Intelligent patient safety management method

Abstract

The invention discloses an intelligent patient safety management method, and relates to the technical field of patient safety management, and the method comprises the steps: collecting motion posture data and physiological parameters outputted by a wearable device worn by a patient, and ultra-wideband positioning data and millimeter wave radar micro-motion signals in an environment, carrying out the time alignment and noise filtering of the data, and carrying out the detection of the motion posture data and physiological parameters. A standardized multi-source sensing data stream is obtained; based on the standardized multi-source perception data stream, extracting a behavior action sequence, a spatial position track and a physiological state change of the patient, and combining the behavior action sequence with a spatial region label to generate a behavior event sequence containing a subject-action-space-time semantic structure; a medical behavior semantic knowledge graph is constructed, and the knowledge graph comprises a patient, a bed body, doors and windows and a nursing station as entity nodes, lying, standing, walking, climbing, approaching and staying as action nodes, and night, heart rate rise and sympathetic activation as state nodes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of patient safety management, and particularly relates to a patient intelligent safety management method. BACKGROUND

[0002] Patient safety management technology refers to a series of technical means and methods aiming to guarantee the safety of patients in hospitals or other medical environments. Therefore, how to utilize advanced technical means to improve the intelligent level and safety of patient safety management has become one of the problems to be solved at present.

[0003] In the field of patient safety management, the traditional method performs action recognition based on a single threshold value or simple rules, cannot handle complex action patterns and individual differences, and most existing technologies focus on the risk assessment of instant behavior, ignoring potential psychophysiological precursors. At the same time, many traditional patient safety management systems adopt fixed alarm thresholds, and fail to dynamically adjust according to the daily rhythm changes and personal health conditions of patients. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a patient intelligent safety management method to solve the problem that the traditional method performs action recognition based on a single threshold value or simple rules, cannot handle complex action patterns and individual differences, and most existing technologies focus on the risk assessment of instant behavior, ignoring potential psychophysiological precursors, and many traditional patient safety management systems adopt fixed alarm thresholds, failing to dynamically adjust according to the daily rhythm changes and personal health conditions of patients.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a patient intelligent safety management method, which comprises:

[0008] Collecting motion posture data and physiological parameters output by a wearable device worn by a patient, and ultra-wideband positioning data and millimeter wave radar micro-motion signals in an environment, time-aligning and noise-filtering the above data to obtain standardized multi-source perception data streams;

[0009] Based on the standardized multi-source perception data streams, extracting a behavior action sequence, a spatial position trajectory and a physiological state change of the patient, combining the behavior action sequence with a spatial region label to generate a behavior event sequence containing subject-action-space-time semantic structure;

[0010] A medical behavior semantic knowledge graph is constructed, the knowledge graph contains patients, beds, doors and windows, nursing stations as entity nodes, lying, standing, walking, climbing, approaching and staying as action nodes, night, heart rate increase and sympathetic activation as state nodes, and logical associations between entities and actions, actions and spaces, and actions and states are established through relationship edges;

[0011] The behavior event sequence is mapped to the behavior semantic knowledge graph to form a dynamic behavior subgraph, a graph neural network is used for feature aggregation of the dynamic behavior subgraph, real-time heart rate variability and skin conductance level are combined, and a behavior risk score at the current time is calculated;

[0012] Based on the time sequence association of historical physiological data and risk events, a long short-term memory network model is constructed, the skin conductance response rate, heart rate power ratio and behavior frequency in the past preset time length are input, and the risk occurrence probability is output as a psychophysiological precursor prediction result;

[0013] The behavior risk score and the psychophysiological precursor prediction result are weighted and fused to generate a comprehensive risk index, voice reminders and lighting enhancement are triggered when the comprehensive risk index reaches a first threshold, and emergency alarms are triggered and pushed to a nursing terminal when the comprehensive risk index reaches a second threshold;

[0014] The response time and event result of each alarm are recorded, and the recorded data are used to update the knowledge graph path weight, the prediction model parameter and the individual behavior baseline.

[0015] As a preferred scheme of the patient intelligent safety management method, the motion posture data and physiological parameters output by the wearable device worn by the patient, and the ultra-wideband positioning data and millimeter wave radar micro-motion signals in the environment are collected, time alignment and noise filtering are performed on the above data, and standardized multi-source perception data streams are obtained, and the specific steps are as follows:

[0016] The original motion posture data sequence output by the wearable device in real time is denoised by using a sliding window average filtering algorithm to obtain a smoothed posture angle sequence;

[0017] The original heart rate signal collected by the wearable device is subjected to high-frequency noise suppression by using a Butterworth low-pass filter to obtain filtered heart rate time series data;

[0018] The position coordinate data stream output by the ultra-wideband positioning system and the micro-motion speed signal stream output by the millimeter wave radar are time-stamped and aligned to eliminate the sampling time delay deviation between the multi-source devices, and a time-synchronized position-micro-motion joint data stream is obtained;

[0019] The normalized data components x' of unified dimension are obtained by normalizing the smoothed attitude angle sequence, the filtered heart rate time series data and the time-synchronized position-micro-motion joint data stream, and the expression is:

[0020]

[0021] Wherein, x is the original data value, x min is the minimum value of the data type in the preset time window, x max is the maximum value of the data type in the preset time window, and x' is the normalized data value.

[0022] All standardized data components are aligned and combined according to the time stamp to form a standardized multi-source perception data stream containing time stamp, normalized attitude angle, normalized heart rate, normalized position coordinate and normalized micro-motion speed.

[0023] As a preferred scheme of the patient intelligent safety management method, wherein: based on the standardized multi-source perception data stream, the behavior action sequence, the spatial position trajectory and the physiological state change of the patient are extracted, the behavior action sequence is combined with the spatial region label to generate a behavior event sequence containing subject-action-space-time semantic structure, and the specific steps are:

[0024] The normalized attitude angle sequence in the standardized multi-source perception data stream is classified and recognized by using threshold segmentation and dynamic time warping algorithm, the occurrence time and duration interval of lying, standing, walking and climbing actions are recognized, and the behavior action sequence is obtained.

[0025] The normalized position coordinate sequence in the standardized multi-source perception data stream is smoothed and predicted by using Kalman filtering algorithm, and the positioning jump noise is eliminated, and a continuous and smooth spatial position trajectory is obtained.

[0026] The normalized heart rate time series data in the standardized multi-source perception data stream is analyzed in frequency domain by using short-time Fourier transform, the heart rate variability indexes including low frequency power and high frequency power are calculated, and the physiological state change sequence is obtained.

[0027] Each action instance in the behavior action sequence is matched with the position coordinate of the corresponding time point in the spatial position trajectory, the position coordinate is mapped to the spatial region label of bed area, corridor area, doorway area and bathroom area according to the preset spatial region division rule, and the four-tuple behavior event record containing patient ID, action type, spatial region and start and end time is generated.

[0028] All behavior event records are arranged in time sequence to form a behavior event sequence.

[0029] As a preferred scheme of the patient intelligent safety management method, wherein: the medical behavior semantic knowledge graph is constructed, the knowledge graph contains patients, beds, doors and windows, nursing stations as entity nodes, lying, standing, walking, climbing, approaching and staying as action nodes, night, heart rate increase and sympathetic activation as state nodes, and the logical association between entities and actions, actions and spaces, and actions and states is established through relationship edges, and the specific steps are:

[0030] The basic elements of the knowledge graph are defined in the form of triples, and the entity nodes include patients, beds, doors and windows, and nursing stations;

[0031] The action nodes include lying, standing, walking, climbing, approaching and staying;

[0032] The state nodes include night, heart rate increase and sympathetic activation;

[0033] The relationship edges are used to connect entities and actions, and the relationship type is defined as execution;

[0034] The relationship edges are used to connect actions and spaces, and the relationship type is defined as occurrence;

[0035] The relationship edges are used to connect actions and states, and the relationship type is defined as trigger;

[0036] The relationship edges are used to connect states and time, and the relationship type is defined as being in;

[0037] All triples are stored in a graph database to construct a medical behavior semantic knowledge graph.

[0038] As a preferred scheme of the patient intelligent safety management method, wherein: the behavior event sequence is mapped to the behavior semantic knowledge graph to form a dynamic behavior subgraph, the graph neural network is used to aggregate the features of the dynamic behavior subgraph, the real-time heart rate variability and skin conductance level are combined to calculate the behavior risk score at the current time, and the specific steps are:

[0039] Each four-tuple behavior event record in the behavior event sequence is mapped to a subgraph instance in the knowledge graph;

[0040] The patient ID, walking, corridor area and time are mapped to a subgraph containing nodes E p , A w , the corridor area and the relationship edge, all behavior events in the time window [T-Δt, T] are mapped to the subgraph, and a dynamic behavior subgraph G T centered on the current time T is formed;

[0041] The graph convolution network GCN is used to process the dynamic behavior subgraph G TFeature aggregation is performed, and the initial feature vector of each node is obtained from a preset embedding table, and the aggregation process is expressed as:

[0042]

[0043] wherein, is the feature representation of node v at the kth layer, is the neighbor node set of node v, W (k) is the trainable weight matrix of the kth layer, and σ is the ReLU activation function, is the feature representation of node u at the (k-1)th layer, and after K-layer aggregation, the final feature vector of the patient node E p is obtained

[0044] The real-time heart rate variability index P of the current time T is obtained from the standardized multi-source perception data stream L / P H The ratio r and the skin conductance level g are spliced into a joint feature vector f.

[0045] The joint feature vector f is input into a fully connected neural network layer, and the behavior risk score s is calculated, and the expression is:

[0046]

[0047] wherein, w is a weight vector, b is a bias term, and s indicates that the higher the behavior risk is.

[0048] As a preferred scheme of the patient intelligent safety management method, wherein: based on the time sequence association of historical physiological data and risk events, a long short-term memory network model is constructed, the skin conductance response change rate, the heart rate power ratio and the behavior frequency in the past preset time length are input, and the risk occurrence probability is output as the psychophysiological precursor prediction result, and the specific steps are:

[0049] From the historical data, physiological and behavioral data segments within a preset time window before the occurrence of a risk event are extracted, each segment containing time series skin conductance response values, heart rate power ratio values and behavior category labels;

[0050] The skin conductance response change rate sequence Δg t = |g t -g t-1 | is calculated, the heart rate power ratio sequence r t = P L (t) / P H (t) is calculated, the behavior occurrence frequency f t in a unit of time is counted, and a three-channel input vector sequence x 1:τ is constituted. ​

[0051] Build a long short-term memory network LSTM model, update the hidden state h t And the cell state c t ;

[0052] The hidden state at the final time is input into the Sigmoid output layer to calculate the risk occurrence probability, and the expression is:

[0053]

[0054] Where w is the output layer weight vector, b is the bias term, and p represents the probability of risk occurrence in the future short time.

[0055] As a preferred scheme of the patient intelligent safety management method, the behavior risk score and the psychophysiological precursor prediction result are weighted and fused to generate a comprehensive risk index, voice reminders and lighting enhancement are triggered when the comprehensive risk index reaches a first threshold value, and emergency alarms are triggered and pushed to a nursing terminal when the comprehensive risk index reaches a second threshold value, and the specific steps are:

[0056] The behavior risk score s and the psychophysiological precursor prediction result p are weighted and fused by using a nonlinear fusion function to calculate the comprehensive risk index, and the expression is:

[0057]

[0058] Where α, β, and γ are adjustable fusion parameters.

[0059] Set the first threshold value T1, when c≥T1, start the voice reminder module to play the preset warning language, and control the ward lighting system to increase the brightness;

[0060] Set the second threshold value T2, when c≥T2, activate the emergency alarm module, and send alarm information containing patient ID, current time, comprehensive risk index, and recent behavior event summary to the preset nursing terminal through the wireless network.

[0061] As a preferred scheme of the patient intelligent safety management method, the response time and event result of each alarm are recorded, and the recorded data are used to update the knowledge graph path weight, prediction model parameter, and individual behavior baseline, and the specific steps are:

[0062] After each alarm is triggered, the time interval δt from the alarm to the confirmation of the response by the nursing staff is recorded, and the event result label l is input by the nursing staff after the event is handled;

[0063] The δt, l, comprehensive risk index c at the alarm time, dynamic behavior subgraph, and LSTM model input sequence x 1:τ are stored in the event log database.

[0064] For knowledge graph path weight update, if the event result l is effective intervention and the response time is within the standard, the relationship edge weight on the key path leading to the current alarm is increased, and vice versa if it is a false alarm.

[0065] For LSTM model parameter update, the newly recorded valid samples are added to the training set, and the small batch gradient descent method is used to fine-tune the model weight.

[0066] For individual behavior baseline update, the behavior frequency and physiological index mean of the patient in the recent non-risk period are counted to adjust the normalization benchmark of the LSTM model input, so that the model adapts to individual differences.

[0067] In a second aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program, when executed by the processor, implements any step of the patient intelligent safety management method according to the first aspect of the present application.

[0068] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the patient intelligent safety management method according to the first aspect of the present application.

[0069] The present application has the following advantages: by performing action recognition, trajectory reconstruction and physiological state analysis on the posture, position and heart rate information in the standardized perception data stream respectively, and combining with the spatial region label to generate a subject-action-space-time four-tuple behavior event sequence, and then mapping it to a medical behavior semantic knowledge graph to form a dynamic behavior subgraph, the structured conversion from the original signal to the high-level semantics is realized, the understanding ability of the system to the behavior intention of the patient in the complex medical scene is enhanced, and the context perception accuracy of the risk behavior is significantly improved. By performing topological feature aggregation on the dynamic behavior subgraph through the graph neural network, the behavior pattern embedding vector of the patient in the medical environment is extracted, and the real-time heart rate variability and skin conductance level are fused to input the nonlinear classifier to calculate the behavior risk score, and an LSTM model is constructed to analyze the historical physiological and behavior time sequence pattern to predict the psychological and physiological precursor, realizing the dual-channel cooperative evaluation of the instant behavior risk and the potential psychological precursor, and significantly improving the foresight and accuracy of the risk warning. BRIEF DESCRIPTION OF DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0071] Figure 1 The flow chart of the patient intelligent safety management method in the embodiment 1. DETAILED DESCRIPTION

[0072] In order to make the above objectives, characteristics and advantages of the present application more apparent, specific embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0073] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. The present application, however, can be practiced in a variety of ways beyond the specific embodiments described herein without departing from the scope of the present application, and it is understood that it covers all techniques falling within the scope of the application sufficient to understand the present application.

[0074] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is independent or alternative to other embodiments.

[0075] Embodiment, refer to Figure 1 For the embodiment of the present application, the embodiment provides a patient intelligent safety management method, comprising the following steps:

[0076] S1, collect the motion posture data and physiological parameters output by the wearable device worn by the patient, and the ultra-wideband positioning data and millimeter wave radar micro-motion signals in the environment, time align and filter the noise of the above data, and obtain standardized multi-source perception data stream;

[0077] Further, a sliding window average filtering algorithm is used to denoise the original motion posture data sequence output by the wearable device in real time, and a smoothed posture angle sequence is obtained.

[0078] A Butterworth low-pass filter is used to suppress high-frequency noise of the original heart rate signal collected by the wearable device, and filtered heart rate time series data is obtained.

[0079] A cross-correlation time synchronization method is used to time stamp align the position coordinate data stream output by the ultra-wideband positioning system and the micro-motion speed signal stream output by the millimeter wave radar, so as to eliminate the sampling time delay deviation between the multi-source devices, and obtain a time-synchronized position-micro-motion joint data stream.

[0080] The smoothed posture angle sequence, the filtered heart rate time series data, and the time-synchronized position-micro-motion joint data stream are normalized, and the normalization uses the minimum-maximum scaling method to obtain standardized data components x' of uniform dimensions in each dimension, and the expression is:

[0081]

[0082] wherein x is the original data value, x min is the minimum value of the data type within the preset time window, x max is the maximum value of the data type within the preset time window, and x' is the normalized data value;

[0083] All the normalized data components are aligned and combined according to the time stamp to form a standardized multi-source perception data stream containing the time stamp, normalized posture angle, normalized heart rate, normalized position coordinate, and normalized micro-motion speed;

[0084] It should be noted that the sliding window average filtering and the Butterworth low-pass filtering used in this step are customized for denoising according to the characteristics of the motion posture and physiological signals, respectively, avoiding the loss of key physiological features by general filtering methods;

[0085] The cross-correlation time synchronization method can adaptively estimate the time delay offset between different devices, significantly improving the consistency of multi-source data in the time dimension; the minimum-maximum normalization method dynamically updates the extreme values based on the sliding time window, adapts to the daily fluctuations of patient behavior and physiological state, ensures that the standardization process does not lose effective change trends, and thus provides a stable, reliable, and comparable input data basis for subsequent high-level semantic analysis.

[0086] S2, based on the standardized multi-source perception data stream, extracting the behavior action sequence, spatial position trajectory, and physiological state change of the patient, combining the behavior action sequence with the spatial region label to generate a behavior event sequence containing the subject-action-space-time semantic structure;

[0087] Furthermore, threshold segmentation and dynamic time warping algorithm are used to classify and recognize the normalized posture angle sequence in the standardized multi-source perception data stream, identify the occurrence time and duration of lying, standing, walking, and climbing actions, and obtain the behavior action sequence;

[0088] Kalman filtering algorithm is used to smooth and predict the normalized position coordinate sequence in the standardized multi-source perception data stream, eliminating positioning jump noise and obtaining a continuous and smooth spatial position trajectory;

[0089] Short-time Fourier transform is used to analyze the frequency domain of the normalized heart rate time series data in the standardized multi-source perception data stream, calculate the heart rate variability indicators including low-frequency power and high-frequency power, and obtain the physiological state change sequence;

[0090] Each action instance in the behavior action sequence is matched with the position coordinates of the corresponding time point in the spatial position trajectory, the position coordinates are mapped into the spatial region labels of the bed region, the corridor region, the doorway region and the bathroom region according to the preset spatial region division rule, and a four-tuple behavior event record containing the patient ID, the action type, the spatial region and the start and end time is generated;

[0091] All the behavior event records are arranged in chronological order to form a behavior event sequence;

[0092] It should be noted that the threshold segmentation combined with the dynamic time warping algorithm can achieve high-precision action recognition under the premise of tolerating action execution speed differences, and is particularly suitable for the actual situation of slow or incoherent actions of the elderly; Kalman filtering not only smooths the positioning trajectory, but also can make reasonable prediction when the signal is lost for a short time, and improve the continuity of spatial perception;

[0093] Short-time Fourier transform is used to extract low-frequency LF and high-frequency HF components in heart rate variability, which is a recognized method for evaluating autonomic nervous function status, and can reflect the emotional tension or physiological stress level of the patient; The four-tuple behavior event is generated by fusing the action and the spatial and time labels, and a minimum behavior unit with complete semantics is constructed, which provides a structured input for subsequent knowledge graph mapping.

[0094] S3, construct a medical behavior semantic knowledge graph, the knowledge graph includes patients, bed bodies, doors and windows, nursing stations as entity nodes, lying, standing, walking, climbing, approaching and staying as action nodes, night, heart rate increase and sympathetic activation as state nodes, and logical associations between entities and actions, actions and spaces, and actions and states are established through relationship edges;

[0095] Further, the basic elements of the knowledge graph are defined in the form of triples, the entity nodes include patients, bed bodies, doors and windows, and nursing stations;

[0096] The action nodes include lying, standing, walking, climbing, approaching and staying;

[0097] The state nodes include night, heart rate increase and sympathetic activation;

[0098] The relationship edges are used to connect the entities and the actions, and the relationship type is defined as execution;

[0099] The relationship edges are used to connect the actions and the spaces, and the relationship type is defined as occurring in;

[0100] The relationship edges are used to connect the actions and the states, and the relationship type is defined as triggering;

[0101] The relationship edges are used to connect the states and the time, and the relationship type is defined as being in;

[0102] Store all triples into the graph database to build the medical behavior semantic knowledge graph;

[0103] It should be noted that the medical behavior semantic knowledge graph explicitly expresses the logical relationship between entities, actions and states in the form of triples, so that the system has interpretable behavior reasoning ability;

[0104] The design of relationship types such as execution, occurrence, triggering and being in line with the logic of clinical nursing can naturally express typical risk scenarios such as patients performing climbing action at night which may trigger sympathetic activation;

[0105] The graph structure is easy to extend and add new entity or action type, and has good maintainability and scalability;

[0106] Storing the knowledge graph into the graph database supports efficient query and subgraph matching, and provides underlying support for real-time dynamic behavior modeling.

[0107] S4, map the behavior event sequence to the behavior semantic knowledge graph to form a dynamic behavior subgraph, use graph neural network to aggregate the features of the dynamic behavior subgraph, combine the real-time heart rate variability and skin conductance level, and calculate the behavior risk score at the current time;

[0108] Further, each four-tuple behavior event record in the behavior event sequence is mapped to a subgraph instance in the knowledge graph;

[0109] Map patient ID, walking, corridor area and time to a subgraph containing node E p , A w , corridor area and relationship edge, merge the subgraphs mapped by all behavior events within the time window [T-Δt, T] to form a dynamic behavior subgraph G T centered at the current time T;

[0110] Use graph convolution network GCN to aggregate the features of the dynamic behavior subgraph G T , the initial feature vector of each node is obtained from the preset embedding table, and the aggregation process expression is:

[0111]

[0112] wherein, is the feature representation of node v at the kth layer, is the neighbor node set of node v, W (k) is the trainable weight matrix of the kth layer, and σ is the ReLU activation function, is the feature representation of node u at the k-1th layer, after K layers of aggregation, the final feature vector of patient node E p is obtained

[0113] obtain real-time heart rate variability index P at current time T from standardized multi-source perception data stream L / P H ratio r and skin conductance level g, will r, g are spliced into joint feature vector f;

[0114] The joint feature vector f is input into the fully connected neural network layer to calculate the behavior risk score s, and the expression is:

[0115]

[0116] where w is the weight vector, b is the bias term, and s indicates that the higher the behavior risk is;

[0117] It should be noted that the behavior event sequence is mapped to the dynamic behavior subgraph, which realizes the conversion from discrete events to context-related behavior patterns, enabling the system to identify complex behavior chains; the graph convolution network GCN effectively captures the topological dependency between patient behavior and environmental entities through multi-layer neighbor aggregation, generating behavior embedding vectors rich in semantics;

[0118] Combining real-time heart rate variability and skin conductance level and other autonomic nervous indicators enhances the sensitivity of the risk score to emotional fluctuations and physiological stress;

[0119] The fully connected neural network layer realizes multi-dimensional feature fusion through nonlinear transformation, and the output behavior risk score can reflect the comprehensive risk level of the current behavior in a specific context.

[0120] S5, based on the time sequence correlation of historical physiological data and risk events, a long short-term memory network model is constructed, which inputs the skin conductance response rate, heart rate power ratio and behavior frequency in the past preset time, and outputs the risk occurrence probability as the psychophysiological precursor prediction result;

[0121] Further, from the historical data, physiological and behavioral data segments containing a preset time window before the occurrence of risk events are extracted, each segment containing time series skin conductance response values, heart rate power ratio and behavior category labels;

[0122] Calculate the skin conductance response rate sequence Δg t = |g t -g t-1 |, calculate the heart rate power ratio sequence r t = P L (t) / P H (t), and count the behavior frequency f t in a unit of time to form a three-channel input vector sequence x 1:τ ;

[0123] Build a long short-term memory network LSTM model, update the hidden state h t and cell state c t ;

[0124] The final hidden state is input into the Sigmoid output layer to calculate the risk occurrence probability, expressed as:

[0125]

[0126] where w is the output layer weight vector, b is the bias term, and p represents the probability of risk occurrence in the near future;

[0127] It should be noted that the long short-term memory network LSTM is good at capturing long-term dependencies in time series and is suitable for modeling the physiological and behavioral evolution rules in the minutes before the occurrence of risk events; the skin conductance response change rate reflects the dynamic process of emotional arousal, the heart rate power ratio reflects the change of sympathetic nervous activity, and the behavior frequency reveals the increasing trend of activity abnormalities, which constitute a multi-dimensional representation of psychophysiological precursors;

[0128] By training the LSTM model with historical risk event annotation data, it can learn individualized or common precursor patterns, and the output risk occurrence probability provides early warning for potential crises that have not yet manifested as high-risk actions, making up for the lag of behavior analysis.

[0129] S6, weight and fuse the behavior risk score and psychophysiological precursor prediction results to generate a comprehensive risk index, trigger voice reminders and lighting enhancement when the comprehensive risk index reaches a first threshold, and trigger emergency alerts and push to the nursing terminal when the comprehensive risk index reaches a second threshold;

[0130] Further, a nonlinear fusion function is used to weight and fuse the behavior risk score s and the psychophysiological precursor prediction result p to calculate the comprehensive risk index, expressed as:

[0131]

[0132] where a, b, and g are adjustable fusion parameters;

[0133] Set the first threshold T1, when c≥T1, start the voice reminder module to play the preset warning language, and control the ward lighting system to increase the brightness;

[0134] Set the second threshold T2, when c≥T2, activate the emergency alarm module, and send alarm information containing patient ID, current time, comprehensive risk index, and recent behavior event summary to the preset nursing terminal through the wireless network;

[0135] It should be noted that the nonlinear fusion function has a synergistic amplification effect, when either the behavioral risk score ss or the psychophysiological precursor probability is higher, the comprehensive risk index increases significantly, reflecting the clinical safety principle that any channel warning needs attention;

[0136] This function form avoids the rigid problem of weight distribution in linear weighting, and is more flexible and robust;

[0137] The hierarchical response mechanism realizes the matching of intervention intensity and risk level, which not only reduces the interference to patients' night rest, but also ensures timely response to high-risk situations, and improves the practicability and user experience of the system.

[0138] S7, record the response time and event result of each alarm, and update the knowledge graph path weight, prediction model parameter and individual behavior baseline using the recorded data;

[0139] Further, after each alarm is triggered, record the time interval δt from the alarm to the confirmation of the response by the nursing staff, and input the event result label l by the nursing staff after the event is handled;

[0140] Put δt, l, the comprehensive risk index c at the alarm time, the dynamic behavior subgraph, and the LSTM model input sequence x 1:τ into the event log database;

[0141] For knowledge graph path weight update, if the event result l is effective intervention and the response time is within the standard, the relationship edge weight on the key path leading to this alarm is increased, otherwise if it is a false alarm, it is reduced.

[0142] For LSTM model parameter update, add the newly recorded valid samples to the training set, and use the small batch gradient descent method to fine-tune the model weight;

[0143] For individual behavior baseline update, statistics the behavior frequency and physiological index mean of this patient in the recent non-risk period, which is used to adjust the normalization benchmark of LSTM model input, so that the model adapts to individual differences;

[0144] It should be noted that the alarm feedback mechanism establishes a closed-loop association between system output and actual clinical effect by recording response time and event result;

[0145] The dynamic update of knowledge graph path weight enables the system to strengthen effective warning paths and suppress false alarms, and gradually optimize the behavior semantic reasoning logic;

[0146] The incremental training of the LSTM model supports online learning of new emerging risk patterns, and improves the prediction accuracy;

[0147] The continuous updating of individual behavior and physiological baseline enables the system to adapt to changes in patient daily rhythm, avoid misjudgment caused by fixed threshold, and realize truly personalized intelligent safety management.

[0148] The embodiment also provides a computer device suitable for the patient intelligent safety management method, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the patient intelligent safety management method proposed in the above embodiment.

[0149] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0150] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to realize the patient intelligent safety management method proposed in the above embodiment. The storage medium can be realized by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk or an optical disk.

[0151] To sum up, the application realizes the structured conversion from the original signal to the high-level semantics by respectively performing action recognition, trajectory reconstruction and physiological state analysis on the posture, position and heart rate information in the standardized perception data stream, combining with the spatial region label to generate the subject-action-space-time four-tuple behavior event sequence, and mapping the four-tuple behavior event sequence to the medical behavior semantic knowledge graph to form a dynamic behavior subgraph, enhances the understanding ability of the system to the behavior intention of the patient in the complex medical scene, significantly improves the context perception accuracy of the risk behavior, aggregates the topological features of the dynamic behavior subgraph through the graph neural network, extracts the behavior mode embedding vector of the patient in the medical environment, and fuses the real-time heart rate variability and skin conductance level to input the nonlinear classifier to calculate the behavior risk score, simultaneously constructs the LSTM model to analyze the historical physiological and behavior time sequence mode to predict the psychological physiological precursor, realizes the dual-channel cooperative evaluation of the instant behavior risk and the potential psychological precursor, and significantly improves the foresight and accuracy of the risk warning.

[0152] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A method for intelligent patient safety management, characterized in that: include: The motion posture data and physiological parameters output by the wearable device worn by the patient, as well as the ultra-wideband positioning data and millimeter-wave radar micro-motion signals in the environment, are collected. The above data are time-aligned and noise filtered to obtain a standardized multi-source sensing data stream. Based on the standardized multi-source sensing data stream, the patient's behavioral action sequence, spatial location trajectory and physiological state changes are extracted. The behavioral action sequence is combined with spatial region labels to generate a behavioral event sequence containing a subject-action-space-time semantic structure. A semantic knowledge graph of medical behavior is constructed. The knowledge graph includes patients, beds, doors and windows, and nursing stations as entity nodes, lying down, standing, walking, climbing, approaching and staying as action nodes, and night, increased heart rate and sympathetic activation as state nodes. Logical associations between entities and actions, actions and space, and actions and states are established through relation edges. The sequence of behavioral events is mapped to the behavioral semantic knowledge graph to form a dynamic behavioral subgraph. A graph neural network is used to perform feature aggregation on the dynamic behavioral subgraph. Combined with real-time heart rate variability and skin conductance level, the behavioral risk score at the current moment is calculated. Based on the temporal correlation between historical physiological data and risk events, a long short-term memory network model is constructed. The model takes into account the skin conductance response rate, heart rate-power ratio, and behavioral frequency within a preset time period in the past, and outputs the probability of risk occurrence as a prediction result of psychophysiological precursors. The behavioral risk score and the psychological and physiological aura prediction results are weighted and fused to generate a comprehensive risk index. When the comprehensive risk index reaches the first-level threshold, a voice reminder and lighting enhancement are triggered. When it reaches the second-level threshold, an emergency alarm is triggered and pushed to the nursing terminal. Record the response time and event result for each alarm, and use the recorded data to update the knowledge graph path weights, prediction model parameters, and individual behavior baselines.

2. The intelligent patient safety management method as described in claim 1, characterized in that: The process involves collecting motion posture data and physiological parameters output from the wearable device worn by the patient, as well as ultra-wideband positioning data and millimeter-wave radar micro-motion signals from the environment. Time alignment and noise filtering are then applied to this data to obtain a standardized multi-source sensing data stream. The specific steps are as follows: A sliding window averaging filter algorithm is used to denoise the raw motion posture data sequence output in real time by the wearable device to obtain a smoothed posture angle sequence. A Butterworth low-pass filter was used to suppress high-frequency noise in the raw heart rate signal collected by the wearable device, resulting in filtered heart rate time-series data. The cross-correlation time synchronization method is used to align the time stamps of the position coordinate data stream output by the ultra-wideband positioning system and the micro-motion velocity signal stream output by the millimeter-wave radar to eliminate the sampling delay deviation between multiple source devices and obtain a time-synchronized position-micro-motion joint data stream. The smoothed attitude angle sequence, filtered heart rate time series data, and time-synchronized position-micromotion joint data stream were normalized using the minimum-maximum scaling method to obtain standardized data components x′ with unified dimensions for each dimension. The expression is as follows: Where x is the original data value, x min x represents the minimum value of this data type within a preset time window. max x′ represents the maximum value of this data type within a preset time window, where x′ is the normalized data value. All standardized data components are aligned and merged according to timestamps to form a standardized multi-source sensing data stream containing timestamps, normalized attitude angles, normalized heart rates, normalized position coordinates, and normalized micro-motion velocities.

3. The intelligent patient safety management method as described in claim 2, characterized in that: Based on the standardized multi-source sensory data stream, the patient's behavioral action sequence, spatial location trajectory, and physiological state changes are extracted. The behavioral action sequence is then combined with spatial region labels to generate a behavioral event sequence containing a subject-action-space-temporal semantic structure. The specific steps are as follows: Threshold segmentation and dynamic time warping algorithms are used to classify and identify normalized posture angle sequences in standardized multi-source sensing data streams, identifying the occurrence time and duration of lying, standing, walking, and climbing actions to obtain behavioral action sequences. The Kalman filter algorithm is used to smooth and predict the trajectory of the normalized position coordinate sequence in the standardized multi-source sensing data stream, eliminating positioning jump noise and obtaining a continuous and smooth spatial position trajectory. Short-time Fourier transform was used to perform frequency domain analysis on normalized heart rate time-series data in a standardized multi-source sensing data stream to calculate heart rate variability indices, including low-frequency power and high-frequency power, and to obtain a physiological state change sequence. Each action instance in the action sequence is matched with the location coordinates of the corresponding time point in the spatial location trajectory. According to the preset spatial area division rules, the location coordinates are mapped to spatial area labels such as bed area, corridor area, door area, and toilet area, generating a four-tuple behavior event record containing patient ID, action type, spatial area, and start and end time. All behavioral event records are arranged in chronological order to form a behavioral event sequence.

4. The intelligent patient safety management method as described in claim 3, characterized in that: The construction of a medical behavior semantic knowledge graph includes patient, bed, door / window, and nursing station as entity nodes; lying down, standing, walking, climbing, approaching, and staying as action nodes; and nighttime, elevated heart rate, and sympathetic activation as state nodes. Logical relationships between entities and actions, actions and space, and actions and states are established through relational edges. The specific steps are as follows: The basic elements of the knowledge graph are defined using a triplet format, with entity nodes including patients, beds, doors and windows, and nursing stations; Action nodes include lying down, standing, walking, climbing, approaching, and staying; State nodes include nighttime, elevated heart rate, and sympathetic activation; Entities and actions are connected using relation edges, and the relation type is defined as execution. Connect actions and spaces using relational edges, defining the relation type as "occurring in"; Actions and states are connected using relational edges, and the relation type is defined as trigger. A relational edge is used to connect state and time, and the relation type is defined as "at"; Store all triples in a graph database to construct a semantic knowledge graph of medical behavior.

5. The intelligent patient safety management method as described in claim 4, characterized in that: The steps are as follows: Mapping the behavioral event sequence to the behavioral semantic knowledge graph to form a dynamic behavioral subgraph; using a graph neural network to perform feature aggregation on the dynamic behavioral subgraph; and combining real-time heart rate variability and skin conductance levels to calculate the behavioral risk score at the current moment. Map each quadruple behavior event record in the behavior event sequence to a subgraph instance in the knowledge graph; Map patient ID, walking distance, corridor area, and time to a node E. p A w The subgraphs of the corridor area and relation edges are merged to form a dynamic behavior subgraph G centered at the current time T. T ; Using Graph Convolutional Network (GCN) to process dynamic behavior subgraphs G T Feature aggregation is performed. The initial feature vector of each node is obtained from a pre-defined embedding table. The aggregation process is expressed as follows: in, Let v be the feature representation of node v at the k-th layer. Let W be the set of neighboring nodes of node v. (k) Let be the trainable weight matrix of the k-th layer, and σ be the ReLU activation function. The feature representation of node u at layer k-1 is aggregated through layer K to obtain patient node E. p The final feature vector Obtain the real-time heart rate variability index P at the current time T from the standardized multi-source sensing data stream. L / P H The ratio r to the skin conductance level g will r and g are concatenated to form a joint feature vector f; Input the joint feature vector f into a fully connected neural network layer to calculate the behavioral risk score s, as expressed by: Where w is the weight vector, b is the bias term, and s represents higher behavioral risk.

6. The intelligent patient safety management method as described in claim 5, characterized in that: The method involves constructing a long short-term memory network model based on the temporal correlation between historical physiological data and risk events. The model takes into account the skin conductance response rate, heart rate-to-power ratio, and behavioral frequency over a preset time period as input, and outputs the probability of risk occurrence as a predictive result of psychophysiological precursors. The specific steps are as follows: Extract physiological and behavioral data segments from historical data that are within a preset time window before the occurrence of risk events. Each segment contains time-series skin conductance response values, heart rate-to-power ratio values, and behavioral category labels. Calculate the skin electrical response rate sequence Δg t =|g t -g t-1 | Calculate the heart rate-power ratio sequence r t =P L (t) / P H (t), the frequency of the behavior occurring per unit time f t This constitutes a three-channel input vector sequence x 1:τ ; Construct a Long Short-Term Memory (LSTM) network model and update the hidden state h. t and unit state c t ; The hidden state at the final moment is input into the Sigmoid output layer to calculate the probability of the risk occurring, expressed as: Where w is the output layer weight vector, b is the bias term, and p represents the probability of risk occurring in the near future.

7. The intelligent patient safety management method as described in claim 6, characterized in that: The steps involve weighted fusion of the behavioral risk score and the prediction results of psychophysiological precursors to generate a comprehensive risk index. When the comprehensive risk index reaches the first-level threshold, a voice reminder and enhanced lighting are triggered; when it reaches the second-level threshold, an emergency alarm is triggered and pushed to the nursing terminal. A nonlinear fusion function is used to weight and fuse the behavioral risk score s and the psychophysiological precursor prediction result p to calculate the comprehensive risk index, the expression of which is: Wherein, α, β, and γ are adjustable fusion parameters; Set a first-level threshold T1. When c≥T1, activate the voice reminder module to play a preset warning message and control the ward lighting system to increase brightness. Set a secondary threshold T2. When c≥T2, activate the emergency alarm module and send an alarm message containing the patient ID, current time, comprehensive risk index, and a summary of recent behavioral events to the preset nursing terminal via the wireless network.

8. The intelligent patient safety management method as described in claim 7, characterized in that: The system records the response time and event outcome for each alarm, and uses this recorded data to update the knowledge graph path weights, prediction model parameters, and individual behavior baselines. The specific steps are as follows: After each alarm is triggered, the time interval δt from the alarm being issued to the nursing staff's confirmation of the response is recorded, and the nursing staff enters the event result label l after the event is handled. Input the comprehensive risk index c at the alarm time (δt, l, and the alarm time), the dynamic behavior subgraph, and the LSTM model into the sequence x. 1:τ Store in the event log database; For knowledge graph path weight updates, if the event result l is an effective intervention and the response time is within the standard, the weight of the relation edge on the critical path that caused this alarm will be increased; otherwise, it will be decreased if it is a false alarm. To update the LSTM model parameters, add the newly recorded valid samples to the training set and fine-tune the model weights using mini-batch gradient descent. To update the individual's behavioral baseline, the frequency of the patient's recent behavior and the mean of physiological indicators during non-risk periods are statistically analyzed. This data is used to adjust the normalized baseline of the LSTM model input, allowing the model to adapt to individual differences.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent patient safety management method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent patient safety management method according to any one of claims 1 to 8.

Images