelligent visual sorting and AR navigation guidance, these instructions are seamlessly embedded into the nurses' actual workflow, which not only realizes global, real-time, and visual control of the ward status, but also realizes a paradigm shift in nursing tasks from passive response to proactive prediction, intelligent sorting, and precise guidance, thereby improving the efficiency, safety, and accuracy of nursing work.

[0007] Secondly, this application provides a hospital ward visualization management system based on multi-source data fusion, which adopts the following technical solution: The data acquisition and standardization module is used to collect monitoring data from IoT devices in hospital wards, data from clinical information systems, and data entered by nurses' terminals. It generates a four-dimensional tag for each data entry, which includes patient identification, bed identification, timestamp, and data type, and outputs a tagged data package. The data hierarchical transmission scheduling module is used to perform hierarchical compression and transmission scheduling on the tagged data packets according to a preset data type priority matrix, and generate hierarchical data streams; The patient status feature calculation module is used to perform spatiotemporal alignment processing on the hierarchical data stream based on the patient identity identifier and bed identifier, and extract vital sign statistical features, behavioral frequency features, treatment progress features and environmental assessment features, and output a standardized feature vector set. The risk fusion assessment module is used to call individual medical attributes from a pre-configured patient profile database, dynamically configure feature weight allocation strategies, and perform weighted fusion operations on the standardized feature vector set to generate a comprehensive risk score table with weight labels. The context-aware alarm module is used to input the comprehensive risk score table into a preset context-aware rule engine, match preset multi-condition alarm thresholds, and generate a hierarchical alarm instruction set. The visualization task scheduling module is used to drive the visualization system according to the hierarchical alarm instruction set to generate a dynamic focused view of the target bed and a nursing task queue; The auxiliary execution and feedback module is used to generate navigation paths and operation instructions based on the nursing task queue through augmented reality devices, receive execution completion signals from nurse terminals, and update patient status data to the clinical information system.

[0008] Thirdly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect.

[0009] In summary, this application includes at least one of the following beneficial technical effects: by assigning a unified spatiotemporal identifier and hierarchical transmission mechanism to multi-source data, it lays an efficient and reliable data foundation; by utilizing personalized weighted fusion and context-aware rules, it transforms raw data into a highly individualized and interpretable comprehensive risk profile and generates clear hierarchical nursing instructions; finally, by leveraging intelligent visualization sorting and AR navigation, it seamlessly embeds these instructions into the nurses' workflow, realizing a paradigm shift from passively responding to alarms to proactively predicting risks, intelligently sorting, and accurately guiding nursing tasks, thereby comprehensively improving ward management efficiency, nursing safety, and operational accuracy. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the first process of a hospital ward visualization management method based on multi-source data fusion, which is one embodiment of this application.

[0011] Figure 2 This is a second flowchart of a hospital ward visualization management method based on multi-source data fusion, which is one embodiment of this application.

[0012] Figure 3This is a schematic diagram of the third process of a hospital ward visualization management method based on multi-source data fusion, which is one embodiment of this application.

[0013] Figure 4 This is a schematic diagram of the fourth process of a hospital ward visualization management method based on multi-source data fusion, which is one embodiment of this application.

[0014] Figure 5 This is a schematic diagram of the fifth process of a hospital ward visualization management method based on multi-source data fusion, according to one embodiment of this application.

[0015] Figure 6 This is a schematic diagram of the sixth process of a hospital ward visualization management method based on multi-source data fusion, which is one embodiment of this application.

[0016] Figure 7 This is a schematic diagram of the seventh process of a hospital ward visualization management method based on multi-source data fusion, which is one embodiment of this application.

[0017] Figure 8 This is a schematic diagram of the eighth process of a hospital ward visualization management method based on multi-source data fusion, which is one embodiment of this application. Detailed Implementation

[0018] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-8 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0019] This application discloses a method for visual management of hospital wards based on multi-source data fusion.

[0020] Reference Figure 1 A hospital ward visualization management method based on multi-source data fusion, specifically including: Step S101: Collect monitoring data from IoT devices in the hospital ward, data from the clinical information system, and data entered by nurses' terminals, and generate a four-dimensional tag for each data item containing patient identification, bed identification, timestamp, and data type, and output a tagged data package; The data sources within the hospital wards are heterogeneous and dispersed: IoT devices (such as mattresses, monitors, infusion pumps, and sensors) generate continuous, time-series physical signal streams; the Clinical Information System (HIS / EMR) stores structured patient attributes and medical records; and nurse terminals input discrete, clinically-judgment-based operational events.

[0021] Next, a four-dimensional tag containing patient identification, bed identification, timestamp, and data type is attached to each piece of collected data, constructing a standardized metadata framework that spans all data sources. Patient identification (such as hospital number) is the gold standard for linking all medical information; bed identification provides the spatial location of the data, serving as the anchor point for physical beds in digital space; timestamps ensure that all events and state changes can be sequentially arranged and analyzed, forming the basis for time-series analysis and causal relationship inference; data types (such as "heart rate," "remaining IV fluid volume," and "number of times turned over") specify the semantic meaning of the data, enabling subsequent targeted analysis and calculation.

[0022] In some embodiments, for example, a heart rate data from the bed monitor is labeled (Patient A, Bed 3, 2023-10-26 14:05:30, Heart Rate), which is associated with Patient A's medical record in the HIS system and the body movement data of Bed 3 reported by the smart mattress.

[0023] Step S102: According to the preset data type priority matrix, perform hierarchical compression and transmission scheduling on the tagged data packets to generate hierarchical data streams; Specifically, during the data transmission process from edge devices to the central server, network bandwidth and computing resources are limited, while different types of medical data have significantly different requirements for real-time performance, reliability, and accuracy. Based on the clinical urgency and value density of the data, a differentiated resource allocation strategy is implemented. The pre-defined data type priority matrix serves as the rule base for this resource allocation strategy, defining the transmission frequency and compression strategies corresponding to different data types.

[0024] In some embodiments, for extremely high-risk events (such as cardiac arrest alarms), real-time lossless transmission is employed because any delay or information distortion could have irreversible consequences. This requires the highest priority network channel and prohibits data compression. For critical vital sign abnormalities, a differential compression method is used to transmit changes periodically (e.g., every 10 seconds). This ensures near real-time monitoring of the status while avoiding the transmission of large amounts of repetitive, unchanged values, thus saving bandwidth. For treatment process data (such as remaining fluid volume), the transmission frequency is increased once it reaches a critical threshold (e.g., below 15%), and decreased within a safe range, achieving dynamic adjustment of attention levels. For environmental parameters and other data that change slowly and have relatively low clinical urgency, a low-frequency transmission of moving averages strategy is adopted to minimize network load while meeting environmental monitoring requirements.

[0025] Step S103: Based on patient identification and bed identification, perform spatiotemporal alignment processing on the hierarchical data stream, and extract vital sign statistical features, behavioral frequency features, treatment progress features and environmental assessment features, and output a standardized feature vector set. Among them, the spatiotemporal alignment process aligns and interpolates all data belonging to the same patient (regardless of the device from which it comes) on the timeline, ensuring that the patient's vital signs, behavior, treatment, and environmental data are synchronously available at the same analysis point in time, thereby constructing a multimodal fusion digital state snapshot of the patient at that moment.

[0026] In this embodiment, vital sign statistical characteristics (such as mean, standard deviation, and trend within a sliding window) reflect the patient's physiological stability and changes over a short period. Behavioral frequency characteristics (such as number of times the patient turns over and duration of time out of bed) are behavioral pattern indicators abstracted from primitive action events, which are important for preventing pressure ulcers and falls. Treatment progress characteristics (such as remaining infusion time) are predictive indicators dynamically calculated from real-time data (drip rate, remaining volume), replacing fixed-time alarms and providing greater accuracy. Environmental assessment characteristics (such as mapping noise decibel values ​​to comfort scores) transform physical parameters into quantitative assessments that may affect patient recovery.

[0027] Step S104: Call the individual medical attributes in the pre-configured patient profile database, dynamically configure the feature weight allocation strategy, and perform weighted fusion operation on the standardized feature vector set to generate a comprehensive risk score table with weight labels. Specifically, a patient's risk is not determined by a single indicator, and the severity of risk predicted by the same indicator varies among patients with different clinical conditions. The individual medical attributes stored in the patient profile database (such as postoperative recovery period, high risk of falls, history of respiratory diseases, and history of allergies) are a digital description of the patient's static vulnerability.

[0028] In this embodiment, the dynamic configuration feature weight allocation strategy is based on these individual medical attributes, assigning different importance weights to different feature dimensions in the risk assessment model. For example, for a patient in the "postoperative recovery period," subtle fluctuations in vital signs may indicate serious complications, thus assigning a higher weight to vital sign characteristics (e.g., 0.55), while the impact of environmental comfort may be relatively minor. For a patient marked as "high risk of falls," the weight of their "behavioral frequency characteristics" (e.g., frequent and sudden getting out of bed) is significantly increased (e.g., 0.50).

[0029] Understandably, this dynamic weighting mechanism allows the risk assessment model to adapt to diverse clinical scenarios, resulting in more individualized and clinically relevant assessments. The weighted fusion operation involves summing the obtained standardized feature vectors according to this dynamic strategy, ultimately outputting a comprehensive risk score with weighted labels. This score is not merely a numerical value; it also implicitly indicates the patient's traits (weighted labels) upon which it was calculated, making risk attribution possible.

[0030] Step S105: Input the comprehensive risk score table into the preset context-aware rule engine, match the preset multi-condition alarm threshold, and generate a hierarchical alarm instruction set; The tiered alarm instruction set includes risk level, triggering reason, and target bed location information; Specifically, quantitative risk scores are combined with pre-defined, multi-condition clinical knowledge rules to generate specific, tiered intervention instructions. The context-aware rule engine is a set of "if-then" rules embedded with a large amount of clinical experience and operating procedures. These rules are not simply single-threshold judgments (such as heart rate > 120), but multi-condition alarm thresholds, capable of characterizing more complex risk situations.

[0031] For example, the rule "Generate a red alert when the overall risk score > 0.7" is used to capture extremely high-risk emergencies with multiple overlapping factors. The rule "Generate a fall warning when the vital signs score > 0.5 and the behavior frequency score > 0.6" precisely describes a specific high-risk fall scenario involving unstable vital signs and agitated behavior. A more intelligent rule, such as "Generate an IV drip reminder when the remaining treatment time is < 3 minutes and the distance between the target bed and the nurse is > 15 meters," integrates treatment progress and spatial resource scheduling information, determining not only the need for IV changing but also the urgent need to arrange for someone to change the IV at a more distant bed. After matching these rules, the engine generates a tiered alarm instruction set that includes the risk level, triggering reason, and target bed location, providing nurses with a clear situational awareness: where the problem occurred, how serious the problem is, and what the possible causes are.

[0032] Step S106: Drive the visualization system according to the hierarchical alarm instruction set to generate a dynamic focused view of the target bed and a nursing task queue; The tiered alarm instruction set serves as a crucial bridge connecting the back-end intelligent analysis system and the front-end nursing execution system. It is pushed to the visualization system in real time, driving updates to the nurse station's large screen, generating a dynamic focused view for the target bed, and updating the list of tasks to be processed according to risk level. By transforming data-driven alarm information into an efficient visualization interface that conforms to human-computer interaction and clinical workflow, and by driving the visualization system based on the tiered alarm instruction set, the displayed content is dynamically determined by the back-end risk calculation, rather than being statically formatted.

[0033] In this embodiment, when a high-risk event is identified, the system automatically highlights, enlarges, or displays detailed information about the bed (such as real-time vital signs curves, infusion progress, and critical alarms) on the nurse station screen, helping nurses grasp the key points within seconds. Simultaneously, the system automatically generates a nursing task queue based on the risk level (such as high, medium, or low) in the instructions.

[0034] In step S107, based on the nursing task queue, a navigation path and operation instructions are generated through augmented reality devices, and the execution completion signal is received from the nurse terminal. The patient status data is then updated to the clinical information system.

[0035] Specifically, spatial computing technology is used to accurately overlay information onto real-world work scenarios and establish a digital closed loop for operations. Based on nursing task queues, augmented reality (AR) devices generate navigation paths and operational instructions, providing nurses with an optimal path from the nurses' station to the target bed, taking into account the real-time layout of the ward (such as avoiding temporary obstacles). Upon arrival, key operational information (such as medications to be administered, dosages, and operational procedures) is overlaid on the real bed or equipment in the form of a virtual layer, reducing the time and errors associated with searching and verifying information.

[0036] Subsequently, the nurse's terminal sends a completion signal, marking the end of an intervention. This signal triggers an update of the patient's status data to the clinical information system, such as clearing the alarm status and automatically or automatically generating a nursing record. This ensures consistency between the system's perceived status and the patient's actual status, and also accumulates data for medical quality traceability and process optimization.

[0037] In the above implementation, data is given a unified spatiotemporal identifier and clinical semantics, and hierarchical transmission is implemented according to data type, laying an efficient and reliable data foundation. Then, through spatiotemporal alignment, multi-dimensional feature extraction, and especially dynamic weighted fusion based on individual patient profiles, the raw data is transformed into a comprehensive risk profile with high individualization and clinical interpretability. Then, using a context-aware rule engine, the quantified risk is transformed into clearly graded and operable nursing instructions. Finally, through intelligent visual sorting and AR navigation guidance, these instructions are seamlessly embedded into the nurses' actual workflow, which not only realizes global, real-time, and visual control of the ward status, but also realizes a paradigm shift in nursing tasks from passive response to proactive prediction, intelligent sorting, and precise guidance, thereby improving the efficiency, safety, and accuracy of nursing work.

[0038] Reference Figure 2 As a consistent implementation of step S102, the step of performing hierarchical compression and transmission scheduling on tagged data packets according to a preset data type priority matrix to generate a hierarchical data stream includes: Step S201: Receive a tagged data packet containing patient identification, bed identification, timestamp, and data type; In this process, patient identification and bed identification together constitute the unique spatial anchor of data in the business logic, ensuring that all subsequent compression and scheduling operations are managed at the specific patient and bed level, thus guaranteeing the accuracy of data attribution. The timestamp provides a temporal label for the event, serving as the fundamental basis for judging data freshness, calculating changes, and ultimately aligning multi-source data spatiotemporally on the server side. The data type is the key criterion for the triage decision in this step, indicating whether the data is related to ECG alarms, remaining IV fluid volume, or ambient temperature, thus providing a classification basis for subsequent clinically value-based triage processing.

[0039] Step S202: According to the preset data type priority matrix, mark the transmission priority level for each data item in the tagged data packet; In the context of medical IoT, the frequency and speed of data generation may far exceed the available network bandwidth and the instant processing capacity of the central processing node. If all data is transmitted indiscriminately, it will lead to network congestion, causing emergency alarm information that is truly related to patient safety to lose its value due to queuing delays.

[0040] In this embodiment, the introduced data type priority matrix is ​​a pre-defined rule mapping table based on clinical knowledge. Essentially, it quantifies and classifies the importance of different data types in clinical decision support. For example, cardiac arrest alarms are defined as the highest priority because they correspond to extreme critical situations requiring a response within seconds, and any transmission delay is unacceptable. Environmental temperature and humidity data are defined as low priority because they are mainly used for long-term trend analysis and comfort optimization, with lower real-time requirements, allowing for a certain degree of delay and aggregation. Data on infusion volume below the threshold and time spent out of bed exceeding the limit are defined as the second highest priority. Data on abnormal turning frequency and abnormal blood oxygen saturation are defined as medium priority.

[0041] Understandably, by querying this matrix, a transmission priority level is dynamically assigned to each input tagged data packet. This qualitative classification assigns a "privileged" or "ordinary" identity to each piece of data, guiding the system to make scheduling choices that are both ethically sound and efficient when competing for resources.

[0042] Step S203: Generate compressed data blocks based on the transmission priority level matching compression algorithm; After prioritizing the data, a compression coding strategy adapted to the clinical importance and characteristics of the data is adopted to significantly reduce the volume of data to be transmitted while preserving as much information value as possible. Compression is not the goal, but a means to balance information fidelity and transmission efficiency.

[0043] For example, for data marked with the highest priority, such as cardiac arrest alarms, the information value lies in the integrity and immutability of the event; any distortion of information could lead to serious consequences. Therefore, a lossless compression algorithm is matched to it, aiming to remove only statistical redundancy from the data while ensuring bit-level accurate restoration, generating the original data block and ensuring the absolute authenticity and reliability of the alarm.

[0044] For second-highest priority continuous monitoring data, such as infusion volume, the core value lies in the changing rather than repetitive static values. Therefore, a differential compression algorithm is used. Its principle is to transmit only the change (Δ value) between the current value and the value of the previous transmission cycle, omitting redundant sampling points that have not changed or have only minor changes, generating highly refined change Δ value data blocks. This achieves a significant data compression ratio without losing key trend information.

[0045] For medium-priority data used for pattern analysis rather than exact value reproduction, such as body motion frequencies, lossy compression algorithms can be used to extract and transmit parameters that reflect their statistical characteristics (such as mean, variance, and peak values) to generate statistical feature data blocks, thereby sacrificing local waveform details in exchange for a significant reduction in the overall data volume.

[0046] For low-priority environmental data, a moving average aggregation algorithm is used to smooth and aggregate the data over a longer time window, transmitting only representative aggregated values ​​(such as the average value) within the window. This filters out high-frequency noise and instantaneous fluctuations, generating aggregated value data blocks, thus minimizing the amount of data while meeting macro-monitoring needs.

[0047] Step S204: Schedule compressed data blocks according to the transmission cycle corresponding to the priority level and output hierarchical data stream.

[0048] After data compression is completed, time-sharing scheduling is implemented based on the clinical real-time requirements of the data to further optimize network bandwidth utilization and ensure that highly urgent information receives an immediate transmission channel.

[0049] For example, a "real-time transmission" strategy can be set for the highest priority data, which means that once such data packets are generated, they can be sent immediately by seizing the network channel without waiting, thereby achieving the theoretical minimum transmission delay and meeting the instantaneous requirements of emergency rescue scenarios.

[0050] Setting a transmission cycle of "once every 10 seconds" for high-priority data strikes a balance between the need for continuous monitoring of the patient's condition and reducing network load, ensuring that critical information such as infusion status can be updated in real time.

[0051] Setting a longer period of "transmitting once every 30 seconds" for medium-priority data acknowledges that needs such as behavioral pattern analysis are not sensitive to second-level real-time performance, allowing data to be accumulated and preprocessed more fully on the edge.

[0052] Setting a long cycle of "transmitting once every 5 minutes" for low-priority data is a strategic way to reduce the frequency of data with slow changes and low value density, such as temperature and humidity, to prevent them from occupying valuable bandwidth at high frequencies.

[0053] Ultimately, all the data blocks that have undergone differentiated compression and are scheduled according to their own rhythms intertwine and converge on the timeline, forming a unified hierarchical data stream. High-priority data blocks enjoy absolute transmission priority and a more frequent appearance frequency within this stream.

[0054] In the above implementation, standardized tagged data packets are received to ensure the clarity of the semantics of the processed objects. Then, a data type priority matrix with built-in clinical knowledge is used to assign an importance level reflecting its clinical value to each data point. On this basis, differentiated compression algorithms are innovatively matched to data of different levels, realizing a gradient processing from lossless compression that ensures bit integrity to lossy aggregation that extracts the essence of features. This reduces the data volume while preserving the clinical value of the information to the maximum extent. Finally, by implementing transmission cycle scheduling that strictly corresponds to the priority, the utilization pattern of network bandwidth is further optimized in the time dimension, ensuring that life-threatening alarm data can seize the channel in a near-zero latency manner, while routine monitoring data is transmitted in an orderly and low-frequency manner.

[0055] In practical applications, this technical solution resolves the core contradiction between "network congestion caused by massive concurrent data transmission" and "the requirement for extremely low transmission latency for critical alarm information" in the medical Internet of Things. It provides a stable, efficient, and clinically significant graded data supply for backend data fusion, real-time analysis, and risk warning, and is the key communication foundation for the entire smart ward management system to achieve real-time and precision.

[0056] Reference Figure 3 As one implementation of step S103, the steps of performing spatiotemporal alignment processing on the hierarchical data stream based on patient identification and bed identification, and extracting vital sign statistical features, behavioral frequency features, treatment progress features, and environmental assessment features to output a standardized feature vector set include: Step S301: Receive a hierarchical data stream containing patient identification, bed identification, timestamp, and data type; The received hierarchical data stream is not the original, mixed bit stream, but a data sequence with clear priority and payload structure after being processed by the preceding hierarchical compression and transmission scheduling steps.

[0057] Step S302: Based on patient identification and bed identification, perform spatiotemporal alignment processing on the hierarchical data stream to generate a time-series data set grouped by bed identification; Data from different devices (such as monitors, smart mattresses, infusion pumps, and sensors) have different sampling frequencies, transmission cycles, and reporting times, resulting in discrete and asynchronous data points in the initial state. These data cannot be directly correlated and analyzed, so spatiotemporal alignment processing is required.

[0058] Specifically, firstly, using "bed identifiers" as spatial anchors, all data belonging to the same physical bed (i.e., the same patient), regardless of the equipment they originate from, are merged and grouped. This integrates fragmented vital signs, behaviors, treatments, and environmental information into a patient-centric, multi-dimensional data cluster in a spatial dimension. Secondly, within each group, data is arranged chronologically based on "timestamps," and interpolation and other techniques are used to fill in data gaps caused by different sampling rates on a unified time grid, forming a continuous, equally spaced time-series data set. This process allows the system to obtain a relatively complete and synchronous multimodal state snapshot of the patient at any aligned time point, enabling subsequent calculations of features with temporal contextual significance (such as trends and frequencies).

[0059] Step S303: Obtain vital sign data from the time series data set, calculate the statistical feature values ​​within the preset time window, and generate vital sign statistical features; Among them, vital signs data (such as heart rate, blood pressure, and blood oxygen) are continuous physiological signals. Their clinical value lies not only in instantaneous readings, but also in their trends, stability, and fluctuation patterns over time.

[0060] Specifically, by employing the sliding window time-series analysis method, statistical features characterizing the short-term state and changing patterns of vital signs are extracted from a continuous, potentially noisy, stream of vital signs. This transcends single-point alarms and enables trend-based early warning. The preset time window is a fixed-length interval (e.g., a 5-minute window) that slides continuously along a time axis. For all data points within the window, statistical measures are calculated; for example, the mean reflects the overall level within that time period, and the standard deviation reflects the dispersion or volatility of the data. The window slides forward continuously as new data arrives, resulting in a series of continuous statistical feature sequences.

[0061] For example, calculating the mean and standard deviation of heart rate within a 5-minute window can determine whether a patient's heart rate is stable, slowly rising, rapidly declining, or experiencing abnormal, drastic fluctuations over a period of time. This window-based statistical characteristic is better at capturing potential, gradual physiological deterioration trends than isolated heart rate "too high" or "too low" alarms, and is also more effective at filtering false alarms caused by single measurement errors or brief interference.

[0062] Step S304: Parse the behavior monitoring data from the time series data set, calculate the frequency of action triggers per unit time, and generate behavior frequency features; Among them, patient behavioral data (such as turning over and getting out of bed) are usually discrete event sequences. By transforming these discrete events into quantifiable behavioral pattern indicators with clinical interpretation significance, the patient's activity status and potential risks can be objectively assessed.

[0063] Specifically, these events were aggregated and measured over time by calculating "frequency statistics within a unit of time." For example, "the number of times a patient turns over within a unit of hour" yields "turning frequency," a feature that quantifies a patient's ability to move independently. A low frequency may indicate excessive sedation, weakness, or an increased risk of pressure sores, while an abnormally high frequency may indicate agitation or discomfort. Similarly, for "getting out of bed events," not only were their occurrences recorded, but their "duration" was also calculated or a "getting out of bed risk index" was generated. This can distinguish between brief sitting up at the bedside and prolonged unsupervised getting out of bed activities, the latter of which carries a significant risk of falls.

[0064] By transforming the original behavioral event stream into these statistical characteristics such as frequency and duration, the understanding of patient behavior is elevated from qualitative perception to quantitative analysis, providing objective and comparable data indicators for assessing their safety status, self-care ability, and comfort.

[0065] Step S305: Extract treatment progress data from the time series data set, calculate the remaining treatment time, and generate treatment progress features; For continuous treatment processes (such as intravenous infusion), while real-time process status (such as current remaining volume and current drip rate) is important, nurses are more concerned with the prediction of future status, such as "how long will it take to finish".

[0066] In this embodiment, static, current process monitoring data can be dynamically calculated and transformed into a predictive time indicator, thereby achieving a shift from passive monitoring to proactive early warning. For example, by dividing the "remaining infusion volume" by the "real-time drip rate," the system can dynamically estimate the remaining time (in minutes) for the infusion to be completed at the current flow rate, i.e., the remaining treatment duration. This allows the system to issue an early warning not only when the remaining volume falls below a fixed threshold (e.g., 10ml), but also when "3 minutes are expected to remain," giving nurses time to prepare medication and go to the bedside.

[0067] Step S306: Extract environmental parameter data from the time series data set, calculate the environmental comfort score by weighted combination, and generate environmental assessment features; Among them, environmental parameters (such as noise, temperature and humidity) are objective physical quantities, but their impact on patients is subjective and non-linear.

[0068] In this embodiment, an objective physical measurement value can be converted into a standardized score that reflects its impact on patient comfort or rehabilitation environment through a preset mapping function, thereby enabling environmental data to be comprehensively evaluated on the same scale as clinical data.

[0069] Taking noise as an example, simple decibel (dB) values ​​lack direct guiding significance for clinicians. Through a well-designed mapping function (e.g., with an optimal value as the comfort score decreases non-linearly as deviation increases), noise levels can be converted into a "comfort score" between 0 and 100. This score quantifies how friendly the environmental conditions are to the patient. A high score indicates a suitable environment, while a low score suggests the environment may interfere with the patient's rest, sleep, or recovery.

[0070] Understandably, incorporating such environmental characteristics into the overall assessment system reflects the "patient-centered" holistic care philosophy, enabling the system to not only focus on the patient's physiological and treatment indicators, but also to identify clinical risks (such as agitation, high blood pressure, and sleep disorders) that may be caused or exacerbated by environmental discomfort (such as continuous high noise), thereby achieving a more comprehensive risk insight.

[0071] Step S307: Combine the vital signs statistical features, behavioral frequency features, treatment progress features, and environmental assessment features into a feature vector set with unified dimensions, and perform standardization scaling processing to output a standardized feature vector set.

[0072] In this process, these heterogeneous features from different dimensions (physiology, behavior, treatment, environment) are standardized and vectorized to construct a unified format digital patient status representation that can be efficiently processed by downstream machine learning models or rule engines.

[0073] Specifically, the feature values ​​are standardized. For example, vital sign statistics (such as mean heart rate and standard deviation) are normalized to the [0,1] interval to eliminate the influence of dimensions; behavioral frequency features are converted into a uniform unit (such as times / hour); and time features may retain their original values. Subsequently, these processed feature values ​​are systematically combined into a standardized feature vector. Each dimension of this vector represents a refined, quantified state of the patient in a specific aspect.

[0074] In the above implementation, dedicated feature extraction strategies were designed for four core clinical data categories: vital signs, behavioral events, treatment progress, and environmental parameters. These strategies included: extracting sliding window statistics revealing trends and stability for continuous vital signs; extracting frequency characteristics quantifying patterns and risks for discrete behavioral events; extracting prospective remaining time predictions for the treatment process; and extracting comfort scores reflecting subjective feelings for environmental parameters. Finally, these heterogeneous features were standardized and vectorized, outputting a standardized feature vector set. This transformed raw sensor readings and event records into high-level digital status indicators that directly support clinical decision-making, laying the foundation for accurate and intelligent risk assessment and early warning in the entire smart ward management system.

[0075] Reference Figure 4 As one implementation of step S104, the steps of calling individual medical attributes from a pre-configured patient profile database, dynamically configuring feature weight allocation strategies, and performing weighted fusion operations on the standardized feature vector set to generate a comprehensive risk score table with weight labels include: Step S401: Call the pre-configured patient profile database to obtain the set of individual medical attribute identifiers associated with the current patient, including at least one of the following: postoperative recovery period identifier, fall risk identifier, respiratory disease identifier, and allergy identifier; The patient profile database is not real-time monitoring data, but rather a knowledge base storing key medical background and historical information about patients. The stored individual medical attribute identifiers, such as postoperative recovery period identifiers, fall risk identifiers, and respiratory disease identifiers, are digital summaries of the patient's current specific clinical stage, past high-risk factors, or underlying disease conditions.

[0076] Specifically, these markers indicate fundamental differences in risk sensitivity among different patient groups. For example, for a patient who has just undergone major surgery, the stability of their vital signs is the primary concern; while for an elderly patient with a history of falls or weak muscles, activities related to getting out of bed are the focus of risk monitoring.

[0077] Step S402: Based on the individual medical attribute identifier set, match the basic weight configuration template from the preset weight strategy library, and dynamically configure the feature weight allocation strategy according to the attribute combination rules. Specifically, the pre-defined weighted strategy library is a set of predefined rules that essentially maps different individual medical attribute identifiers to a set of feature weight coefficients. For example, for patients identified as being in the "postoperative recovery period," the strategy might assign a higher weight (e.g., 0.50-0.60) to "vital sign statistics," because subtle fluctuations in vital signs at this stage are often the earliest and most sensitive signs of complications, and their clinical value is extremely high. At the same time, a moderate weight (e.g., 0.20-0.30) might be assigned to "behavioral frequency characteristics" to monitor the recovery of their activity level, but its urgency is slightly lower than that of vital signs.

[0078] For patients identified as "fall risk," the strategy reverses the weighting, assigning the highest weight (e.g., 0.45-0.55) to "behavioral frequency characteristics" (e.g., frequency and duration of getting out of bed), because preventing accidental falls is the primary safety goal; the weight of vital signs characteristics is correspondingly reduced. For patients with "respiratory diseases," the strategy significantly increases the weight of "environmental assessment characteristics" (e.g., air quality, comfort score) (e.g., 0.35-0.45), because environmental factors have a particularly direct impact on the condition of these patients.

[0079] In this embodiment, the preset weight strategy library stores initial weight templates for a single basic identifier (such as only a "fall risk identifier"), which provides a benchmark reference. However, in actual clinical scenarios, patients often have multiple attribute identifiers, and their risks are not a simple summation of the independent risks of each identifier, but rather involve complex interactions and weighting effects. Therefore, the inventiveness of this technology is mainly reflected in "dynamic configuration according to attribute combination rules".

[0080] For example, a patient exhibiting both "postoperative recovery markers" and "respiratory disease markers" will have significantly higher sensitivity for "vital signs" compared to a patient with only one marker. However, "environmental assessment features" may require careful consideration due to medication effects to avoid over-alarming. The "attribute combination rules" in this step encode this set of rules or algorithmic models of complex clinical experience. By matching and interpreting these rules, the system can dynamically adjust and rebalance the basic weight template, ultimately generating a "feature weight allocation strategy" tailored to the patient that most sensitively reflects their complex risk pattern. This process simulates the thought process of a senior clinical expert making a comprehensive judgment, giving the weight allocation context-awareness and adaptability.

[0081] Step S403: Extract vital sign statistical features, behavioral frequency features, treatment progress features, and environmental assessment features from the standardized feature vector set, and perform weighted fusion calculation according to the feature weight allocation strategy to obtain a comprehensive risk score; The comprehensive risk score is calculated as Σ(eigenvalue × dynamic weight coefficient).

[0082] Specifically, after determining the personalized weighting strategy for the current patient, four types of feature dimensions with clear risk indication significance were extracted from the standardized feature vector set: vital sign statistics quantify physiological stability, behavioral frequency features reflect activity capacity and accident risk, treatment progress features are associated with treatment compliance and effect, and environmental assessment features characterize external safety factors.

[0083] Subsequently, a feature weight allocation strategy that reflects individual specificity is used to perform weighted fusion on the four types of feature values ​​extracted to characterize the patient's current specific state, that is, to sum the products of each feature value and the corresponding dynamic weight coefficient.

[0084] Understandably, features with high weighting will have a greater impact on the final comprehensive score due to fluctuations in their values; features with low weighting will have a relatively limited impact. For example, for postoperative patients, even small changes in vital signs can lead to a significant increase in the comprehensive risk score due to their high weighting; while changes in environmental features, due to their low weighting, have little impact on the overall score. Conversely, for patients with respiratory diseases, the same changes in environmental features will be significantly amplified due to their high weighting.

[0085] For patients identified as high-risk for falls, the weight of their behavioral frequency characteristics is dynamically increased, giving greater influence to behavioral data such as "frequent bed-leaning" or "abnormal toilet-going frequency" in the final score, thus more accurately capturing their primary risks. This dynamic weighting mechanism ensures that the final comprehensive risk score is not a rigid statistical value, but a highly personalized risk quantification indicator that deeply integrates the patient's static attribute profile with dynamic real-time vital signs.

[0086] Step S404: Associate the comprehensive risk score with the patient's identity and bed identification to generate and output a comprehensive risk score table containing risk classification labels.

[0087] Specifically, the calculated abstract score values ​​are strongly bound to specific clinical entities and spatial locations. Associating them with patient identification ensures that the score results belong to a uniquely identified patient individual, achieving precise assessment down to the individual level. Associating them with bed identification locates the risk in physical space, providing direct action guidance for nurses' stations or mobile inspections.

[0088] Finally, a comprehensive risk score table is generated and risk classification labels are added. For example, if the final high-risk score is mainly driven by high values ​​and high weights of behavioral frequency features, the risk classification label can be marked as "behavioral risk"; if it is mainly caused by abnormal vital sign statistical features, it can be marked as "vital sign risk". This risk classification label is a business encapsulation and semantic interpretation of the numerical score.

[0089] The above implementation modifies the traditional static, fixed-weight risk assessment model. Through an attribute-driven dynamic configuration mechanism, the system can intelligently adjust the focus of different assessment dimensions for patients with different traits and risk combinations, thereby improving the accuracy and timeliness of high-risk patient identification. This technical solution formalizes complex clinical experience into rules, ensuring not only the consistency and objectivity of risk assessment results but also providing solid technical support for real-time risk warnings through millisecond-level dynamic response capabilities.

[0090] Reference Figure 5 As one implementation of step S105, the step of inputting the comprehensive risk scoring table into a preset context-aware rule engine, matching preset multi-condition alarm thresholds, and generating a hierarchical alarm instruction set includes: Step S501: Parse the comprehensive risk scoring table to obtain the risk classification labels and comprehensive risk scores contained therein; The risk classification label directly reveals the main contradiction or primary concern of the current risk situation, which determines which specialized clinical response rule base the system should subsequently invoke. For example, if the label is parsed as behavioral risk, the system will focus its decision on rules related to patient activity safety. Simultaneously, the parsing of the comprehensive risk score yields the quantification of the risk level, a continuous variable used for precise threshold determination within the matching rules.

[0091] Step S502: Based on the risk classification label and comprehensive risk score, match the threshold rules in the preset multi-condition alarm rule base to generate the corresponding risk level code; This step is the core of the context-aware rule engine's reasoning. Its logical principle involves comparing and logically calculating the risk type and quantitative score against a rule knowledge base that encodes rich clinical experience and operational procedures, ultimately outputting a risk level code representing urgency and priority. The pre-defined multi-condition alarm rule base is not a simple single-condition rule like "alarm if the score is greater than X," but rather includes "multi-condition" rules capable of characterizing complex clinical situations.

[0092] For example, a "behavioral risk rule" might be defined as "triggered when the behavioral frequency score is >0.6 and the environmental score is <0.4." The "and" indicates that the patient is not only frequently active (high behavioral score), but the environment may also be dimly lit or noisy (low environmental score), and the combination of these two factors significantly increases the risk of falls. This multi-condition rule reduces false alarms caused by short-term fluctuations in a single indicator.

[0093] Understandably, the rule engine matches the parsed risk classification labels and scores against the preconditions of all rules in the rule base. When a matching rule is found, it is triggered, and a risk level code is generated based on the rule's definition. This level code is a discrete level symbol (such as Level 1, Level 2), mapping consecutive risk scores to a finite, defined priority ladder. Level 1 typically represents the most urgent, life-threatening alert, while subsequent levels represent other risks with decreasing urgency but requiring attention.

[0094] In some embodiments, the risk level coding is defined as: Level 1 (triggered by vital sign risk rule), Level 2 (triggered by behavioral risk rule or treatment risk rule), and Level 3 (triggered by environmental risk rule).

[0095] Step S503: Based on the risk classification label, call the preset text mapping rule library to generate the trigger reason description text; The system incorporates a pre-defined text mapping rule base, which maps different risk classification labels to concise, professional clinical terms. For example, when the risk classification label is "treatment risk," the system automatically generates the text "insufficient remaining fluid volume"; when the label is "vital sign risk," it may generate "abnormal blood pressure fluctuations" or "decreased blood oxygen saturation." This step translates the complex backend processes of multi-source data fusion, weighted calculation, and rule matching into easily understandable problem statements familiar to nurses in clinical work. This allows nurses to have a preliminary assessment and mental preparation for potential problems while on their way to the target bed, thereby improving response efficiency and the targeted nature of preparation.

[0096] Step S504: Associate the bed identifiers in the comprehensive risk scoring table to generate target bed location information; Specifically, bed identifiers are extracted from the comprehensive risk scoring table and enriched into precise location information that can be used for navigation, ensuring that decision-making instructions are spatially executable. Association means using this bed identifier as an index to query a richer spatial information database. In this embodiment, the target bed location information includes ward number, bed sequence number, ward three-dimensional spatial coordinates, and the distance to the nearest nurse station.

[0097] Specifically, the ward number (e.g., 305) and bed number (e.g., 2) are parsed from the "bed identifier," which is the basic logical location. Further, the "ward 3D map database" can be accessed to map the bed identifier to specific 3D spatial coordinates (X, Y, Z) in the building information model. In addition, the distance from the bed to the nearest nurse station or emergency equipment point can be calculated. The final generated target bed location information is a composite spatial information package, which makes the alarm command no longer an abstract signal, but an event point with precise coordinates in the physical world of the ward.

[0098] Step S505: Combine the risk level code, trigger cause description text, and target bed location information into a structured data packet and output a graded alarm instruction set.

[0099] Among them, the risk level code determines the order and urgency of the instruction in the subsequent nursing task queue; the trigger reason description text provides nurses with an immediate action context; and the target bed location information indicates the destination of the action.

[0100] In the above implementation, a comprehensive risk scoring table is matched with a pre-set multi-condition alarm rule base capable of depicting complex clinical scenarios, enabling accurate identification of situations involving coupled multiple risk factors and generating clear risk level codes. Based on this, the readability and operational guidance of alarm information are improved by automatically converting machine-readable risk classification labels into trigger cause description text that medical staff can intuitively understand. Simultaneously, by deeply associating bed identifiers and generating rich target bed location information, a solid foundation is laid for subsequent spatial visualization and path navigation. Finally, the risk level, cause description, and precise location information are combined into a hierarchical alarm instruction set for output, improving the timeliness, accuracy, and efficiency of nursing response.

[0101] Reference Figure 6 As a further implementation of the hospital ward visualization management method, after the step of generating a dynamic focused view of the target bed and a nursing task queue based on the hierarchical alarm instruction set driving the visualization system, the method further includes: Step S601: Based on the risk level distribution in the nursing task queue, calculate the comprehensive load index of each nurse zone in real time; Each task in the nursing task queue is accompanied by a risk level distribution (such as high risk, medium risk, low risk) determined by a comprehensive risk score.

[0102] In this embodiment, the calculation of the comprehensive workload index is not a simple statistical count of tasks, but rather an execution of a "risk-resource consumption" coupled model. This model assigns different basic workload weights to tasks with different risk levels (e.g., a weight of 5 for high-risk tasks, representing the high level of focus, complex operation, and emergency response characteristics required; 3 for medium-risk tasks; and 1 for low-risk tasks). These weights are derived from the analysis of historical nursing operation time costs, skill requirements, and psychological stress data.

[0103] Furthermore, the system analyzes in real time the total weight of all pending tasks within each nurse's zone (i.e., a pre-defined physical bed area managed by a specific nurse or team), such as the comprehensive load index = Σ(task risk value × urgency coefficient). The system also introduces an urgency decay factor for each task; after a task is generated, its load contribution increases non-linearly over time (simulating the risk of changes in patient condition) until it is marked as "in progress" or "completed," at which point its contribution drops to zero. Therefore, the comprehensive load index is a scalar that dynamically changes over time. It accurately quantifies the peak instantaneous workload faced by nurses in each zone, comprised of unmanaged high-risk tasks, providing an objective, data-driven basis for subsequent cross-zone resource balancing decisions, rather than relying on the subjective experience of nursing supervisors.

[0104] Step S602: Call the target bed location information of the hierarchical alarm instruction set and construct a spatiotemporal density heat map of high-risk areas in the ward by combining it with historical alarm data; Among them, the target bed location information in the graded alarm instruction set provides the precise spatial coordinates of the current high-risk event (such as the bed number being mapped to two-dimensional plane coordinates).

[0105] Specifically, the core technology for constructing a spatiotemporal density heatmap lies in risk field modeling. The system not only processes current alarm points but also incorporates historical alarm data, including all recent (e.g., the past few hours) alarm events in the calculation based on their occurrence time, spatial location, and alarm level. Each historical alarm event is assigned an influence weight that decays exponentially over time, with more recent events having a greater influence.

[0106] Next, a spatial kernel density estimation algorithm (such as Gaussian diffusion centered on each alarm point) is used to transform each discrete alarm event into a spatially continuous "risk halo" with intensity decreasing from the center outwards. By overlaying and fusing all current and historical alarm halos on a digital ward floor plan, a spatiotemporal density heatmap is generated. The color depth (e.g., from blue to red) visually represents the comprehensive risk density of different geographical areas within a specific time window. High-density red areas not only indicate currently high-risk beds but may also reveal systemic risk hotspots such as nursing blind spots, equipment-intensive areas, or areas with clusters of different patient types. This heatmap provides spatial optimization guidance for resource allocation from a macro perspective.

[0107] Step S603: Obtain real-time location data and task execution status flags uploaded by the nurse terminal, and generate a nurse resource distribution map with availability tags; Among these technologies, real-time location data uploaded by nurses' terminals is acquired through Internet of Things (IoT) technologies (such as UWB and Bluetooth beacons), enabling continuous and high-precision tracking of each nurse's specific location within the ward. Simultaneously, the "task execution status flags" uploaded by the terminals reflect the nurses' current work context, such as "idle," "on the way," "nursing," "document processing," and "shift handover."

[0108] In this embodiment, the process of generating a "nurse resource distribution map with availability labels" is a process of multi-source information fusion and state reasoning. The system associates location coordinates, state flags, and nurses' skill profile databases (such as whether they possess specialized qualifications like intravenous puncture and cardiopulmonary resuscitation). Based on a set of predefined rules or a lightweight prediction model, the system calculates a dynamic availability label for each nurse.

[0109] For example, a nurse whose status is "In Care" might have their "estimated availability time" predicted by the system based on the type of their current task and average time consumption, and be marked as "Busy - Available in X minutes." A nurse whose status is "Idle" and who is located at the nurses' station would be marked as "Immediately Available - Fully Capable." This distribution map is essentially a vector layer, where each nurse represents a data point containing multi-dimensional attributes such as spatial coordinates, dynamic availability status, estimated availability time, and skill set, providing a precise and real-time human resource inventory for global optimized scheduling.

[0110] Step S604: Input the comprehensive load index, the spatiotemporal density heat map of high-risk areas in the ward, and the nurse resource distribution map into the reinforcement learning model, iteratively optimize the task allocation strategy, and output a dynamic nurse scheduling instruction set. Specifically, the outputs of the preceding steps (a comprehensive load index that quantifies the pressure of different regions, a spatiotemporal density heat map that depicts the geographical distribution of risks, and a nurse resource distribution map that characterizes the real-time status of human resources) are used as the "state observation space" of the reinforcement learning model, so as to achieve real-time optimal decision-making under multiple objectives and constraints through the model.

[0111] In the embodiments of this application, the "action space" of a reinforcement learning model (e.g., using a proximal policy optimization PPO or a deep deterministic policy gradient DDPG algorithm) is defined as a combination of decisions that assign a specific nursing task to a specific nurse.

[0112] Furthermore, the model is iteratively optimized through continuous interaction with the environment (i.e., real ward operation simulation or historical data playback). Its reward function is designed to balance multiple key objectives, including maximizing the timely handling rate of high-risk tasks (corresponding to rapid response in high-density areas of the heat map), minimizing the load imbalance between nurses' zones (balancing the comprehensive load index), minimizing the ineffective movement distance and time of nurses (combining location information in the resource distribution map), and ensuring the matching degree between tasks and nurses' skills.

[0113] Next, through millions of trial-and-error learning iterations, the model gradually masters strategies for making globally optimal or near-optimal scheduling decisions in complex and dynamically changing ward conditions. The final output of the dynamic nurse scheduling instruction set is a series of structured instructions, such as "Assign nurse A (located at coordinate X, skill matching) to immediately proceed to bed B (high risk, located in a high-density area on the heat map) to handle high-risk task C, which is expected to alleviate the load index in area D."

[0114] Step S605: Update the nursing task queue sorting in the visualization system according to the dynamic nurse scheduling instruction set.

[0115] Specifically, based on the dynamic nurse scheduling instruction set, the nursing task queue in the visualization system is prioritized and its status is updated synchronously.

[0116] Specifically, the system parses scheduling instructions, marks assigned tasks as "assigned" in the public queue, and attaches the executor information. Simultaneously, it may adjust the display order of other unassigned tasks based on new global optimization strategies. For example, when the workload of a task's assigned area increases sharply, or its location is identified as a new high-risk area, the urgency of the task will be dynamically reassessed by the system, and its ranking in the nursing queue will be moved up accordingly. The updated queue is pushed to all relevant personnel in real time through a visualization system (such as the nurse station screen or mobile PDA). This update ensures that the optimal strategy calculated in the previous step can be instantly transformed into a clear and intuitive work plan to guide nurses' actions, thereby continuously and dynamically optimizing the allocation efficiency of ward nursing resources.

[0117] In the above implementation, the risk pressure inherent in nursing tasks is transformed into a dynamic comprehensive load index, realizing the scientific quantification and visualization of workload; furthermore, by integrating spatiotemporal information into a high-risk area spatiotemporal density heat map, discrete alarm events are upgraded to a continuous field that reveals the distribution pattern of risks, giving the system a forward-looking early warning capability; at the same time, with the help of the nurse resource distribution map, the precise digitization of human resources and real-time perceptibility of their status are realized.

[0118] Building upon this foundation, by introducing a reinforcement learning model as the intelligent scheduling hub, multiple complex objectives such as zonal load balancing, rapid risk response, optimal path efficiency, and matching of personnel to job skills can be comprehensively weighed, resulting in a globally optimized dynamic nurse scheduling instruction set. Ultimately, through the dynamic updating of the nursing task queue order using this instruction set, the response speed and handling rate of high-risk nursing events are improved, nurse workload is effectively balanced, and unnecessary travel distances are reduced. This enhances patient safety and nursing quality while also optimizing nurses' work experience and efficiency, forming a highly adaptive and scalable real-time resource scheduling solution for smart wards.

[0119] Reference Figure 7 As a further implementation of the hospital ward visualization management method, the steps of inputting the comprehensive load index, the spatiotemporal density heat map of high-risk areas in the ward, and the nurse resource distribution map into the reinforcement learning model, iteratively optimizing the task allocation strategy, and outputting a dynamic nurse scheduling instruction set include: Step S701: Integrate the comprehensive load index, the spatiotemporal density heat map of high-risk areas in the ward, and the nurse resource distribution map to generate a multi-dimensional input feature vector set; Specifically, the system preprocesses and encodes three types of data: the load index vector is directly used as a feature component; the heat map matrix is ​​converted into a vector representing local and global risk patterns through spatial grid expansion or convolution feature extraction; and the nurse resource distribution map is parsed, encoding the location coordinates, real-time status indicators (such as busy or idle), skill tags, and other information of each nurse into a fixed-length feature vector.

[0120] Subsequently, tensor concatenation or feature stacking techniques are used to align and concatenate these processed vectors along the feature dimension, forming a unified, high-dimensional, multi-dimensional input feature vector set. This process not only unifies the data format but, more importantly, preserves key information contained in the original data, such as spatial relationships (from heatmaps), pressure distribution (from load indices), and resource topology (from resource distribution maps), through feature combination.

[0121] Step S702: Construct a reinforcement learning state space based on a multi-dimensional input feature vector set, which includes load index vector components, heat map matrix components, and resource distribution map components; This step involves mapping the state space to clinical semantics, transforming raw data into a machine-understandable representation of the decision-making environment. The state space, a core concept in reinforcement learning, defines the dimensional framework for the model's perception of the system's dynamics. The construction process can employ a component decoupling encoding strategy. Load index vector components: retain the original vector structure and directly map to the pressure dimension of the state space (such as the load values ​​of partitions 1-10 corresponding to dimensions 1-10). Heatmap matrix components: The two-dimensional matrix is ​​expanded into a continuous vector in row-major order (e.g., dimensions 11-1034 correspond to heatmap pixel values), and spatial location encoding is injected (e.g., row and column indices are appended to feature values). Resource distribution map components: Extract key attributes (coordinates, availability, skills) of nurse vectors as independent state variables (e.g., dimensions 1035-1234), and process discrete labels through one-hot encoding.

[0122] It should be noted that the design of the state space follows the principle of clinical interpretability: each dimension corresponds to an actual physical meaning (e.g., the 5th dimension of the state vector = zone 5 load, the 300th dimension = risk density of ward (3,4) location), making policy decisions transparent. In addition, the state space introduces temporal difference coding, using the difference between the current state and the previous state as an additional dimension to capture the dynamic change trend of the system.

[0123] Step S703: Call the preset policy evaluation function to calculate the expected return value of the current task allocation policy based on the reinforcement learning state space, which includes a weighted combination of risk handling time efficiency gain and path cost penalty. Among them, the strategy evaluation function is a pre-designed set of mathematical criteria used to evaluate the long-term expected benefits, i.e., the expected return value, of adopting a certain task allocation strategy (i.e., a set of decisions on "which tasks to assign to which nurses") in a specific state space.

[0124] In this embodiment, the strategy evaluation function uses the Bellman equation to calculate the state value, and the expected return is specified as a weighted combination of risk management timeliness gain and path cost penalty. Risk management timeliness gain aims to maximize patient safety benefits; it quantifies the value of rapidly responding to high-risk tasks. For example, a high-risk task involving an ECG monitoring alarm has a high gain if handled quickly, while delayed handling may lead to a worsening of the condition, resulting in zero or even negative gain. Path cost penalty aims to optimize operational efficiency; it quantifies the time and physical cost incurred by nurses due to ineffective movement caused by scheduling decisions.

[0125] By combining these two factors in a weighted manner, the policy evaluation function establishes a manageable balance between the sometimes conflicting objectives of "patient safety" and "nurse efficiency." The model learns through this function that the optimal policy does not always assign the nearest task to the nearest nurse, but rather makes a globally optimal or near-optimal allocation decision after comprehensively weighing task urgency, nurse skill matching, and movement costs.

[0126] Step S704: Execute the strategy improvement function to iteratively update the parameter weights of the current task allocation strategy based on the expected return value, and generate an optimized task allocation strategy. Among them, the policy improvement function is the algorithm engine that drives the update of model parameters (such as policy gradient method, proximal policy optimization, etc.), and its working principle is based on the interactive learning loop of trial and error and feedback.

[0127] Specifically, in the current state space, the model (agent) outputs a scheduling action (such as assigning nurse A to handle task X) based on its current task allocation strategy (defined by a series of internal parameter weights). After this action is executed in a simulated environment or a real system, it produces a result, and the policy evaluation function calculates the immediate and long-term expected reward value brought by the action. The policy improvement function then uses this reward value as a feedback signal to iteratively update the internal parameter weights of the model through complex mathematical optimization algorithms (such as gradient ascent).

[0128] If a certain decision pattern (e.g., "prioritizing the allocation of high-risk tasks to nurses with matching skills and close proximity") yields a high reward, then the weight of the parameter reinforcing this decision pattern will be increased; conversely, the weight of decision patterns that lead to low rewards or penalties will be weakened. Through massive iterative updates, the model's decision strategy is continuously adjusted and optimized, gradually approaching the optimal or near-optimal task allocation strategy that can stably output high-reward scheduling schemes under various complex conditions.

[0129] Step S705: Analyze the optimized task allocation strategy, calculate the mapping relationship between nurse identifiers and nursing task identifiers, and generate a dynamic nurse scheduling instruction set.

[0130] The optimized task allocation strategy is usually represented by a probability distribution or value function within the model. It indicates the advantages and disadvantages of various allocation methods under a given state. The analysis is to decode this strategy into a deterministic assignment relationship.

[0131] In some embodiments, the mapping relationship between nurse identifiers and nursing task identifiers can be solved using an allocation algorithm (such as the Hungarian algorithm, priority-based greedy matching, etc.). This algorithm aims to maximize overall reward or conform to a strategy-oriented approach, finding the most suitable nurse identifier (such as "NURSE-025", uniquely identifying a nurse) for each pending nursing task identifier (e.g., "TASK-202503151103", uniquely identifying a specific nursing task). Finally, a dynamic nurse scheduling instruction set is generated, which is a series of structured instructions, such as "Instruction 001: Dispatch nurse [NURSE-025] to perform task [TASK-202503151103] before [expected arrival time]".

[0132] Understandably, this scheduling instruction set is dynamic because it is continuously generated and updated through the cyclical execution of the above steps as the state space (load, risk, resources) changes in real time.

[0133] In the above implementation, a comprehensive load index characterizing work pressure, a spatiotemporal density heatmap depicting risk profiles, and a nurse resource distribution map reflecting human resource conditions are deeply integrated and represented at a higher level to construct a state space capable of accurately depicting the complex situation in the ward. Furthermore, a strategy evaluation function balancing the timeliness of risk management and path costs establishes a quantitative value orientation for scheduling decisions. The reinforcement learning mechanism, through continuous interaction and iteration, enables the scheduling strategy to evolve autonomously, infinitely approaching the globally optimal solution while ensuring patient safety.

[0134] Ultimately, the system parses the abstract optimization strategy into a specific and executable set of dynamic nurse scheduling instructions in real time, improving the response speed and accuracy of handling high-risk nursing events, scientifically balancing nurses' workload, and effectively reducing unnecessary movement and resource idleness. Thus, while significantly improving patient safety and clinical nursing quality, it optimizes the efficiency of human resource allocation and nurses' work experience, building an adaptive, evolvable, and globally optimal real-time dynamic scheduling core engine for modern smart hospitals.

[0135] Reference Figure 8 As a further implementation of the hospital ward visualization management method, the management method also includes: Step S801: Parse the data source identifier and protocol version number in the tagged data packet to generate a heterogeneous data source lineage map; The input tagged data packets embed data source identifiers (such as "bedside monitor-bed 7-model XYZ") and "protocol version numbers" (such as "HL7 v2.6"), which are the digital genes of the data packets. The process of generating heterogeneous data source lineage maps is to use graph theory models to abstractly model complex systems.

[0136] In this embodiment, the system abstracts each independent data source (such as a specific medical device, clinical information system, or electronic medical record server) as a node in a graph. The flow, transformation, and dependencies of data between different sources (e.g., vital sign data being transmitted from a monitor to a data integration platform and then accessed by a clinical decision support system) are abstracted as edges connecting these nodes. By parsing the protocol version number, the weight or attributes of the edges can be further quantified; for example, data streams using newer protocol versions may be assigned higher credibility weights.

[0137] Based on this, the resulting heterogeneous data source lineage map is a dynamic and visualized relationship network that clearly reveals which systems within the hospital generate data, how the data flows, and the dependencies and influences between different data sources.

[0138] Step S802: Associate the hierarchical alarm instruction set with the heterogeneous data source lineage map, perform alarm source tracing analysis, and output the alarm-data source mapping relationship table; When a system generates a tiered alarm command (such as a "high-risk low blood pressure alarm"), traditional systems often only know the alarm content but not which abnormal data from which data source triggered it. This step aims to establish such a precise causal chain by associating the alarm command with the heterogeneous data source lineage map and performing alarm source tracing analysis.

[0139] Specifically, the core technology of this step is to perform reverse graph traversal or causal reasoning algorithms along the edges of the pedigree graph. Starting from the final data feature or event that triggers the alarm, the system traces the generation and calculation process of the data feature in reverse according to the transformation and dependency relationships recorded in the pedigree graph, tracing back layer by layer until it locates the most original one or several data sources and the original data points they generated. This allows the system to find the source (the specific device or system that generated the abnormal reading) through clues (data flow path).

[0140] Ultimately, the output alarm-data source mapping table is the result document of this traceability analysis. It clearly records the relationship between each alarm and its underlying data source in a structured manner (such as recording alarm ID, associated data source identifier, confidence level, etc.).

[0141] Step S803: Based on the alarm-data source mapping relationship table and the execution completion signal fed back by the nurse terminal, calculate the alarm effectiveness index corresponding to each data source; The alarm effectiveness index is a comprehensive indicator that measures the clinical value and credibility of alarms triggered by a data source. The calculation relies on two key inputs: first, an alarm-data source mapping table, which specifies the objects to be evaluated (i.e., which alarms are handled by which data source); and second, execution completion signals from nurse terminals, which are direct feedback from the clinical frontline regarding the authenticity and value of the alarm. For example, a nurse confirms and processes an alarm, or marks it as a "false alarm."

[0142] In this embodiment, the system counts all alarms triggered by each data source and classifies these alarms based on nurses' feedback signals (e.g., valid true positives, invalid false alarms, no response, etc.). The alarm validity index is a quantitative function result that integrates multiple dimensions. It not only considers the proportion of alarms verified as true and valid by nurses (i.e., positive prediction value) out of the total number of alarms triggered by the data source, but may also incorporate factors such as alarm response time, success rate of nurses' handling results, and confidence of data flow path obtained through bloodline analysis for weighted calculation.

[0143] Specifically, the formula for calculating the alarm effectiveness index is as follows: ; Where N_total represents the total number of alarms triggered by the data source, and N_valid represents the number of valid alarms confirmed by nurses; the weighting coefficient α can be configured to 0.6, the weighting coefficient β can be configured to 0.3, and the clinical tuning parameter γ can be configured to 0.1. Σ confidence represents the sum of the path confidence of all confirmed valid alarms (N_valid) traced back to the original data source through the heterogeneous data source lineage map.

[0144] Understandably, the alarm effectiveness index is not a simple statistical ratio, but a comprehensive quality score that integrates accuracy, timeliness, and clinical acceptance. A data source with a high alarm effectiveness index means that the alarms it generates are accurate and reliable, and can effectively drive clinical intervention; conversely, a low index means that the data source has high noise and a high false alarm rate, which is one of the main causes of alarm fatigue.

[0145] Step S804: Dynamically adjust the weight allocation parameters of the data type priority matrix according to the alarm effectiveness index; Among these, the data type priority matrix is ​​a key control parameter table that defines the resource priority and weight that different sources and types of data can occupy in subsequent processing links (such as computation, compression, and transmission). In traditional static systems, the weights of this matrix are preset and fixed. This step, however, dynamically adjusts the weights of this matrix by using an alarm effectiveness index that reflects the reliability of each data source in real time.

[0146] Specifically, for a data source with a consistently high alarm validity index (such as a well-maintained and accurately calibrated monitor), the system automatically increases the weight of its data in the priority matrix. This means that the data stream generated by this data will receive higher computational priority, more sufficient transmission bandwidth, or higher quality compression guarantees in subsequent processing, thereby ensuring that high-value information can be delivered to the decision-making end faster and with higher fidelity. Conversely, for a data source with a low validity index (such as a device with aging sensors that frequently produces artifacts), the system dynamically lowers its weight, limiting the impact of its low-quality data on limited system resources (such as network bandwidth, storage, and processor computing power). This dynamic adjustment mechanism allows the entire system to continuously focus its superior resources on high-quality, high-value data streams, achieving optimal resource input and output.

[0147] Step S805: Feed the adjusted data type priority matrix back to the hierarchical compression and transmission scheduling steps to reconstruct the hierarchical data stream generation logic.

[0148] The system feeds back this updated matrix in real time to the front-end module responsible for data stream processing, namely the "tiered compression and transmission scheduling step." Based on this dynamic priority strategy, this step immediately reconstructs its tiered data stream generation logic, and the data stream processing behavior will adaptively change.

[0149] For example, for data sources that are given high weight, their data streams may use lossless or high-quality compression algorithms and be placed at the front of the transmission queue to ensure low-latency and high-fidelity transmission; for data sources that are given lower weight, their data streams may use lossy compression to save bandwidth or be scheduled for transmission during non-critical periods.

[0150] Understandably, through this reconstruction, the entire system's processing of data streams is no longer static and uniform, but rather becomes a dynamic and intelligent scheduling guided by the real-time clinical value of the data source (measured by the alarm effectiveness index). This forms a complete optimization loop: the quality of the data stream affects alarm effectiveness, the evaluation results of alarm effectiveness, in turn, dynamically adjust the priority strategy for data stream processing, and the optimized strategy further improves the processing quality of high-value data streams, thereby further improving alarm effectiveness. This cycle repeats itself, continuously self-optimizing.

[0151] In the above implementation, a heterogeneous data source lineage map is constructed, breaking down data silos between devices and the system. Through precise alarm source tracing analysis, clinical alarm phenomena are rigidly correlated with specific underlying data sources, achieving precise problem localization. The execution feedback from nurse terminals is introduced as a monitoring signal, and the calculated alarm effectiveness index provides an objective and quantitative clinical measure of the reliability of each data source. Based on this index, the data type priority matrix is ​​dynamically adjusted, enabling the system's resource allocation strategy to self-evolve according to data quality, achieving intelligent focus of resources on high-value data streams. Finally, by feeding the optimized strategy back to the data stream generation logic and reconstructing it, a closed loop from evaluation to execution is completed, ensuring the real-time implementation of the optimized strategy.

[0152] In practical applications, this technical solution not only suppresses alarm fatigue caused by false alarms from low-quality data sources and improves nurses' response efficiency and trust in high-risk alarms, but also effectively improves the utilization efficiency of network bandwidth and computing resources by optimizing data flow scheduling, and builds an adaptive smart healthcare data management ecosystem that can self-evaluate, self-adjust, and continuously evolve towards higher data quality and better clinical efficacy.

[0153] This application also discloses a hospital ward visualization management system based on multi-source data fusion.

[0154] A hospital ward visualization management system based on multi-source data fusion, specifically including: The data acquisition and standardization module is used to collect monitoring data from IoT devices in hospital wards, data from clinical information systems, and data entered by nurses' terminals. It generates a four-dimensional tag for each data entry, which includes patient identification, bed identification, timestamp, and data type, and outputs a tagged data package. The data hierarchical transmission scheduling module is used to perform hierarchical compression and transmission scheduling on tagged data packets according to a preset data type priority matrix, and generate hierarchical data streams; The patient status feature calculation module is used to perform spatiotemporal alignment processing on hierarchical data streams based on patient identity and bed identification, and extract vital sign statistical features, behavioral frequency features, treatment progress features and environmental assessment features, and output a standardized feature vector set. The risk fusion assessment module is used to call individual medical attributes from a pre-configured patient profile database, dynamically configure feature weight allocation strategies, and perform weighted fusion operations on a standardized feature vector set to generate a comprehensive risk score table with weight labels. The context-aware alarm module is used to input the comprehensive risk score table into the preset context-aware rule engine, match the preset multi-condition alarm thresholds, and generate a hierarchical alarm instruction set. The visualization task scheduling module is used to drive the visualization system based on the hierarchical alarm instruction set, generating a dynamic focused view of the target bed and a nursing task queue. The auxiliary execution and feedback module is used to generate navigation paths and operation instructions based on nursing task queues through augmented reality devices, receive execution completion signals from nurse terminals, and update patient status data to the clinical information system.

[0155] The hospital ward visualization management system based on multi-source data fusion according to the embodiments of this application can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.

[0156] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0157] This application also discloses a computer-readable storage medium.

[0158] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the hospital ward visualization management methods based on multi-source data fusion.

[0159] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0160] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for visual management of hospital wards based on multi-source data fusion, characterized in that, The management method includes: Collect monitoring data from IoT devices in hospital wards, data from clinical information systems, and data entered by nurses' terminals, and generate a four-dimensional tag for each data entry containing patient identification, bed identification, timestamp, and data type, and output a tagged data package; Based on a preset data type priority matrix, hierarchical compression and transmission scheduling are performed on the tagged data packets to generate hierarchical data streams; Based on the patient identification and bed identification, the hierarchical data stream is spatiotemporally aligned, and vital sign statistical features, behavioral frequency features, treatment progress features and environmental assessment features are extracted to output a standardized feature vector set. The system calls upon individual medical attributes from a pre-configured patient profile database, dynamically configures feature weight allocation strategies, and performs weighted fusion operations on the standardized feature vector set to generate a comprehensive risk score table with weight labels. The comprehensive risk scoring table is input into a preset context-aware rule engine, matched with preset multi-condition alarm thresholds, and a hierarchical alarm instruction set is generated. The visualization system is driven by the hierarchical alarm instruction set to generate a dynamic focused view of the target bed and a nursing task queue. Based on the nursing task queue, navigation paths and operation instructions are generated through augmented reality devices, and execution completion signals are received from nurse terminals to update patient status data to the clinical information system.

2. The hospital ward visualization management method based on multi-source data fusion according to claim 1, characterized in that, The steps for generating a hierarchical data stream by performing hierarchical compression and transmission scheduling on the tagged data packets according to a preset data type priority matrix include: Receive tagged data packets containing patient identification, bed identification, timestamp, and data type; According to the preset data type priority matrix, each data item in the tagged data packet is marked with a transmission priority level; Based on the transmission priority level matching compression algorithm, compressed data blocks are generated; The compressed data blocks are scheduled according to the transmission cycle corresponding to the priority level, and a hierarchical data stream is output.

3. The hospital ward visualization management method based on multi-source data fusion according to claim 2, characterized in that, Based on the patient identification and bed identification, the steps of performing spatiotemporal alignment processing on the hierarchical data stream, extracting vital sign statistical features, behavioral frequency features, treatment progress features, and environmental assessment features, and outputting a standardized feature vector set include: Receive hierarchical data streams containing patient identification, bed identification, timestamps, and data types; Based on the patient identification and bed identification, the hierarchical data stream is spatiotemporally aligned to generate a time-series data set grouped by bed identification. Vital sign data are obtained from the time series data set, and statistical feature values ​​within a preset time window are calculated to generate vital sign statistical features. The behavior monitoring data is parsed from the time series data set, and the frequency of action triggers per unit time is calculated to generate behavior frequency features; Treatment progress data is extracted from the time-series data set, and the remaining treatment time is calculated to generate treatment progress features; Environmental parameter data are extracted from the time-series data set, and environmental comfort scores are calculated by weighted combination to generate environmental assessment features. The vital signs statistical features, behavioral frequency features, treatment progress features, and environmental assessment features are combined into a feature vector set with unified dimensions, and then standardized scaling processing is performed to output a standardized feature vector set.

4. The hospital ward visualization management method based on multi-source data fusion according to claim 1, characterized in that, The steps of calling individual medical attributes from a pre-configured patient profile database, dynamically configuring feature weight allocation strategies, and performing weighted fusion operations on the standardized feature vector set to generate a comprehensive risk score table with weighted labels include: Call the pre-configured patient profile database to obtain the set of individual medical attribute identifiers associated with the current patient, including at least one of the following: postoperative recovery period identifier, fall risk identifier, respiratory disease identifier, and allergy identifier; Based on the individual medical attribute identifier set, a basic weight configuration template is matched from the preset weight strategy library, and a feature weight allocation strategy is dynamically configured according to the attribute combination rules. Extract vital sign statistical features, behavioral frequency features, treatment progress features, and environmental assessment features from the standardized feature vector set, and perform weighted fusion operation according to the feature weight allocation strategy to obtain a comprehensive risk score; The comprehensive risk score is associated with the patient's identity and bed number to generate and output a comprehensive risk score table containing risk classification labels.

5. The hospital ward visualization management method based on multi-source data fusion according to claim 4, characterized in that, The steps of inputting the comprehensive risk scoring table into a preset context-aware rule engine, matching it with preset multi-condition alarm thresholds, and generating a tiered alarm instruction set include: The comprehensive risk scoring table is analyzed to obtain the risk classification labels and comprehensive risk scores contained therein; Based on the risk classification labels and comprehensive risk scores, the threshold rules in the preset multi-condition alarm rule base are matched to generate the corresponding risk level code. Based on the risk classification label, a preset text mapping rule library is invoked to generate a description text of the triggering reason; By associating the bed identifiers in the comprehensive risk scoring table, target bed location information is generated; The risk level code, trigger reason description text, and target bed location information are combined into a structured data packet, and a graded alarm instruction set is output.

6. A hospital ward visualization management method based on multi-source data fusion according to any one of claims 1 to 5, characterized in that, After the step of generating a dynamic focused view of the target bed and a nursing task queue based on the hierarchical alarm instruction set driving the visualization system, the following is also included: Based on the risk level distribution in the nursing task queue, the comprehensive load index of each nurse zone is calculated in real time. The target bed location information of the hierarchical alarm instruction set is invoked, and a spatiotemporal density heat map of high-risk areas in the ward is constructed by combining historical alarm data. Acquire real-time location data and task execution status flags uploaded by nurses' terminals, and generate a nurse resource distribution map with availability tags; The comprehensive load index, the spatiotemporal density heat map of high-risk areas in the ward, and the nurse resource distribution map are input into the reinforcement learning model to iteratively optimize the task allocation strategy and output a dynamic nurse scheduling instruction set. Update the nursing task queue sorting in the visualization system according to the dynamic nurse scheduling instruction set.

7. A hospital ward visualization management method based on multi-source data fusion according to claim 6, characterized in that, The steps of inputting the comprehensive load index, the spatiotemporal density heat map of high-risk areas in the ward, and the nurse resource distribution map into the reinforcement learning model, iteratively optimizing the task allocation strategy, and outputting a dynamic nurse scheduling instruction set include: By integrating the comprehensive load index, the spatiotemporal density heat map of high-risk areas in the ward, and the nurse resource distribution map, a multi-dimensional input feature vector set is generated. A reinforcement learning state space is constructed based on the multidimensional input feature vector set, which includes load index vector components, heat map matrix components, and resource distribution map components. The preset strategy evaluation function is invoked to calculate the expected return value of the current task allocation strategy based on the reinforcement learning state space, which includes a weighted combination of risk handling time efficiency gain and path cost penalty. The execution strategy improvement function iteratively updates the parameter weights of the current task allocation strategy based on the expected return value, and generates an optimized task allocation strategy. The optimized task allocation strategy is analyzed, the mapping relationship between nurse identifiers and nursing task identifiers is calculated, and a dynamic nurse scheduling instruction set is generated.

8. The hospital ward visualization management method based on multi-source data fusion according to claim 1, characterized in that, The management method also includes: Parse the data source identifier and protocol version number in the tagged data packet to generate a heterogeneous data source lineage map; Associate the hierarchical alarm instruction set with the heterogeneous data source lineage map, perform alarm source tracing analysis, and output an alarm-data source mapping relationship table; Based on the alarm-data source mapping table and the execution completion signal fed back by the nurse terminal, calculate the alarm effectiveness index corresponding to each data source; Based on the alarm effectiveness index, the weight allocation parameters of the data type priority matrix are dynamically adjusted. The adjusted data type priority matrix is ​​fed back to the hierarchical compression and transmission scheduling steps to reconstruct the hierarchical data stream generation logic.

9. A hospital ward visualization management system based on multi-source data fusion, characterized in that, The management system includes: The data acquisition and standardization module is used to collect monitoring data from IoT devices in hospital wards, data from clinical information systems, and data entered by nurses' terminals. It generates a four-dimensional tag for each data entry, which includes patient identification, bed identification, timestamp, and data type, and outputs a tagged data package. The data hierarchical transmission scheduling module is used to perform hierarchical compression and transmission scheduling on the tagged data packets according to a preset data type priority matrix to generate hierarchical data streams; The patient status feature calculation module is used to perform spatiotemporal alignment processing on the hierarchical data stream based on the patient identity identifier and bed identifier, and extract vital sign statistical features, behavioral frequency features, treatment progress features and environmental assessment features, and output a standardized feature vector set. The risk fusion assessment module is used to call individual medical attributes from a pre-configured patient profile database, dynamically configure feature weight allocation strategies, and perform weighted fusion operations on the standardized feature vector set to generate a comprehensive risk score table with weight labels. The context-aware alarm module is used to input the comprehensive risk score table into a preset context-aware rule engine, match preset multi-condition alarm thresholds, and generate a hierarchical alarm instruction set. The visualization task scheduling module is used to drive the visualization system according to the hierarchical alarm instruction set to generate a dynamic focused view of the target bed and a nursing task queue; The auxiliary execution and feedback module is used to generate navigation paths and operation instructions based on the nursing task queue through augmented reality devices, receive execution completion signals from nurse terminals, and update patient status data to the clinical information system.

10. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images