A method and system for identifying abnormal behaviors in patient monitoring using deep learning algorithms

Abstract

This invention provides a method and system for identifying abnormal patient behaviors using deep learning algorithms, relating to the field of medical monitoring technology. First, it acquires raw activity data streams from multiple sensor units within the target monitoring area, collecting data on patients and medical staff. This constructs a spatiotemporal sequence of patient activity trajectories and a spatiotemporal sequence of medical staff operation trajectories. A spatiotemporal correlation mapping process is then performed to generate an interaction event triggering relationship chain. Next, an abnormal behavior identification model is invoked to perform a quantitative analysis of abnormal behavior pattern deviations within the interaction event triggering relationship chain, obtaining abnormal behavior type labels and a set of temporal abnormal behavior fragments. Based on these results, the propagation path of abnormal behavior is deduced, generating a set of spatial diffusion features of abnormal behavior. Finally, personalized monitoring intervention strategy instructions are generated based on this information and sent to the mobile terminal devices of medical staff. This invention can identify abnormal patient behaviors, analyze their propagation patterns, and provide precise intervention strategies for medical staff.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, and more specifically, to a method and system for identifying abnormal patient monitoring behaviors using deep learning algorithms. Background Technology

[0002] In the field of medical monitoring, the identification and intervention of abnormal patient behavior is a crucial link in ensuring patient safety and improving the quality of medical care. Traditional patient monitoring methods mainly rely on manual observation and regular rounds by medical staff. These methods are not only labor-intensive but also difficult to achieve real-time and comprehensive monitoring of patient behavior. With the development of sensor technology, sensor-based patient monitoring systems have gradually emerged, collecting patient activity data by deploying various sensors in the monitored area.

[0003] However, most existing sensor-based monitoring methods focus solely on the patient's own activity data, neglecting the impact of healthcare workers' actions within the monitored area. In reality, various complex situations can arise during patient-career interactions, closely related to the occurrence of abnormal patient behavior. For example, improper actions by healthcare workers may trigger discomfort in the patient, leading to abnormal behavior; or the patient's abnormal behavior may be a result of some stimulus during interaction with healthcare workers. However, current technologies cannot effectively integrate patient and healthcare worker activity data, accurately identify abnormal behavior patterns during interactions, or deeply analyze the triggering and propagation pathways of abnormal behavior. Consequently, they cannot provide healthcare workers with precise and personalized monitoring and intervention strategies, limiting the actual effectiveness of medical monitoring. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for identifying abnormal patient monitoring behaviors using a deep learning algorithm, the method comprising:

[0005] The raw patient activity data stream is acquired from multiple sensor units deployed within the target monitoring area. The raw patient activity data stream includes time-series activity data units generated by the monitored object within the target monitoring area during a continuous monitoring period, as well as medical operation data units generated by medical staff performing monitoring tasks within the target monitoring area during the continuous monitoring period.

[0006] Based on the time-series activity data units in the original patient activity data stream, a spatiotemporal sequence of the activity trajectory of the monitored object in the target monitoring area is constructed. At the same time, based on the medical and nursing operation data units in the original patient activity data stream, a spatiotemporal sequence of the operation trajectory of medical staff in the target monitoring area is constructed.

[0007] The spatiotemporal sequence of the activity trajectory and the spatiotemporal sequence of the operation trajectory are subjected to monitoring spatiotemporal correlation mapping to generate an interaction event triggering relationship chain between the monitored object and the medical staff. The interaction event triggering relationship chain includes multiple interaction event nodes arranged in chronological order and directed association edges connecting the interaction event nodes. The interaction event nodes correspond to the spatial location overlap event or operation contact event of the monitored object and the medical staff at a specific time point. The directed association edges correspond to the time interval parameter and spatial distance change parameter between the interaction event nodes.

[0008] A pre-built abnormal behavior identification model is invoked to perform abnormal behavior pattern deviation quantification analysis on the interaction event triggering relationship chain, thereby obtaining the abnormal behavior type label of the monitored object within the continuous monitoring period and the time-series abnormal behavior segment set corresponding to the abnormal behavior type label. The time-series abnormal behavior segment set contains multiple abnormal behavior data segments extracted from the time-series activity data unit and the interaction event node identifier associated with each abnormal behavior data segment in the interaction event triggering relationship chain.

[0009] Based on the set of time-series abnormal behavior segments and the interaction event triggering relationship chain, the abnormal behavior inducement propagation path inference processing is performed to generate an abnormal behavior spatial diffusion feature set containing the spatial propagation direction sequence and spatial propagation speed change rate of the abnormal behavior data segments in the target monitoring area.

[0010] Based on the abnormal behavior type label, the abnormal behavior spatial diffusion feature set, and the temporal abnormal behavior fragment set, a personalized monitoring intervention strategy instruction is generated for the monitored object, and the personalized monitoring intervention strategy instruction is sent to the mobile terminal device worn by the medical staff to trigger the corresponding monitoring operation prompt.

[0011] Furthermore, embodiments of the present invention also provide a patient monitoring abnormal behavior recognition system incorporating deep learning algorithms, comprising:

[0012] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the above-described method for identifying abnormal patient monitoring behavior incorporating a deep learning algorithm by executing the machine-executable instructions.

[0013] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions stored in a computer-readable storage medium, wherein a processor of a patient monitoring abnormal behavior recognition system incorporating a deep learning algorithm reads the machine-executable instructions from the computer-readable storage medium, and the processor executes the machine-executable instructions, causing the patient monitoring abnormal behavior recognition system incorporating a deep learning algorithm to execute the aforementioned patient monitoring abnormal behavior recognition method incorporating a deep learning algorithm.

[0014] Based on the above, by comprehensively acquiring the original activity data streams of patients and medical staff within the target monitoring area, a spatiotemporal sequence of patient activity trajectories and a spatiotemporal sequence of medical staff operation trajectories are constructed. These are then subjected to spatiotemporal correlation mapping to generate an interaction event triggering relationship chain. This allows for the capture of interactions between patients and medical staff, including events such as spatial overlap and operational contact, as well as the time intervals and spatial distance changes between these events. A pre-built abnormal behavior recognition model is then invoked to perform abnormal behavior pattern deviation quantification analysis on the interaction event triggering relationship chain. This accurately identifies the patient's abnormal behavior type labels and the corresponding set of temporal abnormal behavior fragments, clarifying the specific time of occurrence of the abnormal behavior. Based on the time and data range, and using a set of temporal abnormal behavior fragments and interactive event triggering relationship chains, the propagation path of abnormal behavior is deduced and processed to generate a set of spatial diffusion characteristics of abnormal behavior. This allows for a deeper understanding of the propagation patterns of abnormal behavior within the monitored area, including changes in propagation direction and speed. Finally, based on the abnormal behavior type label, the set of spatial diffusion characteristics of abnormal behavior, and the set of temporal abnormal behavior fragments, personalized monitoring and intervention strategy instructions are generated and sent to the mobile terminal devices worn by medical staff. This provides medical staff with accurate and timely monitoring operation prompts, effectively improving the targeting and timeliness of medical monitoring, reducing the risks caused by abnormal patient behavior, and improving the overall quality and efficiency of medical monitoring. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the execution flow of the patient monitoring abnormal behavior identification method combined with deep learning algorithm provided in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of exemplary hardware and software components of a patient monitoring abnormal behavior recognition system that combines deep learning algorithms, provided in an embodiment of the present invention. Detailed Implementation

[0017] Figure 1 This is a flowchart illustrating a method for identifying abnormal patient monitoring behaviors using a deep learning algorithm, provided in one embodiment of the present invention. A detailed description follows.

[0018] Step S110: Obtain the raw patient activity data stream collected by multiple sensing units deployed in the target monitoring area. The raw patient activity data stream includes time-series activity data units generated by the monitored object in the target monitoring area during the continuous monitoring period and medical operation data units generated by medical staff performing monitoring tasks in the target monitoring area during the continuous monitoring period.

[0019] In this embodiment, the target monitoring area is set as ward 301 of the internal medicine department of a hospital, where various sensing devices are deployed to collect relevant data. The continuous monitoring period is set to 24 hours from 00:00 on May 10, 2024 to 24:00 on May 10, 2024. The monitored patient is patient A in this ward, and the medical staff performing the monitoring tasks include nurse A, nurse B, and doctor C. The raw patient activity data stream is obtained through the collaborative collection of multiple sensing units, and this data will be used for subsequent construction of activity trajectories and identification of abnormal behaviors.

[0020] For example, in step S111: multiple infrared motion sensing units deployed within the target monitoring area are activated. Each infrared motion sensing unit covers a sub-area within the target monitoring area. The infrared motion sensing unit detects changes in the infrared radiation energy of the monitored object within the sub-area. When the change in infrared radiation energy exceeds a preset energy change threshold, an infrared trigger signal is generated, which includes the sensor unit identifier, the energy change trigger time point, and the energy change amplitude.

[0021] In ward 301, eight infrared motion sensing units, numbered IR001 to IR008, are evenly installed on the ceiling. Each sensing unit covers a sub-area of ​​approximately 6 square meters, enabling comprehensive monitoring of the entire ward. Each infrared motion sensing unit detects the infrared radiation energy within its sub-area at a frequency of 10 times per second. The preset energy change threshold is set at 30% of the baseline energy value, which is the average ambient infrared radiation energy collected when the ward is unoccupied. When patient A moves within the ward and enters a specific sub-area, the infrared radiation energy detected by the infrared motion sensing unit in that area will change. For example, when patient A gets up from the bed and walks towards the window, passing through the sub-areas covered by IR003 and IR005, these two sensing units will sequentially detect an energy change exceeding 30%, generating infrared trigger signals. Each infrared trigger signal contains information such as "IR003, 2024-05-10 08:15:23, 45%", where "IR003" is the sensor unit identifier, "2024-05-10 08:15:23" is the energy change trigger time point, and "45%" is the energy change amplitude.

[0022] Step S112: Simultaneously activate multiple pressure distribution sensing units deployed within the target monitoring area. The pressure distribution sensing units are integrated into the bed mattress, wheelchair cushion, and key walkway area on the ground. The pressure distribution sensing units collect pressure distribution images of the monitored object applied to the bed mattress, wheelchair cushion, and key walkway area on the ground. The pressure distribution images include multiple pressure sensing points and the pressure value corresponding to each pressure sensing point.

[0023] The mattresses of the beds in Ward 301 integrate a 20×20 pressure sensor array, while the wheelchair cushions integrate a 10×10 pressure sensor array. Key aisle areas on the floor (from the bed to the restroom and from the bed to the door) are covered with pressure sensor strips 0.8 meters wide and 3 meters and 4 meters long respectively. These strips consist of a 5×100 pressure sensor array. The pressure distribution sensing units collect pressure distribution images at a frequency of 5 times per second. When patient A is lying in the bed, the pressure distribution image from the mattress displays the pressure values ​​of the sensor arrays corresponding to the areas in contact with patient A's body; for example, the pressure value is higher in the back area and lower in the limb areas. When patient A is sitting in the wheelchair, the pressure distribution image from the wheelchair cushion shows the pressure values ​​of the sensor arrays corresponding to the buttocks and legs. When patient A walks in the aisle, the floor pressure sensor strips sequentially record the pressure distribution image as each foot contacts the ground, with the pressure value of each sensor array changing with each step.

[0024] Step S113: Initially correlate the infrared trigger signal and the pressure distribution image according to their respective generation time points to generate the basic activity data part in the time-series activity data unit.

[0025] In this embodiment, the timestamps of the infrared trigger signal and the pressure distribution image can be precisely aligned, with the timestamp accuracy down to the millisecond level. For each infrared trigger signal's energy change trigger time point, the corresponding pressure distribution image is searched within a 500-millisecond time range before and after it. For example, if IR003 detects an energy change at 08:15:23.350 on 2024-05-10, the pressure distribution image for the time period from 08:15:22.850 to 08:15:23.850 on 2024-05-10 can be searched. If, during this time period, the pressure distribution image of the bed mattress shows a sudden decrease in pressure, while the pressure distribution image of the floor pressure sensor shows a new pressure value, it indicates that Patient A may get up from the bed and start walking. At this time, the infrared trigger signal and the corresponding pressure distribution image are associated as the basic activity data part in the time-series activity data unit, and recorded as "2024-05-10 08:15:23, IR003 triggered, bed pressure decreases, floor aisle pressure appears, activity type: getting up and walking".

[0026] Step S114: Simultaneously activate multiple audio acquisition units deployed within the target monitoring area, acquire environmental audio signals within the target monitoring area through the audio acquisition units, perform voice activity detection processing on the environmental audio signals, and separate voice segments containing the voices of the monitored object and voice segments containing the voices of medical staff from the environmental audio signals.

[0027] Four audio acquisition units were installed in ward 301, located in the four corners of the room, with a sampling frequency of 44.1 kHz and a sampling bit depth of 16 bits. The audio acquisition units continuously collected ambient audio signals and processed them using a Gaussian mixture model-based speech activity detection algorithm. This algorithm analyzes features such as short-time energy and zero-crossing rate of the audio signal to determine whether the current audio segment contains speech activity. When speech activity is detected, the speaker's voiceprint features are further extracted using speech recognition technology. The system pre-stores voiceprint templates for patient A, nurses A and B, and doctor C. By matching the extracted voiceprint features with the voiceprint templates, speech segments containing patient A's voice and speech segments containing medical staff's voices are separated. For example, when nurse A says "It's time to take your temperature" to patient A, the audio acquisition unit captures this audio, processes it, identifies it as a medical staff member's voice segment, and labels it as nurse A.

[0028] Step S115: Perform speech recognition processing on the speech segment containing the voices of medical staff, extract the text content of the speech segment, and parse out the keywords related to medical operations from the text content. The keywords related to medical operations include keywords for vital sign measurement, intravenous injection, wound care, and body position adjustment.

[0029] For the separated speech segments containing medical staff's voices, a deep neural network-based speech recognition model is used to convert the speech signal into text content. This speech recognition model has been trained on a speech corpus specific to hospital scenarios and has a high accuracy rate in recognizing medical terminology. After obtaining the text content, keywords related to medical procedures are parsed using keyword matching. For example, when the text content is "Now I'll measure your blood pressure and temperature," keywords such as "measure," "blood pressure," and "temperature" are extracted, corresponding to vital sign measurement operation keywords. When the text content is "Intravenous injection is required," the keyword "intravenous injection" is extracted, corresponding to intravenous injection operation keywords.

[0030] Step S116: Based on the parsed keywords related to medical and nursing operations and the generation time of the voice segment, generate a preliminary operation record in the medical and nursing operation data unit. The preliminary operation record includes candidate tags for operation type and candidate time points for operation occurrence.

[0031] In this embodiment, the parsed keywords can be matched with a preset operation type mapping table to determine the candidate tags for the operation type. The preset mapping relationships are as follows: "measurement," "blood pressure," and "body temperature" correspond to "vital sign measurement operation"; "intravenous injection" corresponds to "intravenous injection operation"; "wound care" corresponds to "wound care operation"; "turning over" and "adjusting body position" correspond to "body position adjustment operation". The generation time of the voice segment is the candidate time point for the operation. For example, for a voice segment parsed with the keywords "measurement," "blood pressure," and "body temperature" and generated at 2024-05-10 09:30:15, the initial operation record generated is "Operation type candidate tag: vital sign measurement operation, operation occurrence candidate time point: 2024-05-10 09:30:15".

[0032] Step S117: Obtain the acceleration data sequence collected by the mobile terminal device worn by medical staff through its built-in accelerometer, perform action pattern recognition processing on the acceleration data sequence, and identify the specific action type performed by the medical staff. The specific action type includes walking, standing, bending, pushing a cart, and handheld device operation.

[0033] The mobile terminal devices (such as smart bracelets) worn by medical staff A, B, and C have built-in triaxial accelerometers that collect acceleration data sequences at a sampling frequency of 50Hz. This embodiment preprocesses the acceleration data sequences, including noise and baseline drift removal, and then extracts time-domain features (such as mean, variance, and peak value) and frequency-domain features (such as spectral energy distribution). A support vector machine (SVM) classification algorithm is used to perform action pattern recognition on the extracted features. This algorithm has been trained on a large number of acceleration data samples containing different actions of medical staff. For example, when nurse A walks towards patient A's bedside, the acceleration data sequence collected by her mobile terminal device shows periodic changes, which the algorithm identifies as a walking action; when nurse A bends over to place the blood pressure monitor for patient A, the acceleration data sequence shows specific peak and trough changes, which are identified as a bending action.

[0034] Step S118: Cross-validate the identified specific action type with the operation type candidate label in the preliminary operation record. When the specific action type matches the typical action pattern corresponding to the operation type candidate label, confirm that the preliminary operation record is valid and generate a medical operation data unit containing the confirmed operation type code and the candidate time point of the operation occurrence.

[0035] In this embodiment, typical action patterns corresponding to candidate tags for each operation type can be preset. For example, typical action patterns for vital sign measurement operations include standing (the nurse stands beside the patient's bed) and bending over (placing the measuring instrument); typical action patterns for intravenous injection operations include pushing a cart (pushing a treatment cart), standing, and operating a handheld device (operating a syringe). When the candidate tag for the operation type in the preliminary operation record is "vital sign measurement operation," and the corresponding specific action type of the medical staff includes standing and bending over, the two match, confirming the validity of the preliminary operation record. Then, a unique operation type code is assigned to this operation type, such as "001" representing vital sign measurement operation, generating a medical staff operation data unit, such as "Operation type code: 001, operation occurrence time: 2024-05-10 09:30:15."

[0036] Step S120: Construct a spatiotemporal sequence of the activity trajectory of the monitored object within the target monitoring area based on the temporal activity data units in the original patient activity data stream, and simultaneously construct a spatiotemporal sequence of the operation trajectory of medical staff within the target monitoring area based on the medical operation data units in the original patient activity data stream.

[0037] After acquiring the original patient activity data stream, the spatial location and time information of the temporal activity data units were analyzed and processed to construct a spatiotemporal sequence of Patient A's activity trajectory in ward 301. This sequence reflects Patient A's location changes over 24 hours. Simultaneously, for the medical staff operation data units, the operation location and time information were also analyzed to construct a spatiotemporal sequence of the medical staff's operation trajectory within that area, demonstrating the location and time distribution of the operations performed by the medical staff.

[0038] Step S121: Parse the time-series activity data units in the original patient activity data stream, and extract the monitored object identifier, data acquisition timestamp, and spatial location coordinates of the monitored object at the data acquisition timestamp contained in each time-series activity data unit. The spatial location coordinates are calculated based on the output signal of the positioning sensor unit deployed in the target monitoring area.

[0039] The time-series activity data units in the original patient activity data stream are stored in a structured data format, with each data unit containing multiple fields. The monitored object identifier is used to uniquely identify the patient; in this embodiment, patient A's identifier is "P001". The data acquisition timestamp is accurate to the second, in the format "YYYY-MM-DDHH:MM:SS". Spatial location coordinates are calculated by UWB (Ultra-Wideband) positioning sensor units deployed in the ward, using a Cartesian coordinate system with the lower left corner of the ward as the origin (0, 0), the x-axis along the length of the room, and the y-axis along the width of the room, in meters. The positioning sensor units calculate the distance from the monitored object to each sensor by measuring the signal time-of-flight, and then calculate the spatial location coordinates using triangulation. For example, a time-series activity data unit is “P001, 2024-05-10 08:10:00, (3.5, 2.8)”, which means that the spatial coordinates of patient A at 08:10:00 on May 10, 2024 are (3.5, 2.8) meters.

[0040] Step S122: Sort all the extracted spatial coordinates containing the identifier of the monitored object in ascending order according to the data acquisition timestamp to obtain the first spatial coordinate sequence of the monitored object changing with time during the continuous monitoring period.

[0041] All extracted spatial coordinates containing the patient A identifier "P001" are arranged in ascending order of data collection timestamps. For example, starting from 2024-05-10 00:00:00, the spatial coordinates of each time point are arranged sequentially to form the first spatial coordinate sequence. This sequence is an ordered list, where each element contains a timestamp and the corresponding coordinates, such as "[(2024-05-10 00:00:00, (3.0, 2.5)), (2024-05-10 00:01:00, (3.0, 2.5)), ..., (2024-05-10 23:59:00, (3.0, 2.5))]", reflecting the location of patient A at each time point within 24 hours.

[0042] Step S123: Perform trajectory point interpolation encryption processing on the first spatial position coordinate sequence. Insert multiple interpolated position coordinates between the spatial position coordinates corresponding to two adjacent data acquisition timestamps according to a preset time interpolation density to obtain a second spatial position coordinate sequence containing the original spatial position coordinates and the interpolated position coordinates.

[0043] The preset time interpolation density is one point every 10 seconds, meaning that 5 interpolation coordinates need to be inserted between two adjacent original data acquisition timestamps (assuming an interval of 1 minute). A linear interpolation algorithm is used. For two adjacent original coordinate points (t1, (x1, y1)) and (t2, (x2, y2)), where t2-t1=60 seconds, interpolation coordinates need to be inserted at 5 time points: t1+10 seconds, t1+20 seconds, t1+30 seconds, t1+40 seconds, and t1+50 seconds. The interpolation coordinates are calculated as follows: for time point t=t1+Δt (Δt is a multiple of 10 seconds), x=x1+(x2-x1)*(Δt / (t2-t1)), y=y1+(y2-y1)*(Δt / (t2-t1)). For example, if the original coordinates are (2024-05-10 08:10:00, (3.5, 2.8)) and (2024-05-10 08:11:00, (4.5, 2.8)), then the interpolated coordinates inserted at 2024-05-10 08:10:10 are x=3.5+(4.5-3.5)*(10 / 60)=3.5+0.1667≈3.6667, y=2.8+(2.8-2.8)*(10 / 60)=2.8, i.e. (3.6667, 2.8). Through this interpolation process, a second spatial location coordinate sequence is obtained. This sequence has more frequent time intervals and can more accurately reflect the patient's activity trajectory.

[0044] Step S124: Construct a set of movement speed vectors of the monitored object based on the time attribute corresponding to each spatial position coordinate in the second spatial position coordinate sequence. Each movement speed vector in the set of movement speed vectors is calculated by dividing the displacement vector between two adjacent spatial position coordinates by the corresponding time interval.

[0045] In the second spatial position coordinate sequence, each coordinate point has a corresponding timestamp. The displacement vector between two adjacent coordinate points (t_i, (x_i, y_i)) and (t_{i+1}, (x_{i+1}, y_{i+1})) is calculated as (x_{i+1}-x_i, y_{i+1}-y_i), with a time interval Δt = t_{i+1}-t_i (in seconds). Therefore, the x-component of the velocity vector is (x_{i+1}-x_i) / Δt, and the y-component is (y_{i+1}-y_i) / Δt, in meters per second. For example, if the time interval between two adjacent coordinate points is 10 seconds and the displacement vector is (0.1667, 0) meters, then the velocity vector is (0.1667 / 10, 0 / 10) = (0.01667, 0) meters per second. Arrange the velocity vectors calculated from all adjacent coordinate points in chronological order to form a set of velocity vectors.

[0046] Step S125: Associate and store the second spatial position coordinate sequence and the set of movement speed vectors to generate the spatiotemporal sequence of the activity trajectory of the monitored object within the target monitoring area. The spatiotemporal sequence of the activity trajectory includes each spatial position coordinate in the second spatial position coordinate sequence and the instantaneous movement speed vector and instantaneous movement direction angle associated with that spatial position coordinate.

[0047] Each coordinate point in the second spatial position coordinate sequence is associated with the corresponding velocity vector in the set of movement velocity vectors, and the instantaneous movement direction angle is calculated simultaneously. The instantaneous movement direction angle is calculated using the x and y components of the velocity vector, with the formula θ = arctan2(y_speed, x_speed), where θ is in radians, and then converted to degrees (0° to 360°). For example, when the velocity vector is (0.01667, 0) m / s, y_speed is 0, x_speed is positive, so the direction angle is 0°, indicating movement along the positive x-axis. The spatiotemporal sequence of the activity trajectory is stored in the form of a list. Each element contains a timestamp, spatial location coordinates, instantaneous velocity vector, and instantaneous direction angle, such as "[(2024-05-10 08:10:00, (3.5, 2.8), (0, 0), 0°), (2024-05-10 08:10:10, (3.6667, 2.8), (0.01667, 0), 0°), ...]".

[0048] Step S126: Synchronously parse the medical and nursing operation data units in the original patient activity data stream, and extract the medical and nursing personnel identifier, operation execution timestamp, and operation position coordinates of the medical and nursing personnel at the operation execution timestamp contained in each medical and nursing operation data unit. The operation position coordinates are calculated based on the output signal of the positioning sensing unit or the built-in positioning module of the medical and nursing personnel's handheld terminal device.

[0049] The medical and nursing operation data units in the original patient activity data stream also contain multiple fields, such as medical staff identifiers like "N001" for Nurse A, "N002" for Nurse B, and "D001" for Doctor C. The operation execution timestamp format is consistent with the data acquisition timestamp. There are two ways to obtain the operation location coordinates: for operations performed at a fixed location in the ward, they are calculated by the UWB positioning sensor unit in the ward; for operations performed by medical staff using handheld terminal devices (such as tablets), they are calculated by the GPS and Bluetooth positioning modules built into the terminal device, mainly relying on Bluetooth positioning in indoor environments. For example, if Nurse A measures the blood pressure of Patient A at 09:30:15 on 2024-05-10, the operation location coordinates obtained through UWB positioning are (3.2, 2.6) meters, and the corresponding medical and nursing operation data unit is "N001, 2024-05-10 09:30:15, (3.2, 2.6), 001", where "001" is the operation type code.

[0050] Step S127: Sort all extracted operation location coordinates containing medical staff identifiers in ascending order according to the operation execution timestamp to obtain the third spatial location coordinate sequence of the medical staff changing over time during the continuous monitoring period.

[0051] For each medical staff member, the extracted operation location coordinates containing their identifier are arranged in order of operation execution timestamp from earliest to latest. For example, the third-space location coordinate sequence of Nurse A is "[(2024-05-10 08:00:00, (1.0, 1.0)), (2024-05-10 09:30:15, (3.2, 2.6)), ...]", where (1.0, 1.0) may be the location coordinates of the nurse station.

[0052] Step S128: Perform trajectory point smoothing filtering on the third spatial position coordinate sequence to eliminate abnormal coordinate jump points caused by instantaneous fluctuations in the positioning signal, and generate a smoothed fourth spatial position coordinate sequence.

[0053] A moving average filtering algorithm is used to process the third spatial coordinate sequence, with a window size of 5. For each point in the sequence, the average of the coordinates of the two points before and after it (a total of 5 points) is taken as the smoothed coordinate of that point. For example, if a segment in the original third spatial coordinate sequence is "(t1, (x1, y1)), (t2, (x2, y2)), (t3, (x3, y3)), (t4, (x4, y4)), (t5, (x5, y5)),...", then the smoothed coordinates corresponding to t3 are ((x1+x2+x3+x4+x5) / 5, (y1+y2+y3+y4+y5) / 5). This method effectively eliminates sudden coordinate jumps caused by positioning signal interference, making the trajectory smoother and obtaining the fourth spatial coordinate sequence.

[0054] Step S129: Based on the time attribute corresponding to each operation position coordinate in the fourth spatial position coordinate sequence and the operation type code contained in the medical operation data unit, construct the spatial position mapping relationship of each medical operation performed by the medical staff during the movement, and generate an operation trajectory spatiotemporal sequence containing the operation position coordinate, the time point of arrival at the operation position coordinate, the operation type code performed at the operation position coordinate, and the time point of departure from the operation position coordinate.

[0055] For each operational position coordinate in the fourth spatial position coordinate sequence, the arrival time is determined by combining the operation execution timestamp in the medical and nursing operation data unit. The departure time is determined by analyzing subsequent coordinate changes. When the position coordinate of the medical staff moves away from the current operational position coordinate by a certain distance (e.g., 1 meter), the corresponding timestamp is the departure time. The operation type code is directly extracted from the medical and nursing operation data unit. For example, if Nurse A arrives at the position (3.2, 2.6) meters away at 09:30:15 on 2024-05-10 and performs a vital sign measurement operation with operation type code "001", and leaves the position at 09:35:20 on 2024-05-10, the corresponding record in the spatiotemporal sequence of the operation trajectory would be "Operation position coordinates: (3.2, 2.6), arrival time: 2024-05-10 09:30:15, operation type code: 001, departure time: 2024-05-10 09:35:20".

[0056] Step S130: Perform monitoring spatiotemporal association mapping processing on the activity trajectory spatiotemporal sequence and the operation trajectory spatiotemporal sequence to generate an interaction event triggering relationship chain between the monitored object and the medical staff. The interaction event triggering relationship chain includes multiple interaction event nodes arranged in chronological order and directed association edges connecting the interaction event nodes. The interaction event nodes correspond to the spatial location overlap event or operation contact event of the monitored object and the medical staff at a specific time point. The directed association edges correspond to the time interval parameter and spatial distance change parameter between the interaction event nodes.

[0057] By performing temporal and spatial correlation analysis between the spatiotemporal sequence of patient A's activity trajectory and the spatiotemporal sequence of medical staff's operation trajectory, the overlapping spatial locations or operational contact between the two at specific time points are identified. These situations are used as interaction event nodes, and the time interval and spatial distance changes between nodes are calculated as parameters of directed association edges, thereby constructing an interaction event triggering relationship chain to clearly demonstrate the interaction process between patients and medical staff.

[0058] Step S131: Obtain the first spatial location coordinates of the monitored object at a specific time point contained in the activity trajectory spatiotemporal sequence and the second spatial location coordinates of the medical staff at the same time point contained in the operation trajectory spatiotemporal sequence. The same time point is determined by performing a global time axis alignment operation between the data acquisition timestamp and the operation execution timestamp.

[0059] The global timeline alignment operation uses the hospital's unified clock system as a reference, converting the data acquisition timestamps of the activity trajectory spatiotemporal sequence and the operation execution timestamps of the operation trajectory spatiotemporal sequence to the time under this reference clock. Then, at 1-second intervals, specific time points are selected on the global timeline, such as integer moments per second (e.g., 00:00:00, 00:00:01, ...). For each specific time point, the first spatial position coordinates of patient A corresponding to that time point are found from the activity trajectory spatiotemporal sequence (if no directly recorded coordinates are available, they are obtained through interpolation), and the second spatial position coordinates of the medical staff corresponding to that time point are found from the operation trajectory spatiotemporal sequence. For example, at the specific time point 09:30:00 on 2024-05-10, patient A's first spatial position coordinates are (3.5, 2.8) meters, and nurse A's second spatial position coordinates are (3.2, 2.6) meters.

[0060] Step S132: Calculate the Euclidean spatial distance between the first spatial coordinates and the second spatial coordinates to obtain the instantaneous spatial distance between the monitored object and the medical staff at the same time point.

[0061] The Euclidean formula for calculating spatial distance is d = √[(x2-x1)² + (y2-y1)²], where (x1, y1) are the first spatial coordinates and (x2, y2) are the second spatial coordinates. For example, if patient A's coordinates are (3.5, 2.8) meters and nurse A's coordinates are (3.2, 2.6) meters, then the instantaneous spatial distance d = √[(3.2-3.5)² + (2.6-2.8)²] = √[(-0.3)² + (-0.2)²] = √(0.09 + 0.04) = √0.13 ≈ 0.3606 meters.

[0062] Step S133: Compare the instantaneous spatial distance value with a preset interaction trigger distance threshold. When the instantaneous spatial distance value is less than or equal to the interaction trigger distance threshold, determine that the monitored object and the medical staff have a spatial position overlap event at the same time point, and generate a spatial overlap event record containing the same time point, the first spatial position coordinates, the second spatial position coordinates, and the instantaneous spatial distance value.

[0063] The preset interaction trigger distance threshold is set to 1 meter based on the ward environment and monitoring needs. When the calculated instantaneous spatial distance value is less than or equal to 1 meter, it is considered that patient A and medical staff overlap in spatial position. For example, the calculated instantaneous spatial distance value is approximately 0.3606 meters, which is less than 1 meter. Therefore, it is determined that a spatial overlap event occurred at 09:30:00 on 2024-05-10, and a spatial overlap event record is generated as "Time point: 2024-05-10 09:30:00, First spatial position coordinates: (3.5, 2.8), Second spatial position coordinates: (3.2, 2.6), Instantaneous spatial distance value: 0.3606 meters".

[0064] Step S134: Synchronously parse the operation type code corresponding to each operation position coordinate in the spatiotemporal sequence of the operation trajectory, and extract the set of medical and nursing operation types that directly affect the monitored object from the operation type code. The set of medical and nursing operation types that directly affect the monitored object includes vital sign measurement operation type, intravenous injection operation type, wound care operation type, and body position adjustment operation type.

[0065] Each record in the spatiotemporal sequence of the operation trajectory contains an operation type code. A pre-defined list of operation type codes directly affecting the monitored individual is provided; for example, "001" corresponds to vital sign measurement, "002" to intravenous injection, "003" to wound care, and "004" to postural adjustment. By traversing the spatiotemporal sequence of the operation trajectory, records whose operation type codes belong to this list are extracted, forming a set of medical and nursing operation types. For instance, if the spatiotemporal sequence of the operation trajectory contains records with operation type codes "001" and "002," the operation types corresponding to these records belong to the set of medical and nursing operation types directly affecting the monitored individual.

[0066] Step S135: Determine the occurrence time of each medical operation that directly affects the monitored object based on the operation execution timestamp, obtain the first spatial position coordinates of the monitored object corresponding to the occurrence time, and calculate the operation contact space distance between the first spatial position coordinates of the monitored object corresponding to the occurrence time and the operation position coordinates of the medical operation.

[0067] For each operation in the set of medical and nursing operation types, its occurrence time is the operation execution timestamp. The first spatial position coordinates of patient A corresponding to this occurrence time point are found in the spatiotemporal sequence of the activity trajectory. Then, the Euclidean spatial distance between this coordinate and the operation position coordinates is calculated, which is the operation contact spatial distance. For example, if the occurrence time of an intravenous injection operation is 2024-05-10 14:20:00, the operation position coordinates are (3.3, 2.7) meters, and the first spatial position coordinates of patient A at this time are (3.3, 2.7) meters, then the operation contact spatial distance d = 0 meters.

[0068] Step S136: When the operation contact space distance is less than or equal to a preset operation contact distance threshold, it is determined that the monitored object and the medical staff have an operation contact event at the time of occurrence, and an operation contact event record containing the time of occurrence, the first spatial location coordinates, the operation location coordinates, and the operation type code is generated.

[0069] The preset threshold for the operational contact distance is set to 0.5 meters, because operations that directly affect the monitored object usually require the two to be very close. When the operational contact spatial distance is less than or equal to 0.5 meters, an operational contact event is determined to have occurred. For example, the operational contact spatial distance for the above intravenous injection operation is 0 meters, which is less than 0.5 meters. Therefore, the operational contact event record is generated as "Occurrence time: 2024-05-10 14:20:00, First spatial position coordinates: (3.3, 2.7), Operation position coordinates: (3.3, 2.7), Operation type code: 002".

[0070] Step S137: Merge and sort the spatial overlap event records and the operation contact event records according to the time points they contain to obtain an initial interaction event sequence. The initial interaction event sequence contains multiple interaction event units arranged in chronological order, and each interaction event unit corresponds to a spatial overlap event or an operation contact event.

[0071] Collect all spatial overlap event records and operational contact event records, extract the time point of each record, and then merge and sort the records in ascending order of time. For example, if the time point of the spatial overlap event record is 2024-05-10 09:30:00 and the time point of the operational contact event record is 2024-05-10 14:20:00, then the initial interaction event sequence will arrange these two events in chronological order. If the time points of the two events are the same, the operational contact event will be prioritized.

[0072] Step S138: Calculate the time interval parameter based on the time points of two adjacent interactive event units in the initial interactive event sequence, and calculate the spatial distance change parameter based on the spatial position coordinates corresponding to the two adjacent interactive event units.

[0073] For two adjacent interactive event units in the initial interactive event sequence, the time interval parameter is the difference between the time of the later event and the time of the earlier event, in seconds. The calculation of the spatial distance change parameter requires first determining the spatial position of each interactive event unit. For spatially overlapping events, the first spatial position coordinates of the monitored object are taken; for operational contact events, the first spatial position coordinates of the monitored object are also taken. Then, the Euclidean spatial distance between the spatial position coordinates of two adjacent events is calculated as the spatial distance change parameter. For example, if the time points of two adjacent events in the initial interactive event sequence are t1 and t2, and their corresponding spatial position coordinates are (x1, y1) and (x2, y2), then the time interval parameter Δt = t2 - t1, and the spatial distance change parameter Δd = √[(x2 - x1)² + (y2 - y1)²].

[0074] Step S139: Using each interactive event unit in the initial interactive event sequence as an interactive event node, and the directed connection line between two adjacent interactive event nodes as a directed association edge, the time interval parameter and the spatial distance change parameter are attached to the corresponding directed association edge to generate the interactive event triggering relationship chain between the monitored object and the medical staff.

[0075] Each interactive event unit in the initial interactive event sequence is abstracted as an interactive event node. Each node contains information such as the event type (spatial overlap or operational contact), time point, and spatial coordinates. Adjacent nodes are connected by directed edges, with the direction pointing from the earlier node to the later node. Time interval parameters and spatial distance variation parameters are labeled on these directed edges. For example, if node A is a spatial overlap event at 09:30:00 on 2024-05-10, and node B is an operational contact event at 14:20:00 on 2024-05-10, the directed edge connecting A and B is labeled with a time interval parameter of 19800 seconds (5 hours and 30 minutes) and a spatial distance variation parameter (assuming node A's coordinates are (3.5, 2.8) and node B's coordinates are (3.3, 2.7), then Δd ≈ 0.2236 meters). This constructs the interactive event triggering chain.

[0076] Step S1310: After generating the spatial overlap event record and the operation contact event record, obtain the operation type code contained in the operation contact event record, and search for the corresponding standard operation duration from the preset operation duration database according to the operation type code.

[0077] The pre-defined operation duration database stores the standard operation duration corresponding to each operation type code. This data is based on statistical analysis of long-term operation records from the hospital. For example, the standard operation duration for operation type code "001" (vital sign measurement operation) is 5 minutes, "002" (intravenous injection operation) is 15 minutes, "003" (wound care operation) is 20 minutes, and "004" (posture adjustment operation) is 3 minutes. When an operation contact event record is generated, the operation type code, such as "002", is extracted, and the corresponding standard operation duration of 15 minutes is found in the database.

[0078] Step S1311: Calculate the theoretical end time of the operation contact event based on the occurrence time in the operation contact event record and the duration of the standard operation.

[0079] The theoretical end time is equal to the occurrence time plus the standard operation duration. For example, if the occurrence time of the operation contact event is 2024-05-10 14:20:00 and the standard operation duration is 15 minutes, then the theoretical end time is 2024-05-10 14:35:00.

[0080] Step S1312: Within a preset time window before and after the theoretical end time point, detect whether the spatial position coordinates of the medical staff in the spatiotemporal sequence of the operation trajectory are continuously within a preset distance range near the operation position coordinates in the operation contact event record, and detect whether the spatial position coordinates of the monitored object in the spatiotemporal sequence of the activity trajectory are continuously within a preset distance range near the operation position coordinates.

[0081] The preset time window is 5 minutes before and 5 minutes after the theoretical end time, with a preset distance range of 1 meter. For example, if the theoretical end time is 14:35:00, the time window is from 14:30:00 to 14:40:00. Within this time window, the system continuously monitors whether nurse A's spatial coordinates are within 1 meter of the operation position coordinates (3.3, 2.7), and simultaneously monitors whether patient A's spatial coordinates are also within this range.

[0082] Step S1313: If the spatial coordinates of the medical staff and the spatial coordinates of the monitored object are continuously within a preset distance range near the operation position coordinates within a preset time window before and after the theoretical end time point, then the operation contact event is determined to be fully executed, and a complete execution mark is added to the operation contact event record.

[0083] If, within the time window from 14:30:00 to 14:40:00, the spatial coordinates of Nurse A and Patient A are consistently within 1 meter of each other (3.3, 2.7), it indicates that the intravenous injection procedure was successfully performed. Therefore, add a "Complete Execution Mark: Yes" to the contact event record for this procedure.

[0084] Step S1314: If the spatial coordinates of the medical staff or the spatial coordinates of the monitored object leave a preset distance range near the operation position coordinates before the theoretical end time is reached, the operation contact event is determined to be interrupted, and an interruption mark and the interruption time point are added to the operation contact event record.

[0085] Suppose that at 14:32:00, nurse A leaves due to an emergency, and her spatial location coordinates exceed the range of 1 meter within (3.3, 2.7). At this time, the theoretical end time of 14:35:00 has not yet been reached. Therefore, the intravenous injection operation is determined to be interrupted, and "Interruption execution flag: Yes, interruption time: 2024-05-10 14:32:00" is added to the operation contact event record.

[0086] Step S1315: Based on the interruption execution flag and interruption occurrence time of the operation contact event record, update the corresponding interaction event node attributes in the interaction event triggering relationship chain, and attach an operation interruption information field to the interaction event node. The operation interruption information field includes an interruption identifier and an interruption time.

[0087] Update the attributes of the interactive event node corresponding to the operation contact event that was interrupted above, and add an operation interruption information field, such as "Operation interruption information: {Interruption ID: Yes, Interruption time: 2024-05-10 14:32:00}".

[0088] Step S1316: In the subsequent abnormal behavior pattern deviation quantification analysis process, the interactive event node containing the operation interruption information field is taken as the key analysis object, and its initial weight coefficient in the feature input tensor of the abnormal behavior recognition model is increased.

[0089] In the abnormal behavior recognition model, higher initial weight coefficients are set for interactive event nodes that contain operation interruption information fields. For example, the initial weight coefficient of normal nodes is 1.0, while the initial weight coefficient of these nodes is set to 1.5, so that these interruption events that may be related to abnormal behavior are given more attention during model training and inference.

[0090] Step S140: Call the pre-built abnormal behavior recognition model to perform abnormal behavior pattern deviation quantification analysis on the interaction event triggering relationship chain, and obtain the abnormal behavior type label of the monitored object in the continuous monitoring period and the time-series abnormal behavior segment set corresponding to the abnormal behavior type label. The time-series abnormal behavior segment set contains multiple abnormal behavior data segments extracted from the time-series activity data unit and the interaction event node identifier associated with each abnormal behavior data segment in the interaction event triggering relationship chain.

[0091] The pre-built abnormal behavior recognition model is constructed based on deep learning technology. It can perform in-depth analysis of the triggering relationship chain of interactive events and identify abnormal behavior patterns. Through the analysis of this model, the abnormal behavior types of patient A during the 24-hour monitoring period are determined, and corresponding abnormal behavior data segments are extracted from the time-series activity data units to form a set of time-series abnormal behavior fragments.

[0092] Step S141: Input the interaction event triggering relationship chain into the input layer of the abnormal behavior recognition model. The interaction event triggering relationship chain contains the sequence of interaction event nodes and the time interval parameter and spatial distance change parameter carried by the directed associated edge, which together constitute the feature input tensor of the abnormal behavior recognition model.

[0093] The sequence of interactive event nodes in the interactive event triggering relationship chain is arranged chronologically. Each node contains features such as event type, time point, and spatial location coordinates. Directed edges carry time interval parameters and spatial distance variation parameters. This information is converted into a tensor form acceptable to the model. For example, each node is represented as a feature vector, containing event type encoding (0 for spatially overlapping events, 1 for operational contact events), time encoding of the time point (e.g., converting time to seconds in a day), x and y components of the spatial location coordinates, and whether there is an operational interruption indicator (1 for interruption, 0 for no interruption). The parameters of the directed edges serve as the connection features between nodes. The above node features and edge features together form the feature input tensor, whose shape may be (number of nodes, node feature dimension) or (number of edges, edge feature dimension).

[0094] Step S142: The feature input tensor is spatially structured using the feature encoder of the abnormal behavior recognition model. The feature encoder contains multiple stacked graph convolutional network layers. Each graph convolutional network layer performs neighborhood feature aggregation on the interactive event node to generate a high-order spatial neighborhood feature representation of each interactive event node in the current layer. The high-order spatial neighborhood feature representation integrates the topological connection relationship between the interactive event node and its neighboring interactive event nodes in the interactive event triggering relationship chain, as well as the parameter information carried by the directed association edge.

[0095] The feature encoder employs three stacked graph convolutional network layers. The input to each layer is the node feature matrix and adjacency matrix of the previous layer. The adjacency matrix is ​​constructed based on the directed edges of the relationship chain triggered by interaction events. An element A[i][j] of 1 indicates a directed edge from node i to node j, otherwise it is 0. Each graph convolutional network layer performs neighborhood feature aggregation on nodes. Specifically, for each node i, the features of all its neighboring nodes j are aggregated, considering the influence of edge features (time interval parameters and spatial distance variation parameters) on the aggregation weights. For example, edges with smaller time interval parameters and smaller spatial distance variation parameters have greater weights in the aggregation of their corresponding neighboring node features. Through this method, the features of each node not only contain its own information but also integrate information from neighboring nodes and edge parameters, generating a high-order spatial neighborhood feature representation.

[0096] Step S143: Perform feature concatenation on the high-order spatial neighborhood feature representations of different orders output by multiple graph convolutional network layers to obtain the multi-scale spatial structure embedding vector of each interactive event node.

[0097] The three graph convolutional network layers output first-, second-, and third-order high-order spatial neighborhood feature representations, respectively, with each order having a dimension of 64. These three order feature representations are concatenated along their respective feature dimensions, resulting in a multi-scale spatial structure embedding vector for each node with a dimension of 64 × 3 = 192. For example, the first-order feature representation is vector V1, the second-order is V2, and the third-order is V3; the concatenated multi-scale spatial structure embedding vector is [V1, V2, V3].

[0098] Step S144: The time series encoder of the abnormal behavior recognition model performs temporal dependency modeling on the multi-scale spatial structure embedding vector according to the temporal order of the interactive event nodes. The time series encoder includes a bidirectional long short-term memory network structure. The bidirectional long short-term memory network structure generates a temporal context feature vector of the current interactive event node containing forward and backward temporal context information based on the multi-scale spatial structure embedding vector of the current interactive event node and the hidden state vectors of the previous and next time moments.

[0099] The time-series encoder employs a bidirectional long short-term memory (LSTM) network structure with two hidden layers, each containing 128 hidden units. Multi-scale spatial structure embedding vectors are input into the bidirectional LSM network in chronological order of the interactive event nodes. For each node, the forward LSM network calculates the forward hidden state based on the features of all nodes preceding it, and the backward LSM network calculates the backward hidden state based on the features of all nodes following it. The forward and backward hidden states are then concatenated to obtain a temporal context feature vector containing both forward and backward temporal context information, with a dimension of 128 × 2 = 256.

[0100] Step S145: Input the temporal context feature vector into the abnormal behavior recognition model's abnormal pattern classifier. The abnormal pattern classifier includes a multilayer perceptron network structure. The multilayer perceptron network structure performs nonlinear transformation processing on the temporal context feature vector and outputs the probability distribution vector of the monitored object belonging to multiple preset abnormal behavior types at each time point corresponding to each interaction event node.

[0101] The anomaly pattern classifier is a three-layer multilayer perceptron network. The first layer has an input dimension of 256 (the dimension of the temporal context feature vector) and an output dimension of 128; the second layer has an input dimension of 128 and an output dimension of 64; the third layer has an input dimension of 64 and an output dimension equal to the preset number of abnormal behavior types (e.g., 5 abnormal behavior types, then the output dimension is 5). ReLU activation is used between each layer, and the last layer uses Softmax activation, converting the output into a probability distribution vector. Each element of this probability distribution vector corresponds to the probability of an abnormal behavior type, and the sum of all elements is 1. For example, an output probability distribution vector of [0.1, 0.8, 0.05, 0.03, 0.02] indicates that the probability of belonging to the second abnormal behavior type at this time point is 0.8.

[0102] Step S146: Determine the candidate abnormal behavior type label for the time point based on the abnormal behavior type corresponding to the maximum probability value in the probability distribution vector, and merge the time points where the candidate abnormal behavior type labels appear consecutively on the time axis to form multiple candidate abnormal behavior time periods.

[0103] For each interactive event node, the abnormal behavior type with the highest probability value in the probability distribution vector is selected as the candidate abnormal behavior type label for that time point. For example, if the abnormal behavior type corresponding to the highest probability value of 0.8 in the probability distribution vector is "unauthorized departure from bed," then the candidate abnormal behavior type label for that time point is "unauthorized departure from bed." Then, time points on the timeline that consecutively exhibit the same candidate abnormal behavior type label are searched and merged into a single candidate abnormal behavior time period. For example, if time points from t1 to t5 are all labeled as "unauthorized departure from bed," then t1 to t5 are merged into a single candidate abnormal behavior time period.

[0104] Step S147: For each candidate abnormal behavior time period, extract the first interactive event node identifier corresponding to the start time point and the second interactive event node identifier corresponding to the end time point of the candidate abnormal behavior time period.

[0105] Each interactive event node has a unique identifier, such as "E001" or "E002". For a candidate abnormal behavior time period, find the interactive event node corresponding to its start time point and record its first interactive event node identifier, such as "E005"; find the interactive event node corresponding to its end time point and record its second interactive event node identifier, such as "E010".

[0106] Step S148: Based on the first interactive event node identifier and the second interactive event node identifier, extract the original data segment within the corresponding time period from the time-series activity data unit in the original patient activity data stream, generate an abnormal behavior data segment in the time-series abnormal behavior segment set, and use the candidate abnormal behavior type label corresponding to the candidate abnormal behavior time period as the abnormal behavior type label of the abnormal behavior data segment.

[0107] In the temporal activity data units of the original patient activity data stream, based on the time point corresponding to the first interaction event node identifier and the time point corresponding to the second interaction event node identifier, all data units between these two time points are extracted to form an abnormal behavior data segment. For example, if the time point corresponding to the first interaction event node identifier "E005" is t1 and the time point corresponding to the second interaction event node identifier "E010" is t2, then the temporal activity data units between t1 and t2 are extracted as the abnormal behavior data segment, and their abnormal behavior type is labeled as "unauthorized departure from bed".

[0108] Step S149: Perform the same extraction and labeling operation on all candidate abnormal behavior time periods to obtain a set of time-series abnormal behavior segments containing multiple abnormal behavior data segments and the abnormal behavior type label corresponding to each abnormal behavior data segment. At the same time, establish the association between each abnormal behavior data segment and the interactive event node identifier corresponding to the time point it contains.

[0109] Iterate through all candidate abnormal behavior time periods, repeating steps S147 and S148 to obtain multiple abnormal behavior data segments, each with a corresponding abnormal behavior type label. Simultaneously, record the interaction event node identifiers corresponding to the time points contained in each abnormal behavior data segment. For example, if an abnormal behavior data segment contains time points t1 to t5, the corresponding interaction event node identifiers are E005 to E010. Establish this association for subsequent analysis.

[0110] Step S1410: Before the feature encoder of the abnormal behavior recognition model performs spatial structure encoding on the feature input tensor, the interaction event nodes contained in the interaction event triggering relationship chain are classified by node type. The interaction event nodes corresponding to the spatial overlapping events are marked as first-class nodes, and the interaction event nodes corresponding to the operation contact events are marked as second-class nodes.

[0111] Before the feature encoder processes the feature input tensor, the type of each node in the interaction event triggering chain is determined. If the event corresponding to a node is a spatial overlap event, it is marked as a first-type node and represented by the number "0"; if it is an operation contact event, it is marked as a second-type node and represented by the number "1". For example, node E001 is a spatial overlap event and is marked as a first-type node; node E002 is an operation contact event and is marked as a second-type node.

[0112] Step S1411: Assign different initial feature mapping matrices to the first type of nodes and the second type of nodes respectively. The dimension of the initial feature mapping matrix of the first type of nodes matches the number of parameters contained in the spatial overlap event record, and the dimension of the initial feature mapping matrix of the second type of nodes matches the number of parameters contained in the operation contact event record.

[0113] The spatial overlap event record includes four parameters: time point, first spatial location coordinates, second spatial location coordinates, and instantaneous spatial distance value. Therefore, the initial feature mapping matrix for the first type of node has a dimension of 4×64 (mapping 4-dimensional features to 64-dimensional features). The operation contact event record includes four parameters: occurrence time point, first spatial location coordinates, operation location coordinates, and operation type encoding. Therefore, the initial feature mapping matrix for the second type of node also has a dimension of 4×64. These two initial feature mapping matrices are learnable parameters of the model and will be adjusted based on the data during training.

[0114] Step S1412: Map the spatial overlapping event record to the initial embedding vector of the first type of node using the initial feature mapping matrix of the first type of node, and map the operation contact event record to the initial embedding vector of the second type of node using the initial feature mapping matrix of the second type of node.

[0115] For the first type of node, the four parameters from its spatial overlap event record are combined into a 4-dimensional vector, which is then multiplied by the initial feature mapping matrix (4×64) of the first type of node to obtain a 64-dimensional initial embedding vector for the first type of node. Similarly, for the second type of node, the four parameters from its operation contact event record are combined into a 4-dimensional vector, which is then multiplied by the initial feature mapping matrix (4×64) of the second type of node to obtain a 64-dimensional initial embedding vector for the second type of node.

[0116] Step S1413: When performing neighborhood feature aggregation operation in the graph convolutional network layer, different aggregation weight parameters are used according to the node type of the neighborhood nodes. For neighborhood aggregation of the first type of nodes, the first aggregation weight matrix is ​​used, and for neighborhood aggregation of the second type of nodes, the second aggregation weight matrix is ​​used, so that the high-order spatial neighborhood feature representation can distinguish the differences in the impact of different types of interactive events on abnormal behavior patterns.

[0117] When a graph convolutional network layer aggregates neighborhood features, it determines the type of each neighboring node. If the neighboring node is a first-type node (spatial overlapping event), its features are weighted using a first aggregation weight matrix (64×64); if the neighboring node is a second-type node (operational contact event), its features are weighted using a second aggregation weight matrix (64×64). In this way, different types of neighboring nodes contribute differently to the features of the current node, enabling the high-order spatial neighborhood feature representation to reflect the differences in the impact of different interaction event types.

[0118] Step S1414: After generating the temporal context feature vector through the time series encoder, the temporal context feature vector is input into the attention mechanism module built into the abnormal behavior recognition model. The attention mechanism module calculates the attention score between the temporal context feature vector of each interactive event node and the preset global abnormal behavior pattern vector.

[0119] The pre-defined global abnormal behavior pattern vector is a 256-dimensional vector, obtained by average pooling the temporal context feature vectors of a large number of abnormal behavior samples. The attention mechanism module calculates the cosine similarity between the temporal context feature vector (256-dimensional) of each node and the global abnormal behavior pattern vector (256-dimensional), which is used as the attention score. The formula for calculating the cosine similarity is: Attention Score = (Temporal Context Feature Vector · Global Abnormal Behavior Pattern Vector) / (||Temporal Context Feature Vector|| × ||Global Abnormal Behavior Pattern Vector||), where "·" represents the vector dot product and "||||" represents the L2 norm of the vector.

[0120] Step S1415: The temporal context feature vector is weighted and summed according to the attention score to generate a global abnormal behavior representation vector that integrates information from all interaction event nodes.

[0121] Multiply the temporal context feature vector of each node by its corresponding attention score, and then sum all the weighted vectors to obtain the global anomalous behavior representation vector. For example, if there are n nodes, and the temporal context feature vector of each node is V_i, and the attention score is a_i, then the global anomalous behavior representation vector V_global = sum(a_i * V_i) for i = 1 to 1.

[0122] Step S1416: Concatenate the global abnormal behavior representation vector with the temporal context feature vector of each interactive event node to generate an enhanced temporal context feature vector, and input it into the abnormal pattern classifier for classification.

[0123] The global anomalous behavior representation vector (256-dimensional) and the temporal context feature vector of each node (256-dimensional) are concatenated along the feature dimension to obtain the enhanced temporal context feature vector with a dimension of 256 + 256 = 512. This enhanced vector is then input into the anomalous pattern classifier for classification processing to improve classification accuracy.

[0124] Step S150: Based on the set of time-series abnormal behavior segments and the interaction event triggering relationship chain, perform abnormal behavior induction and propagation path deduction processing to generate an abnormal behavior spatial diffusion feature set containing the spatial propagation direction sequence and spatial propagation speed change rate of the abnormal behavior data segments within the target monitoring area.

[0125] By combining a set of temporally abnormal behavior fragments and interactive event triggering relationship chains, the spatial propagation path of abnormal behavior is analyzed, and its propagation direction and speed changes are determined, thereby generating a set of spatial diffusion characteristics of abnormal behavior.

[0126] Step S151: Select the first abnormal behavior data segment from the set of time-series abnormal behavior segments as the initial abnormal behavior seed segment, and parse the first interactive event node identifier set associated with the initial abnormal behavior seed segment. The first interactive event node identifier set contains the unique identifiers of all interactive event nodes within the time period corresponding to the initial abnormal behavior seed segment.

[0127] The abnormal behavior data segments in the time-series abnormal behavior segment set are arranged in chronological order, and the first abnormal behavior data segment is the initial abnormal behavior seed segment. For example, the time period corresponding to this segment is from t1 to t5, and the associated interactive event node identifiers are E005, E006, E007, E008, and E009. These identifiers form the first interactive event node identifier set.

[0128] Step S152: Based on each interactive event node identifier in the first set of interactive event node identifiers, extract the spatial coordinates of the corresponding interactive event node from the interactive event triggering relationship chain to generate an initial set of abnormal behavior spatial seed points.

[0129] For each identifier in the first set of interactive event node identifiers, such as E005, find the spatial coordinates (x5, y5) of the node from the interactive event triggering relationship chain, and collect all these coordinates to form the initial abnormal behavior spatial seed point set, such as "[(x5, y5), (x6, y6), (x7, y7), (x8, y8), (x9, y9)]".

[0130] Step S153: Perform convex hull calculation on the initial abnormal behavior space seed point set to obtain the smallest convex polygon region containing all points in the initial abnormal behavior space seed point set, and mark the smallest convex polygon region as the influence domain of the initial abnormal behavior space.

[0131] The Graham scan algorithm is used to calculate the convex hull of the initial anomalous behavior space seed point set. First, the point with the smallest y-coordinate in the set is selected as the starting point. If multiple points have the same y-coordinate, the point with the smallest x-coordinate is chosen. Then, the other points are sorted according to their polar angle with the starting point, and points with the same polar angle are sorted according to their distance from the starting point. Next, points are added to the convex hull sequentially, while simultaneously checking if the line segment formed by the current point and the two preceding points turns left. If it turns right, the middle point is removed. This process continues until all points have been processed, yielding the vertices of the convex hull. The smallest convex polygon region formed by connecting these vertices is the influence domain of the initial anomalous behavior space.

[0132] Step S154: Select the next abnormal behavior data segment that is temporally adjacent to the initial abnormal behavior seed segment from the set of temporal abnormal behavior segments as the target segment for propagating abnormal behavior, and parse the second set of interactive event node identifiers associated with the target segment for propagating abnormal behavior.

[0133] In the set of time-series abnormal behavior segments, the first abnormal behavior data segment after the initial abnormal behavior seed segment is the target segment for propagating abnormal behavior. The interactive event node identifiers associated with this segment are parsed to form a second set of interactive event node identifiers, such as "[E010, E011, E012]".

[0134] Step S155: Based on each interactive event node identifier in the second set of interactive event node identifiers, extract the spatial coordinates of the corresponding interactive event node from the interactive event triggering relationship chain to generate a set of spatial target points for propagating abnormal behavior.

[0135] Similar to step S152, extract the spatial coordinates of each identifier in the second set of interactive event node identifiers to generate a set of spatial target points for propagating abnormal behavior, such as "[(x10, y10), (x11, y11), (x12, y12)]".

[0136] Step S156: Perform convex hull calculation on the target point set of the propagation abnormal behavior space to obtain the smallest convex polygon region containing all points in the target point set of the propagation abnormal behavior space, and mark the smallest convex polygon region as the influence domain of the propagation abnormal behavior space.

[0137] The Graham scanning algorithm, the same as in step S153, is used to calculate the convex hull of the target point set in the propagation anomaly behavior space to obtain the influence domain of the propagation anomaly behavior space.

[0138] Step S157: Calculate the spatial displacement vector between the geometric center coordinates of the initial abnormal behavior spatial influence domain and the geometric center coordinates of the propagation abnormal behavior spatial influence domain. The direction of the spatial displacement vector is the spatial propagation direction of the initial abnormal behavior seed fragment to the propagation abnormal behavior target fragment. The magnitude of the spatial displacement vector divided by the time interval between the initial abnormal behavior seed fragment and the propagation abnormal behavior target fragment is the spatial propagation speed.

[0139] The coordinates of the geometric center point of the initial anomalous behavior spatial influence domain are obtained by calculating the average of the coordinates of the convex hull vertices, i.e., ((x5+x6+x7+x8+x9) / 5, (y5+y6+y7+y8+y9) / 5. Similarly, the coordinates of the geometric center point of the propagating anomalous behavior spatial influence domain are calculated as ((x10+x11+x12) / 3, (y10+y11+y12) / 3). The spatial displacement vector is (center2x-center1x, center2y-center1y), and its direction is calculated using the arctangent function. The spatial propagation speed is the magnitude of the spatial displacement vector (√[(center2x-center1x)²+(center2y-center1y)²]) divided by the time interval between two anomalous behavior segments (target segment start time - seed segment end time).

[0140] Step S158: The spatial propagation direction and the spatial propagation speed are used as the first set of spatial diffusion feature parameters and stored in association in the corresponding records of the initial abnormal behavior seed fragment and the propagation abnormal behavior target fragment.

[0141] The calculated spatial propagation direction angle and spatial propagation speed values ​​are stored and associated with the initial abnormal behavior seed segment and the propagation abnormal behavior target segment, for example, "Initial segment ID: S001, Target segment ID: S002, Spatial propagation direction: 30°, Spatial propagation speed: 0.05 m / s".

[0142] Step S159: Continue to select subsequent abnormal behavior data segments from the set of time-series abnormal behavior segments as new propagation abnormal behavior target segments, and repeatedly execute the operations of parsing the set of interactive event node identifiers, generating the spatial influence domain of propagation abnormal behavior, calculating the spatial displacement vector and spatial propagation speed, until all abnormal behavior data segments in the set of time-series abnormal behavior segments have been traversed to obtain multiple sets of spatial diffusion characteristic parameters.

[0143] Following the chronological order, each anomalous behavior data segment in the set of temporal anomalous behavior segments is taken as the target segment for propagating anomalous behavior. The above calculation is performed with the previous segment to obtain multiple sets of spatial diffusion characteristic parameters, such as "S002→S003: direction 45°, speed 0.06 m / s; S003→S004: direction 60°, speed 0.04 m / s;...".

[0144] Step S1510: Arrange the spatial propagation directions in the multiple sets of spatial diffusion characteristic parameters in chronological order to generate a spatial propagation direction sequence of the abnormal behavior data segment within the target monitoring area.

[0145] Arrange the spatial propagation direction angles in multiple sets of spatial diffusion characteristic parameters according to the time sequence of the corresponding abnormal behavior segments to form a spatial propagation direction sequence, such as "[30°, 45°, 60°, ...]".

[0146] Step S1511: Based on the spatial propagation velocity values ​​arranged in chronological order among the multiple sets of spatial diffusion characteristic parameters, calculate the rate of change of spatial propagation velocity between adjacent time periods to generate a spatial propagation velocity change rate sequence.

[0147] The formula for calculating the rate of change of spatial propagation speed is (current speed - previous speed) / previous speed × 100%. For example, if the previous speed is 0.05 m / s and the current speed is 0.06 m / s, then the rate of change is (0.06 - 0.05) / 0.05 × 100% = 20%. The rate of change of each adjacent speed is calculated in chronological order to generate a sequence of spatial propagation speed change rates, such as "[20%, -33.3%, ...]".

[0148] Step S1512: The spatial propagation direction sequence and the spatial propagation speed change rate sequence together constitute the spatial diffusion feature set of the abnormal behavior.

[0149] By combining the spatial propagation direction sequence and the spatial propagation velocity change rate sequence, a set of spatial diffusion characteristics of anomalous behavior is formed, which fully describes the spatial propagation direction and velocity changes of anomalous behavior.

[0150] For example, the method may further include: step S1513: after generating the initial abnormal behavior spatial influence domain and the propagation abnormal behavior spatial influence domain, obtaining the indoor spatial layout map of the target monitoring area, wherein the indoor spatial layout map marks the spatial boundary coordinates of the bed area, corridor area, nurse station area, medical device storage area and entrance / exit area.

[0151] The indoor spatial layout map of the target monitoring area is a digital map pre-stored in the system, in which each area has a clearly defined spatial boundary coordinate. For example, the boundary coordinates of the bed area are (2.0, 2.0), (4.0, 2.0), (4.0, 4.0), (2.0, 4.0), and the boundary coordinates of the corridor area are (0.0, 2.0), (2.0, 2.0), (2.0, 4.0), (0.0, 4.0), etc.

[0152] Step S1514: Overlay the initial abnormal behavior spatial influence domain and the propagation abnormal behavior spatial influence domain onto the indoor space layout map, and identify the indoor space area types covered by the initial abnormal behavior spatial influence domain and the indoor space area types covered by the propagation abnormal behavior spatial influence domain.

[0153] By comparing coordinates, it is determined which areas in the interior spatial layout diagram have overlapping boundary coordinates between the convex polygonal region of the initial abnormal behavior's spatial influence domain and these areas. For example, if most of the initial abnormal behavior's spatial influence domain lies within the boundary coordinate range of the bed area, then the area type it covers is identified as the bed area. Similarly, the area type covered by the spatial influence domain of the spreading abnormal behavior, such as the corridor area, is identified.

[0154] Step S1515: Based on the changes in the identified indoor space area types, adjust the granularity of the spatial propagation direction sequence. When the spatial propagation direction sequence points from the bed area to the corridor area, refine the corresponding spatial propagation direction as a diffusion direction towards the corridor area. When the spatial propagation direction sequence points from the corridor area to the adjacent bed area, refine the corresponding spatial propagation direction as a diffusion direction towards the adjacent bed area.

[0155] If the initial abnormal behavior spatial influence domain covers the bed area, and the propagation abnormal behavior spatial influence domain covers the corridor area, then the corresponding directional angle of 30° in the spatial propagation direction sequence is refined as "diffusion direction towards the corridor area (30°)". If the propagation abnormal behavior spatial influence domain covers the adjacent bed area, it is marked as "diffusion direction towards the adjacent bed area (45°)".

[0156] Step S1516: Simultaneously, obtain the electronic medical record information of the monitored subjects corresponding to each bed area within the target monitoring area. The electronic medical record information includes the monitored subjects' age information, diagnosed disease information, and cognitive impairment assessment score information.

[0157] Obtain electronic medical record information from other patients in ward 301 (such as beds 2 and 3) from the hospital's electronic medical record system. For example, the patient in bed 2 is 75 years old, diagnosed with hypertension, and has a cognitive impairment assessment score of 65 (out of 100, with lower scores indicating poorer cognitive function).

[0158] Step S1517: Based on the occurrence of the diffusion direction towards the adjacent bed area in the spatial propagation direction sequence, extract the cognitive impairment assessment score information of the monitored object corresponding to the adjacent bed area, and compare the cognitive impairment assessment score information with the cognitive impairment assessment score information of the current monitored object.

[0159] When the spatial propagation direction sequence shows "diffusion direction towards adjacent bed area", the adjacent bed area is identified as bed number 2, and its cognitive impairment assessment score of 65 points is extracted. The cognitive impairment assessment score of the currently monitored subject (patient A) is 80 points, and the 65 points are compared with the 80 points.

[0160] Step S1518: If the cognitive impairment assessment score of the monitored object corresponding to the adjacent bed area is lower than the preset cognitive function threshold, then add a risk propagation marker to the spatial propagation direction sequence in the abnormal behavior spatial diffusion feature set.

[0161] The preset cognitive function threshold is 70 points. Patients in adjacent bed areas scored 65 points (below 70 points) on the cognitive impairment assessment; therefore, a "risk propagation marker: yes" was added to the record corresponding to that direction in the spatial propagation direction sequence.

[0162] Step S1519: Associate and store the risk propagation marker with the set of spatial diffusion features of abnormal behavior, and when generating personalized monitoring intervention strategy instructions in the future, give priority to preventive intervention for the monitored subjects in adjacent bed areas corresponding to the spatial propagation direction sequence with risk propagation markers, and add the preventive intervention operation instructions to the personalized monitoring intervention strategy instructions.

[0163] The risk propagation marker is associated with the corresponding record in the set of abnormal behavior spatial diffusion features and stored accordingly, for example, "diffusion direction to adjacent bed area (45°), risk propagation marker: yes". When generating personalized monitoring intervention strategy instructions, preventive intervention operation instructions for the patient in bed 2 are added for this direction with the risk propagation marker, such as increasing the number of rounds and conducting cognitive function assessments.

[0164] Step S160: Generate a personalized monitoring intervention strategy instruction for the monitored object based on the abnormal behavior type label, the abnormal behavior spatial diffusion feature set, and the temporal abnormal behavior fragment set, and send the personalized monitoring intervention strategy instruction to the mobile terminal device worn by the medical staff to trigger the corresponding monitoring operation prompt.

[0165] Based on information such as the type of abnormal behavior, spatial diffusion characteristics, and temporal fragments of abnormal behavior, a personalized monitoring and intervention strategy instruction was formulated for Patient A. This instruction specified in detail the operations, times, and locations that medical staff need to perform, and then sent to the mobile terminal devices of medical staff through the hospital's internal network to remind them to perform the corresponding monitoring operations.

[0166] Step S161: Analyze the start and end times of each abnormal behavior data segment contained in the time-series abnormal behavior segment set, and calculate the duration parameter of each abnormal behavior data segment based on the start and end times.

[0167] For each anomalous behavior data segment in the set of time-series anomalous behavior segments, extract its start and end times. The duration parameter is equal to the end time minus the start time, in minutes. For example, if the start time of a certain anomalous behavior data segment is 2024-05-10 10:00:00 and the end time is 2024-05-10 10:15:00, then the duration parameter is 15 minutes.

[0168] Step S162: Obtain the spatial propagation direction sequence and spatial propagation speed change rate sequence contained in the abnormal behavior spatial diffusion feature set, compare the speed change rate value in the spatial propagation speed change rate sequence with the preset speed change rate threshold, and identify abnormal propagation acceleration points where the speed change rate value exceeds the speed change rate threshold and abnormal propagation deceleration points where the speed change rate value is lower than the speed change rate threshold.

[0169] The preset threshold for the rate of change of velocity is ±20%. Each value in the spatial propagation velocity rate of change sequence is compared to ±20%. Points with a rate of change greater than 20% are considered abnormal propagation acceleration points, and points with a rate of change less than -20% are considered abnormal propagation deceleration points. For example, 25% in the rate of change sequence is an abnormal propagation acceleration point, and -30% is an abnormal propagation deceleration point.

[0170] Step S163: Based on the spatial propagation directions corresponding to the abnormal propagation acceleration point and the abnormal propagation deceleration point in the spatial propagation direction sequence, generate a dynamic change map of the abnormal behavior propagation situation. The dynamic change map of the abnormal behavior propagation situation marks the spatial location areas where the propagation speed of the abnormal behavior changes significantly within the target monitoring area.

[0171] On the indoor spatial layout map of the target monitoring area, determine and mark the location areas of the acceleration and deceleration points of abnormal propagation based on their spatial propagation directions and time points. For example, mark one location area as "10:05, acceleration point, direction 30°" and another location area as "10:10, deceleration point, direction 45°", thus forming a dynamic change map of the abnormal behavior propagation trend.

[0172] Step S164: Call the preset intervention strategy rule base. The intervention strategy rule base stores the mapping relationship between various abnormal behavior type labels and various basic intervention strategy templates. Each basic intervention strategy template includes a list of medical and nursing operations to be performed, a recommended execution time window for performing the medical and nursing operation, and a list of medical devices required to perform the medical and nursing operation.

[0173] In the intervention strategy rule base, the abnormal behavior type label "unauthorized departure from bed" corresponds to the basic intervention strategy template T001. This template includes a list of medical and nursing operations (such as "go to check on the patient's condition" and "guide the patient back to the bed"), a recommended execution time window (within 5 minutes after the abnormal behavior occurs), and a list of medical devices (such as "flashlight" and "communication devices").

[0174] Step S165: Match the corresponding basic intervention strategy template from the intervention strategy rule base according to the abnormal behavior type label, and use it as the initial intervention strategy template.

[0175] Patient A's abnormal behavior type is labeled "unauthorized departure from bed". The basic intervention strategy template T001 is matched from the intervention strategy rule base and used as the initial intervention strategy template.

[0176] Step S166: Compare and fuse the recommended execution time window in the initial intervention strategy template with the duration parameter of the abnormal behavior data segment and the time point of the abnormal propagation acceleration point marked in the dynamic change map of the abnormal behavior propagation situation, adjust the start and end times of the recommended execution time window, and generate a personalized execution time window so that the personalized execution time window covers the time period before the abnormal propagation acceleration point.

[0177] The recommended execution time window for the initial intervention strategy template T001 is within 5 minutes of the occurrence of the abnormal behavior. The duration of the abnormal behavior data segment is 15 minutes, and the time point of the abnormal propagation acceleration point is 3 minutes after the occurrence of the abnormal behavior. The recommended execution time window is adjusted to within 2 minutes of the occurrence of the abnormal behavior to 1 minute before the time point of the abnormal propagation acceleration point; that is, the personalized adjusted execution time window is 2 minutes to 2 minutes after the occurrence of the abnormal behavior, ensuring that the intervention is completed before the acceleration point.

[0178] Step S167: Based on the spatial propagation direction sequence in the abnormal behavior spatial diffusion feature set, predict the possible propagation direction of the abnormal behavior in the future, and combine the spatial location area marked in the dynamic change map of the abnormal behavior propagation situation, extract the adjacent monitoring bed identifier or adjacent functional area identifier in the possible propagation direction from the spatial layout database of the target monitoring area, add the adjacent monitoring bed identifier or adjacent functional area identifier to the medical device equipment list in the initial intervention strategy template, and generate a personalized medical device equipment list that includes the extended monitoring range.

[0179] The spatial propagation direction sequence predicts that the future propagation direction may be towards the adjacent bed 2 area. The identifier "B002" for bed 2 is extracted from the spatial layout database and added to the medical device equipment list. Considering the need for contact with the patient in bed 2, "disposable gloves" are also added to the list, generating a personalized medical device equipment list.

[0180] Step S168: The personalized execution time window, the personalized medical device list, and the list of medical and nursing operation actions in the initial intervention strategy template are encapsulated to generate the personalized monitoring intervention strategy instruction. The personalized monitoring intervention strategy instruction contains multiple sub-instruction units arranged in chronological order. Each sub-instruction unit corresponds to a medical and nursing operation action, the specific time point for performing the medical and nursing operation action, the medical device equipment required to perform the medical and nursing operation action, and the target spatial location coordinates for performing the medical and nursing operation action.

[0181] The personalized execution time window (e.g., from 10:02:00 on 2024-05-10 to 10:04:00 on 2024-05-10), the personalized list of medical devices (flashlight, communication equipment, disposable gloves), and the list of medical and nursing operation actions are encapsulated into multiple sub-instruction units. For example, sub-instruction unit 1: "Operation: Go to check the patient's condition, specific time: 2024-05-10 10:02:00, required equipment: flashlight, communication equipment, target location coordinates: (3.5, 2.8)"; sub-instruction unit 2: "Operation: Guide the patient back to the bed, specific time: 2024-05-10 10:03:00, required equipment: communication equipment, target location coordinates: (3.0, 2.5)"; sub-instruction unit 3: "Operation: Check the patient's condition in bed 2, specific time: 2024-05-10 10:04:00, required equipment: disposable gloves, target location coordinates: (5.0, 2.5)".

[0182] Step S169: The personalized monitoring intervention strategy instruction is sent to the mobile terminal device worn by the medical staff through the hospital's internal wireless communication network. After receiving the personalized monitoring intervention strategy instruction, the mobile terminal device parses the multiple sub-instruction units and displays the corresponding medical staff operation prompts, required medical device prompts, and target spatial location navigation information for each sub-instruction unit in chronological order on the display screen, so as to trigger the medical staff to perform the corresponding monitoring operation according to the personalized monitoring intervention strategy instruction.

[0183] The hospital's internal wireless communication network uses Wi-Fi 6 technology to send personalized monitoring and intervention strategy instructions to a mobile terminal device (such as a smartwatch or tablet) worn by nurse A. Upon receiving the instruction, the mobile terminal device parses multiple sub-instruction units and displays them on the screen in chronological order. For example, it first displays "10:02: Go to (3.5, 2.8) to check on the patient's condition, carrying a flashlight and communication equipment," and provides a navigation path from the current location to the target location; then at 10:03, it displays the prompt information for the next sub-instruction unit, and so on, guiding nurse A to perform the corresponding monitoring operations.

[0184] In one exemplary embodiment, a patient monitoring abnormal behavior recognition system incorporating deep learning algorithms is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2As shown, this patient monitoring abnormal behavior recognition system incorporating deep learning algorithms includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a patient monitoring abnormal behavior recognition method incorporating deep learning algorithms. The display unit is used to generate a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the shell of the patient monitoring abnormal behavior recognition system that combines deep learning algorithms, or an external keyboard, touchpad, or mouse, etc.

[0185] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for identifying abnormal patient monitoring behaviors using deep learning algorithms, characterized in that, The method includes: Acquire raw patient activity data streams collected by multiple sensor units deployed within the target monitoring area; Based on the time-series activity data units in the original patient activity data stream, a spatiotemporal sequence of the activity trajectory of the monitored object in the target monitoring area is constructed. At the same time, based on the medical and nursing operation data units in the original patient activity data stream, a spatiotemporal sequence of the operation trajectory of medical staff in the target monitoring area is constructed. The spatiotemporal sequence of the activity trajectory and the spatiotemporal sequence of the operation trajectory are subjected to monitoring spatiotemporal correlation mapping to generate an interaction event triggering relationship chain between the monitored object and medical staff. The pre-built abnormal behavior recognition model is invoked to perform abnormal behavior pattern deviation quantitative analysis on the interaction event triggering relationship chain, so as to obtain the abnormal behavior type label of the monitored object in the continuous monitoring period and the set of time-series abnormal behavior fragments corresponding to the abnormal behavior type label. Based on the set of time-series abnormal behavior segments and the interaction event triggering relationship chain, the abnormal behavior inducement propagation path inference processing is performed to generate an abnormal behavior spatial diffusion feature set containing the spatial propagation direction sequence and spatial propagation speed change rate of the abnormal behavior data segments in the target monitoring area. Based on the abnormal behavior type label, the abnormal behavior spatial diffusion feature set, and the time-series abnormal behavior fragment set, a personalized monitoring intervention strategy instruction is generated for the monitored object, and the personalized monitoring intervention strategy instruction is sent to the mobile terminal device worn by the medical staff to trigger the corresponding monitoring operation prompt. The step of performing spatiotemporal correlation mapping processing on the activity trajectory spatiotemporal sequence and the operation trajectory spatiotemporal sequence to generate an interaction event triggering relationship chain between the monitored object and medical staff includes: The first spatial coordinates of the monitored object at a specific time point contained in the activity trajectory spatiotemporal sequence and the second spatial coordinates of the medical staff at the same time point contained in the operation trajectory spatiotemporal sequence are obtained. The same time point is determined by performing a global time axis alignment operation between the data acquisition timestamp and the operation execution timestamp. Calculate the Euclidean spatial distance between the first spatial coordinates and the second spatial coordinates to obtain the instantaneous spatial distance between the monitored object and the medical staff at the same time point; The instantaneous spatial distance value is compared with a preset interaction trigger distance threshold. When the instantaneous spatial distance value is less than or equal to the interaction trigger distance threshold, it is determined that the monitored object and the medical staff have an overlapping spatial position event at the same time point, and a spatial overlap event record containing the same time point, the first spatial position coordinates, the second spatial position coordinates, and the instantaneous spatial distance value is generated. The operation type code corresponding to each operation position coordinate in the spatiotemporal sequence of the operation trajectory is analyzed synchronously, and the set of medical and nursing operation types that directly affect the monitored object is extracted from the operation type code. The set of medical and nursing operation types that directly affect the monitored object includes vital sign measurement operation type, intravenous injection operation type, wound care operation type, and body position adjustment operation type. The occurrence time of each medical and nursing operation that directly affects the monitored object is determined based on the operation execution timestamp, and the first spatial position coordinates of the monitored object corresponding to the occurrence time point are obtained. The operation contact space distance between the first spatial position coordinates of the monitored object corresponding to the occurrence time point and the operation position coordinates of the medical and nursing operation is calculated. When the spatial distance of the operation contact is less than or equal to a preset operation contact distance threshold, it is determined that the monitored object and the medical staff have an operation contact event at the time of occurrence, and an operation contact event record containing the time of occurrence, the first spatial location coordinates, the operation location coordinates, and the operation type code is generated; The spatial overlap event records and the operation contact event records are merged and sorted according to the time points they contain to obtain an initial interaction event sequence. The initial interaction event sequence contains multiple interaction event units arranged in chronological order, and each interaction event unit corresponds to a spatial overlap event or an operation contact event. The time interval parameter is calculated based on the time points of two adjacent interactive event units in the initial interactive event sequence, and the spatial distance change parameter is calculated based on the spatial position coordinates of the two adjacent interactive event units. Each interactive event unit in the initial interactive event sequence is taken as an interactive event node, and the directed connection line between two adjacent interactive event nodes is taken as a directed association edge. The time interval parameter and the spatial distance change parameter are attached to the corresponding directed association edge to generate an interactive event triggering relationship chain between the monitored object and the medical staff.

2. The method for identifying abnormal patient monitoring behavior using a deep learning algorithm according to claim 1, characterized in that, The step of constructing a spatiotemporal sequence of the monitored object's activity trajectory within the target monitoring area based on the temporal activity data units in the original patient activity data stream, and simultaneously constructing a spatiotemporal sequence of the medical staff's operation trajectory within the target monitoring area based on the medical operation data units in the original patient activity data stream, includes: The time-series activity data units in the original patient activity data stream are parsed, and the monitored object identifier, data acquisition timestamp, and spatial location coordinates of the monitored object at the data acquisition timestamp are extracted from each time-series activity data unit. The spatial location coordinates are calculated based on the output signal of the positioning sensor unit deployed in the target monitoring area. Based on the data acquisition timestamp, all extracted spatial location coordinates containing the identifier of the monitored object are sorted in ascending order to obtain the first spatial location coordinate sequence of the monitored object changing over time during the continuous monitoring period. The first spatial location coordinate sequence is subjected to trajectory point interpolation encryption processing. Multiple interpolated location coordinates are inserted between the spatial location coordinates corresponding to two adjacent data acquisition timestamps according to a preset time interpolation density to obtain a second spatial location coordinate sequence containing the original spatial location coordinates and the interpolated location coordinates. The movement speed vector set of the monitored object is constructed based on the time attribute corresponding to each spatial position coordinate in the second spatial position coordinate sequence. Each movement speed vector in the movement speed vector set is calculated by dividing the displacement vector between two adjacent spatial position coordinates by the corresponding time interval. The second spatial position coordinate sequence and the set of movement speed vectors are associated and stored to generate the spatiotemporal sequence of the activity trajectory of the monitored object within the target monitoring area. The spatiotemporal sequence of the activity trajectory includes each spatial position coordinate in the second spatial position coordinate sequence and the instantaneous movement speed vector and instantaneous movement direction angle associated with that spatial position coordinate. The medical and nursing operation data units in the original patient activity data stream are parsed synchronously. The medical and nursing personnel identifier, operation execution timestamp, and operation position coordinates of the medical and nursing personnel at the operation execution timestamp are extracted from each medical and nursing operation data unit. The operation position coordinates are calculated based on the output signal of the positioning sensing unit or the built-in positioning module of the medical and nursing personnel's handheld terminal device. Based on the operation execution timestamp, all extracted operation location coordinates containing medical staff identifiers are sorted in ascending order to obtain the third spatial location coordinate sequence of the medical staff changing over time during the continuous monitoring period. The third spatial position coordinate sequence is subjected to trajectory point smoothing filtering to eliminate abnormal coordinate jump points caused by instantaneous fluctuations in the positioning signal, thereby generating a smoothed fourth spatial position coordinate sequence. Based on the time attribute corresponding to each operation position coordinate in the fourth spatial position coordinate sequence and the operation type code contained in the medical operation data unit, a spatial position mapping relationship of each medical operation performed by the medical staff during the movement is constructed, and an operation trajectory spatiotemporal sequence containing operation position coordinates, the time point of arrival at the operation position coordinate, the operation type code performed at the operation position coordinate, and the time point of departure from the operation position coordinate is generated.

3. The method for identifying abnormal patient monitoring behavior using a deep learning algorithm according to claim 1, characterized in that, The pre-built abnormal behavior recognition model is invoked to perform abnormal behavior pattern deviation quantification analysis on the interaction event triggering relationship chain, obtaining the abnormal behavior type label of the monitored object within the continuous monitoring period and the set of time-series abnormal behavior fragments corresponding to the abnormal behavior type label, including: The interaction event triggering relationship chain is input into the input layer of the abnormal behavior recognition model. The sequence of interaction event nodes contained in the interaction event triggering relationship chain, as well as the time interval parameter and spatial distance change parameter carried by the directed associated edge, together constitute the feature input tensor of the abnormal behavior recognition model. The feature input tensor of the abnormal behavior recognition model is spatially structured by a feature encoder. The feature encoder contains multiple stacked graph convolutional network layers. Each graph convolutional network layer performs neighborhood feature aggregation on the interactive event node to generate a high-order spatial neighborhood feature representation of each interactive event node in the current layer. The high-order spatial neighborhood feature representation integrates the topological connection relationship between the interactive event node and its neighboring interactive event nodes in the interactive event triggering relationship chain, as well as the parameter information carried by the directed associated edges. The high-order spatial neighborhood feature representations of different orders output by multiple graph convolutional network layers are concatenated to obtain the multi-scale spatial structure embedding vector of each interactive event node. The time-series encoder of the abnormal behavior recognition model performs temporal dependency modeling on the multi-scale spatial structure embedding vector according to the temporal order of the interactive event nodes. The time-series encoder includes a bidirectional long short-term memory network structure. The bidirectional long short-term memory network structure generates a temporal context feature vector of the current interactive event node containing forward and backward temporal context information based on the multi-scale spatial structure embedding vector of the current interactive event node and the hidden state vectors of the previous and next time moments. The temporal context feature vector is input into the abnormal behavior recognition model's abnormal pattern classifier. The abnormal pattern classifier includes a multilayer perceptron network structure. The multilayer perceptron network structure performs nonlinear transformation processing on the temporal context feature vector and outputs the probability distribution vector of the monitored object belonging to a preset variety of abnormal behavior types at each time point corresponding to each interaction event node. Based on the abnormal behavior type corresponding to the maximum probability value in the probability distribution vector, the candidate abnormal behavior type label for that time point is determined, and the time points where the candidate abnormal behavior type label appears consecutively on the time axis are merged to form multiple candidate abnormal behavior time periods. For each candidate abnormal behavior time period, extract the first interactive event node identifier corresponding to the start time point and the second interactive event node identifier corresponding to the end time point of the candidate abnormal behavior time period. Based on the first interactive event node identifier and the second interactive event node identifier, the original data segments within the corresponding time period are extracted from the time-series activity data unit in the original patient activity data stream, generating an abnormal behavior data segment in the time-series abnormal behavior segment set, and the candidate abnormal behavior type label corresponding to the candidate abnormal behavior time period is used as the abnormal behavior type label of the abnormal behavior data segment. Perform the same truncation and labeling operations on all candidate abnormal behavior time periods to obtain a set of time-series abnormal behavior segments containing multiple abnormal behavior data segments and the abnormal behavior type label corresponding to each abnormal behavior data segment. At the same time, establish the association between each abnormal behavior data segment and the interactive event node identifier corresponding to the time point it contains.

4. The method for identifying abnormal patient monitoring behavior using a deep learning algorithm according to claim 2, characterized in that, The abnormal behavior propagation path inference processing based on the set of temporal abnormal behavior segments and the interaction event triggering relationship chain generates an abnormal behavior spatial diffusion feature set containing the spatial propagation direction sequence and spatial propagation speed change rate of the abnormal behavior data segments within the target monitoring area, including: Select the first abnormal behavior data segment from the set of time-series abnormal behavior segments as the initial abnormal behavior seed segment, and parse the first interactive event node identifier set associated with the initial abnormal behavior seed segment. The first interactive event node identifier set contains the unique identifiers of all interactive event nodes within the time period corresponding to the initial abnormal behavior seed segment. Based on each interactive event node identifier in the first set of interactive event node identifiers, the spatial coordinates of the corresponding interactive event node are extracted from the interactive event triggering relationship chain to generate an initial set of abnormal behavior spatial seed points. Perform convex hull calculation on the initial abnormal behavior space seed point set to obtain the smallest convex polygon region containing all points in the initial abnormal behavior space seed point set, and mark the smallest convex polygon region as the initial abnormal behavior space influence domain. Select the next abnormal behavior data segment that is temporally adjacent to the initial abnormal behavior seed segment from the set of temporal abnormal behavior segments as the target segment for propagating abnormal behavior, and parse the second set of interactive event node identifiers associated with the target segment for propagating abnormal behavior; Based on each interactive event node identifier in the second set of interactive event node identifiers, the spatial coordinates of the corresponding interactive event node are extracted from the interactive event triggering relationship chain to generate a set of spatial target points for propagating abnormal behavior. Perform convex hull calculation on the target point set of the propagation abnormal behavior space to obtain the smallest convex polygon region containing all points in the target point set of the propagation abnormal behavior space, and mark the smallest convex polygon region as the influence domain of the propagation abnormal behavior space. Calculate the spatial displacement vector between the geometric center point coordinates of the initial abnormal behavior spatial influence domain and the geometric center point coordinates of the propagation abnormal behavior spatial influence domain. The direction of the spatial displacement vector is the spatial propagation direction of the initial abnormal behavior seed fragment to the propagation abnormal behavior target fragment. The magnitude of the spatial displacement vector divided by the time interval between the initial abnormal behavior seed fragment and the propagation abnormal behavior target fragment is the spatial propagation speed. The spatial propagation direction and the spatial propagation speed are used as the first set of spatial diffusion feature parameters and are associated and stored in the corresponding records of the initial abnormal behavior seed fragment and the propagation abnormal behavior target fragment; Continue to select subsequent abnormal behavior data segments from the set of time-series abnormal behavior segments as new propagation abnormal behavior target segments, and repeatedly execute the operations of parsing the set of interactive event node identifiers, generating the spatial influence domain of propagation abnormal behavior, calculating the spatial displacement vector and spatial propagation speed, until all abnormal behavior data segments in the set of time-series abnormal behavior segments have been traversed, and multiple sets of spatial diffusion characteristic parameters are obtained. Arrange the spatial propagation directions in the multiple sets of spatial diffusion characteristic parameters in chronological order to generate a spatial propagation direction sequence of the abnormal behavior data segment within the target monitoring area; Based on the spatial propagation velocity values ​​arranged in chronological order from the multiple sets of spatial diffusion characteristic parameters, the rate of change of spatial propagation velocity between adjacent time periods is calculated to generate a spatial propagation velocity change rate sequence. The spatial propagation direction sequence and the spatial propagation speed change rate sequence together constitute the spatial diffusion feature set of the abnormal behavior.

5. The method for identifying abnormal patient monitoring behavior using a deep learning algorithm according to claim 1, characterized in that, The step of generating personalized monitoring and intervention strategy instructions for the monitored object based on the abnormal behavior type label, the abnormal behavior spatial diffusion feature set, and the temporal abnormal behavior fragment set, and sending the personalized monitoring and intervention strategy instructions to the mobile terminal device worn by the medical staff to trigger corresponding monitoring operation prompts, includes: The start and end times of each abnormal behavior data segment contained in the set of time-series abnormal behavior segments are analyzed, and the duration parameter of each abnormal behavior data segment is calculated based on the start and end times. The spatial propagation direction sequence and spatial propagation speed change rate sequence contained in the abnormal behavior spatial diffusion feature set are obtained. The speed change rate value in the spatial propagation speed change rate sequence is compared with a preset speed change rate threshold to identify abnormal propagation acceleration points where the speed change rate value exceeds the speed change rate threshold and abnormal propagation deceleration points where the speed change rate value is lower than the speed change rate threshold. Based on the spatial propagation directions corresponding to the acceleration and deceleration points of the abnormal propagation in the spatial propagation direction sequence, a dynamic change map of the abnormal behavior propagation situation is generated. The dynamic change map of the abnormal behavior propagation situation marks the spatial location areas where the propagation speed of the abnormal behavior changes significantly within the target monitoring area. The system calls a preset intervention strategy rule base, which stores the mapping relationship between various abnormal behavior type labels and various basic intervention strategy templates. Each basic intervention strategy template includes a list of medical and nursing operations to be performed, a recommended execution time window for performing the medical and nursing operation, and a list of medical devices required to perform the medical and nursing operation. Based on the abnormal behavior type label, the corresponding basic intervention strategy template is matched from the intervention strategy rule base and used as the initial intervention strategy template; The recommended execution time window in the initial intervention strategy template is compared and fused with the duration parameter of the abnormal behavior data segment and the time point of the abnormal propagation acceleration point marked in the dynamic change map of the abnormal behavior propagation situation. The start and end times of the recommended execution time window are adjusted to generate a personalized execution time window, so that the personalized execution time window covers the time period before the abnormal propagation acceleration point. Based on the spatial propagation direction sequence in the abnormal behavior spatial diffusion feature set, predict the possible propagation direction of the abnormal behavior in the future, and combine the spatial location area marked in the dynamic change map of the abnormal behavior propagation situation, extract the adjacent monitoring bed identifier or adjacent functional area identifier in the possible propagation direction from the spatial layout database of the target monitoring area, add the adjacent monitoring bed identifier or adjacent functional area identifier to the medical device equipment list in the initial intervention strategy template, and generate a personalized medical device equipment list that includes the extended monitoring range. The personalized execution time window, the personalized medical device list, and the list of medical and nursing operations in the initial intervention strategy template are encapsulated to generate the personalized monitoring intervention strategy instruction. The personalized monitoring intervention strategy instruction contains multiple sub-instruction units arranged in chronological order. Each sub-instruction unit corresponds to a medical and nursing operation, the specific time point for performing the medical and nursing operation, the medical device required to perform the medical and nursing operation, and the target spatial coordinates for performing the medical and nursing operation. The personalized monitoring and intervention strategy instructions are sent to the mobile terminal devices worn by medical staff through the hospital's internal wireless communication network. After receiving the personalized monitoring and intervention strategy instructions, the mobile terminal devices parse the multiple sub-instruction units and display the corresponding medical staff operation prompts, required medical equipment prompts, and target spatial location navigation information for each sub-instruction unit in chronological order on the display screen, so as to trigger the medical staff to perform the corresponding monitoring operations according to the personalized monitoring and intervention strategy instructions.

6. The method for identifying abnormal patient monitoring behavior using a deep learning algorithm according to claim 1, characterized in that, The step of performing spatiotemporal correlation mapping processing on the spatiotemporal sequence of the activity trajectory and the spatiotemporal sequence of the operation trajectory to generate an interaction event triggering relationship chain between the monitored object and medical staff also includes: After generating the spatial overlap event record and the operation contact event record, the operation type code contained in the operation contact event record is obtained, and the corresponding standard operation duration is searched from the preset operation duration database according to the operation type code. Based on the occurrence time in the operation contact event record and the duration of the standard operation, calculate the theoretical end time of the operation contact event; Within a preset time window before and after the theoretical end time, it is detected whether the spatial coordinates of the medical staff in the spatiotemporal sequence of the operation trajectory are continuously within a preset distance range near the operation position coordinates in the operation contact event record, and whether the spatial coordinates of the monitored object in the spatiotemporal sequence of the activity trajectory are continuously within a preset distance range near the operation position coordinates. If the spatial coordinates of the medical staff and the spatial coordinates of the monitored object remain within a preset distance range near the operation position coordinates within a preset time window before and after the theoretical end time point, the operation contact event is determined to be fully executed, and a full execution mark is added to the operation contact event record. If the spatial coordinates of the medical staff or the spatial coordinates of the monitored object leave a preset distance range near the operation position coordinates before the theoretical end time is reached, the operation contact event is determined to be interrupted, and an interruption mark and the interruption time point are added to the operation contact event record. Based on the interruption execution flag and interruption occurrence time of the operation contact event record, update the corresponding interaction event node attribute in the interaction event triggering relationship chain, and attach an operation interruption information field to the interaction event node. The operation interruption information field includes an interruption identifier and an interruption time. In the subsequent abnormal behavior pattern deviation quantification analysis, the interactive event nodes containing the operation interruption information field will be the focus of analysis, and their initial weight coefficients in the feature input tensor of the abnormal behavior recognition model will be increased.

7. The method for identifying abnormal patient monitoring behavior using a deep learning algorithm according to claim 3, characterized in that, The step of calling a pre-built abnormal behavior recognition model to perform abnormal behavior pattern deviation quantification analysis on the interaction event triggering relationship chain also includes: Before the feature encoder of the abnormal behavior recognition model performs spatial structure encoding on the feature input tensor, the interaction event nodes contained in the interaction event triggering relationship chain are classified by node type. The interaction event nodes corresponding to the spatial overlapping events are marked as first-type nodes, and the interaction event nodes corresponding to the operation contact events are marked as second-type nodes. Different initial feature mapping matrices are assigned to the first type of nodes and the second type of nodes respectively. The dimension of the initial feature mapping matrix of the first type of nodes matches the number of parameters contained in the spatial overlap event record, and the dimension of the initial feature mapping matrix of the second type of nodes matches the number of parameters contained in the operation contact event record. The spatial overlapping event record is mapped to the initial embedding vector of the first type of node through the initial feature mapping matrix of the first type of node, and the operation contact event record is mapped to the initial embedding vector of the second type of node through the initial feature mapping matrix of the second type of node. When performing neighborhood feature aggregation in the graph convolutional network layer, different aggregation weight parameters are used according to the node type of the neighborhood nodes. For neighborhood aggregation of the first type of nodes, the first aggregation weight matrix is ​​used, and for neighborhood aggregation of the second type of nodes, the second aggregation weight matrix is ​​used, so that the high-order spatial neighborhood feature representation can distinguish the differences in the impact of different types of interactive events on abnormal behavior patterns. After generating the temporal context feature vector through the time series encoder, the temporal context feature vector is input into the attention mechanism module built into the abnormal behavior recognition model. The attention mechanism module calculates the attention score between the temporal context feature vector of each interactive event node and the preset global abnormal behavior pattern vector. The attention score is used to perform a weighted summation of the temporal context feature vector to generate a global abnormal behavior representation vector that integrates information from all interaction event nodes; The global abnormal behavior representation vector is concatenated with the temporal context feature vector of each interactive event node to generate an enhanced temporal context feature vector, which is then input into the abnormal pattern classifier for classification.

8. A patient monitoring abnormal behavior recognition system combining deep learning algorithms, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the patient monitoring abnormal behavior identification method incorporating a deep learning algorithm according to any one of claims 1 to 7 by executing the machine-executable instructions.

9. A computer program product, characterized in that, The computer program product includes machine-executable instructions stored in a computer-readable storage medium. The processor of the patient monitoring abnormal behavior recognition system incorporating a deep learning algorithm reads the machine-executable instructions from the computer-readable storage medium and executes the machine-executable instructions, causing the patient monitoring abnormal behavior recognition system incorporating a deep learning algorithm to perform the patient monitoring abnormal behavior recognition method incorporating a deep learning algorithm as described in any one of claims 1 to 7.

Images