Intelligent reminder of dosing interruption events and multi-source perception risk intervention system and method

Abstract

The application provides a drug interruption event intelligent reminding and multi-source sensing risk intervention system and method, belongs to the technical field of medical care safety management, aims at the problem that frequent interruption events in the drug administration process significantly increase the risk of medication errors, and realizes the following effects: the drug interruption event intelligent reminding and multi-source sensing risk intervention system collects interruption event signals from personnel, systems and environments in real time through a multi-source sensing module, identifies, classifies and dynamically assesses the risks through an intelligent analysis module, and executes personalized multi-modal reminding, communication shunting, environment control and task scheduling and other intervention strategies through a risk intervention module according to the risk levels, the system integrates a digital twin engine to construct a virtual mapping and predict a high-risk window, adopts federated learning to continuously optimize the model, and uses the blockchain technology to audit and store the key events, and the method can effectively reduce the incidence of medication errors caused by interruptions, and improve the job satisfaction of nursing staff and the safety level of patients.

Description

Technical Field

[0001] This invention proposes an intelligent reminder and multi-source sensing risk intervention system and method for drug administration interruption events, which relates to the field of medical and nursing safety management technology. Background Technology

[0002] Medication safety is a key area of ​​healthcare quality management. Clinical practice and research show that nurses face frequent interruptions during medication administration, which significantly increase the risk of medication errors. The incidence of medication errors increases with the number of interruptions. When nurses encounter interruptions during medication administration, they typically cope with them through task switching, task delays, and multitasking, with task switching being the most common response (69.2%). Interruptions can trigger cognitive impairments in nurses, including loss of attention, memory, or perception, disrupting their thought processes, distracting them, reducing work efficiency, and even leading to medication errors. Furthermore, interruptions often trigger negative emotions in nurses, resulting in decreased cognitive function and task performance, as well as feelings of loss of professional identity and helplessness. Therefore, interruptions in medication administration are a major cause of medication errors and are closely related to the frequency and severity of adverse events. Summary of the Invention

[0003] This invention addresses the problem that frequent interruptions during drug administration significantly increase the risk of medication errors by proposing an intelligent reminder and multi-source perception risk intervention system and method for drug administration interruptions.

[0004] First, a nursing interruption event is defined as a sudden external event that interrupts or delays the current task or distracts the recipient's attention during the provision of ethical nursing services within a prescribed time, role, and environment. Interruption events that lead to adverse nursing events are defined as adverse outcome nursing interruption events. In the context of medication error events, a nursing interruption event occurs when any step in the process of recording, transcribing, verifying, preparing, administering, or supervising medication according to doctor's orders is interrupted, delayed, or distracted by a sudden factor, leading to non-compliance with the medication administration procedure and potentially resulting in a medication error event.

[0005] This invention proposes an intelligent reminder and multi-source sensing risk intervention system for drug administration interruption events, comprising: The multi-source sensing module collects multi-source sensing data in the drug delivery environment in real time through a sensor network, and packages the multi-source sensing data according to its source to form a standardized event stream output. The environmental perception unit deploys ultra-wideband positioning anchors and infrared light curtain detection devices in the drug delivery work area; it monitors the environment by fusing ultra-wideband positioning, infrared light curtain boundary detection and behavior analysis algorithms, and generates a standardized environmental interruption signal containing information such as interference type, location coordinates, duration, intruder identity and heat map risk factors when an interference event is confirmed. The communication management unit manages communication request data during the critical period of drug administration, and uses speech recognition and natural language processing models to identify the urgency of incoming calls and execute dynamic triage strategies. The digital twin engine receives data from the multi-source sensing module, the environmental sensing unit, and the communication management unit, constructs and updates a virtual mapping of the drug delivery environment in real time, dynamically generates a risk heat map, runs simulation prediction to output prediction data, and predicts future high-risk interference windows. The intelligent analysis module receives the predicted data from the digital twin engine, uses a multi-label classification model as the event recognition and classification engine, and calls the dynamic risk assessment engine optimized by the federated learning collaborative network to obtain the risk level. The risk intervention module is used to perform multimodal adaptive reminders through wearable devices and augmented reality interfaces based on the risk level, and uses a dynamic task queue to prioritize and sequentially guide interrupted tasks and core drug administration tasks.

[0006] Furthermore, the sensor network includes: The personnel perception subsystem uses deployed cameras and a computer vision behavior recognition algorithm based on YOLOv5 to automatically identify the behavior of others. At the same time, it monitors the nurse's posture, movement trajectory, operation rhythm and physiological signals through smart name tags or wristbands with integrated inertial measurement units. The system perception subsystem monitors changes in medical orders, patient call lights, and laboratory critical value alarm signals in real time through the hospital information system interface, and integrates the API of the hospital communication system to obtain real-time event streams of incoming telephone calls and internal broadcasts. The environmental sensing subsystem deploys noise and light sensors to monitor changes in ambient noise or lighting.

[0007] Furthermore, the environmental perception unit deploys ultra-wideband anchor points to form an electronic fence, while simultaneously deploying an infrared light curtain to detect unauthorized crossing behavior and using a convolutional neural network to determine the identity of the entrant.

[0008] Furthermore, multimodal adaptive alerts are executed through wearable devices and augmented reality interfaces to trigger risk intervention strategies, including: Low-risk intervention: Non-invasive alerts are sent by vibrations from the nurse's augmented reality glasses or smartwatch, overlaid with color-coded prompts on a visual interface; Intermediate risk intervention: Automatically activate intelligent communication filtering to transfer non-emergency calls and ward calls to voicemail or AI voice assistant, while displaying suspended tasks in a list and suggesting processing order through the task scheduling unit; Advanced risk intervention: Initiate deep protection by projecting warning light strips with smart lights or controlling smart gates to restrict entry, and force the visual and operational focus to switch back to the core drug delivery process by reminding or automatically pausing non-emergency task interfaces.

[0009] Furthermore, the digital twin engine uses a stochastic simulation algorithm to predict high-risk interference windows in the future, models the nursing environment as a discrete state process, defines a state vector that includes the number of nurses, patients, and family members in each area, the length of the pending medical orders queue, and the system alarm status, and defines multiple types of response channels corresponding to nursing behavior events.

[0010] This invention also proposes a method for intelligent reminder and multi-source sensing risk intervention for drug administration interruption events, implemented based on the aforementioned intelligent reminder and multi-source sensing risk intervention system for drug administration interruption events, comprising the following steps: S1: Collect multi-source sensing data in the drug delivery environment in real time through a sensor network, and label and package the multi-source sensing data according to its source to form a standardized event stream output; S2: The drug delivery work area is monitored by the fusion technology of ultra-wideband positioning, infrared light curtain boundary detection and behavior analysis algorithm. When an interference event is confirmed, a standardized environmental interruption signal containing information such as interference type, location coordinates, intruder identity and heat map risk factors is generated. S3: Manage communication request data during critical periods of drug administration, classify incoming voice calls by urgency, and implement dynamic triage strategies based on nurse activity stage tags; S4: Receives standardized event streams, standardized environmental interruption signals, and communication request data; constructs and updates a high-fidelity virtual mapping of the drug delivery environment in real time; dynamically generates a risk heatmap; and runs simulation predictions to output prediction data and predict future high-risk interference windows. S5: Receive the predicted data output from step S4, call the multi-label classification model to identify the interruption event, and input it into the dynamic risk assessment engine optimized by the federated learning collaborative network to calculate the risk level. S6: Based on the risk level, multimodal adaptive reminders are implemented through nurses' wearable devices and augmented reality interfaces, and a dynamic task queue is used to prioritize and sequence interrupted tasks and core drug administration tasks.

[0011] Furthermore, in step S1, the real-time acquisition of multi-source sensing data in the drug delivery environment through the sensor network includes: automatically identifying the behavior of others through deployed cameras using a computer vision behavior recognition algorithm based on YOLOv5, and simultaneously monitoring the nurse's posture, movement trajectory, operating rhythm and physiological signals through a smart badge or wristband with an integrated inertial measurement unit worn by the nurse. The system monitors changes in medical orders, patient call lights, and laboratory emergency alarm signals in real time via the hospital information system interface. It also integrates the hospital communication system API to obtain real-time event streams of incoming telephone calls and internal broadcasts. Noise and light sensors are deployed to monitor changes in ambient noise or lighting.

[0012] Furthermore, in step S4, a high-fidelity virtual mapping of the drug administration environment is constructed and updated in real time through a digital twin step, and a random simulation algorithm is used to predict the high-risk interference window period in the future time period. The state vector defined by the random simulation algorithm includes: the number of nurses, patients, and family members in each area, the length of the pending medical orders queue, and the system alarm status.

[0013] Furthermore, in step S6, a multimodal adaptive alert is executed through the wearable device and the augmented reality interface to trigger a risk intervention strategy, including: Low-risk intervention: Non-invasive alerts are sent by vibrations from the nurse's augmented reality glasses or smartwatch, overlaid with color-coded prompts on a visual interface; Intermediate risk intervention: Automatically activate intelligent communication filtering to transfer non-emergency calls and ward calls to voicemail or AI voice assistant, while displaying suspended tasks in a list and suggesting processing order through the task scheduling unit; Advanced risk intervention: Initiate deep protection by projecting warning light strips with smart lights or controlling smart gates to restrict entry, and force the visual and operational focus to switch back to the core drug delivery process by reminding or automatically pausing non-emergency task interfaces.

[0014] Furthermore, in step S5, the federated learning collaborative network optimization process includes: periodically uploading the risk assessment weights and GNN parameters of the local models deployed on each participating hospital for interruption event identification and risk assessment to the federation aggregation server after homomorphic encryption; using an aggregation algorithm to aggregate the local models of multiple hospitals to generate a global model; and then encrypting and downloading the local models of the local hospitals for updating.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: This invention transforms medication interruption events from reactive response to proactive management. Through multi-source sensing and intelligent analysis, the system can accurately identify the source and risk level of the interruption, avoiding the efficiency losses caused by traditional one-size-fits-all management. Dynamic intervention strategies are implemented differently based on operational criticality, ensuring that core medication administration is not disrupted while maintaining necessary communication channels. The environmental sensing unit upgrades static warning signs into dynamic intelligent protective spaces, significantly improving the safety of the physical environment. The task scheduling unit guides nurses to handle multiple tasks in an orderly manner, reducing cognitive load and errors caused by task transitions. The closed-loop optimization mechanism enables the system to continuously learn and evolve, adapting to the work characteristics of different wards and teams, ultimately effectively reducing the incidence of medication errors and improving nurses' job satisfaction and patient safety. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A schematic diagram of a system for intelligent reminders and multi-source perception risk intervention for drug administration interruption events; Figure 2 Schematic diagram of the layout of the environmental sensing unit; Figure 3 For risk heatmap; Figure 4 A schematic diagram of the AR glasses interface; Figure 5 Flowchart of intervention methods for drug administration interruption event risks. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] The intelligent reminder and multi-source perception risk intervention system for drug administration interruption events includes a multi-source perception module, an environmental perception unit, a communication management unit, a digital twin engine, an intelligent analysis module, and a risk intervention module.

[0020] The multi-source sensing module, serving as the system's data input, is responsible for real-time acquisition of all-dimensional interruption event signals originating from itself, others, the system, and the environment within the drug delivery environment via an IoT sensor network. It then performs preliminary labeling and packaging of all sensing data according to four main sources, forming a standardized event stream output. The IoT sensor network constructed by the multi-source sensing module specifically includes: The personnel perception subsystem achieves full-coverage monitoring through intelligent visual recognition and nurses' own state perception. 4K wide-angle cameras are deployed at key nodes in nurse stations, dispensing rooms, and corridors. An improved YOLOv5-based computer vision behavior recognition algorithm automatically identifies the behavior of others. The backbone network of this algorithm integrates a CBAM attention mechanism and combines 3D-CNN and Bi-LSTM to process temporal behaviors, enabling it to automatically identify various interruption behaviors such as approaching, conversing, or gesturing calls.

[0021] This invention constructs a spatiotemporal dual-branch architecture to realize the detection and classification of target behaviors in video sequences. The technical details of each module are as follows: (1) Spatial feature extraction branch: A spatial feature extraction network is built based on YOLOv5s, and its attention mechanism is enhanced and improved: After each C3 module of the backbone and feature fusion network, a convolutional block attention module (CBAM) is inserted. This module adaptively enhances key spatial features and suppresses redundant feature interference through a serial combination of channel attention and spatial attention, thereby improving feature representation capabilities.

[0022] This branch takes a single-frame RGB image as input. After the improved network extraction, the output dimension is Spatial features of the pyramid S t .

[0023] (2) Spatiotemporal feature fusion and temporal modeling module To capture the temporal correlation characteristics of behavior, a temporal modeling method is constructed by cascading 3D-CNN and Bi-LSTM: A1. Construction of the Spacetime Cube The time window length is set to T=8 frames, the frame sampling interval Δt=0.5s, and a single time window can cover a continuous behavioral segment with a duration of 4s; the spatial feature pyramid S corresponding to the 8 frames within the window is then used to construct the spatial feature pyramid. t Stack along the temporal dimension to construct a spacetime cube .

[0024] A2, 3D-CNN Temporal Coding Instead of the original YOLOv5's FPN+PAN feature pyramid structure, a 3D convolutional layer is used to encode the spatiotemporal cube: By using a 3D convolution kernel with a kernel size of (3,3,3), the number of feature channels is increased from 256 to 512. With a temporal step size of (1,2,2), the spatial dimension of the spatiotemporal features is downsampled, and the encoded spatiotemporal features are output. .

[0025] A3, Bi-LSTM Long-Term Timing Modeling To capture the long-range dependencies of behavioral sequences, a bidirectional long short-term memory network (Bi-LSTM) is introduced to model spatiotemporal features: First, the spatial dimension of the spatiotemporal feature H is flattened to obtain a temporal feature sequence with a dimension of 8×51200; Then, a bidirectional LSTM network with a hidden layer dimension of 256 is used to bidirectionally encode the temporal features, with an output dimension of... The global temporal feature vector Z.

[0026] (3) Multi-task output head Employing a multi-task output structure, it simultaneously achieves object detection and behavior classification: The object detection branch is used to retain the anchor-based detection head of YOLOv5 and realize the regression output of the object bounding box coordinates (x,y,w,h) and detection confidence.

[0027] The behavior classification branch is used to input the global temporal feature vector Z into a multilayer fully connected network. After dimensional transformation (512→128→13), the Softmax activation function is used to classify and recognize 13 types of behaviors. The behavior categories include: approaching, talking, gesturing to call, using the telephone, handing over objects, intruding, loitering, etc.

[0028] The model training data comes from real hospital scenarios, with a sample size of 120,000 frames. Performance metrics show that the behavior recognition accuracy reaches 92.7%, and the average recognition latency is 230 milliseconds per frame.

[0029] Meanwhile, nurses wear smart name tags or wristbands with integrated inertial measurement units. These devices integrate a 6-axis accelerometer and gyroscope IMU inertial measurement unit and a UWB tag to monitor their posture, movement trajectory, operating rhythm, and physiological signals such as heart rate variability and skin conductance in real time. The positioning accuracy of the movement trajectory is ±15 centimeters.

[0030] The system perception subsystem monitors electronic signals such as doctor's order changes, patient call lights turning on, and laboratory critical value alarms in real time through the hospital information system interface, and integrates the API of the hospital communication system to obtain real-time event streams such as incoming calls and internal broadcasts. In specific implementation, the system perception subsystem captures electronic signals such as doctor's order changes, patient call lights, laboratory critical values, and PACS image viewing requests in real time through HL7 or FHIR standard protocols, with a delay time less than 500 milliseconds. At the same time, it accesses the hospital PBX phone system and the internal broadcast API to obtain event stream data such as incoming call numbers, call durations, and broadcast priorities.

[0031] The sensing subsystem deploys noise sensors and light sensors to monitor sudden increases in environmental noise decibels or lighting changes. Specifically, it uses MEMS noise sensors with a measurement range of 30 to 120 decibels and an accuracy of plus or minus 1.5 decibels, and illuminance sensors with a measurement range of 0 to 20,000 lux to monitor sudden decibel changes. When the change amount is greater than 10 decibels, an alarm is triggered, as well as abnormal lighting such as flickering or sudden dimming. The sensing subsystem also connects intelligent LED light strips, electronic displays, and acoustic dispelling devices through Modbus or OPC-UA protocols to achieve environmental control linkage.

[0032] The environmental perception unit is specifically used to monitor and identify the medication work area, and achieves high-precision determination through the triple fusion technology of ultra-wideband positioning, infrared light curtain boundary detection, and behavior analysis algorithms. As Figure 2 shown, the environmental perception unit arranges four ultra-wideband anchors at the entrances of the pharmacy and the high-risk drug preparation room to form an electronic fence to track personnel coordinates in real time. At the same time, it deploys a 940-nanometer infrared light curtain with a response time less than 50 milliseconds, which can detect unauthorized crossing behaviors, and combines the above data to use a lightweight convolutional neural network to judge the identity of the entrants. The determination accuracy rate reaches 98.3%, and the false alarm rate is 2.1%.

[0033] When the digital twin engine predicts that the family member's visitation path will intersect with the nurse's workflow, the environmental perception unit can receive predictive instructions, trigger a yellow warning three seconds in advance, and project a gentle yellow light curtain through the LED light strip, achieving a leap from perception and reaction to prediction and prevention.

[0034] The hierarchical response strategy executes differential control according to the risk level. When the risk is low, only a prompt message of "Please wait a moment during medication preparation" is displayed on the electronic display. When the risk is medium, a synchronized sound and light warning is activated and a voice message of "You have entered the no-interruption area, please be quiet" is played. When the risk is high, the access control of the no-interruption area is automatically locked and security personnel are called. Once a disturbance event is confirmed, the unit immediately generates a standardized environmental interruption signal containing information such as the disturbance type, location coordinates, duration, identity of the intruder, and heat map risk factors, and sends it to the intelligent analysis module.

[0035] The communication management unit manages communication requests during critical medication administration periods through an intelligent communication gateway and a rule engine, employing a dual-core architecture of rule engine and AI intent recognition. This unit integrates eight communication interfaces, including IP phones, call lights, mobile applications, internal broadcasts, and walkie-talkies, enabling omnichannel access. For incoming voice calls, a speech recognition model transcribes the calls and then uses a lightweight natural language processing model to determine urgency, classifying them into four categories: urgent, important, general, and spam. The communication management unit receives nurse activity stage tags from the intelligent analysis module. During critical stages such as verification or medication preparation, a communication silence mode is activated, and a dynamic triage strategy is implemented based on the classification results: emergency calls are allowed but hands-free calling is mandatory and automatically disconnected after 15 seconds to minimize interference; important calls are transferred to the AI ​​voice assistant and a structured summary is pushed to the nurse's augmented reality glasses; general calls receive an automatic SMS reply that the nurse is administering medication and will reply later; spam calls are directly blocked and recorded in a blockchain blacklist. The unit also records communication blocking success rates, the number of missed emergency calls, and nurse satisfaction scores, generating a daily communication management report.

[0036] The digital twin engine receives data from multi-source sensing modules, environmental sensing units, and communication management units to construct and update in real time a high-fidelity virtual mapping of the drug delivery environment, including personnel coordinates, equipment status, and area occupancy. Based on building information modeling and real-time sensor data, this digital twin engine constructs a 3D spatial model, forming a multi-layered twin model containing nurses, patients, equipment, and area status, updated once per second. It also integrates ultra-wideband coordinates, inertial measurement unit attitude data, hospital information system events, and heatmap data to drive the twin's status updates.

[0037] The digital twin engine dynamically generates a risk heat map, displaying personnel flow in a 10-meter by 10-meter grid within the twin. When the personnel density exceeds 5 people per square meter, it is displayed in red; when the density is between 3 and 5 people, it is displayed in yellow; and when it is less than 3 people, it is displayed in green. Heat map factors are used to represent the color level (red, yellow, green) or the numerical risk weight corresponding to each grid in the risk heat map generated by the digital twin engine. Heat map factors directly participate in risk assessment.

[0038] The digital twin engine uses simulation to predict and output future high-risk interference windows. It employs a stochastic simulation algorithm to predict risk periods 15 to 30 minutes in the future. For example, based on input conditions such as the current number of nurses in the pharmacy, the number of family members waiting in the corridor, and the number of medical orders awaiting review in the system, the system can output predictions such as the pharmacy entering a high-risk period during a specific time and suggest deploying additional auxiliary nurses. Each time a high-risk event is triggered, the engine automatically saves a snapshot of the twin's state, including personnel location, environmental parameters, and task queues, generating a unique snapshot identifier for auditing and backtracking.

[0039] Stochastic simulation algorithms model the nursing environment as a discrete state process, defining state vectors. : ; Where M is the total number of areas, such as pharmacy, ward, corridor, etc., and m∈{1,…,M}; Let N, P, and F represent the number of nurses (N), patients (P), and family members (F) within region m, respectively. Q(t) is the length of the current queue of medical orders to be processed; A(t)∈{0,1} is the system alarm status, 0=normal, 1=critical alarm.

[0040] The initial state vector X(0) is determined by reading the real-time snapshot of the digital twin engine at the current moment. It includes key system state parameters such as the real-time number of people in each area, the length of the medical order queue, the equipment occupancy status, and the on-duty status of medical staff, providing a realistic initial system state for the simulation. Table 1 shows the mapping of six types of reaction channels and their corresponding rate constants.

[0041] Table 1

[0042] Parameter calibration method: Calculated using historical UWB trajectory data; Data fitting from the call system: The exponential distribution of historical arrival time intervals was fitted using maximum likelihood estimation; Regression via interrupted event logs: The goal is to obtain from multiple In the middle, find one This causes the model to predict the number of interruptions. The sum of squared errors between the actual number of interruptions and the minimum number of interruptions is minimized.

[0043] Calculate the rate constant A for all reaction channels. I And calculate the overall reaction rate A0: ; I is the sequence number of the reaction channel; The time interval for the next event is generated based on the exponential distribution. : ; Where r1 is the first uniformly distributed random number in the interval [0,1].

[0044] Select the A reaction occurs, satisfying: ; Where r2 is a second uniformly distributed random number in the interval [0,1].

[0045] The cumulative probability interval is calculated, and a second uniformly distributed random number r2 is used to determine which reaction channel is triggered; this ensures that events with a high probability of occurrence have a greater chance of being selected.

[0046] The above algorithm can accurately simulate the random occurrence of discrete events such as personnel flow and medical order generation in the nursing environment, thereby predicting high-risk interference windows in the future. In a specific embodiment, at 9:00 AM, the current status of the dispensing room is predicted as follows: Area 1 dispensing room: 1 nurse, 0 family members, Q=3, that is, 3 pending medical orders, which are also 3 medications to be dispensed; Area 2 Corridor: 2 family members awaiting visitation; Area 3 Nursing Station: 1 nurse processes medical orders; A randomized simulation algorithm was used to perform 1000 trajectory simulations: the transfer rate of family members from the corridor to the pharmacy was 0.8 times / hour; the probability of a family member breaking into the pharmacy during the time period of 9:05-9:08 was 68%. like Figure 3 The image shown is a risk heatmap, and the system has determined that region 1 is a high-risk interference window.

[0047] The intelligent analysis module uses the BERT-GNN multi-label classification model as the event recognition and classification engine. The BERT branch, based on domain-pre-trained ClinicalBERT, processes text information such as medical orders and call recording transcriptions and outputs semantic embeddings. The GNN branch constructs a spatial relationship graph of nurse-patient-device-region topology. The graph node features include UWB coordinates, task type, and heatmap factors. GraphSAGE is used to aggregate neighbor information. The number of sampling layers in GraphSAGE is set to 2, with 10 neighbor nodes aggregated in the first layer and 5 in the second layer. The aggregation function is the mean aggregator. The hidden layer dimension is set to 256.

[0048] The fusion output layer ultimately outputs interruption reason labels in 4 major categories and 32 subcategories. Testing on 120,000 labeled data points covering 8 hospitals showed that the model achieved an overall accuracy of 89.6% and a macro-average F1 score of 0.873, outperforming the single BERT model by 6.2 percentage points.

[0049] Specifically, this embodiment uses the BERT-GNN multi-label classification model to understand and model multi-source text information in nursing scenarios. The specific implementation method is as follows: Text feature extraction module (BERT branch): Employs a clinicalBERT-base model fine-tuned for the medical field, consisting of 12 Transformer layers with a hidden layer dimension of 768. The input is a concatenated text sequence: [CLS] + medical order text + [SEP] + call transcript text + [SEP] + broadcast content, with a maximum sequence length of 256. The first vector of the last hidden state, i.e., the [CLS] position, is extracted as the text feature vector. .

[0050] Heterogeneous Graph Construction Module (GNN Branch): Constructs a heterogeneous nursing scenario graph G=(V,E) to represent the multi-entity relationships in the nursing scenario, where the node set V contains four types of entities, each carrying multi-dimensional features.

[0051] The nurse node's features include: UWB coordinates (x, y, z) and IMU pose quaternion (q). w ,q x ,q y ,q z The current task encodes a one-hot vector; Patient nodes, characterized by: fall risk score, allergy history code, and nursing level; Device nodes, characterized by: device type code, operating status, and alarm level; Regional nodes, characterized by: region type (dispensary / ward / corridor), current personnel density, and ambient noise level (decibels). The edge set E contains three types of associated edges, used to represent the interaction relationships between entities: Spatial edge, the Euclidean distance between nodes is less than 2m; Logical edge, nurse-patient binding relationship; Communication side, telephone call relationship.

[0052] The 768-dimensional text feature vector output by the BERT branch and the 256-dimensional graph feature vector output by the GNN branch are concatenated and fused through a fully connected layer, resulting in a feature dimension of 1024.

[0053] Risk Assessment and Federated Learning Optimization: The intelligent analysis module calls upon the dynamic risk assessment engine optimized by the federated learning collaborative network to calculate the risk value using an enhanced dynamic risk assessment formula. : ; Normalized frequency coefficients ; Normalized source risk weights ; Key coefficients in the normalization stage ;

[0054] In the formula, F represents the frequency coefficient, with a baseline value of 0.31, calculated based on one interruption occurring every 3.2 dosing procedures, and dynamically adjusted according to the number of real-time interruptions in the past hour; S represents the source risk weight, determined by expert assessment: 3.5 for family inquiries, 4.2 for telephone calls, 5.0 for critical value alarms, and 5.0 for system failures; C represents the stage criticality coefficient: 5.0 for the verification stage, 4.5 for the medication preparation stage, 3.0 for the infusion stage, and 1.5 for the observation stage; t is the time decay factor, i.e., the duration of the interruption. The time accumulation coefficient, ranging from 0.1 to 0.3, increases the risk value by 10% to 30% for every minute of interruption. This aligns with the linear accumulation of nurses' cognitive load with the duration of interruption. H is a spatial heatmap factor, output in real time by the twin, with a value of 1.3 in the red area, 1.1 in the yellow area, and 1.0 in the green area. P is a predictive factor, output by the federated learning GNN model, representing the probability of interference in the next 15 minutes, ranging from 0.8 to 1.2.

[0055] The risk grading standard is as follows: a score greater than 0 and less than or equal to 2.5 indicates low risk (only recorded); a score greater than 2.5 and less than or equal to 5.0 indicates medium risk (prompt alert); and a score greater than 5.0 indicates high risk (triggering mandatory intervention and initiating blockchain evidence storage). The risk grading standard is determined through ROC curve analysis and can be dynamically adjusted in different departments.

[0056] The intelligent analysis module participates in the construction of a horizontal federated learning system in the medical consortium blockchain. On the first day of each month, the risk assessment weights and GNN parameters of the local model are uploaded to the consortium aggregation server after being homomorphically encrypted. An aggregation algorithm is used to aggregate the models of 10 to 15 hospitals to generate a global model. On the third day, the model is encrypted and downloaded to the local machine for updates. The original data does not leave the hospital throughout the process to comply with HIPAA and GDPR requirements. This collaborative network improves the recognition rate of rare interruption events by 41%, and the model AUC increases from 0.84 to 0.91.

[0057] The risk intervention module provides multimodal adaptive reminders based on the risk level and the nurse's individual condition through wearable devices and augmented reality interfaces. It also uses a dynamic task queue to prioritize and sequentially guide interrupted tasks and core drug administration tasks.

[0058] Specifically, multimodal physiological emotion inference is achieved through personalized reminder units. Heart rate variability time-domain RMSSD and frequency-domain LF / HF indices and skin conductance responses are collected using PolarH10 chest straps or EmpaticaE4 wristbands. Facial micro-expression units are captured at 30fps using the front-facing camera of AR glasses. The MobileFaceNet model is used to infer fatigue and anxiety levels. The task load index is calculated by combining the current task queue length and task complexity weight in the twin, forming a dynamic weighted scoring mechanism: the reminder method score is equal to w1 multiplied by the emotion fit plus w2 multiplied by the cognitive load influence plus w3 multiplied by the personal preference history, where w1 to w3 are adjustable in real time and the initial values ​​are 0.4, 0.4, and 0.2, respectively. The PPO reinforcement learning algorithm is used to learn and optimize from historical acceptance rate data.

[0059] Preferably, the multimodal adaptive reminder provides a multimodal reminder library including visual, tactile, and auditory reminders. Visual reminders project a non-focused amber icon onto the upper right corner of the field of vision via AR glasses for less than 1.5 seconds. Tactile reminders use a smart bracelet with a custom vibration mode to distinguish between low-frequency short vibrations for general reminders and high-frequency long vibrations for emergency reminders, and support 10 levels of intensity adjustment. Auditory reminders play whisper-level voice messages via bone conduction headphones at a volume of 45 decibels, triggered only when the nurse is stationary. A task scheduling unit initiates dynamic task queue management when it detects that a nurse has received more than two task requests simultaneously or is attempting to switch tasks. An improved CURE algorithm is used to calculate the task priority score, which is equal to urgency multiplied by 0.6, patient safety relevance multiplied by 0.3, and estimated time multiplied by 0.1. The urgency value is 10 for criticality, 9 for medication verification, and 3 for family inquiries. Patient safety relevance is assessed based on the patient's fall risk and allergy history in the twin. The estimated time is dynamically updated based on eye-tracking of the AR glasses to infer task progress. Figure 4 As shown, the interface presents a semi-transparent task list for AR glasses. The current core task is highlighted in green, while interrupted tasks are sorted by priority and displayed in amber. It provides one-click postponement or gesture operation for delegating to others.

[0060] like Figure 5 As shown, the intervention method for drug administration interruption event risk based on the above system includes the following steps: S1. Collect multi-source sensing data in the drug delivery environment in real time through a sensor network, and label and package the multi-source sensing data according to its source to form a standardized event stream output.

[0061] Cameras are deployed at key nodes, and a spatiotemporal dual-branch architecture behavior recognition algorithm based on YOLOv5 is used to automatically identify 13 types of interruption behaviors of others, such as approaching and talking. At the same time, nurses wear smart name tags or wristbands with integrated inertial measurement units (IMUs) to monitor their own posture, movement trajectory, operation rhythm and heart rate variability and other physiological signals in real time.

[0062] Through standard protocol interfaces such as HL7 or FHIR, it can monitor electronic signals in the hospital information system in real time, such as changes in medical orders, patient call lights, and laboratory critical value alarms; at the same time, it integrates communication system APIs to obtain real-time event streams such as incoming telephone calls and internal broadcasts.

[0063] Deploy MEMS noise sensors and illuminance sensors to continuously monitor sudden changes in noise levels and abnormal lighting conditions in the environment.

[0064] Finally, all multi-source sensing data is initially labeled and packaged according to its source to form a standardized JSON-Schema format event stream output, ensuring that end-to-end latency is minimized.

[0065] S2. Monitor and identify the status of the drug delivery work area. The determination is achieved through the fusion technology of ultra-wideband positioning, infrared light curtain boundary detection and behavior analysis algorithm. When an interference event is confirmed, a standardized environmental interruption signal containing interference type, location coordinates, intruder identity and heat map risk factor information is generated.

[0066] This system specifically monitors and identifies the status of physical drug administration work areas, such as dispensing rooms. This step utilizes ultra-wideband anchor points to form an electronic fence, deploys infrared light curtains to detect boundary crossings, and integrates a lightweight convolutional neural network for entrant identification, achieving high-precision spatial status awareness. The system incorporates hierarchical response logic: when the digital twin engine predicts a potential interference path, it can trigger a yellow alert in advance; when an intrusion actually occurs, differentiated control is implemented based on the risk level, ranging from low-risk electronic screen prompts to medium-risk audio-visual warnings, and finally to high-risk access control locking and security calls. Once interference is confirmed, a standardized environmental interruption signal is immediately generated, containing information such as the interference type, location coordinates, intruder identity, and heatmap risk factors.

[0067] S3 manages communication request data during critical periods of drug administration, classifies incoming voice calls by urgency, and implements dynamic triage strategies based on nurse activity stage tags.

[0068] All communication requests during the critical drug administration period are managed through an intelligent communication gateway and rules engine. This step integrates eight types of communication interfaces, including IP telephony, paging lights, and mobile applications, to achieve omnichannel access.

[0069] The processing flow includes: voice recognition and transcription of incoming calls, followed by the use of a lightweight natural language processing model to determine their urgency level and classify them into four categories: urgent, important, general, and spam. The system receives activity stage tags from the intelligent analysis module and automatically activates a communication silence mode during critical stages such as verification or medication dispensing. Based on the classification results, it executes dynamic triage strategies, such as: allowing emergency calls but forcing hands-free calling and suspending them after 15 seconds; transferring important calls to the AI ​​voice assistant and pushing summaries to AR glasses; automatically replying to general calls via SMS; and directly blocking spam calls.

[0070] S4 receives standardized event streams, standardized environmental interruption signals, and communication request data; constructs and updates a high-fidelity virtual mapping of the drug delivery environment in real time; dynamically generates a risk heatmap; and runs simulation predictions to output future high-risk interference windows.

[0071] The data from the aforementioned steps are received and integrated to construct and update a high-fidelity virtual mapping of the drug delivery environment every second. This virtual mapping includes multi-level information such as personnel coordinates, equipment status, and area occupancy.

[0072] Dynamically generated risk heatmaps display personnel density risk in a 10m x 10m grid format; for example, a density exceeding 5 people per square meter is displayed in red. Another core function is simulation prediction: using a stochastic simulation algorithm, the nursing environment is modeled as a discrete-state process. A state vector is defined, including the number of nurses, patients, and family members in each area, the length of the pending medical orders queue, and alarm status. Multiple response channels, such as "nurse movement" and "family member entry," are defined to simulate nursing behavioral events, thereby predicting high-risk interference windows in the next 15 to 30 minutes.

[0073] S5. Receive the predicted data output from step S4, call the multi-label classification model to identify the interruption event, and input it into the dynamic risk assessment engine optimized by the federated learning collaborative network to calculate the risk level.

[0074] The predicted data from step S4 is received and subjected to deep understanding and risk assessment. This step first employs a BERT-GNN multi-label classification model for event identification and classification: the BERT branch processes text information based on ClinicalBERT, while the GNN branch constructs a heterogeneous graph of nurses, patients, equipment, and regions to aggregate spatial relationship features, ultimately outputting interruption reason labels in 4 major categories and 32 subcategories. Subsequently, a dynamic risk assessment engine optimized by a federated learning collaborative network is invoked, and an enhanced dynamic risk assessment formula is used to calculate the risk value. : ; Normalized frequency coefficients ; Normalized source risk weights ; Key coefficients in the normalization stage ; Wherein, F is the frequency coefficient, S is the source risk weight, C is the stage criticality coefficient, H is the spatial heatmap factor, and P is the predictive factor, with the risk level determined according to risk grading standards. The parameters of this dynamic risk assessment engine are periodically encrypted, aggregated, and updated through a horizontal federated learning system participating in the medical consortium blockchain to improve model performance.

[0075] The optimization process of the federated learning collaboration network includes: regularly uploading the risk assessment weights and GNN parameters of the local models deployed at each participating hospital for interruption event identification and risk assessment to the consortium aggregation server after homomorphic encryption; using an aggregation algorithm to aggregate the local models of multiple hospitals to generate a global model; and then encrypting and downloading the global model to the local hospital for updating.

[0076] Specifically, the Paillier additive homomorphic encryption algorithm is used. Each participating hospital generates a public-private key pair locally and uses the public key to encrypt the model parameters (gradients or weights) to generate ciphertext. The server can only perform addition operations on the ciphertext but cannot decrypt it to obtain the original data; the communication protocol uses gRPC or HTTPS for secure transmission.

[0077] The aggregation algorithm is as follows: ; in, Let the local model parameters of the k-th hospital in round t be... Let n be the sample size of the hospital, and n be the total sample size.

[0078] The local model consists of two parts: The BERT-GNN multi-label classification model is a local event recognition and classification engine. It adopts a two-branch architecture: one branch is based on the pre-trained ClinicalBERT model in the medical field, which is used to process unstructured text information such as medical orders and phone recordings to extract semantic features; the other branch is a graph neural network (GNN), which is used to construct a heterogeneous relationship graph of the nursing scene, including nodes such as nurses, patients, equipment, and areas. It enhances the context of event understanding by aggregating spatial and logical relationship information. The features of the two branches are finally fused to output specific interruption reason labels.

[0079] The dynamic risk assessment engine is the core of the local model's quantitative assessment. It receives classified event information and integrates heatmap factors, predictive factors, etc. from the digital twin engine to calculate real-time risk values ​​by enhancing the dynamic risk assessment formula.

[0080] In the federated learning process, this complete composite model, which runs and is trained locally within the hospital, is referred to as the local model. The knowledge it learns, namely the model parameters, is used for collaborative optimization across hospitals.

[0081] The trainable weight parameters of the graph neural network (GNN) branch in the local model. In the BERT-GNN model architecture, the GNN branch is responsible for modeling the graph structure of the care environment. Specifically: First, the nursing scenario is abstracted as a heterogeneous network graph. In the graph, nodes V represent entities such as nurses, patients, equipment, and areas, and edges E represent relationships such as spatial proximity, nursing binding, or communication between them. Each node is associated with features such as coordinates, task type, and equipment status.

[0082] Then, through multi-layered information passing and aggregation operations, this GNN branch allows each node to incorporate information from its neighboring nodes, thereby learning the complex dependencies and contextual features between entities in the graph. This process relies on a large number of adjustable weight matrices and bias terms, forming the GNN parameters. The GNN parameters determine how the model extracts effective features from the spatial and relational topology of the nursing scenario, which is crucial for accurately identifying complex disruption events caused by personnel interactions, spatial proximity, etc. In the parameter aggregation step of federated learning, these weight parameters of each hospital's local GNN branch are encrypted and uploaded to generate a more powerful global graph model.

[0083] S6. Based on the risk level and the nurse's individual status, multimodal adaptive reminders are executed through wearable devices and augmented reality interfaces, and a dynamic task queue is used to prioritize and sequentially guide interrupted tasks and core drug administration tasks.

[0084] Based on the risk level determined in step S5, interventions are implemented in conjunction with personalized state perception of nurses. This step collects heart rate variability and skin conductance responses using wearable devices, and captures facial micro-expressions using AR glasses cameras. Models such as MobileFaceNet are used to infer nurses' fatigue and anxiety levels, and cognitive load is calculated using a task queue to form a dynamic weighted scoring mechanism to select the optimal reminder method. Finally, multimodal adaptive reminders are implemented through augmented reality interfaces, audio-visual devices, and other channels, and dynamic priority and sequential guidance are applied to interrupted tasks and core medication administration tasks to help nurses efficiently handle distractions.

[0085] In a preferred embodiment, a corresponding risk intervention strategy is triggered based on the obtained real-time risk level.

[0086] Low-risk intervention: The system sends non-invasive alerts via gentle vibrations from the nurse's augmented reality glasses or smartwatch, overlaid with soft color blocks on the visual interface. This mild alert aims to help nurses maintain focus and alleviate negative emotions that may accumulate from frequent minor interruptions, thereby improving their coping abilities and safety awareness.

[0087] Intermediate risk intervention: The system automatically activates intelligent communication filtering. When a nurse is performing medication verification, incoming calls and ward calls will be automatically transferred to voicemail or answered by an AI voice assistant, informing the caller that the nurse is in the middle of a critical medication administration procedure and should wait, thus transferring and diverting interruptions. Simultaneously, the task scheduling unit clearly displays all pending tasks in a list within the nurse's field of vision, suggesting a processing order to help nurses transition smoothly from multitasking and reduce cognitive load.

[0088] Advanced Risk Intervention: The system initiates in-depth protection. On one hand, it coordinates with environmental controls, such as projecting red warning light strips around the medication preparation station using smart lighting, or temporarily restricting entry through smart gates, effectively reinforcing the boundaries of the fixed, uninterrupted zone. On the other hand, the system strongly reminds and suggests, or automatically pauses non-urgent paperwork tasks on the nurse's workstation, forcibly switching the visual and operational focus back to the core medication administration process, thus preventing medication errors caused by task switching or distraction.

[0089] Example 1 Intelligent intervention for interruptions in intravenous infusion preparation At 9 a.m. one morning, an N2-level responsible nurse was preparing intravenous medications for several patients in the ward's medication preparation room. The medication preparation room had been set up as a monitored area by the system, and the nurse's tasks included verifying medical orders, dispensing medications, mixing and preparing the medications, and affixing labels.

[0090] 1. Real-time data acquisition Once the nurse begins her procedure, the multi-source sensor network deployed in the dispensing room activates. A wide-angle camera on the ceiling, combined with computer vision algorithms, detects that only the nurse is present in the dispensing room, meeting the initial uninterrupted zone condition. The nurse's smart badge, integrated with ultra-wideband positioning and posture sensors, shows that she is stationary in front of the dispensing station, her attention focused on the medical orders on the computer screen. At this moment, the sensor network captures two concurrent signals: first, the nurses' station phone rings, a signal from the hospital's communication system event stream; second, a sound sensor outside the dispensing room detects a patient's family member loudly inquiring about the examination location via voiceprint recognition. The system integrates these signals in real time, marking them as potential interruption event streams from the system and from others and the environment.

[0091] 2. Intelligent Analysis and Risk Assessment The intelligent analysis module immediately processes the aforementioned data stream. First, it quickly analyzes the family member's inquiry using natural language processing technology, determining it to be a non-emergency medical issue. Simultaneously, based on a temporal behavior analysis model, it confirms that the nurse is currently in a critical stage of medication verification and preparation, a stage with a high error rate. The intelligent analysis module then invokes an interruption cause analysis model, categorizing the incoming call as system communication interference and the family member's inquiry as unnecessary conversation. The risk assessment model, based on historical data, assigns a high weight to the disruptive potential of the incoming call, and considering the current critical operational stage and the double interruption within a short period, the overall risk level is determined to be high.

[0092] 3. Implementation of dynamic risk intervention Since the risk level is high, the risk intervention module immediately executes the combined strategy: First, communication diversion. The communication management unit automatically intercepts the incoming call and transfers it to an AI voice assistant for answering and recording. The assistant informs the caller that the nurse dispensing medication is performing a critical operation and asks them to leave a message or call back later, thus directly diverting and diverting the most disruptive source of the interruption.

[0093] Secondly, enhanced environmental isolation. When the environmental sensing unit is triggered, the smart LED light strip at the entrance of the medication preparation room instantly turns red and flashes slowly, while the electronic display above the door shows the message "Do not disturb during critical operations." Simultaneously, the system uses the nurses' station's broadcast module to gently and specifically address the family area, directing them to the nurses' station for inquiries and indicating that the medication preparation area is temporarily closed. This effectively strengthens the physical and visual boundaries of the uninterrupted area, reducing environmental interference.

[0094] Third, attention guidance. On the augmented reality glasses worn by nurses, the currently configured medication list and patient information window are highlighted with a bright border, and a striking red warning sign appears at the edge of the lens, reminding them to focus on the current check during a high-risk interruption. The task scheduling unit suspends all other non-critical tasks, forcing their cognitive resources to concentrate on the task at hand, preventing task switching or inefficient multitasking due to interruptions, and reducing the risk of medication errors.

[0095] 4. Recording and Optimization All data from this incident was encrypted and uploaded to the security big data platform. The data includes the source of the interruption, the time of occurrence, the nurse's stage of operation at the time, the risk level assessed by the system, the intervention measures taken, and the final result: the nurse successfully completed the configuration without interference and without errors. The post-incident data analysis engine compared this incident with massive amounts of historical data, discovering that the interruption pattern of family members inquiring about non-urgent matters at the pharmacy entrance frequently occurred in the morning and often overlapped with incoming phone calls. Based on this, the system automatically optimized the model, potentially triggering environmental voice prompts earlier or adjusting risk level thresholds in similar time periods and situations in the future. Simultaneously, this case of successfully intercepting calls without incurring complaints will be incorporated into the knowledge base as an effective strategy to continuously improve nurses' response capabilities and safety awareness, and to refine overall safety management strategies.

[0096] Example 2 Interruption Management in Oral Medication Dispensing and Patient Education Scenarios One afternoon at 3 p.m., a nurse brought a mobile nursing cart to the ward to distribute oral medications to patients and planned to conduct medication education. The mobile nursing cart integrates a multi-source sensing module, including positioning tags, an onboard camera, and a communication interface.

[0097] 1. Real-time data acquisition After the nurse entered the ward, she began verifying the patient's identity. The vehicle's camera detected another patient's family member approaching the cart, seemingly inquiring about discharge procedures. Simultaneously, the hospital's information system pushed a new medical order, requiring the immediate discontinuation of a certain oral medication for a patient in the current ward. The nurse's smart bracelet detected a slight increase in her heart rate, reflecting a state of distraction. The system categorized these signals as a multi-source interruption event stream originating from others, the system, and herself.

[0098] 2. Intelligent Analysis and Risk Assessment The intelligent analysis module identified the current stage as bedside verification during medication administration. This stage directly involves matching patient identity with medication and is highly sensitive to risk. The module categorized family inquiries as requests for information from others, changes in medical orders as system updates, and heart rate changes as fluctuations in personal attention. The risk assessment engine determined that changes in medical orders were high-priority system signals, family inquiries were moderate interference, and changes in personal physiological indicators suggested that cognitive load was already at a critical level. A comprehensive assessment determined the current risk level to be moderate.

[0099] 3. Implementation of dynamic risk intervention For intermediate-level risks, the risk intervention module implements three strategies. Regarding communication diversion and information integration, the communication management unit queues family inquiries, displaying a "please wait" message on the mobile nursing cart's screen. Simultaneously, the system pushes new medical orders to the nurse's augmented reality glasses in a highlighted format, displaying only key information, stopping medications, and providing a voice announcement to ensure nurses are immediately informed of critical changes without having to manually check the system. Regarding task scheduling and sequence guidance, the task scheduling unit generates a task list on the nursing cart's screen. The first priority is handling medical order changes, the second priority is completing the current dispensing, and the third priority is responding to family inquiries. The system suggests nurses pause dispensing and return to the nursing station to handle medical order changes, avoiding continuing dispensing tasks with inconsistent information. Regarding environmental alerts, the status indicator light on the mobile nursing cart changes from green to yellow, visually indicating to those around the nurse that they are performing a task requiring focused attention, prompting others to minimize distractions.

[0100] 4. Record keeping and strategy optimization A key characteristic of this incident was the simultaneous occurrence of system information updates and on-site personnel interference. Data analysis revealed that nurses were significantly more likely to delay tasks when changes to medical orders occurred during the medication administration phase compared to other phases. Based on this, the system optimized its task scheduling algorithm, prioritizing task suspension over continuation in such situations. Simultaneously, the system adjusted its family consultation triage strategy, providing richer self-service information on the mobile nursing cart screen to guide families in obtaining answers independently and reducing direct interference with nurses.

[0101] Example 3 Multi-source interruption collaborative processing in emergency situations One evening during the night shift, a nurse was preparing to administer emergency medication to a critically ill patient via intravenous injection in the intensive care unit when the patient's condition changed, triggering an alarm on the monitor.

[0102] 1. Real-time data acquisition Once the nurse entered bedside medication administration mode, the multi-source sensing module immediately enhanced its sensitivity. The monitor alarm signal was captured as a system emergency signal. Simultaneously, a family member of another critically ill patient urgently called the nurse through the call system, stating that their patient was experiencing breathing difficulties. Eye-tracking data from the nurse's smart glasses showed her gaze rapidly switching between the two patients, reflecting a highly stressed decision-making state. Environmental sensors detected abnormal noise from people running in the ward corridor. These signals constituted a high-density multi-source disruption event stream.

[0103] 2. Intelligent analysis and risk assessment.

[0104] The intelligent analysis module determined that the patient was in a critical stage of receiving life-saving medication, and any interruption could have fatal consequences. The system categorized the monitor alarm as a critical system alarm, which has the highest priority and is considered essential information rather than a disruptive interruption. The family call was categorized as an emergency request for help from another person, also carrying a high-risk weight. Eye-tracking data indicated that the caregiver was showing signs of attention fragmentation. The risk assessment engine immediately determined the current risk level to be high and triggered the emergency response mode.

[0105] 3. Implementation of dynamic risk intervention Advanced risk intervention strategies prioritize patient safety and the effectiveness of nursing decision-making. Regarding intelligent routing of emergency communications, the communication management unit identifies family calls as emergency calls for help, without blocking them, but immediately transfers the call information and patient bed number to the nurse's smart glasses, displaying it at the edge of their field of vision with minimal visual interference. Simultaneously, it automatically notifies the on-call nurse to handle the situation, achieving rapid triage of emergency calls. For purifying the resuscitation environment, the environmental perception unit, in conjunction with the corridor lighting system, creates a visual isolation light curtain in the nurse's bed area, prompting other personnel to detour. The ward broadcast system automatically announces that the resuscitation area should be kept quiet to reduce environmental noise interference. In terms of enhanced cognitive support, the personalized reminder unit transmits a low-frequency vibration pattern through a smart bracelet to help nurses maintain a focused rhythm. The augmented reality glasses lock the current patient information interface, blocking non-urgent document task pop-ups. The task scheduling unit only displays tasks directly related to the resuscitation, such as medication injection rate and vital sign monitoring frequency, while all other tasks are temporarily frozen. Regarding team collaboration activation, the system automatically pushes resuscitation event notifications to the head nurse and on-call physician, displaying the current interruption source and nurse workload status, supporting dynamic allocation of team resources.

[0106] 4. Record-keeping and Continuous Improvement Mechanism This incident involved handling multi-source disruptions in a genuine emergency, with data analysis focusing on how the system distinguishes between disruptive disruptions and essential information. Post-incident analysis showed that intelligent routing and team notification functions significantly reduced response time. Based on this, the system optimized its algorithms in emergency mode, enhancing the accuracy of alarm recognition for life support equipment such as monitors and ventilators, and refining the priority determination rules for emergency calls. Simultaneously, cognitive support strategies were optimized, adding respiratory monitoring and stress alerts in high-pressure situations to further ensure the quality of nursing staff's decision-making and patient safety.

[0107] The above three embodiments cover three typical nursing scenarios: routine configuration, bedside distribution, and emergency rescue, respectively, fully demonstrating the adaptability, intelligence, and practicality of the system and method of the present invention under different levels of complexity and risk.

[0108] To scientifically verify the actual effectiveness of risk level intervention measures, this invention designed and implemented a systematic preliminary experimental and clinical simulation test, covering the entire chain of verification from the simulation laboratory to the real clinical environment.

[0109] In the nursing simulation laboratories of three top-tier hospitals, this invention recruited 150 registered nurses from different qualifications and departments to conduct a standardized simulated drug administration task test over a period of three months. The experiment employed a double-blind, randomized controlled design, dividing the nurses into two groups of 75 each. The intervention group received customized alerts based on risk level when the system triggered risk interventions, including low-risk text prompts, medium-risk voice prompts, and high-risk strong prompts; the control group only performed the routine procedures. The simulated tasks rigorously replicated real-world scenarios, including drug verification, preparation, and dispensing, and incorporated typical interruption events such as telephone calls, family inquiries, and system prompts, generating a total of 12,000 simulated operation events.

[0110] Key experimental data showed that the medication error rate in the intervention group, defined as dosage errors, drug confusion, and missed dispensing, significantly decreased from a baseline average of 4.8% to 1.4%, a reduction of 70.8% (p-value less than 0.001). The average interruption response time for nurses decreased from 48.3 seconds to 31.8 seconds, a reduction of 34.2%. Anxiety scores, measured using the Hamilton Anxiety Rating Scale, decreased by 37.5%, indicating that the intervention effectively alleviated task stress.

[0111] In a real clinical setting, this invention analyzed six months of follow-up data. After applying high-risk interventions such as strong reminders to force a change in visual focus, the incidence of medication errors decreased from 2.7% to 1.6%, a reduction of 41.3%. Error types concentrated in high-risk stages such as medication verification saw a reduction of 68.9%. A nurses' safety awareness survey was conducted monthly using standardized questionnaires. Results showed that 85% of nurses reported that the interventions significantly improved process focus, 78% of nurses demonstrated more standardized operating procedures in subsequent assessments, and the completion rate of verification steps increased from 72% to 91%.

[0112] Statistical analysis employed t-tests and chi-square tests, confirming that all results were highly statistically significant (p < 0.01), and that there were no significant differences in baseline characteristics between the experimental and control groups (p > 0.05). Furthermore, the adaptability of the intervention was further validated through simulation of interruption scenarios. In 95% of high-risk scenarios, such as simultaneous changes in medical orders and medication verification, the strong reminder mechanism reduced the time for nurses to refocus their visual attention on the core process to an average of 8.7 seconds, compared to 22.3 seconds in the control group, effectively eliminating medication errors caused by distraction. These data not only demonstrate that the intervention effectively reduces the risk of medication errors but also quantify its positive impact on nurses' coping abilities and safety awareness, with a 34% improvement in response speed and 85% of nurses proactively reporting improvements. In subsequent hospital deployments, this data was incorporated into system optimization iterations to ensure that the intervention strategy is always based on empirical evidence.

[0113] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart reminder and multi-source sensing risk intervention system for drug administration interruption events, characterized in that, include: The multi-source sensing module collects multi-source sensing data in the drug delivery environment in real time through a sensor network, and packages the multi-source sensing data according to its source to form a standardized event stream output. The environmental perception unit deploys ultra-wideband positioning anchors and infrared light curtain detection devices in the drug delivery work area; it monitors the environment by fusing ultra-wideband positioning, infrared light curtain boundary detection and behavior analysis algorithms, and generates a standardized environmental interruption signal containing information such as interference type, location coordinates, duration, intruder identity and heat map risk factors when an interference event is confirmed. The communication management unit manages communication request data during the critical period of drug administration, and uses speech recognition and natural language processing models to identify the urgency of incoming calls and execute dynamic triage strategies. The digital twin engine receives data from the multi-source sensing module, the environmental sensing unit, and the communication management unit, constructs and updates a virtual mapping of the drug delivery environment in real time, dynamically generates a risk heat map, runs simulation prediction to output prediction data, and predicts future high-risk interference windows. The intelligent analysis module receives the predicted data from the digital twin engine, uses a multi-label classification model as the event recognition and classification engine, and calls the dynamic risk assessment engine optimized by the federated learning collaborative network to obtain the risk level. The risk intervention module is used to perform multimodal adaptive reminders through wearable devices and augmented reality interfaces based on the risk level, and uses a dynamic task queue to prioritize and sequentially guide interrupted tasks and core drug administration tasks.

2. The intelligent reminder and multi-source sensing risk intervention system for drug administration interruption events according to claim 1, characterized in that, The sensor network includes: The personnel perception subsystem uses deployed cameras and a computer vision behavior recognition algorithm based on YOLOv5 to automatically identify the behavior of others. At the same time, it monitors the nurse's posture, movement trajectory, operation rhythm and physiological signals through smart name tags or wristbands with integrated inertial measurement units. The system perception subsystem monitors changes in medical orders, patient call lights, and laboratory critical value alarm signals in real time through the hospital information system interface, and integrates the API of the hospital communication system to obtain real-time event streams of incoming telephone calls and internal broadcasts. The environmental sensing subsystem deploys noise and light sensors to monitor changes in ambient noise or lighting.

3. The intelligent reminder and multi-source sensing risk intervention system for drug administration interruption events according to claim 1, characterized in that, The environmental perception unit deploys ultra-wideband anchor points to form an electronic fence, while simultaneously deploying an infrared light curtain to detect unauthorized crossing behavior and using a convolutional neural network to determine the identity of the entrant.

4. The intelligent reminder and multi-source sensing risk intervention system for drug administration interruption events according to claim 1, characterized in that, Multimodal adaptive alerts are triggered via wearable devices and augmented reality interfaces to initiate risk intervention strategies, including: Low-risk intervention: Non-invasive alerts are sent by vibrations from the nurse's augmented reality glasses or smartwatch, overlaid with color-coded prompts on a visual interface; Intermediate risk intervention: Automatically activate intelligent communication filtering to transfer non-emergency calls and ward calls to voicemail or AI voice assistant, while displaying suspended tasks in a list and suggesting processing order through the task scheduling unit; Advanced risk intervention: Initiate deep protection by projecting warning light strips with smart lights or controlling smart gates to restrict entry, and force the visual and operational focus to switch back to the core drug delivery process by reminding or automatically pausing non-emergency task interfaces.

5. The intelligent reminder and multi-source sensing risk intervention system for drug administration interruption events according to claim 4, characterized in that, The digital twin engine uses a stochastic simulation algorithm to predict high-risk interference windows in the future, models the nursing environment as a discrete state process, defines a state vector that includes the number of nurses, patients, and family members in each area, the length of the pending medical orders queue, and the system alarm status, and defines multiple types of response channels corresponding to nursing behavior events.

6. A method for intelligent reminder and multi-source sensing risk intervention for drug administration interruption events, implemented based on the intelligent reminder and multi-source sensing risk intervention system for drug administration interruption events as described in any one of claims 1-5, characterized in that, Includes the following steps: S1: Collect multi-source sensing data in the drug delivery environment in real time through a sensor network, and label and package the multi-source sensing data according to its source to form a standardized event stream output; S2: The drug delivery work area is monitored by the fusion technology of ultra-wideband positioning, infrared light curtain boundary detection and behavior analysis algorithm. When an interference event is confirmed, a standardized environmental interruption signal containing information such as interference type, location coordinates, intruder identity and heat map risk factors is generated. S3: Manage communication request data during critical periods of drug administration, classify incoming voice calls by urgency, and implement dynamic triage strategies based on nurse activity stage tags; S4: Receives standardized event streams, standardized environmental interruption signals, and communication request data; constructs and updates a high-fidelity virtual mapping of the drug delivery environment in real time; dynamically generates a risk heatmap; and runs simulation predictions to output prediction data and predict future high-risk interference windows. S5: Receive the predicted data output from step S4, call the multi-label classification model to identify the interruption event, and input it into the dynamic risk assessment engine optimized by the federated learning collaborative network to calculate the risk level. S6: Based on the risk level, multimodal adaptive reminders are implemented through nurses' wearable devices and augmented reality interfaces, and a dynamic task queue is used to prioritize and sequence interrupted tasks and core drug administration tasks.

7. The intelligent reminder and multi-source sensing risk intervention method for drug administration interruption events according to claim 6, characterized in that, In step S1, the real-time collection of multi-source sensing data in the drug delivery environment through the sensor network includes: automatically identifying the behavior of others through deployed cameras using a computer vision behavior recognition algorithm based on YOLOv5; and monitoring the nurse's posture, movement trajectory, operation rhythm and physiological signals through a smart badge or wristband with an integrated inertial measurement unit worn by the nurse. The system monitors changes in medical orders, patient call lights, and laboratory emergency alarm signals in real time via the hospital information system interface. It also integrates the hospital communication system API to obtain real-time event streams of incoming telephone calls and internal broadcasts. Noise and light sensors are deployed to monitor changes in ambient noise or lighting.

8. The intelligent reminder and multi-source sensing risk intervention method for drug administration interruption events according to claim 6, characterized in that, In step S4, a high-fidelity virtual mapping of the drug administration environment is constructed and updated in real time through a digital twin process. A stochastic simulation algorithm is used to predict high-risk interference windows in the future. The state vector defined by the stochastic simulation algorithm includes: the number of nurses, patients, and family members in each area, the length of the pending medical orders queue, and the system alarm status.

9. The intelligent reminder and multi-source sensing risk intervention method for drug administration interruption events according to claim 6, characterized in that, In step S6, a multimodal adaptive alert is executed through the wearable device and the augmented reality interface to trigger a risk intervention strategy, including: Low-risk intervention: Non-invasive alerts are sent by vibrations from the nurse's augmented reality glasses or smartwatch, overlaid with color-coded prompts on a visual interface; Intermediate risk intervention: Automatically activate intelligent communication filtering to transfer non-emergency calls and ward calls to voicemail or AI voice assistant, while displaying suspended tasks in a list and suggesting processing order through the task scheduling unit; Advanced risk intervention: Initiate deep protection by projecting warning light strips with smart lights or controlling smart gates to restrict entry, and force the visual and operational focus to switch back to the core drug delivery process by reminding or automatically pausing non-emergency task interfaces.

10. The intelligent reminder and multi-source sensing risk intervention method for drug administration interruption events according to claim 6, characterized in that, In step S5, the federated learning collaborative network optimization process includes: periodically uploading the risk assessment weights and GNN parameters of the local models deployed on each participating hospital for interruption event identification and risk assessment to the federation aggregation server after homomorphic encryption; using an aggregation algorithm to aggregate the local models of multiple hospitals to generate a global model; and then encrypting and downloading the global model to the local hospital for updating.

Images