AI-enabled nursing operation intelligent guidance earphone system

Abstract

This invention relates to the field of medical information technology and discloses an AI-enabled intelligent guidance headset system for nursing operations. The system includes a perception and interaction module, a data processing and fusion module, a cognition and decision-making core module, and a model and knowledge base. The system integrates multi-source data such as nurses' verbal descriptions and physiological parameters to construct a unified evidence vector. Through the cognition and decision-making core module, probabilistic inference is performed based on a dynamic Bayesian network model to update the posterior probability distribution of hidden risk states. The core of this system lies in its calculation of the uncertainty quantification value of this distribution and comparison with a preset threshold to determine whether to perform direct guidance or proactive inquiry. In proactive inquiry, the system determines key inquiry information by calculating the expected information gain, forming a closed loop of human-computer interaction. This invention solves the problems of rigid guidance and inability to handle uncertainty in existing technologies, achieving accurate risk identification and dynamic intelligent decision-making, thus improving nursing safety and intelligence levels.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to an AI-enabled intelligent guidance headphone system for nursing procedures. Background Technology

[0002] In clinical nursing, especially in high-risk scenarios such as postoperative monitoring, nurses need to continuously monitor multiple physiological indicators and clinical signs of patients in order to detect potential complications in a timely manner. To assist in this process, existing clinical information systems, bedside monitoring equipment, and some streamlined electronic verification tools have been utilized to some extent. These technologies have played a role in improving data recording efficiency and providing alarms for extreme values ​​of individual indicators.

[0003] However, current assistive technologies still have significant limitations in handling complex clinical situations. The information provided by existing systems is often fragmented and isolated, requiring nurses to integrate data from different devices and combine it with subjective observations, which increases cognitive load and relies on personal experience. At the same time, the guidance logic of these systems is mostly based on fixed rules or procedural processes, lacking the ability to dynamically assess the evolution of risk under the combined effects of multiple variables, and making it difficult to identify early risks pointed to by multiple minor abnormalities.

[0004] More importantly, the current interaction model between technology and nurses is one-way. The system can only passively present data or trigger alarms, but when faced with incomplete information or vague observations from nurses (such as "the patient's complexion is slightly poor"), the system cannot understand this uncertainty, let alone actively ask questions to obtain key information to clarify the ambiguous situation. This lack of interaction prevents the system from deeply participating in the nurse's clinical reasoning process and makes it difficult to provide effective support at critical decision-making points. Therefore, there is an urgent need for a guidance technology that can intelligently integrate multi-source information, understand and quantify uncertainty, and engage in closed-loop interaction with nurses to meet the higher requirements of safety and decision-making quality in modern clinical nursing. Summary of the Invention

[0005] The technical problem that this invention aims to solve is that existing nursing guidance technologies or devices are mostly one-way, procedural information broadcasts, lacking the ability to perceive and understand the dynamic changes in clinical scenarios in real time. Especially when faced with incomplete or ambiguous clinical situations, they cannot provide intelligent decision support and it is difficult to achieve an effective human-computer interaction closed loop to assist nurses in making key judgments.

[0006] To address the aforementioned technical challenges, this invention provides an AI-enabled intelligent guidance headset system for nursing procedures. This system integrates real-time data from multiple sources, dynamically assesses clinical risks through probabilistic reasoning, and introduces a dual-branch interaction strategy based on uncertainty quantification. This allows the system to provide direct instruction guidance when the judgment is clear, and to proactively initiate key information inquiries when the judgment is unclear, thereby forming a closed loop of human-machine collaboration and significantly improving the safety and intelligence level of nursing procedures.

[0007] The first aspect of this invention provides an AI-enabled intelligent guidance headset system for nursing procedures. This system physically includes a computing unit, an intelligent guidance headset terminal, and one or more collaborative sensing devices. The software system running on the computing unit includes:

[0008] The perception and interaction module, serving as the interface between the system and the user, is used to handle audio interactions with the intelligent guide headset terminal, including receiving the user's voice commands and spoken information, and broadcasting the guide commands generated by the system.

[0009] The data processing and fusion module is used to connect and process multi-source heterogeneous data from the sensing and interaction module, collaborative sensing devices, and external information systems, and to construct a uniformly formatted and time-synchronized evidence vector within each preset time slice.

[0010] Models and knowledge bases are used to persistently store the models and knowledge required for the system to run, including dynamic Bayesian network models pre-built for different care scenarios, as well as threshold parameters and rules required for various decisions.

[0011] The core module of cognition and decision-making, as the core processing unit of the system, has the core technical solution of receiving the unified evidence vector and executing a set of logic that includes real-time reasoning, uncertainty quantification and two-branch decision-making.

[0012] Specifically, after receiving the evidence vector, the cognition and decision-making core module first performs probabilistic inference based on a dynamic Bayesian network model loaded from the model and knowledge base that corresponds to the current nursing scenario. This inference calculates or updates the posterior probability distribution of one or more pre-defined, unobservable hidden risk states within the nursing scenario. This inference process can employ a forward algorithm, the core of which is to recursively update the system's belief in hidden risk states based on the following relationship:

[0013] ;

[0014] in, It is a time film The set of state variables It is the deadline slice The sequence of observational evidence, It is a state transition model. It is an observation model.

[0015] After obtaining the posterior probability distribution of the hidden risk state Subsequently, the system needs to assess the clarity of its own judgment. To do this, the module calculates the uncertainty quantification value of the posterior probability distribution. In one embodiment, this uncertainty quantification value is calculated using Shannon information entropy, with the following formula:

[0016] ;

[0017] in, This is the quantified value of the uncertainty we are looking for. It is a state of hidden risk. The total number of possible discrete states, It is its first There are several possible state values. The higher the entropy value, the more uncertain the system is about the current hidden risk state.

[0018] This module compares the calculated uncertainty quantification value with a preset uncertainty threshold loaded from the model and knowledge base, and performs different branching operations based on the comparison result:

[0019] If the uncertainty quantification value is less than or equal to the uncertainty threshold, it indicates that the system's judgment based on the current evidence has a high degree of confidence, and a direct guidance operation is executed. In this operation, the module further compares the posterior probability of the hidden risk state with a preset risk threshold. If it exceeds the risk threshold, a clear risk warning instruction is generated; otherwise, a regular operation guidance instruction is generated.

[0020] If the uncertainty quantification value exceeds the uncertainty threshold, indicating that the system cannot make a reliable judgment, an active inquiry operation is executed. The core of this operation lies in the system's decision on which information to inquire about to most effectively reduce uncertainty. To this end, this module calculates the expected information gain of each piece of evidence in the key missing evidence set. For any potential evidence... The formula for calculating its expected information gain is:

[0021] ;

[0022] in, For the expected information gain, It is the information entropy before the inquiry. It is assumed that new evidence has been observed. The expected conditional entropy is then calculated. The system selects the evidence that maximizes the expected information gain as the optimal inquiry target and generates inquiry instructions based on this target.

[0023] Through the above mechanism, the present invention system constructs a dynamic, closed-loop interactive process. It not only provides guidance but also recognizes the limitations of its own cognitive abilities and compensates for insufficient information through intelligent interaction with the user, thereby making more reliable judgments.

[0024] A second aspect of this invention provides an AI-enabled intelligent guidance method for nursing operations. This method is implemented through the aforementioned system, and its core steps include:

[0025] Continuously collect and process multi-source real-time data, including user voice parsing data and physiological parameter data, to construct a unified evidence vector in each time slice;

[0026] The unified evidence vector is received, and probabilistic inference is performed based on a preset dynamic Bayesian network model to update the posterior probability distribution of the hidden risk state.

[0027] Calculate the quantified value of the uncertainty of the posterior probability distribution;

[0028] The uncertainty quantification value is compared with a preset uncertainty threshold, and a two-branch operation is performed accordingly: if the uncertainty quantification value is greater than the threshold, an active inquiry operation is performed, key missing evidence is determined by calculating the expected information gain and an inquiry instruction is generated; if the uncertainty quantification value is less than or equal to the threshold, a direct guidance operation is performed, risk warning or regular guidance instruction is generated by judging the risk threshold.

[0029] Finally, the generated instructions are converted into voice and broadcast to guide the user's operation.

[0030] This invention provides an AI-enabled intelligent guidance headset system for nursing procedures. It offers the following advantages:

[0031] 1. This invention integrates multi-source heterogeneous data from nurses' oral accounts, collaborative sensing devices, and external information systems into a unified evidence vector through a data processing and fusion module. The cognition and decision-making core module performs real-time probabilistic inference on this vector based on a dynamic Bayesian network model, enabling dynamic quantification and assessment of hidden risk states that cannot be directly observed. When the posterior probability of a hidden risk state exceeds a preset risk threshold, the system automatically generates a risk warning instruction, thereby assisting nurses in timely detection of problems in the early stages of risk occurrence and avoiding delays or oversights that may result from relying solely on manual rounds and subjective judgment.

[0032] 2. The core module of cognition and decision-making in this invention quantifies the uncertainty of the system's own judgment by calculating the Shannon information entropy of the posterior probability distribution. When the quantified uncertainty value exceeds a preset uncertainty threshold, the system will perform an active inquiry operation, determining and proposing key questions that can most effectively reduce uncertainty by calculating the expected information gain. This mechanism transforms the system from a one-way instruction broadcaster into an intelligent collaborative partner capable of perceiving its own cognitive limitations and actively seeking clarification, effectively solving the problem of decision-making difficulties when information is ambiguous or incomplete in clinical practice.

[0033] 2. This invention uses time slices as units to continuously receive evidence vectors and update its understanding of the scenario. Each guidance or inquiry decision in the core cognitive and decision-making module is the result of comprehensive reasoning based on all available evidence up to the current moment. This makes the guidance process of this invention dynamic and data-driven, capable of adaptively adjusting the content and direction of the interaction according to real-time changes in the patient's vital signs and immediate feedback from nurses, thereby providing truly personalized guidance tailored to the specific clinical situation. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the functional architecture of an AI-enabled intelligent guidance earphone system for nursing operations according to an embodiment of the present invention;

[0035] Figure 2 This is a flowchart illustrating an AI-enabled intelligent guidance method for nursing operations according to an embodiment of the present invention.

[0036] The module consists of: 100, Perception and Interaction Module; 200, Data Processing and Fusion Module; 300, Cognition and Decision-Making Core Module; and 400, Model and Knowledge Base. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0038] See attached document Figure 1 , Figure 1This is a functional architecture diagram of an AI-enabled intelligent guidance headset system for nursing procedures according to an embodiment of the present invention. The system provided by the present invention physically includes a computing unit, an intelligent guidance headset terminal communicatively connected to the computing unit, and one or more collaborative sensing devices communicatively connected to the computing unit. The computing unit may be a server deployed in the cloud or an edge computing server deployed within a medical institution. The intelligent guidance headset terminal integrates a microphone array, an audio output unit, and a wireless communication module. The collaborative sensing devices include wearable sensors for collecting patient physiological data or IoT conversion modules for acquiring the status of medical devices.

[0039] Based on the aforementioned hardware architecture, the functional architecture of the system of the present invention, namely the software system running on the computing unit, may include: a perception and interaction module 100, a data processing and fusion module 200, a cognition and decision-making core module 300, and a model and knowledge base 400.

[0040] The perception and interaction module 100 serves as the interface for the system to interact with the user and the environment. Its input end is connected to the microphone array of the intelligent guidance headset terminal to receive the nurse's voice commands and verbal information; its output end is connected to the audio output unit of the intelligent guidance headset terminal to broadcast guidance instructions to the nurse. This module integrates functional units for speech recognition, natural language understanding, and speech synthesis, used to convert external unstructured speech information into internally processable structured data, and to convert text commands generated internally into speech output.

[0041] The data processing and fusion module 200 connects to the sensing and interaction module 100, collaborative sensing devices, and an external hospital information system. This module receives real-time data streams from multiple heterogeneous sources, including structured text data from the sensing and interaction module 100, continuous physiological parameter data from the collaborative sensing devices, and patient medical record data from the hospital information system. The module performs timestamp alignment, data cleaning, and format normalization on the received multi-source data, and ultimately constructs a unified evidence vector containing all currently available information within each preset time slice for use by subsequent modules.

[0042] The cognition and decision-making core module 300 is the core processing unit of the system of this invention. Its input end is connected to the data processing and fusion module 200, and its internal logic processing relies on the model and knowledge base 400. Based on the received evidence vector, the cognition and decision-making core module 300 performs real-time probabilistic inference to dynamically update the posterior probability distribution of preset hidden risk states in the nursing scenario. Further, this module calculates the information entropy of the inference result to quantify its uncertainty, and based on the comparison between the uncertainty value and a preset threshold, decides to implement either direct guidance or proactive inquiry strategies. When implementing the proactive inquiry strategy, this module determines the key information that needs to be inquired from the user by calculating the expected information gain and generates corresponding inquiry instructions.

[0043] The Model and Knowledge Base 400 is a persistent storage unit bidirectionally connected to the Cognition and Decision-Making Core Module 300. It stores pre-built dynamic Bayesian network models for different nursing scenarios, including network topology and conditional probability parameters. Furthermore, the Model and Knowledge Base 400 also stores various threshold parameters, clinical rules, and knowledge templates for generating guidance instructions and inquiry questions required for system operation. The Cognition and Decision-Making Core Module 300 needs to load the corresponding models and parameters from this library when performing reasoning and decision-making.

[0044] See attached document Figure 2 , Figure 2 This is a flowchart illustrating an AI-enabled intelligent guidance method for nursing procedures according to an embodiment of the present invention. The present invention provides an intelligent guidance method for nursing procedures, which may include the following steps:

[0045] S201: Receive the nursing task initiation command and, based on the nursing scenario identified by the command, load the corresponding dynamic Bayesian network model from the model and knowledge base 400. This step involves the perception and interaction module 100 receiving the nurse's initial voice command and passing the parsed task information to the cognition and decision-making core module 300, which then loads the model.

[0046] S202, continuously collect multi-source real-time data, and process and fuse the data to construct an evidence vector for each time slice. This step is performed by the data processing and fusion module 200, and the multi-source real-time data includes speech parsing data from the perception and interaction module 100, physiological parameter data from collaborative sensing devices, and patient data from external information systems.

[0047] S203 receives the evidence vector and performs probabilistic inference based on the loaded dynamic Bayesian network model to update the posterior probability distribution of one or more pre-defined hidden risk states in the care scenario. This step is performed by the cognition and decision-making core module 300.

[0048] S204, calculate the information entropy of the posterior probability distribution to obtain a quantified value of the uncertainty of the reasoning result for the hidden risk state. This step is performed by the cognitive and decision-making core module 300.

[0049] S205, compare the uncertainty quantification value with a preset uncertainty threshold, and perform different branch operations based on the comparison result. If the uncertainty quantification value is less than or equal to the uncertainty threshold, proceed to step S206; if the uncertainty quantification value is greater than the uncertainty threshold, proceed to step S207.

[0050] S206, execute the direct guidance operation. This operation includes: determining whether the posterior probability of the hidden risk state exceeds a preset risk threshold; if so, generating a risk warning instruction; otherwise, generating a regular operation guidance instruction. The generated instruction is ultimately output in voice form through the perception and interaction module 100.

[0051] S207, Perform an active inquiry operation. This operation includes: determining a key missing evidence that can minimize uncertainty by calculating the expected information gain; generating an inquiry command based on the key missing evidence; and outputting it in speech form through the perception and interaction module 100.

[0052] In step S208, the system receives the user's voice feedback on the interrogative command, parses the feedback information, and incorporates it as new evidence into the evidence vector for the next time slice. It then returns to step S203. This step constitutes a closed loop in human-computer interaction, reducing the uncertainty of reasoning by actively acquiring information until the system can perform direct guidance operations.

[0053] To specifically implement the functions of this invention, the perception and interaction module 100 may include an upstream data processing unit, a voice information parsing unit, and an instruction synthesis and broadcasting unit in a concrete implementation. The upstream data processing unit acquires the raw audio signal from the microphone array of the smart guide headset terminal and performs signal enhancement and purification processing. The voice information parsing unit receives the processed audio signal and converts it into structured intent and entity data. The instruction synthesis and broadcasting unit receives text instructions from other modules of the system, converts them into voice signals, and outputs them through the headset audio unit.

[0054] The upstream data processing unit is designed to collect nurses' voices from the medical environment and suppress environmental noise and interference to the maximum extent possible, outputting a high-quality audio stream suitable for subsequent speech recognition. To achieve this, the unit integrates processing logic such as multi-microphone beamforming, acoustic echo cancellation, and environmental noise suppression.

[0055] In one embodiment, the unit utilizes multiple microphones physically distributed on the smart guide headset terminal to form a microphone array. For the multiple audio signals acquired by this microphone array, the beamforming unit performs spatial filtering processing to enhance the target speech signal from the direction of the nurse's mouth while attenuating interfering sound sources from other directions. One specific implementation is delay-summation beamforming. The output signal of this method... It can be expressed by the following formula:

[0056] ;

[0057] in, Indicates time, It represents the total number of microphones in the microphone array. It is the microphone index. It is the first Each microphone in time The collected signals. It is applied to the first The gain weight of each microphone signal can be used to adjust the sensitivity differences between different microphones or to form a specific spatial response pattern. It is applied in the first step to align the time difference of sound waves from the target sound source arriving at different microphones. The delay on the microphone signal is calculated based on the geometry of the microphone array and the estimated direction of the sound source.

[0058] Since the system of this invention broadcasts voice commands through headphones, to prevent the microphone from picking up these broadcast sounds and mistaking them for the nurse's input, the upstream data processing unit also includes an acoustic echo cancellation unit. This unit receives the reference audio signal that the system is about to broadcast and, based on the relationship between this signal and the actual signal acquired by the microphone, establishes an adaptive acoustic echo path model. During system operation, this unit uses this model to predict the echo components in the microphone signal and subtracts them from the original acquired signal, thereby eliminating interference from the system's own output to the input.

[0059] Even after beamforming and echo cancellation, the audio signal may still contain background noise that cannot be removed by spatial filtering, such as the continuous buzzing sound from medical devices or diffuse noise in the environment. Therefore, the upstream data processing unit further includes a noise suppression unit. This unit can employ spectral subtraction or a deep learning-based noise reduction algorithm. Taking spectral subtraction as an example, this unit first estimates and smoothly updates the power spectrum of the noise signal during the inactive intervals of the speech signal. Then, during the speech activity, each frame of the audio signal is transformed to the frequency domain, the estimated noise power spectrum is subtracted from the signal's power spectrum, and the result is inversely transformed back to the time domain, thereby achieving the purpose of suppressing steady-state background noise. For the specific implementation of speech recognition and natural language understanding in this unit, those skilled in the art can use existing deep learning models or services; their specific architecture and training methods are well-known technologies in the field and will not be elaborated upon here.

[0060] The high-quality audio stream output from the upstream data processing unit is fed into the speech information parsing unit. The core function of this unit is to further parse the unstructured text sequence converted by the speech recognition (ASR) engine into structured data that the system can understand. This process specifically includes two parallel tasks: intent recognition and entity extraction.

[0061] In one specific embodiment, the speech information parsing unit employs a joint model based on a pre-trained language model to simultaneously perform intent recognition and entity extraction. The input to this joint model is the text sequence output by the ASR engine. The output of the model is the intent classification result and the entity label sequence.

[0062] Intent recognition is treated as a text classification task. The system predefines a set of intents, which includes all possible intent categories that nurses may express during nursing operations, such as initiating a nursing task, reporting patient observations, querying medical orders, and confirming completion of an operation. The joint model includes an intent classification pathway, which maps the overall semantic representation of the input text sequence to the predefined set of intents and outputs a confidence score for each intent category. The category with the highest score is selected as the final intent recognition result.

[0063] Entity extraction is treated as a sequence labeling task. The system predefines a set of entity labels, which contains key information types in the nursing scenario, such as: patient identification (e.g., bed number), operation items (e.g., dressing change), observation indicators (e.g., skin color), and observation results (e.g., paleness). The joint model includes an entity extraction pathway, which analyzes each lexical in the input text sequence and assigns it an entity label. A specific labeling system is the BIO (Begin, Inside, Outside) format, where B - entity type indicates the beginning of an entity, I - entity type indicates the middle or end of an entity, and O indicates that the lexical does not belong to any entity. For example, for the text "Change the drainage tube for bed 15", the label sequence output by the entity extraction pathway should be "O B-patient identification I-patient identification B-operation item I-operation item I-operation item".

[0064] To achieve the above functionality, the underlying layer of this joint model can employ a shared pre-trained language model, such as the BERT architecture. The input text sequence is first processed by this shared model to generate word-level vector representations containing rich contextual information. Subsequently, an intent classification head and an entity extraction head are connected to the output of this shared model. The intent classification head uses the vector representations of specific words (such as the [CLS] word) to compute the probability distribution of intent categories through one or more fully connected layers and a Softmax activation function. The entity extraction head then inputs the sequence of vector representations of all words into a linear layer and a Conditional Random Field (CRF) layer. The introduction of the CRF layer allows the model to learn the transition constraints between entity labels (e.g., an I-patient identifier label is more likely to be preceded by a B-patient identifier label than a B-operation item label), thereby improving the accuracy of entity boundary recognition.

[0065] Finally, the voice information parsing unit combines the identified intent and extracted entities into a structured data object, such as a JSON-formatted object, and then passes it to the data processing and fusion module 200 for use in subsequent steps.

[0066] When the cognition and decision-making core module 300 or other modules of the system generate instructions that need to be conveyed to nurses, these instructions are sent in the form of structured text objects to the instruction synthesis and broadcasting unit in the perception and interaction module 100. This unit is used to convert this text information into high-quality speech and broadcast it through the audio output unit of the intelligent guidance headset terminal.

[0067] In one embodiment, the instruction synthesis and broadcasting unit receives a structured text object containing not only the text string of the content to be broadcast but also priority metadata of the instruction. This unit includes an instruction scheduling unit that maintains a priority queue. When a new instruction is received, the instruction scheduling unit inserts it into the appropriate position in the queue based on its priority metadata. High-priority instructions, such as emergency risk warnings, are placed at the head of the queue, interrupting currently broadcast low-priority instructions; while routine guidance instructions enter the queue sequentially to await broadcast. This mechanism ensures that the most critical information is delivered first.

[0068] After retrieving the instruction text to be played from the instruction queue, a text-to-speech (TTS) engine converts it into an audio waveform. To ensure the naturalness and clarity of the played speech, this invention employs a two-stage deep learning TTS architecture that includes an acoustic model and a vocoder.

[0069] The first stage of this architecture is the acoustic model, which transforms the input text sequence into an intermediate acoustic representation, such as a Mel spectrogram. This acoustic model can be a sequence-to-sequence network based on an attention mechanism. The encoder part of this network encodes the input character or phoneme sequence into a hidden state sequence containing contextual information; the decoder part, guided by the attention mechanism, generates Mel spectrograms frame by frame, where the attention mechanism ensures that the generated spectrogram frames are temporally aligned with the corresponding parts of the input text.

[0070] The second stage of this architecture is the vocoder, which synthesizes the Mel spectrograms generated by the acoustic model into the final audible audio waveform. To generate high-fidelity speech, this vocoder can be a stream-based generative model or an autoregressive model. The vocoder uses the Mel spectrogram as conditional input, directly modeling and generating from the sampled points of the original audio waveform. Speech synthesized in this way shows significant improvements in both sound quality and naturalness compared to traditional concatenation or parametric synthesis methods.

[0071] For the specific model training method of this text-to-speech engine, those skilled in the art can use publicly available large-scale speech datasets for end-to-end training. The specific network structure and optimization method are well-known technologies in this field and will not be elaborated here.

[0072] Finally, the digital audio waveform signal generated by the vocoder is converted from digital to analog and then played by the audio output unit of the smart guide earphone terminal, thus completing a complete voice guidance broadcast process.

[0073] The data processing and fusion module 200 is designed to process and integrate multi-source, heterogeneous, and asynchronous data streams generated by upstream modules and external systems into structured information with a unified format and synchronized time, enabling the cognition and decision-making core module 300 to perform effective reasoning. In specific implementations, this module may include a multi-source data access unit, a data preprocessing unit, and a unified evidence vector construction unit.

[0074] The multi-source data access unit is used to establish communication links with different data sources and parse the incoming data. For data from the sensing and interaction module 100, this unit receives the structured data objects output by the module through its internal interface. For data from collaborative sensing devices, this unit receives data packets via wireless communication methods such as Bluetooth Low Energy and Wi-Fi, according to the device's communication protocol, and parses the specific physiological parameter values ​​according to preset protocol specifications. For data from the hospital information system, this unit obtains the patient's static or semi-static medical record information through a secure API interface. To ensure the accuracy of subsequent processing, this unit associates a timestamp with each received data item to ensure the comparability of data from different sources in the time dimension.

[0075] After receiving the raw data with synchronization timestamps, the data preprocessing unit cleans and normalizes it. This preprocessing unit includes a data cleaning unit to handle missing and outlier values. For short-term data loss due to signal loss or transmission errors, this unit can employ different interpolation strategies based on the data type. For example, linear interpolation is used for physiological data with relatively stable changes (such as body temperature), while for highly volatile data (such as heart rate), previous value imputation is used for short-term missing values. For outliers that exceed the normal physiological range or exhibit drastic changes compared to neighboring data points, this unit can use moving window-based statistical methods, such as the 3-sigma principle, for detection, and treat the identified outliers as missing values.

[0076] The data preprocessing unit also includes a data normalization unit, used to convert data of different types and dimensions into a uniform format. For continuous numerical data (e.g., heart rate, skin temperature), this unit uses the min-max normalization method to linearly map it to an interval. The transformation formula is as follows:

[0077] ;

[0078] in, These are the original measurements. and These are the lower and upper limits of the normal range of variation for this physiological parameter, as preset in the clinical knowledge base. This is a normalized numerical value. For discrete, qualitative data (such as "skin color" dictated by a nurse), this unit uses one-hot encoding to vectorize it. For example, if "skin color" has three possible values: "normal," "pale," and "flushed," then "pale" will be converted into a three-dimensional vector. In this way, all types of input information are converted into a uniform numerical format.

[0079] The preprocessed data is then fed into a unified evidence vector construction unit, which assembles all available information into a fixed-dimensional evidence vector within a preset discrete time slice. Using 300 as the core module of cognition and decision-making at a given time point Input.

[0080] This unit operates on a fixed time period (e.g., every 5 seconds), which defines a time slice. At the end of each time slice, the unit collects and processes all preprocessed data items that arrived within that time period. For continuous data sources (such as heart rate sensors) that may generate multiple data points within a single time slice, the unit employs an aggregation strategy to generate a unique representative value for that time slice. One specific aggregation strategy is to use the latest valid value received within that time slice as the observation value of the variable at that moment. For event-based data (such as a nurse's single verbal observation), if it occurred within the time slice, its corresponding encoded value is set to valid; otherwise, it is set to the default state.

[0081] This evidence vector The dimensions and structure are predefined, corresponding to the set of observable evidence variables defined by the dynamic Bayesian network model in the core module 300 of cognition and decision-making. Strict correspondence. Each dimension or set of dimensions in the vector uniquely maps to a specific evidence variable. For example, the first dimension of the vector might correspond to the normalized heart rate value, the second dimension to the normalized skin temperature value, and the third to fifth dimensions might collectively form a one-hot encoded vector to represent the observed skin color.

[0082] To ensure the robustness of the system, this unit also includes a missing value handling mechanism. If, within a certain time slice, no valid observation data is received for a certain expected evidence variable, the unit will not leave the corresponding dimension of the vector empty. One specific approach is to use a previous value imputation strategy. That is, if, within the current time slice... If the value of a variable cannot be observed, then that variable is in the evidence vector. The value in will be inherited from the previous time slice. Vector of evidence The value in the system. In the initial stage of system startup, if no historical values ​​are available, the preset clinical neutral value is used to fill the gap.

[0083] Through the aggregation, mapping, and padding processes described above, this unified evidence vector construction unit generates a complete evidence vector at the end of each time slice. This vector comprehensively represents the state of the external world that the system can perceive at that moment. After its generation, the evidence vector is immediately transmitted to the cognition and decision-making core module 300 to drive the next round of probabilistic reasoning and decision-making.

[0084] The cognition and decision-making core module 300 is the core information processing module of the system of this invention, used to perform dynamic scene understanding, risk assessment, and interactive decision-making. This module receives a unified evidence vector from the data processing and fusion module 200 and outputs specific instructions to drive the perception and interaction module 100. In specific implementations, this module may include a real-time reasoning unit, an uncertainty quantification unit, an active inquiry strategy unit, and a guidance instruction generation unit.

[0085] The real-time inference unit functions by inferring, in real-time, pre-defined, and unobservable hidden risk states within a nursing scenario based on a continuously input sequence of evidence vectors. This unit receives data from the data processing and fusion module 200 at each time slice. Generated evidence vector It performs probabilistic inference calculations based on a dynamic Bayesian network model loaded from the model and knowledge base 400, which is specific to the current nursing scenario.

[0086] The goal of this reasoning process is to obtain the complete sequence of observational evidence up to the current moment. Under the given conditions, calculate the set of hidden state variables. The posterior probability distribution, also known as the belief state of the system, is used. In one embodiment, the real-time inference unit employs a forward algorithm to recursively update this belief state. Each iteration of the algorithm consists of two steps: prediction and update. Its core computational process can be described by the following proportional relationship:

[0087] ;

[0088] In this formula:

[0089] Representative in time slice At that time, it is the set of all state variables (including hidden variables and evidence variables) in the model.

[0090] Represents the time slice from the initial time slice 1 to the current time slice. The evidence vector observation sequence.

[0091] It is a set of state variables The specific combination of values ​​in the previous time slice.

[0092] In the previous time segment The calculated posterior probability distribution, i.e. the belief state of the previous moment, is used as the input for this recursive calculation.

[0093] It is a transition model defined by a dynamic Bayesian network, representing the change in the system state from the previous time step. Evolves to the value at the current time. The probability of.

[0094] It is an observation model defined by a dynamic Bayesian network, representing the current system state as... Under these conditions, evidence was observed. The probability of.

[0095] This indicates that the left-hand side term is proportional to the right-hand side term. After calculation, normalization is required to ensure that the sum of probabilities is 1.

[0096] Specifically, the unit first performs a prediction step, which involves taking the values ​​of all possible states from the previous time step. Perform weighted summation The unit calculates the prior probability distribution of the system state at the current time slice, without considering any new evidence. Then, it performs an update step, utilizing the new evidence from the current time slice. and observation model The prior probability is then corrected to obtain the joint posterior probability distribution at the current time. .

[0097] After calculating the joint posterior probability distribution, the real-time inference unit can obtain the marginal posterior probability distributions of any one or more hidden risk states of interest by performing marginalization calculations on this distribution. For example, for hidden risk states... Its final state of belief is The result will be output to the uncertainty quantification unit for subsequent decision-making.

[0098] The posterior probability distribution calculated by the real-time inference unit is passed to the uncertainty quantification unit. This unit provides the system with a numerical value regarding the confidence level of the current inference result, enabling the system to self-assess the clarity of its understanding of the scene.

[0099] In one specific embodiment, the uncertainty quantification unit achieves this function by calculating the Shannon information entropy of the posterior probability distribution. For any hidden risk state... In fact, the output of the real-time inference unit is its current state based on all evidence. Given the posterior probability distribution, if the hidden state has... Mutually exclusive discrete values Then its information entropy It can be calculated using the following formula:

[0100] ;

[0101] In this formula:

[0102] This is the quantized value of the uncertainty we are looking for, in bits.

[0103] It is the index of the hidden state value.

[0104] It is a state of hidden risk. The total number of possible discrete states.

[0105] yes The There are 10 possible state values.

[0106] In a given sequence of evidence Under the conditions, The state is The posterior probability is derived directly from the calculation results of the real-time inference unit.

[0107] According to the conventions of information theory, when At that time, the item The value is 0.

[0108] The calculated entropy value provides a direct basis for the system's subsequent decision-making logic. A higher entropy value means that the posterior probability distribution is relatively uniform, that is, multiple possible states have a non-negligible probability. In this case, the system's judgment of the hidden state has a high degree of uncertainty. Conversely, a lower entropy value (close to 0) means that the probability quality is concentrated on one or a very few states, indicating that the system can determine the hidden state with a high degree of confidence based on the current evidence.

[0109] Finally, the unit outputs the calculated entropy value, which represents the degree of uncertainty of the current reasoning result, to the active inquiry strategy unit and the guidance instruction generation unit as key inputs for their decision-making.

[0110] The entropy value output by the uncertainty quantification unit is fed into the proactive inquiry strategy unit. When this entropy value exceeds the preset uncertainty threshold in the model and knowledge base (400), the proactive inquiry strategy unit is activated. Its function is to proactively identify key information that can most effectively reduce this uncertainty when the system's judgment of the current scenario has high uncertainty, and decide to initiate an inquiry targeting this information.

[0111] In one specific embodiment, the unit achieves this function by calculating and comparing the expected information gain of different potential pieces of evidence. The unit first obtains a pre-defined set of key missing evidence for the current nursing scenario from the model and knowledge base 400. This set includes all evidentiary variables that the system believes can be obtained by actively asking nurses questions and that have a significant impact on risk assessment, such as the patient's skin color and whether they complain of dyspnea.

[0112] For this set Each potential evidence variable in This unit simulates the situation after acquiring the evidence and calculates the expected reduction in information entropy, i.e., the information gain. This addresses the hidden risk state of the target. Regarding potential evidence Expected information gain The calculation formula is:

[0113] ;

[0114] In this formula:

[0115] It is the posterior entropy calculated by the uncertainty quantification unit based on all currently known evidence.

[0116] Based on the known current evidence Under these conditions, new evidence was observed. After that, in hidden state The expected conditional entropy. The calculation of this term requires... The weighted average of all possible values ​​is expanded as follows:

[0117] ;

[0118] in:

[0119] Potential evidence variables The set of all possible values. For example, if If the color is "skin color", then the set is {normal, pale, flushed}.

[0120] Based on currently known evidence, regarding the evidence to be observed. The value is The predicted probability. This probability can be obtained by performing a prediction step on the current belief state.

[0121] It is a hypothetical posterior entropy. This unit simulates the interpretation of observations. Add it to the evidence set and rerun the inference update to obtain a new posterior probability distribution. Then, its information entropy is calculated based on this distribution.

[0122] This proactive inquiry strategy unit will traverse the set of key missing evidence. The unit examines all variables and calculates the expected information gain for each. Then, it selects the evidence variable that yields the maximum information gain. The target of this investigation is selected through a process that can be represented as follows:

[0123] ;

[0124] In this way, the system strategically selects the question most likely to clarify the current ambiguity. Ultimately, the unit will use the selected optimal inquiry target evidence variable... The identifier is output to the boot instruction generation unit.

[0125] The guidance instruction generation unit in the cognition and decision-making core module 300 gathers the processing results from the aforementioned units, including the posterior probability distribution output by the real-time inference unit, the entropy value output by the uncertainty quantification unit, and the optimal inquiry target output by the proactive inquiry strategy unit under specific conditions. This unit is the final executor of the system's decision-making logic, used to determine the interaction strategy to be adopted at the current moment based on these inputs and generate specific instruction text.

[0126] The core of this unit is a two-branch decision-making logic, which is based on the entropy value calculated by the uncertainty quantification unit. Compared with the preset uncertainty threshold loaded from the model and knowledge base 400 The comparison results.

[0127] If the entropy value is less than or equal to the uncertainty threshold ( If the system's judgment of the current scenario is high, then the unit executes a direct guidance operation. In this operation branch, the unit further examines the posterior probability of the hidden risk state calculated by the real-time inference unit. For a specific risk state... High-risk values The unit will use its posterior probability With a preset risk threshold Compare. If If a significant clinical risk is identified, the system will generate a risk warning instruction. If the posterior probability of all monitored risk states does not exceed their corresponding risk thresholds, the system considers the current operation to be on track and can generate a routine operation procedure guide or status confirmation instruction.

[0128] If the entropy value is greater than the uncertainty threshold ( If the system's judgment of the current scenario is uncertain, it cannot provide reliable direct guidance. In this case, the unit performs an active inquiry operation. In this operation branch, the unit receives the optimal inquiry target evidence variable that maximizes information gain, determined by the active inquiry strategy unit. Based on this target variable, the unit generates an exploratory instruction aimed at guiding the nurse to observe and provide feedback on this key information.

[0129] Whether generating warnings, routine guidance, or inquiry instructions, this unit employs a template-based text generation method. The model and knowledge base (400) store instruction templates for different instruction types and specific scenarios. These templates are text strings containing placeholders. For example, a risk warning template might be: "Note that patient [bed number] may have [risk name], please check their [related vital signs] immediately." An inquiry instruction template might be: "Please describe the patient's [inquiry target]." Based on the current decision, the guidance instruction generation unit selects an appropriate template from the knowledge base and uses specific information from the current context (e.g., bed number, risk status name, target variable name obtained from the voice information parsing unit, etc.) to fill the placeholders in the template, thereby generating the instruction text.

[0130] Finally, the unit packages the generated instruction text and its priority metadata into a structured instruction object and sends it to the instruction synthesis and broadcasting unit in the perception and interaction module 100 to complete the final voice output.

[0131] The model and knowledge base 400 can physically be one or more databases or file systems deployed on a computing unit. Functionally, this module may include a model repository and a rules and knowledge repository. The model repository is used to persistently store dynamic Bayesian network models pre-built for different care scenarios. The rules and knowledge repository is used to store structured knowledge such as various threshold parameters, instruction templates, and key evidence sets required for system operation.

[0132] The dynamic Bayesian network model used in the cognitive and decision-making core module 300 of this invention was constructed offline before system deployment. This construction process ensures the clinical validity of the model and the authenticity of the data, and specifically includes two stages: model structure definition and model parameter learning.

[0133] In the model structure definition phase, the goal is to establish a graph model topology that accurately reflects the causal and relational relationships among variables in a nursing scenario. This structure consists of a set of nodes representing relevant variables and a set of directed edges representing direct dependencies between variables.

[0134] In one embodiment of the invention, the definition of this structure is primarily constructed based on domain expert knowledge. Specifically, researchers collaborate with senior clinical nursing experts and physicians, using authoritative clinical practice guidelines and medical literature as a foundation, to first identify key variables in a specific nursing scenario (e.g., postoperative bleeding risk monitoring). These variables include: hidden risk states that need to be assessed (e.g., bleeding risk level), evidence variables that can be directly or indirectly observed by the system (e.g., heart rate, blood pressure, drainage fluid color, patient complaints), and operational variables that nurses may perform. Subsequently, experts, based on their clinical experience and medical knowledge, determine the causal relationships between these variables, which are represented in the graph as directed edges from cause nodes to result nodes.

[0135] After the model structure is determined, the model parameter learning phase begins. The goal of this phase is to quantify the dependencies defined in the model, that is, to assign specific values ​​to the conditional probability table (CPT) of each node in the network. Each node's CPT defines the probability distribution of each possible value of that node given any set of values ​​taken by its parent node. To accomplish this step, this invention employs a statistical learning method based on real-world data. Specifically, a large amount of anonymized historical clinical datasets related to the target nursing scenario are extracted from information systems such as electronic medical records in medical institutions. Since these datasets often contain missing data, and the hidden risk states themselves are not directly labeled, this invention uses the expectation-maximization (EM) algorithm to learn the model parameters. The specific implementation of this EM algorithm, which estimates parameters by iteratively executing the E-step (expectation step) and M-step (maximization step), is a well-known technique in the field and will not be elaborated upon here.

[0136] Through the above structural definition and parameter learning, the system constructs a dedicated dynamic Bayesian network model for each preset nursing scenario. These trained models are serialized and stored in a model repository along with their corresponding scenario identifiers, so that the cognition and decision-making core module 300 can quickly load them based on the scenario identifiers when receiving task instructions.

[0137] In addition to the probabilistic model, the rule and knowledge repository in the model and knowledge base 400 are used to store the deterministic knowledge and configuration parameters required for system operation. This knowledge and parameters provide specific quantitative basis and behavioral framework for the judgment and decision-making of the cognitive and decision-making core module 300.

[0138] In one embodiment, the rules and knowledge repository includes a threshold parameter management unit. This unit stores a series of key numerical thresholds associated with specific care scenarios in a structured form, such as a database table or configuration file. Specifically, this unit stores uncertainty thresholds used to trigger different decision logics. The threshold is set based on offline simulation testing combined with assessments by clinical experts, aiming to strike a balance between the reliability of the system's inference results and the proactivity of the interaction. This unit also stores the corresponding risk alarm threshold for each monitorable hidden risk state. This threshold defines the warning system that must be generated when the posterior probability of a high-risk value for a given risk state exceeds this value. Furthermore, this unit also stores the upper and lower limits of the normal value range required for normalization of continuous physiological data in the data processing and fusion module 200. and .

[0139] The rules and knowledge repository also includes an instruction template management unit. This unit stores text templates used to generate various voice instructions. These templates are categorized according to their purpose, primarily including risk warning templates, routine guidance templates, and proactive inquiry templates. Each template is a text string containing one or more placeholders, which are populated with specific dynamic information when the instruction is generated. For example, a proactive inquiry template might be stored as "Please confirm the [inquiry target] status of patient [bed number]." By decoupling the logical structure of the instructions from the specific wording, the system allows clinical staff to easily customize and optimize the guidance scripts without modifying the core program code.

[0140] Furthermore, this rule and knowledge repository are also used to manage contextual knowledge related to proactive inquiry strategies. Specifically, for each care scenario, the repository contains a pre-defined set of key missing evidence. This set explicitly lists the most valuable observational variables to actively explore in a specific scenario when the system faces high uncertainty. For example, in a postoperative bleeding risk monitoring scenario, this set might include variables such as "drainage fluid color" and "dressing exudation." During system runtime, the active exploration strategy unit loads the set of evidence matching the current scenario from this repository and calculates the expected information gain within the scope of this set.

[0141] These thresholds, templates, and rule sets, stored in the rules and knowledge repository, are dynamically loaded into the corresponding processing units when the system initializes a specific care task, thereby ensuring the accuracy of system decisions and scenario adaptability.

[0142] Example:

[0143] This embodiment uses the routine care of a patient who underwent right knee replacement surgery on the first day after surgery as an example to illustrate the specific application process of the intelligent guidance system of the present invention.

[0144] Scene setting:

[0145] Nurse Li, wearing the intelligent guided headset terminal of this invention, prepares to conduct the first postoperative lower limb neurovascular function assessment for Mr. Wang in bed 15 of the orthopedic ward. This assessment is a crucial step in preventing serious postoperative complications such as deep vein thrombosis and compartment syndrome.

[0146] 1. Task Initiation and Model Loading

[0147] Nurse Xiao Li initiated the task via voice: "Start postoperative neurovascular assessment for bed 15."

[0148] The upstream data processing unit of the perception and interaction module 100 acquires the speech, and after beamforming and noise reduction, sends it to the speech information parsing unit. This unit recognizes the speech as text and jointly parses it to find that the intent is "start nursing task", the entity is "bed number: 15" and "task type: postoperative neurovascular assessment".

[0149] The analysis results are passed to the cognitive and decision-making core module 300. Based on the task type "postoperative neurovascular assessment", this module loads a dynamic Bayesian network model, corresponding decision thresholds, and instruction templates that match this scenario into the instruction model and knowledge base 400.

[0150] Simultaneously, the data processing and fusion module 200 begins operation. Its multi-source data access unit obtains basic information about the patient in bed 15 (the type of surgery is right knee replacement) from the hospital information system via API, and begins receiving real-time skin temperature data transmitted from a collaborative sensing device (wireless skin temperature patch) pre-placed on the dorsum of the patient's right foot.

[0151] 2. Initial guidance and data fusion

[0152] After the system completes initialization, the cognition and decision-making core module 300 generates the first routine guidance instruction.

[0153] The guidance instruction generation unit selects a standard guidance template from the knowledge base and generates the instruction text: "Please first observe the skin color of the patient's right lower limb foot."

[0154] The text was sent to the perception and interaction module 100, converted into speech by the instruction synthesis and broadcasting unit, and broadcast to nurse Xiao Li through headphones.

[0155] After observing the patient, Nurse Li verbally reported, "The skin color of the right foot of patient 15 is slightly pale." At the same time, the data processing and fusion module 200 continuously received and processed the skin temperature data, and found that the skin temperature of the affected limb was consistently 0.8 degrees Celsius lower than that of the healthy side.

[0156] This module performs one-heat encoding on "skin color: pale" and integrates the normalized skin temperature data to construct a unified evidence vector for the current time slice. The data is then transmitted to the cognitive and decision-making core module 300.

[0157] 3. Proactive inquiry triggered by uncertainty

[0158] The real-time reasoning unit of the Cognition and Decision-Making Core Module 300 receives evidence vectors. The model performs inference based on a dynamic Bayesian network model. The posterior probability of the hidden risk state "neurovascular damage" in the model has increased, but has not yet reached the risk threshold that would trigger a high-level alarm. .

[0159] However, the description "slightly pale" is subjectively vague, and combined with a persistent slight decrease in skin temperature, this leads to a more dispersed posterior probability distribution of the "neurovascular damage" state in the model. The information entropy calculated by the uncertainty quantification unit... Exceeded the preset uncertainty threshold .

[0160] High uncertainty activated the active inquiry strategy unit. This unit retrieved a set of key missing evidence for the scenario from the knowledge base, including "capillary refill time" and "dorsalis pedis artery pulsation status." By calculating the expected information gain, the system determined that obtaining the objective indicator of "capillary refill time" could most effectively reduce the uncertainty in the current assessment of the "neurovascular damage" state.

[0161] The guidance instruction generation unit receives the optimal query target "capillary refill time" and generates the query instruction: "For further evaluation, please check and report the capillary refill time of the patient's right toe." This instruction is then broadcast to the nurse.

[0162] 4. Risk Identification and Clear Early Warning

[0163] Following instructions, nurse Xiao Li pressed on the nail bed of the patient's right big toe, observed, and reported: "Capillary refill time is greater than 3 seconds."

[0164] The perception and interaction module 100 analyzes the speech and extracts the key information "capillary refill time > 3 seconds".

[0165] The data processing and fusion module 200 encodes this information (e.g., encodes it as a numerical value representing "anomaly") and constructs a new evidence vector. .

[0166] When the new evidence vector is input into the cognitive and decision-making core module 300, the real-time inference unit performs a new round of inference updates. Since "refill time greater than 3 seconds" is a strong anomalous indicator, the posterior probability of the hidden risk state "neurovascular damage" increases sharply, significantly exceeding the risk threshold. At the same time, as judgments become clearer, information entropy... A significant decrease.

[0167] The decision logic of the instruction generation unit determines that the high-risk warning conditions have been met. This unit selects a high-priority risk warning template from the knowledge base and generates a clear instruction with action suggestions: "High-risk warning: The patient's right lower limb may have neurovascular damage. Please immediately notify the on-duty physician and check if the right lower limb bandage is too tight, while also checking the dorsalis pedis artery pulsation."

[0168] The high-priority instruction was immediately broadcast through the headset, ensuring that nurse Xiao Li could take the most critical intervention measures at the first opportunity.

[0169] Through the above embodiments, the system of the present invention can not only provide routine process guidance in postoperative care of orthopedic patients, but more importantly, it can dynamically integrate multi-source information, actively explore to obtain decision-making evidence when uncertainty arises, and finally provide timely and clear warnings and action guidance when high-risk conditions are identified, thereby effectively assisting nurses in preventing the occurrence of serious complications.

[0170] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-enabled intelligent guidance headphone system for nursing procedures, characterized in that: include: One or more collaborative sensing devices; Intelligent guide headphone terminal; as well as The computing unit includes: The perception and interaction module is used to receive user voice information collected by the intelligent guide earphone terminal and convert system-generated instructions into voice for broadcast. The data processing and fusion module is used to receive and process speech parsing data from the perception and interaction module, physiological parameter data from the collaborative sensing device, and patient data from an external information system, in order to construct a unified evidence vector at each time slice. The cognition and decision-making core module is used to receive the unified evidence vector, perform probabilistic inference based on a preset dynamic Bayesian network model to update the posterior probability distribution of the hidden risk state, calculate the uncertainty quantification value of the posterior probability distribution, and decide whether to perform a direct guidance operation or an active inquiry operation based on the comparison result of the uncertainty quantification value and a preset uncertainty threshold. A model and knowledge base are used to store the dynamic Bayesian network model and the uncertainty threshold. The core module for cognition and decision-making is specifically used for: The uncertainty quantification value is obtained by calculating the Shannon information entropy of the posterior probability distribution; If the uncertainty quantification value is greater than the uncertainty threshold, then the active inquiry operation is performed; If the uncertainty quantification value is less than or equal to the uncertainty threshold, then the direct guidance operation is performed; When the cognitive and decision-making core module performs the active inquiry operation, it is specifically used for: Obtain a pre-defined set of key missing evidence from the model and knowledge base; The optimal target evidence is determined by calculating the expected information gain of each piece of evidence in the set for reducing the Shannon information entropy of the posterior probability distribution. Based on the optimal evidence for the inquiry target, an inquiry instruction is generated.

2. The AI-enabled intelligent guidance earphone system for nursing operations according to claim 1, characterized in that, When the cognitive and decision-making core module performs the direct guidance operation, it is specifically used for: The posterior probability of the hidden risk state is compared with a preset risk threshold; If the posterior probability is greater than the risk threshold, a risk warning instruction is generated; If the posterior probability is less than or equal to the risk threshold, then a normal operation guidance instruction is generated.

3. The AI-enabled care operation intelligent guidance earphone system according to claim 1, wherein, The data processing and fusion module includes: A multi-source data access unit is used to establish communication with different data sources and associate timestamps with the received data. The data preprocessing unit is used to perform data cleaning and normalization on the received data. The unified evidence vector construction unit is used to assemble all preprocessed data items into a fixed-dimensional evidence vector corresponding to the set of observable evidence variables of the dynamic Bayesian network model within a preset discrete time slice.

4. The AI-enabled care operation intelligent guidance earphone system according to claim 1, wherein, The perception and interaction module includes: The upstream data processing unit is used to perform beamforming, acoustic echo cancellation, and noise suppression processing on the raw audio signal obtained from the microphone array of the smart guide headphone terminal. The speech information parsing unit is used to perform intent recognition and entity extraction on the processed audio signal in order to convert it into structured speech parsing data.

5. The AI-enabled intelligent guidance earphone system for nursing operations according to claim 1, characterized in that, The perception and interaction module also includes: The instruction synthesis and broadcasting unit includes an instruction scheduling unit and a text-to-speech engine. The instruction scheduling unit is used to queue the received text instructions according to the instruction priority metadata. The text-to-speech engine is used to convert the text instructions into audio waveforms and broadcast them through the audio output unit of the smart guide headset terminal.

6. The AI-enabled intelligent guidance earphone system for nursing operations according to claim 1, characterized in that, The model and the dynamic Bayesian network model stored in the knowledge base are constructed offline in the following way: The topology is based on a domain expert knowledge definition model, which includes hidden risk state nodes, observable evidence variable nodes, and directed edges representing the dependencies between them. Based on historical clinical datasets, the expectation-maximization algorithm is used to learn the conditional probability table parameters of each node in the model.

7. The AI-enabled intelligent guidance earphone system for nursing operations according to claim 1, characterized in that, The model and knowledge base also store instruction templates for generating instructions, as well as a set of key missing evidence for the active inquiry operation; the cognitive and decision-making core module generates the instructions by filling in the instruction templates.

8. An AI-enabled nursing operation intelligent guidance method based on the AI-enabled nursing operation intelligent guidance earphone system according to any one of claims 1-7. Includes the following steps: Receive and process multi-source real-time data to construct a unified evidence vector in each time slice; The unified evidence vector is received, and probabilistic inference is performed based on a preset dynamic Bayesian network model to update the posterior probability distribution of the hidden risk state. Calculate the quantified value of the uncertainty of the posterior probability distribution; The uncertainty quantification value is compared with a preset uncertainty threshold. If the uncertainty quantification value is greater than the uncertainty threshold, an active inquiry operation is performed to generate an inquiry instruction. If the uncertainty quantification value is less than or equal to the uncertainty threshold, a direct guidance operation is performed to generate a risk warning instruction or a regular operation guidance instruction. The generated instructions are converted into speech and read aloud.

Images