An RFID-based intelligent monitoring and analysis management system for inpatients
By constructing a risk area association map and dynamically adjusting the risk area trigger weights, combined with RFID positioning and the patient's pathological state, the problem of lagging location tracking for patients with cognitive impairment in existing technologies has been solved. This enables early identification of potential dangers and personalized intervention, thereby improving the safety management level of hospitalized patients.
Patent Information
- Application Number
- CN202611097545.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-23
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, location tracking systems for hospitalized patients with cognitive impairment have long response times, making it difficult to adjust patrol priorities in a timely manner. Furthermore, they lack clear judgment on patients' behavioral trends, which may cause patients to continue moving after an alarm is triggered, increasing the risk of them accidentally entering dangerous areas.
By constructing a risk area association map, combining patients' cognitive impairment classification data and real-time physiological parameters, calculating the pathological drift baseline value, dynamically adjusting the risk area trigger weight, using RFID positioning data to map patient trajectories, employing MCMC to predict future trajectory distribution, and correcting the transfer probability through a Bayesian incremental strategy, we can achieve early identification and intervention of potential risks to patients.
It enables personalized risk assessment and dynamic adaptation for hospitalized patients, improves the ability to identify potential dangers in advance, reduces the real-time inspection pressure on nursing staff, and enhances the level of intelligence in patient safety management.
Smart Images

Figure CN123120225A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent medical monitoring technology, specifically an RFID-based intelligent monitoring, analysis and management system for hospitalized patients. Background Technology
[0002] The safety management of hospitalized patients is an important part of hospital nursing work. For hospitalized patients with cognitive impairments such as Alzheimer's disease, due to their impaired orientation and memory, they are prone to getting lost in the complex hospital environment and may accidentally enter stairwells, equipment rooms or hospital exits.
[0003] Currently, RFID technology has been applied to the management of hospitalized patients. In related systems, UWB positioning technology can achieve centimeter-level positioning accuracy, and Bluetooth Angle of Arrival (AOA) technology can also provide sub-meter-level positioning services. The above positioning methods can deploy corresponding readers or base stations in the hospital and have patients wear positioning tags to achieve real-time tracking of the patient's location. When the system detects that a patient has entered a preset electronic fence area, it will trigger an alarm signal and notify nursing staff to assist in timely intervention.
[0004] However, in actual clinical applications, nursing staff typically require a certain response time from receiving an alarm signal to arriving at the scene. Furthermore, patients with cognitive impairment may continue to move after the alarm is triggered. Relevant statistics show that during the window period between the alarm being triggered and the arrival of nursing staff, the average distance the patient moves can reach several meters to more than ten meters. This distance is sufficient for the patient to cross corridor intersections or enter adjacent functional areas, thus deviating from the expected location when the nursing staff arrives. In addition, before the alarm is triggered, nursing staff usually have difficulty obtaining reference information related to the patient's current behavioral trends. During routine rounds, they lack clear judgment criteria regarding the risk of the patient approaching dangerous areas, making it difficult to adjust the focus of rounds and allocate attention in a timely manner. Summary of the Invention
[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.
[0006] The technical solution adopted by this invention to solve its technical problem is: Risk Zone Networking Module: Collects RFID reader deployment information to divide risk zones. Based on the alarm frequency, nursing intensity, and dwell time of each zone, the three data are normalized and weighted and summed. The risk level is assessed according to the threshold. The association strength is calculated based on the physical connectivity of the zones, access control logic, and path dependence. The risk zone association graph is constructed with the zones as nodes and the association strength as edges. Pathological anchoring module: Acquires patient cognitive impairment classification data and real-time physiological parameters, extracts pathological features through time-series slice filtering, calculates the degradation rate of orientation, memory and executive function dimensions as the pathological drift benchmark value, and adaptively adjusts the trigger weight of risk area based on this. Historical trajectory nodes are compared with weights one by one, and the median of historical trigger frequency is used as the threshold for screening to generate a set of exclusive risk-sensitive areas and a weight matrix. The trajectory tracking module collects time-series positioning data through an RFID wristband to construct the patient's movement trajectory, maps it to a risk area association map, calculates the risk value of trajectory nodes by combining a set of exclusive risk-sensitive areas and a weight matrix, generates an individualized risk sequence, and extracts the current risk intensity as the initial distribution constraint for MCMC extrapolation based on the sequence change rate. Risk prediction module: Starting from the current trajectory node, combined with the inter-regional transfer probability and movement delay, the MCMC constrained by the pathological drift benchmark value is used to predict the future trajectory distribution within a preset time window, and the sampling ratio of high-risk areas is dynamically adjusted according to the pathological drift benchmark value to generate probability-weighted paths and screen high-risk paths with a probability of crossing the target danger area exceeding the threshold. The early warning triggering module calculates the remaining arrival time series based on the expected arrival time of the high-risk path and real-time location. It performs first-order differencing on the series to obtain the rate of change and second-order differencing to identify the acceleration trend. When both levels of differencing exceed the threshold, the warning level is matched according to the remaining time and an intervention instruction is pushed. At the same time, based on the patient's actual risk-avoidance results, the transfer probability in the risk area association map is corrected through a Bayesian incremental strategy.
[0007] As a further aspect of the present invention: The process of constructing a risk area association diagram is as follows: Obtain the location coordinates of each RFID reader within the hospital and its corresponding physical space area. Divide the space into multiple spatial interval nodes based on the building's functional zoning and ward unit layout, and associate and bind the readers deployed in each interval with the corresponding nodes. Retrieve historical abnormal event records for each spatial interval, count the number of times patients mistakenly enter, stay for longer than the allotted time, move without authorization, and receive nursing interventions, and combine this with the frequency of nursing staff rounds, the average time patients stay, and the functional attributes of the area to calculate the basic risk score for each spatial interval. Based on the basic risk score, the nodes in the spatial interval are marked with their hazard levels, and spatial transfer edges are established according to the physical connectivity between the regions. Obtain hospital access control status information and patient historical movement path data, and calculate the transfer impact coefficient between nodes in different risk areas based on access control opening conditions, passage direction constraints, and high-frequency patient movement paths; A risk area association matrix is constructed based on the hazard level, spatial transfer edge, and transfer impact coefficient, and a risk area association map is generated based on the association strength between nodes in the matrix.
[0008] The process of extracting pathological features through time-series slice filtering is as follows: Obtain the patient's corresponding cognitive impairment level assessment results, records of previous cognitive function changes, and real-time data on heart rate, movement amplitude, activity cycle, and abnormal behavior response; Real-time physiological parameters are sliced according to a preset time window, dividing the continuous physiological change process into multiple state segments, and an adaptive filtering method is used to remove random fluctuation noise in different time segments. Based on the filtered physiological state segments, physiological stability features, activity rhythm change features, and behavioral response lag features were extracted respectively. By synchronously mapping the characteristics of cognitive impairment level, physiological changes, and behavioral changes over time, a multidimensional pathological feature encompassing cognitive, physiological, and behavioral dimensions is constructed.
[0009] The process of generating a dedicated set of risk-sensitive regions and a weight matrix is as follows: Obtain the patient's orientation score, memory retention score, and executive function score within multiple cognitive assessment cycles; Based on the relationship between the changes in scores of each dimension and the time interval between adjacent assessment periods, the rates of orientation decline, memory decline, and executive function decline were calculated respectively. Based on the degradation rate corresponding to each functional dimension and the preset dimension influence coefficient, a fusion calculation is performed to generate the pathological drift baseline value corresponding to the patient's current stage. Based on the pathological drift baseline value, the regional trigger weights in the risk area association map are scaled and adjusted, and the patient's historical abnormal trajectory is matched with the adjusted regional weights to filter regional nodes that exceed the risk sensitivity threshold, thereby generating a patient-specific risk sensitivity area set. Based on the adjustment weights corresponding to the nodes in the risk-sensitive area set, a patient-specific trigger weight matrix is constructed.
[0010] The process of mapping to the risk area association diagram is as follows: Acquire continuous positioning data generated by the patient wearing an RFID wristband, and parse the location coordinates, acquisition timestamp, and signal strength information corresponding to each positioning data point; Calculate the patient's movement direction, movement distance, and movement speed based on the positional change relationship between adjacent positioning points, and generate the corresponding movement vector parameters; By combining the patient's dwell time at each location point, the consecutive location points are connected in chronological order to form a spatiotemporal state chain that includes the relationship between positional changes and temporal evolution. Map the location nodes in the spatiotemporal state chain to the risk area association graph, and determine the risk area node corresponding to each location node through spatial coordinate matching; Based on the patient's specific risk-sensitive area set, the mapped trajectory nodes are filtered to extract risk trajectory nodes that are associated with the patient's current pathological state.
[0011] The process of extracting the current risk intensity as the initial distribution constraint for MCMC inference based on the sequence change rate is as follows: Based on the node number of the risk area where each trajectory node is located, the corresponding weight parameters in the trigger weight matrix are called, and the risk gain value corresponding to each trajectory node is calculated by combining the node dwell time, movement speed and the basic level of the risk area. The risk gain values of each node are arranged according to the patient's movement time sequence to generate an individualized trajectory risk sequence; Calculate the change in risk gain between adjacent time points, and determine the rate of change of the risk sequence based on the relationship between the change in risk gain and the time interval; Extract the risk change rate corresponding to the end node of the current trajectory and use it as the initial risk intensity of the patient's current tendency to move towards the danger zone.
[0012] The process of extrapolating the future trajectory distribution within a preset time window using MCMC with pathological drift baseline constraints is as follows: The initial state node of the Markov chain is determined based on the spatial location corresponding to the current trajectory node, and the initial danger intensity is converted into the initial state probability distribution. Obtain historical transfer frequency, spatial distance, and patient movement delay information between adjacent nodes in the risk area association map, and calculate the basic state transfer probability between nodes; The pathological matching influence coefficient is calculated based on the patient's pathological drift baseline value. The baseline state transfer probability is dynamically corrected based on the influence coefficient to obtain the pathological adaptive transfer probability. A Markov chain sampling process is constructed based on pathological adaptive transfer probability, and a multi-step trajectory state sequence within a preset time window is generated using Monte Carlo random sampling. The probability of different trajectory state sequences is statistically analyzed to form a future trajectory distribution that includes all candidate trajectory paths under time constraints.
[0013] The process of selecting high-risk routes with a probability exceeding a threshold for traversing a target danger zone is as follows: Based on the future trajectory distribution, extract all candidate trajectory paths within the distribution and the risk area node sequence corresponding to each path; The comprehensive risk weight of each candidate path is calculated by combining the number of risk areas traversed by each path, the danger level of each node, and the time spent in each area. Candidate trajectory paths are probabilistically weighted based on the comprehensive risk weight of the path, generating a set of probabilistically weighted paths with risk probability labels; The cumulative probability of each probability-weighted path crossing the target dangerous area node is calculated, and the cumulative probability is compared with the preset dangerous path screening threshold. Paths exceeding the threshold are identified as high-risk paths.
[0014] The process of obtaining the rate of change by performing first-order differencing on the sequence and identifying the accelerating trend by performing second-order differencing is as follows: The estimated arrival time of the target danger zone node corresponding to the high-risk path is obtained, and the dynamic remaining distance between the current position and the target danger zone is calculated by combining the patient's real-time location coordinates, movement speed and movement direction. The remaining arrival time series is generated based on the spatial change relationship of continuous sampling time. The remaining arrival time series is differentially analyzed by adjacent time points in chronological order. The change in remaining arrival time per unit time is calculated, and the first-order rate of change is determined based on the ratio of the change to the sampling time interval. The first-order rate of change is differentially calculated again within a continuous time window to extract the trend of the first-order rate of change, and the second-order rate of change corresponding to the remaining arrival time series is solved in combination with the time interval. The combined first-order and second-order rates of change characterize the speed change and acceleration trend of the patient's approach to the target danger zone.
[0015] The process of correcting the transition probability in the risk region association graph using a Bayesian incremental strategy is as follows: The system obtains the first-order and second-order rates of change corresponding to the patient's current dangerous approach state. When the first-order rate of change exceeds the preset change threshold and the second-order rate of change exceeds the acceleration threshold, the patient is determined to have entered a rapid approach state. The system then matches the corresponding warning level based on the current remaining arrival time in the remaining arrival time sequence. Based on the warning level, a risk intervention instruction is generated, which includes the patient's current location, target danger area, estimated arrival time and intervention method, and pushed to the nursing terminal. At the same time, the patient's actual movement trajectory, whether the target danger area has been entered and the actual arrival time are collected after the warning is triggered. Based on the discrepancy between the patient's actual risk-avoidance results and the high-risk path prediction results, regional transfer feedback samples are constructed. The original transfer probabilities in the risk area association graph are used as prior probabilities, and Bayesian incremental updates are performed using the newly added feedback samples to calculate the corrected regional node transfer probabilities. The updated transition probabilities are mapped back to the risk area association map, and the sampling weights and trajectory inference parameters of high-risk areas are adjusted synchronously based on the correction results.
[0016] The beneficial effects of this invention are as follows: By integrating RFID spatial positioning, patient pathological state perception, risk area topology modeling, and probabilistic trajectory prediction, the system enables early detection and proactive intervention of abnormal movement behaviors in hospitalized patients, especially those with cognitive impairment. The risk area network module comprehensively analyzes hospital spatial layout, historical risk events, nursing resource distribution, and patient activity patterns to construct a dynamic risk spatial model with regional risk correlations. This transforms risk identification from traditional fixed-area alarms to comprehensive risk assessment based on spatial correlations. The pathological anchoring module incorporates the degree of patient cognitive impairment, changes in physiological state, and pathological drift trends, allowing for personalized adjustments to risk-sensitive areas for different patients. This enables the system to adapt to behavioral pattern shifts caused by changes in patient condition, improving the targeting and dynamic adaptability of risk assessment. The trajectory tracking module maps real-time RFID positioning information to a risk area correlation map and calculates individualized risk sequences based on patient-specific risk weights, enabling the tracking of patient movement processes. This system continuously quantifies potential danger trends to avoid delayed early warnings caused by relying solely on current location. Through a risk prediction module, it uses a MCMC probability prediction mechanism constrained by pathological drift benchmarks to predict the patient's future movement path and dynamically adjusts the sampling ratio of high-risk areas based on changes in the patient's condition, improving the accuracy and effectiveness of risk path prediction. This allows the system to identify potential risks before the patient enters a danger zone. The early warning trigger module performs multi-level change analysis on the remaining arrival time trends in the predicted path, identifying accelerated behavior of patients rapidly approaching danger zones. Combined with actual risk-avoidance results, a Bayesian incremental strategy is used to continuously correct the area transfer probability, giving the risk model self-learning and continuous optimization capabilities. In summary, this technical solution enables a shift from "post-event alarm" to "pre-event prediction and dynamic intervention," improving the ability to detect abnormal patient departures, accidental entry into danger zones, and unexpected events in advance, reducing the real-time patrol pressure on nursing staff, and simultaneously enhancing the intelligence level and risk control accuracy of hospital patient safety management. Attached Figure Description
[0017] The invention will now be further described with reference to the accompanying drawings.
[0018] Figure 1 This is a structural framework diagram of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the method in Embodiment 1 of the present invention. Detailed Implementation
[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0020] Example 1: Please refer to Figure 1As shown in the embodiment of the present invention, an intelligent monitoring and analysis management system for inpatients based on RFID includes: Risk Zone Networking Module: Collects RFID reader deployment information to divide risk zones. Based on the alarm frequency, nursing intensity, and dwell time of each zone, the three data are normalized and weighted and summed. The risk level is assessed according to the threshold. The association strength is calculated based on the physical connectivity of the zones, access control logic, and path dependence. The risk zone association graph is constructed with the zones as nodes and the association strength as edges.
[0021] In the risk area network construction module, the process of building the risk area association diagram is as follows: The hospital obtains the location coordinates and physical space areas of each RFID reader within the hospital. Based on building functional zoning and ward unit layout, multiple spatial interval nodes are divided, and the readers deployed in each interval are associated and bound to the corresponding nodes. Using the hospital's intelligent wiring drawings and on-site surveying and calibration, the location information of all fixed RFID readers in the hospital is collected. The installation location coordinates of each reader are collected using a unified three-dimensional coordinate system within the hospital, and the device number, installation coordinates, floor, and department information are entered into the database. Based on the architectural design drawings, the functional zones of each floor are identified, and areas with clear physical boundaries such as the outpatient hall, emergency passage, inpatient corridor, ward area, intensive care unit, operating room cleanroom, equipment room, fire escape, stairwell, elevator lobby, and entrances are divided into independent spatial intervals. Sub-areas within the same functional zone that are separated by walls, fire doors, etc., are further... The system is further subdivided, with each ward and its buffer zone defined as an independent interval node, using the ward door as the boundary. All intervals are uniformly numbered and their floor, functional attributes, and boundary coordinates are recorded to form a list of spatial interval nodes. The interval to which a card reader belongs is determined based on its installation coordinates, and the device number is written into the association list of the corresponding interval. Multiple card readers within a single interval are associated with the same node to enhance signal coverage redundancy. Card readers located near the interval boundary are also associated with adjacent intervals on both sides. For positioning blind spots caused by signal obstruction, multipath effects, etc., fault-tolerant positioning is achieved through cross-coverage and redundant association of adjacent card readers at the interval boundary. When a patient is in a blind spot, the adjacent interval card readers jointly determine the patient's location and generate a mapping table with interval nodes as the index and the associated card reader number list as the field, which is stored in the database for the trajectory tracking module to call to realize the mapping of positioning data to spatial intervals.
[0022] Historical abnormal event records for each spatial interval were retrieved to count the number of patient trespasses, overstays, unauthorized movements, and nursing interventions. Combined with nursing staff rounds frequency, average patient stay time, and area functional attributes, a basic risk score was calculated for each spatial interval. The hospital's 12-month inpatient spatial activity history database was retrieved to precisely match each defined spatial interval node. Four core risk data categories were statistically analyzed within each interval: the number of unauthorized patient trespasses, the number of overstays exceeding a preset safety threshold (with a standard safety stay threshold of 10 minutes), the number of unauthorized movements deviating from normal activity paths, and the number of proactive interventions, on-site guidance, and emergency treatments by nursing staff for risk issues in that area. Simultaneously, the hospital's nursing scheduling system was linked with patient... The system collects patient behavior statistics, including the frequency of nursing rounds, the number of daily rounds, and the historical average length of stay for all inpatients within each zone. This data is then combined with the functional attributes of the zones for risk-weighted classification. Equipment rooms, fire exits, hospital entrances / exits, and stairwells are defined as high-risk functional areas; ordinary corridors and public activity areas are defined as medium-risk areas; and wards and nurses' stations are defined as low-risk areas. A standardized basic risk scoring formula is pre-set, with the frequency of abnormal events as the core weight. The lower the frequency of nursing rounds, the longer the patient stay, and the higher the risk attribute of the zone, the higher the basic risk score for that zone. Through multi-dimensional data weighted fusion calculation, a standardized basic risk score of 0 to 100 is quantitatively output for each spatial zone node. A higher score indicates a higher inherent safety risk for the zone.
[0023] Based on the basic risk score, the nodes in the spatial intervals are labeled with their hazard levels, and spatial transfer edges are established according to the physical connectivity between the areas. A three-tiered hazard level classification threshold is pre-set: 0-30 points are classified as low risk, 31-60 points as medium risk, and 61-100 points as high risk. Based on the basic risk score of each spatial interval node, the hazard level classification and labeling of all nodes are automatically completed, and differentiated risk labels are configured for nodes of different levels to facilitate subsequent model recognition and visualization. Furthermore, combined with the physical layout characteristics of the hospital building, the physical connectivity between the nodes in each spatial interval is determined. Two adjacent spatial nodes with direct passage connections, no wall obstructions, and accessible by foot are defined as interoperable node pairs. For each interoperable node pair, a dedicated spatial transfer edge is constructed. Each transfer edge represents a patient's ability to move independently between two intervals. Basic connectivity attributes such as physical distance, passage width, and passage difficulty between nodes are recorded. All spatial nodes and transfer edges together constitute the basic framework of the hospital's spatial topology, completely reconstructing all spatial passage paths that patients can move within the hospital.
[0024] The system acquires hospital access control status information and patients' historical movement path data. Based on access control opening conditions, passage direction constraints, and patients' high-frequency movement paths, it calculates the transfer impact coefficient between nodes in different risk areas. It connects in real-time to the hospital's intelligent access control system, collecting data on the opening and closing status, access control conditions, and passage direction restrictions of all hospital access control systems around the clock. It filters out access control types that allow free passage, card access, and prohibited passage for patients, clarifying the corresponding passage constraint rules for each access control, such as isolation access control in wards that can only be opened by medical staff, exit access control that is normally locked, and general corridor access control that allows free two-way passage. Simultaneously, it retrieves big data on patients' historical movement trajectories throughout the year, using trajectory clustering algorithms to statistically analyze various spatial data. The system identifies high-frequency patient movement paths between nodes, selecting the top 80% of mainstream paths based on daily average passage frequency, and eliminating occasional and irregular invalid movement trajectories. Combining access control constraints with high-frequency passage patterns, the system quantifies the transfer influence coefficient between any two connected spatial nodes. For node pairs with bidirectional free passage, high patient passage frequency, and no access control restrictions, a higher transfer influence coefficient is assigned. For node pairs with unidirectional passage, restricted access control, and very few patient passages, the transfer influence coefficient is reduced. For node pairs that are physically connected but normally locked by access control and completely impassable, the transfer influence coefficient is set to zero. This accurately quantifies the ease and probability weight of patient movement and transfer between different areas.
[0025] A risk area association matrix is constructed based on hazard level, spatial transfer edge, and transfer impact coefficient. A risk area association map is then generated based on the association strength between nodes in the matrix. Using all spatial interval nodes with completed risk level labeling as the basic units of the matrix, a two-dimensional square matrix structure with matching dimensions is constructed according to the unified node numbering order. Each matrix element corresponds to all spatial risk nodes within the institute. The preset hazard level quantification value of each node, the connectivity of the spatial transfer edges between nodes, and the transfer impact coefficients calculated between each pair of nodes are multi-dimensionally fused and assigned values. For node pairs with valid spatial transfer edges, the transfer impact coefficient and the hazard level weights of the two endpoints are substituted to complete the matrix element assignment. For nodes without valid spatial transfer edges... Pairs of nodes with physical connectivity or completely restricted access are uniformly assigned a value of zero, ultimately forming a risk area association matrix that fully covers the entire hospital space. This matrix can intuitively represent the access association strength, risk transmission capacity, and patient transfer probability between any two risk areas, fully encompassing the hospital's spatial topology and regional risk characteristics. After the matrix is constructed, non-zero elements in the matrix are used as effective association edges, and the spatial interval nodes corresponding to the matrix row and column indices are used as graph vertices. A force-directed layout algorithm is used to map the spatial coordinates of all vertices and edges. By simulating the mutual repulsion of charged particles and the iterative balance of spring tension, nodes with high association strength are clustered together in the visualization plane, while nodes with low or no association are clustered together. The nodes are initially separated to form a preliminary layout reflecting the actual spatial topology and correlation strength. Based on this, the actual building coordinates of the hospital are overlaid as background constraints, and the node positions are calibrated a second time to ensure that the relative orientation of each node in the graph network is consistent with the actual physical layout of the hospital. The edges of the graph network are assigned differentiated visual codes according to the numerical values of the matrix elements. Edges with correlation strength higher than the overall mean plus two standard deviations are defined as strongly correlated edges, rendered with thick lines and high-saturation colors, and accompanied by dynamic flowing arrows to indicate the direction of movement. Edges with correlation strength between the mean and the mean plus two standard deviations are defined as moderately correlated edges, rendered with medium line widths and regular colors. Edges with correlation strength lower than the mean are defined as weakly correlated edges. Edges are connected using thin dashed lines and light colors. Simultaneously, graph vertices are visually labeled differently based on the risk level of each node: low-risk nodes are marked with green circles, medium-risk nodes with yellow diamonds, and high-risk nodes with red triangles. The basic risk score value for that area is overlaid inside each node. The final result is a risk area association graph containing graph topology, visual hierarchical coding of association strength, visual hierarchical coding of node risk levels, and physical spatial coordinate constraints. The output format is compatible with mainstream graph database imports and front-end web visualization rendering engines, providing standardized spatial model support for subsequent patient trajectory risk mapping, pathological adaptive weight adjustment, and trajectory probability inference.
[0026] Pathological anchoring module: Acquires patient cognitive impairment classification data and real-time physiological parameters, extracts pathological features through time-series slice filtering, calculates the degradation rate of orientation, memory and executive function dimensions as the pathological drift benchmark value, and adaptively adjusts the trigger weight of risk areas accordingly. Historical trajectory nodes are compared with weights one by one, and the median of historical trigger frequency is used as the threshold for screening to generate a set of exclusive risk-sensitive areas and a weight matrix.
[0027] In the pathological anchoring module, the process of extracting pathological features through temporal slice filtering is as follows: The system acquires the patient's corresponding cognitive impairment level assessment results, records of previous cognitive function changes, and real-time data on heart rate, movement amplitude, activity cycle, and abnormal behavior responses; it simultaneously connects to the hospital's cognitive impairment assessment record system and the patient's wearable RFID smart monitoring terminal to complete the synchronous collection and aggregation of multi-source pathological data; the system batch retrieves the cognitive impairment level assessment results stored in the patient's records, uniformly classifies the levels using the hospital's standard cognitive function scoring criteria, and archives records of score changes from previous cognitive function re-examinations and follow-up data, comprehensively preserving information on the patient's long-term cognitive function fluctuations and regression trends; wearable monitoring... The terminal continuously collects physiological and behavioral data of the patient at a fixed sampling frequency. Heart rate data is sampled continuously at a frequency of once per second. The amplitude of movement is captured in real time by the patient's limb swing amplitude and body activity through the built-in three-axis sensor module. The activity cycle is statistically analyzed to show the alternation pattern between the patient's daytime activity period and nighttime rest period. Abnormal behavior response data records the time and duration of abnormal behaviors such as sudden agitation, immobility, and irregular wandering. All collected data is uniformly integrated into the system's backend database, and timestamps are marked and bound to the patient's identity, providing a complete, synchronous, and corresponding original data foundation for subsequent pathological feature analysis.
[0028] Real-time physiological parameters are sliced according to a preset time window, dividing the continuous physiological changes into multiple state segments. An adaptive filtering method is used to remove random fluctuation noise in different time segments. A fixed-length sliding time window is preset as the benchmark for time slicing, and a five-minute duration is uniformly used as the basic slicing unit. At the same time, a one-second sliding step size is set to segment the continuous physiological data stream segment by segment, decomposing the patient's uninterrupted physiological changes throughout the day into several continuous and partially overlapping independent state segments. Each state segment independently carries the patient's heart rate fluctuations, limb movements, activity rhythms, and behavioral changes within the corresponding time period, realizing the discretization and standardization of continuous dynamic physiological processes. For random noise caused by equipment acquisition jitter, environmental signal interference, and instantaneous somatosensory fluctuations within each state segment, the system uses an adaptive weighted filtering algorithm to remove noise. The filtering weight is dynamically adjusted according to the fluctuation amplitude of physiological data in different time periods, and small abnormal values in the stable state are smoothly removed, while the true physiological fluctuations and behavioral changes of the patient are completely preserved. While thoroughly filtering out invalid random noise, the loss of effective pathological features caused by over-filtering is avoided, ensuring the authenticity and validity of the data in each state segment.
[0029] Based on the filtered physiological state segments, physiological stability features, activity rhythm variation features, and behavioral response lag features are extracted respectively. Multi-dimensional feature quantification is performed on each time-series state segment after noise reduction processing to construct three categories of core pathological behavioral features. Regarding physiological stability features, the system statistically analyzes the mean heart rate, heart rate fluctuation variance, and body motion amplitude fluctuation coefficient for each time window. The stability of the patient's physiological state is determined by the magnitude of the numerical fluctuation range; a lower fluctuation coefficient indicates a more stable physiological state, while a higher fluctuation coefficient indicates significant emotional and physical agitation. Regarding activity rhythm variation features, the system... The system statistically analyzes the duration of patient activity, rest duration, and frequency of activity initiation and cessation at different times, compares the differences in activity levels between day and night, quantifies the degree of disruption in patients' daily activity patterns, and accurately captures typical pathological behavioral characteristics of patients with cognitive impairment, such as inverted day-night activity and irregular frequent walking. For the characteristic of delayed behavioral response, the system combines changes in the hospital environment with routine nursing intervention signals to statistically analyze the duration of patients' behavioral responses to changes in the external environment, nursing prompts, and environmental sound and light signals. By analyzing the response delay duration, the system quantifies the degree of decline in patients' perceptual and behavioral execution abilities, achieving a complete and accurate quantitative output of multidimensional pathological characteristics at a single time period.
[0030] This system performs time-synchronized mapping of cognitive impairment level features, physiological change features, and behavioral change features to construct a multidimensional pathological feature set encompassing cognitive, physiological, and behavioral dimensions. Using the timestamps of time-series slices as a unified alignment benchmark, it completes time-synchronized matching of multi-source heterogeneous features, mapping static cognitive impairment level features, dynamic physiological change features, and real-time behavioral change features to the same timeline system. The system retrieves basic cognitive level attribute data of patients for the corresponding time period, associating and binding long-term stable cognitive pathological features with segment-by-segment dynamically changing physiological and behavioral features, eliminating feature misalignment issues caused by inconsistent data collection frequencies and update cycles for different types of data. After achieving precise time alignment, the system performs structured fusion and integration of the three types of features, unifying the quantification standards and output dimensions of each dimension to construct a standardized multidimensional pathological feature dataset. This dataset simultaneously covers the basic degree of cognitive decline in patients, the real-time fluctuation patterns of physiological state, and abnormal changes in behavioral response, providing a comprehensive and three-dimensional representation of the patient's current pathological state and behavioral characteristics. This provides accurate pathological data support for subsequent calculation of pathological drift benchmark values and adaptive adjustment of regional risk weights.
[0031] In the pathological anchoring module, the process of generating a patient-specific set of risk-sensitive regions and a weight matrix is as follows: The system acquires orientation scores, memory retention scores, and executive function scores corresponding to multiple cognitive assessment cycles for patients. It connects to the hospital's cognitive function assessment database to batch retrieve longitudinal follow-up assessment data of hospitalized patients with cognitive impairment, selecting multiple standardized cognitive assessment records with evenly spaced and effective intervals. The assessment dimensions strictly correspond to the clinically accepted cognitive assessment system, uniformly collecting three core quantitative indicators: orientation score, memory retention score, and executive function score. Orientation score characterizes the patient's ability to recognize time, place, and people in their environment; memory retention score quantifies the patient's short-term and long-term memory retention capacity; and executive function score reflects the patient's comprehensive ability to plan, judge, regulate behavior, and adapt to the environment. The system screens patients for at least three valid assessment cycles, including initial baseline assessment data and subsequent periodic review data, ensuring the data span fully covers the patient's recent cognitive change trends. Simultaneously, all score data undergoes outlier cleaning and compliance verification, eliminating invalid samples with assessment errors, missing data, or abnormal fluctuations. The screened valid multi-cycle score data is uniformly archived and bound to patient identity information, providing an accurate and reliable original assessment data source for subsequent regression rate calculations.
[0032] Based on the relationship between the changes in scores and time intervals of each dimension between adjacent assessment periods, the rates of orientation decline, memory decline, and executive function decline are calculated separately. Using each pair of adjacent cognitive assessment periods as the calculation unit, the rate of decline is quantified for each of the three cognitive dimensions. First, the score difference between two adjacent assessments is statistically analyzed for each dimension. The change in score for the corresponding dimension is obtained by subtracting the score of the earlier assessment from the score of the later assessment. In patients with cognitive impairment, this change is negative in the deterioration state, and the larger the absolute value, the more significant the decline in cognitive function. Simultaneously, the actual time interval between two assessments is accurately read, and the unit of measurement is uniformly set to natural days to eliminate calculation bias caused by inconsistent periods. The rate of orientation decline, memory decline, and executive function decline per unit time is obtained by calculating the ratio of the score change to the actual time interval. For multi-period assessment data, the system uses a linear mixed-effects model to fit the long-term trajectory of each dimension, calculates the average rate of decline for each dimension, avoids calculation errors caused by fluctuations in a single assessment, and accurately characterizes the rate of functional decline in different cognitive dimensions of the patient.
[0033] The system calculates the pathological drift baseline value corresponding to the patient's current stage by integrating the degradation rate corresponding to each functional dimension and the preset dimension influence coefficient. Based on the correlation between clinical cognitive impairment and behavioral risk, fixed dimension influence coefficients are pre-configured for each of the three cognitive functional dimensions. These coefficients are calibrated using a large number of clinical samples to distinguish the impact weight of different cognitive dimension degradation on the patient's abnormal movement risk. Orientation degradation corresponds to the highest influence coefficient, while memory retention degradation and executive function degradation are matched with their respective coefficient weights. The system performs weighted multiplication of the calculated degradation rates of the three cognitive dimensions with their corresponding dimension influence coefficients, and then linearly accumulates and fuses all the weighted results to obtain a quantitative result of the patient's comprehensive cognitive degradation. This result is then normalized by combining the patient's age and disease duration with a baseline correction coefficient, mapping the final result to a standard numerical range of zero to one, generating a unique pathological drift baseline value corresponding to the patient's current stage. This baseline value accurately quantifies the degree of drift in the patient's cognitive pathological state; a larger value indicates a faster rate of cognitive decline, greater behavioral uncertainty, and a higher risk of spatial movement.
[0034] The trigger weights of regions in the risk region association map are scaled and adjusted based on the pathological drift baseline value. The patient's historical abnormal trajectory is matched with the adjusted region weights to filter region nodes that exceed the risk sensitivity threshold, generating a patient-specific set of risk sensitive regions. Based on the previously constructed risk region association map, the initial base trigger weights of all spatial region nodes are retrieved. Using the patient's current pathological drift baseline value as a scaling factor, the weights of all region nodes are adaptively and dynamically scaled and corrected. When the patient's pathological drift baseline value is too large, the trigger weights of each risk region are simultaneously amplified to improve the system's risk sensitivity to abnormal patient movement and adapt to the high-risk behavioral characteristics of patients with severe cognitive decline. When the patient's pathological drift baseline value is too low, the trigger weight of the region is appropriately reduced to decrease invalid warning interference in low-risk scenarios. After the weight adjustment is completed, the system retrieves all historical abnormal movement trajectory data of the patient and performs point-by-point matching and verification between all spatial region nodes along the trajectory and the adjusted region weight. The system presets a fixed risk sensitivity threshold and compares the real-time weight value of each region node one by one, filtering out all region nodes whose weight value exceeds the preset threshold. These regions are the high-sensitivity risk regions corresponding to the patient's pathological state. The system deduplicates and integrates the filtered sensitive regions and classifies them in an orderly manner, ultimately forming a set of exclusive risk sensitivity regions that are fully adapted to the individual patient's condition.
[0035] Based on the adjusted weights corresponding to the nodes in the risk-sensitive region set, a patient-specific trigger weight matrix is constructed. Using all spatial nodes within the patient's exclusive risk-sensitive region set as matrix construction units, the nodes are arranged in a uniform numbering order to create a two-dimensional trigger weight matrix adapted to the patient's individual risk characteristics. The matrix's row and column dimensions strictly correspond to the number of sensitive region nodes, and the values of each element within the matrix are sequentially filled with the final trigger weights of the corresponding region nodes after correction by pathological drift baseline values. For ordinary spatial nodes not within the sensitive region set, a default extremely low weight value is uniformly set to mitigate the impact of non-sensitive region risks. Weight normalization calibration is performed simultaneously during matrix construction to ensure a uniform and reasonable distribution of all weight values, avoiding risk assessment imbalances caused by excessively high or low extreme values in a single region. The resulting individualized trigger weight matrix can accurately assign differentiated values to different risk regions and can be dynamically updated according to changes in the patient's pathological state, providing precise individualized weight model support for subsequent trajectory node risk value calculation, individualized risk sequence generation, and risk intensity extrapolation.
[0036] The trajectory tracking module collects time-series positioning data through RFID wristbands to construct the patient's movement trajectory, maps it to a risk area association map, calculates the risk value of trajectory nodes by combining a set of exclusive risk-sensitive areas and a weight matrix, generates an individualized risk sequence, and extracts the current risk intensity as the initial distribution constraint for MCMC extrapolation based on the sequence change rate.
[0037] In the trajectory tracking module, the process of mapping to the risk area association map is as follows: The system acquires continuous positioning data generated by patients wearing RFID wristbands and analyzes the location coordinates, acquisition timestamps, and signal strength information corresponding to each data point. Patients wear active RFID smart positioning wristbands throughout the entire process, with the wristband continuously emitting radio frequency identification signals at a fixed sampling frequency. RFID readers deployed throughout the hospital receive the tag signals in real time and transmit the data back. The system continuously aggregates real-time positioning data streams collected by readers throughout the hospital, forming continuous raw positioning data covering the patient's activities at all times. The system performs structured parsing on each uploaded positioning data packet, accurately extracting core effective data content, including the patient's current planar and vertical coordinates, the high-precision acquisition timestamp corresponding to the positioning signal, and the real-time signal strength value of the received radio frequency signal. The system preprocesses and cleans the raw positioning data, removing invalid positioning data with signal disconnections, instantaneous jumps, and excessively low signal strength. Simultaneously, it performs time alignment and identity binding on the effective data, ensuring that each set of coordinate data, time data, and signal strength data corresponds to each other and is sequentially ordered, providing an accurate and reliable raw positioning data source for subsequent motion parameter calculations and trajectory construction.
[0038] The system calculates the patient's movement direction, distance, and speed based on the positional changes between adjacent positioning points, generating corresponding movement vector parameters. All valid positioning points are arranged in chronological order according to their timestamps. Spatial position difference calculations are performed on each pair of adjacent time-series positioning points, and the real-time movement direction angle of the patient is determined based on the spatial coordinate difference, accurately determining the patient's movement trend towards different areas such as corridors, wards, and stairwells. The system calculates the actual straight-line distance between adjacent positioning points using the Euclidean distance formula, and combines this with the acquisition time interval of two sets of positioning points to calculate the displacement per unit time, obtaining the patient's real-time movement speed. For patients in a stationary state, the system automatically identifies positioning sequences with no coordinate changes and determines the movement speed to be zero. Through point-by-point iterative calculation, corresponding movement direction, distance, and speed parameters are matched for each time-series positioning node, uniformly integrating them to form a complete set of movement vector parameters that includes spatial changes and motion states, accurately quantifying the patient's motion state characteristics at each moment.
[0039] By combining the patient's dwell time at each location point, consecutive location points are connected in chronological order to form a spatiotemporal state chain that includes positional changes and temporal evolution. By statistically analyzing the continuous dwell time at adjacent location points, the patient's dwell time at each location point is determined. When there is no significant shift in location coordinates across multiple consecutive frames and the signal strength is stable, the system automatically accumulates the patient's static dwell time, completely recording the patient's dwell behavior characteristics at different spatial locations. Using a high-precision timestamp as the sole sorting criterion, the system sequentially connects all location nodes carrying movement vector parameters and dwell time parameters, completing the point connection according to the patient's actual movement sequence. Invalid jump points and disconnected data are removed after cleaning, forming a continuous and uninterrupted complete movement trajectory link. This trajectory link is no longer a single static position coordinate sequence, but a multi-dimensional spatiotemporal state chain that simultaneously carries spatial position migration, movement speed changes, dwell time distribution, and temporal evolution. It can completely reproduce the patient's movement process and behavioral state at any time period, realizing the temporal and structured representation of the patient's in-hospital movement behavior.
[0040] The system maps the location nodes in the spatiotemporal state chain to the risk area association map, and determines the risk area node corresponding to each location node through spatial coordinate matching. It retrieves the previously constructed risk area association map, reads the coordinate boundary range and area number information of all spatial interval nodes, and establishes a one-to-one mapping ledger between spatial coordinates and risk area nodes. The system performs spatial matching calculations on all trajectory location nodes in the spatiotemporal state chain one by one, comparing the real-time coordinates of the trajectory nodes with the boundary coordinate range of each risk area. Through an area inclusion determination algorithm, it accurately identifies the spatial interval where the current location node is located, matching the corresponding risk area node number and area risk attribute. For trajectory nodes located at the critical position of two area boundaries, the system combines real-time signal strength change trends and the direction of movement of the trajectory before and after to perform auxiliary correction, avoiding boundary position mismatch problems. Through point-by-point coordinate matching, it achieves precise binding of all trajectory nodes to risk topology areas, allowing each segment of patient movement trajectory to correspond to a specific risk spatial unit, completing the deep integration of patient behavior trajectory and the hospital risk spatial model.
[0041] Based on a patient-specific set of risk-sensitive regions, the mapped trajectory nodes are filtered to extract risk trajectory nodes associated with the patient's current pathological state. The system then retrieves a patient-generated set of risk-sensitive regions, containing all highly sensitive risk region nodes that match the patient's current level of cognitive decline and pathological state. Using this sensitive region set as the filtering criterion, the system iterates through all mapped trajectory nodes, comparing the risk region IDs of the trajectory nodes with the region IDs in the sensitive region set. Trajectory nodes falling within the specific sensitive region are retained, while invalid trajectory nodes corresponding to ordinary low-risk regions are removed. The resulting risk trajectory nodes are key behavioral nodes that trigger individualized high-risk regions during the patient's movement. These nodes accurately reflect high-risk movement trends under the patient's current pathological state, effectively filtering out a large amount of risk-free routine movement trajectory data, significantly reducing the computational load for subsequent risk calculations, and improving the relevance and accuracy of risk assessment. This provides accurate and effective risk trajectory samples for subsequent trajectory node risk value calculations and individualized risk sequence generation.
[0042] In the trajectory tracking module, the process of extracting the current danger intensity as the initial distribution constraint for MCMC inference based on the sequence change rate is as follows: Based on the node number of the risk area where each trajectory node is located, the system calls the corresponding weight parameters in the trigger weight matrix and calculates the risk gain value for each trajectory node by combining the node dwell time, movement speed, and the basic risk level of the risk area. The system then performs risk quantification calculations on each of the selected risk trajectory nodes. First, it reads the corresponding risk area node number for each trajectory node and accurately retrieves the exclusive area trigger weight parameters within the patient's individualized trigger weight matrix based on the number, ensuring that the weight values fully match the patient's current cognitive pathological state. Simultaneously, the system extracts real-time motion and dwell time characteristic data corresponding to the trajectory node, including the patient's actual dwell time in the current risk area, real-time movement speed, and the preset basic risk level score of the risk area. The system employs a multi-factor weighted fusion algorithm, using the area-specific trigger weight as the core benchmark, combined with the quantification rule that longer dwell time leads to higher risk accumulation and faster movement speed leads to a stronger trend towards danger, while also overlaying the inherent risk coefficient of the area's basic risk level, coupling and calculating multi-dimensional parameters. Through standardized weighted calculation, the system obtains the real-time risk gain value of a single trajectory node. This value dynamically represents the patient's instantaneous risk accumulation at the current location, achieving refined and individualized quantitative output of the risk level for each trajectory node.
[0043] The risk gain values of each node are arranged according to the patient's movement time sequence to generate an individualized trajectory risk sequence. Using the high-precision timestamp of trajectory acquisition as the sole time series reference, the risk gain values of all calculated trajectory nodes are reconstructed and arranged in an orderly manner. The system strictly follows the time sequence logic of the patient's actual movement, continuously stringing together the risk gain data of discrete nodes collected at different times to avoid problems such as time sequence disorder and node reversal. During the arrangement process, smooth transition data within short sampling intervals are simultaneously supplemented, the risk gain values during stationary periods are steadily extended, and the risk gain changes during continuous movement periods are continuously connected to eliminate the data discontinuity problem caused by discrete sampling. Finally, a one-dimensional individualized trajectory risk sequence that dynamically changes over time, is continuously ordered, and closely matches the patient's actual movement behavior is generated, fully recording the dynamic fluctuation process of the risk level during the patient's monitoring period, and providing standardized time series data samples for subsequent risk change trend analysis.
[0044] The system calculates the change in risk gain between adjacent time points and determines the risk sequence change rate based on the relationship between the change in risk gain and the time interval. It performs temporal difference operations on the constructed individualized trajectory risk sequence, calculating the difference between the risk gain values of two adjacent time points one by one to obtain the absolute change in risk gain within a unit time interval, accurately reflecting the rise and fall of the patient's risk level at adjacent times. The system synchronously reads the actual sampling time interval between two adjacent nodes, using the real time span as a normalization benchmark to avoid calculation errors caused by the uniform interval assumption under the fixed sampling mode of the equipment. It then calculates the ratio of the change in risk gain to the corresponding time interval to obtain the risk sequence change rate per unit time. A positive change rate represents a continuous increase in the patient's risk level and an escalation of risk-prone behavior, while a negative change rate represents a gradual decrease in the patient's risk level and a weakening of the dangerous trend. The system traverses the entire risk sequence point by point to generate a continuous and complete risk change rate sequence, accurately depicting the dynamic evolution speed of the patient's risk status.
[0045] The system extracts the risk change rate corresponding to the current trajectory's end node and uses it as the initial risk intensity for the patient's current approach to a dangerous area. It then locates the latest acquired trajectory end node in the time-series risk sequence, extracts the real-time risk sequence change rate corresponding to that node, and standardizes this change rate to a quantization range of zero to one, forming an initial risk intensity that characterizes the patient's real-time approach to danger. This risk intensity directly reflects the rate of risk increase in the patient's current movement behavior; a higher value indicates a more significant trend towards a high-risk area and a higher degree of behavioral abnormality. The system imports this initial risk intensity into the subsequent MCMC probability inference model to constrain the assignment logic of the Markov chain's initial state probability distribution, replacing the traditional fixed initial distribution setting. This ensures that the model's initial inference state fully matches the patient's real-time behavioral trend and pathological state. Through real-time dynamic initial constraints, the subsequent trajectory probability inference more closely matches the patient's current true risk situation, effectively improving the targeting and accuracy of short-term future trajectory prediction, and achieving intelligent trajectory prediction under the dual constraints of pathological state and behavioral trend. Risk prediction module: Starting from the current trajectory node, it combines the inter-regional transfer probability and movement delay, uses the pathological drift baseline value to predict the future trajectory distribution within a preset time window, and dynamically adjusts the sampling ratio of high-risk areas based on the pathological drift baseline value to generate probability-weighted paths and screen high-risk paths with a probability of crossing the target danger area exceeding the threshold.
[0046] In the risk assessment module, the process of MCMC (Morbidity and Risk Mitigation) using pathological drift baseline constraints to predict the future trajectory distribution within a preset time window is as follows: The system determines the initial state node of the Markov chain based on the spatial location corresponding to the current trajectory node, and converts the initial risk intensity into an initial state probability distribution. It reads the spatial coordinates of the patient's latest trajectory node, matches it with a risk area association map to complete real-time spatial positioning, determines the patient's current risk area node, defines this area node as the initial state node of the Markov chain inference model, and locks the starting spatial location of the trajectory prediction, ensuring that the inference starting point is perfectly aligned with the patient's actual real-time location. The system retrieves the patient's real-time initial risk intensity calculated in previous steps; this value quantifies the speed and urgency of the patient's current approach to the danger zone. The system uses a standardized probability mapping function to linearly convert the risk intensity values in the zero-to-one interval into a Markov chain initial state probability distribution. Higher risk intensity values indicate a greater initial probability of transitioning to a high-risk area, while lower risk intensity values indicate a greater likelihood of remaining in the current low-risk state. This constructs a dynamic initial probability distribution that perfectly matches the patient's real-time behavioral state, replacing the traditional fixed uniform initial distribution mode, allowing subsequent trajectory inference to adapt to the patient's individualized risk characteristics from the initial stage.
[0047] The system acquires historical transfer frequencies, spatial distances, and patient movement delays between adjacent nodes in the risk area association graph, and calculates the basic state transfer probability between nodes. It then iterates through all pairs of spatially connected adjacent nodes in the risk area association graph, batches retrieves historical patient movement statistics from the entire hospital's database, and calculates the historical transfer frequency between any two adjacent nodes, summarizing the historical transfer frequency statistics between regions. Simultaneously, it reads pre-stored data on the straight-line spatial distance between adjacent nodes and the typical patient movement delay in the risk topology model. The movement delay is the average cross-regional travel time for patients, obtained from a large number of historical samples, used to characterize the ease and time characteristics of regional travel. Using historical transfer frequencies as the base weight, and combining quantitative rules that closer spatial distances result in higher transfer probabilities and shorter movement delays in terms of travel likelihood, the system performs normalized weighted fusion calculations on multi-dimensional statistical parameters to obtain the basic state transfer probability between any adjacent risk area nodes. This basic transfer probability is entirely constructed based on the hospital's spatial layout and big data on patient historical behavior, objectively reflecting the probability patterns of routine cross-regional patient movement within the hospital, and providing benchmark parameters for subsequent individualized adjustments.
[0048] The system calculates the pathological matching influence coefficient based on the patient's pathological drift baseline value, and dynamically corrects the baseline state transition probability based on the influence coefficient to obtain the pathological adaptive transition probability. It retrieves the updated pathological drift baseline value for the patient's current stage, which quantifies the degree of cognitive decline and behavioral loss of control drift. The system converts the pathological drift baseline value into the corresponding pathological matching influence coefficient using a preset pathological association mapping formula. Higher cognitive decline results in a stronger adjustment of the influence coefficient, significantly amplifying the transition probability weight of high-risk areas. Using the calculated pathological matching influence coefficient as a correction factor, the system dynamically scales and corrects the baseline state transition probability between each node. For patients with severe cognitive drift, the probability of transitioning to high-risk areas is appropriately increased, while the probability of stable residence is reduced. For patients with mild cognitive stability, the transition probability is slightly adjusted to maintain the regular movement pattern. After correction, the transition probabilities of all nodes are normalized a second time to ensure that the sum of all downstream transition probabilities of the same node satisfies the probability constraint condition. Finally, the system obtains the pathological adaptive transition probability adapted to the individual patient's pathological state, achieving dynamic adaptation of model parameters to each patient's specific condition.
[0049] A Markov chain sampling process is constructed based on pathological adaptive transition probability, and a Monte Carlo random sampling method is used to generate a multi-step trajectory state sequence within a preset time window. Using pathological adaptive transition probability as the core state transition rule, a first-order Markov chain state transition model adapted to individual patient characteristics is built. Each state jump in the model is strictly executed according to the corrected regional transition probability, ensuring that the trajectory jump pattern conforms to the patient's pathological behavior characteristics. The system presets a fixed-length prediction time window, uniformly using a ten-minute short-term prediction window as the extrapolation interval to meet the timeliness requirements of early clinical intervention. Within the prediction time window, the system uses a Monte Carlo random sampling algorithm for large-scale repeated sampling, completing multi-step state jumps through multiple rounds of random simulation iterations. Each round of sampling can generate a complete future regional jump trajectory sequence, with a single window and a single extrapolation generating thousands of independent trajectory state sequences. During the sampling process, the sampling bias is dynamically adjusted based on the pathological drift baseline value. The higher the degree of pathological drift, the more the model automatically increases the sampling ratio of high-risk areas, generating more simulated trajectories approaching high-risk areas, fully exploring the patient's potential dangerous movement paths.
[0050] The system statistically analyzes the probability of occurrence of different trajectory state sequences to form a future trajectory distribution encompassing all candidate trajectory paths under time constraints. After completing batch sampling and extrapolation within a preset time window, all generated multi-step trajectory state sequences are classified, statistically analyzed, and their probabilities are summarized. Duplicate trajectory paths with identical content and the same sequence of nodes traversing the same area are merged and statistically analyzed, and the sampling frequency of each type of candidate trajectory path is recorded. The system calculates the ratio of the occurrence frequency of each type of path to the total number of samples, normalizes the solution, and obtains the independent occurrence probability of each candidate trajectory path. With the prediction time axis as a constraint, the system integrates and includes all candidate trajectory paths with different probability levels and different travel directions, fully covering various movement scenarios that patients may encounter in the future, such as stationary stay, routine passage, abnormal risk approach, and accidental entry into high-risk areas, ultimately forming a complete future trajectory probability distribution result. This distribution can quantitatively display the multiple possible states of the patient's future movement, providing comprehensive probabilistic data support for subsequent high-risk path screening and risk warning assessment.
[0051] In the hazard assessment module, the process of selecting high-risk paths with a probability exceeding a threshold for traversing a target hazard area is as follows: Based on the future trajectory distribution, all candidate trajectory paths and their corresponding risk area node sequences are extracted. Relying on the future trajectory distribution results generated within a preset time window by the preceding MCMC simulation, the probability distribution set is fully analyzed and decomposed, traversing and collecting all valid candidate trajectory paths generated by simulation sampling, eliminating any omissions of low-probability potential risk paths. For each candidate trajectory path, the system decomposes all spatial area nodes passed sequentially during the path's movement according to the temporal progression order, completely restoring the temporal sequence of risk area nodes corresponding to a single trajectory, and accurately recording the complete spatial link of the trajectory's starting node, intermediate nodes, and ending node. Simultaneously, the system performs deduplication and classification on all candidate paths, merging and archiving duplicate simulation paths with completely identical temporal arrangement of regional nodes, while retaining independent candidate paths with differentiated movement directions and different traversed areas, forming a complete path sample set covering all possible future movements of the patient, providing comprehensive path data support for subsequent multi-dimensional risk weight calculations.
[0052] The system calculates the comprehensive risk weight of each candidate path by combining the number of risk areas traversed, the hazard level of nodes, and the time spent in each area. For each candidate trajectory, a multi-dimensional risk characteristic quantification calculation is performed, relying on the fusion of three core risk factors to solve for the comprehensive risk weight of the path. First, the total number of risk area nodes traversed along a single path is counted; the more high-risk areas traversed, the higher the basic risk of the path. Second, the preset hazard level scores of each traversed area node are retrieved, with higher weight coefficients assigned to high-risk area nodes, and progressively decreasing values assigned to medium- and low-risk areas, accumulating and integrating the risk scores of the entire path. Simultaneously, the simulated time spent in each area during trajectory simulation is considered; the longer the patient spends in a high-risk area, the more significant the risk accumulation effect, and the corresponding path risk weight increases accordingly. The system uses a preset standardized weighted fusion formula to normalize and couple the area quantity factor, hazard level factor, and time spent factor, eliminating the differences in the dimensions of parameters across different dimensions, and finally outputting a unique comprehensive risk weight for each candidate trajectory path, accurately quantifying the overall risk level of a single path.
[0053] Candidate trajectory paths are probabilistically weighted based on a comprehensive risk weight, generating a probability-weighted path set with risk probability identifiers. Using the original sampling probability of each candidate path as a baseline, a secondary weighting correction operation is performed, combining the comprehensive risk weight of the path to achieve bidirectional fusion calibration of probability and risk level. For paths with higher comprehensive risk weights, the system appropriately increases the proportion of their corresponding probability, enhancing the identification priority of high-risk paths in the overall sample set. For common paths with lower risk weights, their probability proportion is slightly weakened to avoid interference from a large number of low-risk paths in the risk assessment results. After weighting correction, the system uniformly normalizes the corrected probabilities of all paths to ensure that the sum of all path probabilities conforms to the probability distribution constraint standard. Simultaneously, each path is bound to a unique risk level label and a corrected probability identifier, ultimately generating a structured, hierarchically clear probability-weighted path set, achieving preliminary stratification between ordinary paths, potentially risky paths, and high-risk paths.
[0054] The system calculates the cumulative probability of each probability-weighted path traversing target danger zones, compares this cumulative probability with a preset danger path screening threshold, and identifies paths exceeding the threshold as high-risk paths. A fixed danger path screening probability threshold is pre-set based on clinical nursing safety standards and hospital risk management needs. This threshold, calibrated with a large number of clinical samples, effectively balances the false alarm and false alarm rates. The system iterates through all candidate paths in the probability-weighted path set, analyzes the travel path of each path individually, and determines whether the path's temporal node sequence contains targets such as stairwells, equipment rooms, hospital entrances / exits, and isolation restricted areas. Dangerous area nodes: For paths passing through target dangerous areas, the system accumulates the weighted cumulative probability of traversing high-risk areas throughout the entire route, accurately quantifying the likelihood of a patient mistakenly entering a dangerous area in the future. The cumulative traversal probability of each path is compared one by one with a preset screening threshold. When the cumulative probability of a path is greater than the set threshold, the path is determined to have a significant risk of escalation and is classified as a high-risk path; otherwise, it is determined to be a safe and normal path. The system finally summarizes all the screened high-risk paths to form an accurate list of future dangerous paths, providing core risk basis for subsequent assessment of remaining arrival time, accelerated trend identification, and tiered early warning push.
[0055] The early warning triggering module calculates the remaining arrival time series based on the expected arrival time of the high-risk path and real-time location. It performs first-order differencing on the series to obtain the rate of change and second-order differencing to identify the acceleration trend. When both levels of differencing exceed the threshold, the warning level is matched according to the remaining time and an intervention instruction is pushed. At the same time, based on the patient's actual risk-avoidance results, the transfer probability in the risk area association map is corrected through a Bayesian incremental strategy.
[0056] In the early warning triggering module, the process of obtaining the rate of change by performing first-order differencing on the sequence and identifying the accelerating trend by performing second-order differencing is as follows: The system obtains the estimated arrival time of the target danger zone node corresponding to the high-risk path, and calculates the dynamic remaining distance between the current location and the target danger zone by combining the patient's real-time location coordinates, movement speed, and movement direction. Based on the spatial variation relationship of continuous sampling times, a remaining arrival time series is generated. The system retrieves the previously screened high-risk path set, reads the target danger zone node information corresponding to each high-risk path, and obtains the initial estimated arrival time obtained from trajectory extrapolation, using this as a benchmark for trend analysis. The system simultaneously updates the patient's RFID wristband positioning data in real time, continuously collecting the patient's current precise planar coordinates, real-time movement speed, and movement direction vector parameters, combined with the target danger zone... The spatial boundary coordinates of the area are used to dynamically calculate the remaining travel distance from the patient's real-time location to the effective boundary of the danger zone using a Euclidean spatial distance algorithm. The system continuously iterates and updates the remaining distance value at a fixed sampling frequency, and dynamically corrects the remaining travel time at each sampling moment by combining the patient's instantaneous movement speed, avoiding the static error caused by the average speed estimation. The system arranges the remaining arrival times corresponding to multiple consecutive sampling moments in chronological order to form a continuous, orderly, and dynamically updated remaining arrival time series, which fully records the real-time decay change process of the remaining time as the patient approaches the danger zone, providing stable and reliable time series data samples for subsequent differential trend analysis.
[0057] The system performs adjacent time-series difference calculations on the remaining arrival time sequence to calculate the change in remaining arrival time per unit time. The first-order rate of change is determined based on the ratio of this change to the sampling time interval. Using the temporal sequence as a benchmark, the system performs adjacent time-series difference calculations on the remaining arrival time sequence, calculating the difference between the remaining arrival time values of two consecutive sampling times to obtain the absolute change in remaining time under continuous time sequence. This directly reflects the reduction in the time taken for patients to approach the danger zone within adjacent sampling intervals. The system reads the actual time interval between two sets of adjacent sampling points, using the actual sampling duration as a normalization benchmark to eliminate calculation bias caused by uneven intervals in the fixed sampling mode of the equipment. The ratio of the remaining time change at each moment to the corresponding sampling time interval is calculated to obtain the first-order rate of change of remaining arrival time per unit time. This first-order rate of change accurately characterizes the average speed at which patients approach the danger zone. A larger rate of change indicates a faster overall speed at which patients approach the high-risk zone, effectively quantifying the basic advancing trend of patients' danger-seeking behavior.
[0058] Within a continuous time window, the first-order rate of change is differentially calculated again to extract its trend. The second-order rate of change corresponding to the remaining arrival time series is then calculated using time intervals. A fixed-duration sliding calculation window is set up, and secondary differential operations are performed on multiple consecutive sets of first-order rate of change sequences within the window to capture the trend of the rate of change. The system iterates through the first-order rate of change values at adjacent times within the sliding window, calculating the difference between the first-order rate of change point by point to obtain the rate change. This is then normalized using the corresponding time-series sampling interval to obtain the second-order rate of change. The second-order rate of change is used to characterize the rate of change of the first-order rate of change, accurately identifying acceleration behavior and sudden changes in the patient's condition during movement. When the patient approaches the danger zone at a constant speed, the second-order rate of change approaches zero. When the patient actively accelerates and rushes towards the danger zone, the second-order rate of change increases significantly. When the patient decelerates or stops, the second-order rate of change reverses and falls back, achieving refined extraction of the dynamic sudden change characteristics of dangerous behavior.
[0059] The system combines first-order and second-order rates of change to characterize the speed changes and acceleration trends of patients approaching the target danger zone. By integrating these two rates of change in a multi-dimensional manner, a dual-dimensional danger-approaching trend identification system is constructed, comprehensively covering the speed and acceleration characteristics of patients approaching the danger zone. The first-order rate of change is used to determine the baseline speed of the patient approaching the high-risk area, distinguishing between slow and rapid approaches. The second-order rate of change is used to capture dynamic acceleration mutations during the patient's movement, identifying high-risk abnormal behaviors such as sudden agitation or abrupt acceleration. Through this dual-indicator joint judgment method, the system avoids the limitation of a single rate indicator in identifying dynamic trends. It can identify both steady, slow, and continuous low-risk approaches, as well as accurately capture sudden, rapid, and high-risk approaches. This comprehensive, three-dimensional, and accurate characterization of the patient's real-time danger-approaching situation provides a reliable quantitative trend basis for subsequent dual-threshold joint early warning judgments, early warning level matching, and nursing intervention instruction delivery.
[0060] In the early warning triggering module, the process of correcting the transition probability in the risk area association graph using a Bayesian incremental strategy is as follows: The system acquires the first-order and second-order rates of change corresponding to the patient's current danger approach state. When the first-order rate of change exceeds a preset threshold and the second-order rate of change exceeds an acceleration threshold, the patient is determined to have entered a rapid approach to danger state, and the corresponding warning level is matched based on the current remaining arrival time in the remaining arrival time series. The system retrieves the remaining arrival time series generated during the patient's approach to the target danger area in real time, and simultaneously extracts the core indicators of the first-order and second-order rates of change obtained from sequence analysis. The system pre-stores clinically calibrated fixed rate of change and acceleration thresholds, which are used to determine the patient's baseline approach speed and instantaneous acceleration trend, respectively, and continuously monitors and compares the two indicators in parallel. When the first-order rate of change exceeds... When the preset change threshold is reached and the second-order rate of change simultaneously exceeds the preset acceleration threshold, the system excludes low-risk scenarios such as slow approach and normal movement, accurately determining that the patient has entered a rapid approach to danger state, and identifying high-risk behaviors such as sudden agitation and active acceleration towards the danger zone. The system reads the latest remaining arrival time at the end of the remaining arrival time series in real time, and matches the corresponding standardized warning level according to the time interval grading rules. The shorter the remaining arrival time and the faster the patient approaches, the higher the warning level is matched. Different warning levels correspond to differentiated intervention priorities and treatment strategies, realizing refined grading of high-risk situations and providing a grade basis for the generation of subsequent intervention instructions.
[0061] Based on the warning level, the system generates risk intervention instructions including the patient's current location, target danger zone, estimated arrival time, and intervention method, and pushes them to the nursing terminal. Simultaneously, it collects information on the patient's actual movement trajectory after the warning is triggered, whether the patient entered the target danger zone, and the actual arrival time. According to the matched warning level, it retrieves a pre-set standardized intervention strategy library. Different warning levels correspond to differentiated intervention reminders and handling requirements: high-level warnings trigger emergency on-site intervention prompts, while medium- and low-level warnings trigger routine patrol intervention prompts. The system integrates the patient's real-time location coordinates, target danger zone number and area attributes, estimated arrival time derived from the trajectory, and the corresponding standard intervention method to automatically generate structured risk intervention instructions. The system pushes these instructions in real-time to the nurse station backend and nursing staff's handheld terminals via the hospital's nursing push link, ensuring that nursing staff receive high-risk information and handling plans for patients as soon as possible. After the warning is pushed, the system continuously and frequently collects the patient's subsequent real-time RFID location trajectory, recording the patient's complete movement path after the warning is triggered. It simultaneously statistically analyzes whether the patient successfully entered the target danger zone, the actual arrival time in the danger zone, and the final location, forming complete closed-loop observation data for warning and intervention, providing real sample support for subsequent model iteration and optimization.
[0062] Based on the discrepancy between the patient's actual risk-avoidance outcome and the predicted high-risk path, regional transfer feedback samples are constructed. The original transfer probability in the risk area association graph is used as the prior probability, and Bayesian incremental updates are performed using the newly added feedback samples to calculate the corrected regional node transfer probability. The high-risk path prediction results obtained from MCMC are compared one by one with the patient's actual movement risk-avoidance results to quantify the discrepancy between the predicted path and the actual travel path, and to determine whether the original transfer probability of the model is too high or too low. The system integrates multi-dimensional data such as regional jump relationships, predicted transfer probability, actual transfer occurrence, and pathological matching coefficients, using a single complete early warning intervention event as a unit. Standardized regional transfer feedback samples are constructed, and each high-risk traversal behavior of a patient generates an independent incremental sample, enabling continuous accumulation and updating of sample data. The system unifies the original historical transfer probabilities of each adjacent node in the risk area association graph as the prior probability of the Bayesian update. Based on the Bayesian incremental learning rule, the observation likelihood of the newly added feedback sample is introduced, and the prior probability distribution is iteratively corrected by combining the sample credibility weight. The entire update process does not require retraining the model with massive historical data, but only completes parameter fine-tuning through a single new sample, effectively adapting to the dynamic changes in patient behavior patterns. Finally, the posterior transfer probability that fits the latest patient movement pattern is calculated, accurately correcting the prediction bias of the original model.
[0063] The updated transition probabilities are mapped back to the risk region association graph, and the sampling weights and trajectory inference parameters of high-risk regions are adjusted synchronously based on the correction results. The posterior transition probabilities obtained from the Bayesian incremental update are mapped one by one back to the transition edges of each node in the risk region association graph, replacing the original initial transition probability parameters of the model, thus completing the dynamic iterative update of the core parameters of the risk topology network. The system synchronously adjusts the core parameters of the MCMC trajectory inference module according to the correction magnitude of the transition probabilities. For high-risk regions with significantly increased transition probabilities after the update, their random sampling weights are appropriately increased to improve the model's probability of capturing high-risk movement paths. For regular regions with decreased transition probabilities after the update, the proportion of invalid sampling is correspondingly reduced to optimize the overall inference efficiency. The system synchronously normalizes the transition probabilities and sampling weights of all nodes to ensure that the sum of regional transition probabilities meets the Markov chain probability constraint conditions, maintaining the stability and accuracy of the model operation. Through continuous incremental sample updates and adaptive parameter adjustments, the closed-loop self-learning optimization of the risk region association model is achieved, allowing the model inference results to continuously adapt to the latest movement behavior characteristics and pathological risk patterns of in-hospital patients, and continuously improving the accuracy of trajectory prediction and risk warning.
[0064] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. An RFID based intelligent monitoring and analysis management system for in-patients, characterized in that: include: Risk Zone Networking Module: Collects RFID reader deployment information to divide risk zones. Based on the alarm frequency, nursing intensity, and dwell time of each zone, the three data are normalized and weighted and summed. The risk level is assessed according to the threshold. The association strength is calculated based on the physical connectivity of the zones, access control logic, and path dependence. The risk zone association graph is constructed with the zones as nodes and the association strength as edges. Pathological anchoring module: Acquires patient cognitive impairment classification data and real-time physiological parameters, extracts pathological features through time-series slice filtering, calculates the degradation rate of orientation, memory and executive function dimensions as the pathological drift benchmark value, and adaptively adjusts the trigger weight of risk areas accordingly. Historical trajectory nodes are compared with weights one by one, and the median of historical trigger frequency is used as the threshold for screening to generate a set of exclusive risk-sensitive areas and a weight matrix. The trajectory tracking module collects time-series positioning data through an RFID wristband to construct the patient's movement trajectory, maps it to a risk area association map, calculates the risk value of trajectory nodes by combining a set of exclusive risk-sensitive areas and a weight matrix, generates an individualized risk sequence, and extracts the current risk intensity as the initial distribution constraint for MCMC extrapolation based on the sequence change rate. Risk prediction module: Starting from the current trajectory node, combined with the inter-regional transfer probability and movement delay, the MCMC constrained by the pathological drift benchmark value is used to predict the future trajectory distribution within a preset time window, and the sampling ratio of high-risk areas is dynamically adjusted according to the pathological drift benchmark value to generate probability-weighted paths and screen high-risk paths with a probability of crossing the target danger area exceeding the threshold. The early warning triggering module calculates the remaining arrival time series based on the expected arrival time of the high-risk path and real-time location. It performs first-order differencing on the series to obtain the rate of change and second-order differencing to identify the acceleration trend. When both levels of differencing exceed the threshold, the warning level is matched according to the remaining time and an intervention instruction is pushed. At the same time, based on the patient's actual risk-avoidance results, the transfer probability in the risk area association map is corrected through a Bayesian incremental strategy.
2. The RFID based in-patient intelligent monitoring, analyzing and managing system as claimed in claim 1, wherein, The process of constructing a risk area association diagram is as follows: Obtain the location coordinates of each RFID reader within the hospital and its corresponding physical space area. Divide the space into multiple spatial interval nodes based on the building's functional zoning and ward unit layout, and associate and bind the readers deployed in each interval with the corresponding nodes. Retrieve historical abnormal event records for each spatial interval, count the number of times patients mistakenly enter, stay for longer than the allotted time, move without authorization, and receive nursing interventions, and combine this with the frequency of nursing staff rounds, the average time patients stay, and the functional attributes of the area to calculate the basic risk score for each spatial interval. Based on the basic risk score, the nodes in the spatial interval are marked with their hazard levels, and spatial transfer edges are established according to the physical connectivity between the regions. Obtain hospital access control status information and patient historical movement path data, and calculate the transfer impact coefficient between nodes in different risk areas based on access control opening conditions, passage direction constraints, and high-frequency patient movement paths; A risk area association matrix is constructed based on the hazard level, spatial transfer edge, and transfer impact coefficient, and a risk area association map is generated based on the association strength between nodes in the matrix.
3. The RFID based intelligent monitoring, analyzing and management system for hospitalized patients as claimed in claim 1 wherein, The process of extracting pathological features through time-series slice filtering is as follows: Obtain the patient's corresponding cognitive impairment level assessment results, records of previous cognitive function changes, and real-time data on heart rate, movement amplitude, activity cycle, and abnormal behavior response; Real-time physiological parameters are sliced according to a preset time window, dividing the continuous physiological change process into multiple state segments, and an adaptive filtering method is used to remove random fluctuation noise in different time segments. Based on the filtered physiological state segments, physiological stability features, activity rhythm change features, and behavioral response lag features were extracted respectively. By synchronously mapping the characteristics of cognitive impairment level, physiological changes, and behavioral changes over time, a multidimensional pathological feature encompassing cognitive, physiological, and behavioral dimensions is constructed.
4. The RFID-based intelligent monitoring and analysis management system for inpatients according to claim 1, characterized in that, The process of generating a dedicated set of risk-sensitive regions and a weight matrix is as follows: Obtain the patient's orientation score, memory retention score, and executive function score within multiple cognitive assessment cycles; Based on the relationship between the changes in scores of each dimension and the time interval between adjacent assessment periods, the rates of orientation decline, memory decline, and executive function decline were calculated respectively. Based on the degradation rate corresponding to each functional dimension and the preset dimension influence coefficient, a fusion calculation is performed to generate the pathological drift baseline value corresponding to the patient's current stage. Based on the pathological drift baseline value, the regional trigger weights in the risk area association map are scaled and adjusted, and the patient's historical abnormal trajectory is matched with the adjusted regional weights to filter regional nodes that exceed the risk sensitivity threshold, thereby generating a patient-specific risk sensitivity area set. Based on the adjustment weights corresponding to the nodes in the risk-sensitive area set, a patient-specific trigger weight matrix is constructed.
5. The RFID-based intelligent monitoring and analysis management system for inpatients according to claim 1, characterized in that, The process of mapping to the risk area association diagram is as follows: Acquire continuous positioning data generated by the patient wearing an RFID wristband, and parse the location coordinates, acquisition timestamp, and signal strength information corresponding to each positioning data point; Calculate the patient's movement direction, movement distance, and movement speed based on the positional change relationship between adjacent positioning points, and generate the corresponding movement vector parameters; By combining the patient's dwell time at each location point, the consecutive location points are connected in chronological order to form a spatiotemporal state chain that includes the relationship between positional changes and temporal evolution. Map the location nodes in the spatiotemporal state chain to the risk area association graph, and determine the risk area node corresponding to each location node through spatial coordinate matching; Based on the patient's specific risk-sensitive area set, the mapped trajectory nodes are filtered to extract risk trajectory nodes that are associated with the patient's current pathological state.
6. The RFID-based intelligent monitoring and analysis management system for inpatients according to claim 1, characterized in that, The process of extracting the current risk intensity as the initial distribution constraint for MCMC inference based on the sequence change rate is as follows: Based on the node number of the risk area where each trajectory node is located, the corresponding weight parameters in the trigger weight matrix are called, and the risk gain value corresponding to each trajectory node is calculated by combining the node dwell time, movement speed and the basic level of the risk area. The risk gain values of each node are arranged according to the patient's movement time sequence to generate an individualized trajectory risk sequence; Calculate the change in risk gain between adjacent time points, and determine the rate of change of the risk sequence based on the relationship between the change in risk gain and the time interval; Extract the risk change rate corresponding to the end node of the current trajectory and use it as the initial risk intensity of the patient's current tendency to move towards the danger zone.
7. The RFID-based intelligent monitoring and analysis management system for inpatients according to claim 1, characterized in that, The process of extrapolating the future trajectory distribution within a preset time window using MCMC with pathological drift baseline constraints is as follows: The initial state node of the Markov chain is determined based on the spatial location corresponding to the current trajectory node, and the initial danger intensity is converted into the initial state probability distribution. Obtain historical transfer frequency, spatial distance, and patient movement delay information between adjacent nodes in the risk area association map, and calculate the basic state transfer probability between nodes; The pathological matching influence coefficient is calculated based on the patient's pathological drift baseline value. The baseline state transfer probability is dynamically corrected based on the influence coefficient to obtain the pathological adaptive transfer probability. A Markov chain sampling process is constructed based on pathological adaptive transfer probability, and a multi-step trajectory state sequence within a preset time window is generated using Monte Carlo random sampling. The probability of different trajectory state sequences is statistically analyzed to form a future trajectory distribution that includes all candidate trajectory paths under time constraints.
8. The RFID-based intelligent monitoring and analysis management system for inpatients according to claim 1, characterized in that, The process of selecting high-risk routes with a probability exceeding a threshold for traversing a target danger zone is as follows: Based on the future trajectory distribution, extract all candidate trajectory paths within the distribution and the risk area node sequence corresponding to each path; The comprehensive risk weight of each candidate path is calculated by combining the number of risk areas traversed by each path, the danger level of each node, and the time spent in each area. Candidate trajectory paths are probabilistically weighted based on the comprehensive risk weight of the path, generating a set of probabilistically weighted paths with risk probability labels; The cumulative probability of each probability-weighted path crossing the target dangerous area node is calculated, and the cumulative probability is compared with the preset dangerous path screening threshold. Paths exceeding the threshold are identified as high-risk paths.
9. The RFID-based intelligent monitoring and analysis management system for inpatients according to claim 1, characterized in that, The process of obtaining the rate of change by performing first-order differencing on the sequence and identifying the accelerating trend by performing second-order differencing is as follows: The estimated arrival time of the target danger zone node corresponding to the high-risk path is obtained, and the dynamic remaining distance between the current position and the target danger zone is calculated by combining the patient's real-time location coordinates, movement speed and movement direction. The remaining arrival time series is generated based on the spatial change relationship of continuous sampling time. The remaining arrival time series is differentially analyzed by adjacent time points in chronological order. The change in remaining arrival time per unit time is calculated, and the first-order rate of change is determined based on the ratio of the change to the sampling time interval. The first-order rate of change is differentially calculated again within a continuous time window to extract the trend of the first-order rate of change, and the second-order rate of change corresponding to the remaining arrival time series is solved in combination with the time interval. The combined first-order and second-order rates of change characterize the speed change and acceleration trend of the patient's approach to the target danger zone.
10. The RFID-based intelligent monitoring and analysis management system for inpatients according to claim 1, characterized in that, The process of correcting the transition probability in the risk region association graph using a Bayesian incremental strategy is as follows: The system obtains the first-order and second-order rates of change corresponding to the patient's current dangerous approach state. When the first-order rate of change exceeds the preset change threshold and the second-order rate of change exceeds the acceleration threshold, the patient is determined to have entered a rapid approach state. The system then matches the corresponding warning level based on the current remaining arrival time in the remaining arrival time sequence. Based on the warning level, a risk intervention instruction is generated, which includes the patient's current location, target danger area, estimated arrival time and intervention method, and pushed to the nursing terminal. At the same time, the patient's actual movement trajectory, whether the target danger area has been entered and the actual arrival time are collected after the warning is triggered. Based on the deviation between the patient's actual risk-avoidance results and the high-risk path prediction results, regional transfer feedback samples are constructed. The original transfer probability in the risk area association graph is used as the prior probability, and Bayesian incremental updates are performed using the newly added feedback samples to calculate the corrected regional node transfer probability. The updated transition probabilities are mapped back to the risk area association map, and the sampling weights and trajectory inference parameters of high-risk areas are adjusted synchronously based on the correction results.
