Chest surgery nursing data intelligent management method based on biological feature recognition

Abstract

The invention relates to the technical field of medical data intelligent management, and discloses a chest surgery nursing data intelligent management method based on biological feature recognition. The method comprises the following steps of: performing hierarchical decoupling and time synchronization on a multi-source monitoring data stream of a patient, and separating out a biological characteristic signal, a nursing event text and a physiological waveform substream; for each substream characteristic, dynamic baseline establishment, semantic anchor point embedding based on biological characteristic context and key waveform fragment characterization processing are respectively carried out. The method comprises the following steps: mining an implicit association mode among substreams by cooperatively analyzing a cross-verification mechanism of a network, and generating an association mode matrix; and performing backtracking causal inference on the basis of the matrix, identifying a causal chain among nursing events, sign changes and physiological abnormalities, and finally constructing a personalized nursing intervention scheme. According to the method, deep semantic fusion and accurate causal association analysis of multi-modal data are realized, and the pertinence and reliability of nursing decision are enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical data management technology, specifically to an intelligent management method for thoracic surgery nursing data based on biometric recognition. Background Technology

[0002] In thoracic surgery clinical nursing scenarios, patient monitoring information exhibits typical multi-source heterogeneous characteristics, mainly including continuous physiological waveforms from devices, text logs recorded by nurses, and signals from various biosensors. Current technologies typically manage these data streams separately. Physiological waveforms trigger independent alarms based on thresholds set by the monitors, text logs are only used for recording and archiving, and biometric data are compared using fixed standards. There is a lack of effective collaboration between these systems, resulting in data that is coarsely synchronized in time and fragmented at the semantic level.

[0003] This approach has limitations. Independent analysis cannot establish an effective correlation between nursing procedures, patient complaints, and changes in objective vital signs. The system struggles to automatically determine whether a single "analgesia" record truly corresponds to the relief of heart rate abnormalities, or to identify which textual descriptions of nursing procedures are frequently accompanied by specific perturbations in the blood oxygen waveform. Early warnings rely on simple triggering of single-dimensional rules, leading to frequent false alarms and missed alarms. The formulation and adjustment of nursing interventions are often based on fragmented information and clinical experience, lacking a coherent chain of causal evidence automatically derived from multimodal data to provide precise support.

[0004] A method is needed to fundamentally change the paradigm of multi-source data fusion analysis. The core issues this method needs to address are: how to interpret unstructured text records within a dynamic physiological context to obtain semantic representations that include biological characteristics; and how to construct an analysis mechanism that enables deep cross-verification and correlation mining of text, waveforms, and signals, thereby automatically discovering clinically significant cross-modal patterns and generating interpretable, personalized nursing intervention logic based on these patterns. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent management method for thoracic surgery nursing data based on biometric recognition, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides an intelligent management method for thoracic surgery nursing data based on biometric recognition, the method comprising:

[0007] Raw monitoring data streams of thoracic surgery patients are received from multiple nursing terminals. These raw monitoring data streams integrate the patients' real-time physiological waveforms, nursing event text logs, and biometric signals collected by biosensors.

[0008] The original monitoring data stream is subjected to layered decoupling processing to separate the biometric signal sub-stream, the nursing event text sub-stream, and the physiological waveform sub-stream, and timestamp synchronization and alignment are performed on each sub-stream;

[0009] For the biometric signal substream, dynamic vital sign baselines are established and updated; for the nursing event text substream, semantic anchor embedding based on biometric context is performed; and for the physiological waveform substream, key waveform segments are clipped and characterized.

[0010] The processed biometric signal sub-stream, nursing event text sub-stream, and physiological waveform sub-stream are input into a collaborative analysis network. The collaborative analysis network uses a cross-verification mechanism to mine the implicit correlation patterns between different sub-stream data and outputs a correlation pattern matrix.

[0011] Based on the aforementioned association pattern matrix, a retrospective causal inference process is initiated, which is used to identify the causal chain between nursing events, changes in biometrics, and abnormal physiological waveforms.

[0012] Based on the structure and strength of the causal chain, a personalized nursing intervention plan is constructed, which includes specific intervention points and parameter adjustment suggestions.

[0013] Preferably, the semantic anchor embedding process based on biometric context is performed on the nursing event text substream, including:

[0014] Each text record in the nursing event text substream is segmented and entity recognition is performed to extract entities related to nursing procedures, medications, and vital sign descriptions.

[0015] From the established biometric knowledge graph of thoracic surgery nursing, retrieve the set of biometric concepts that have a predefined association with the entity;

[0016] The retrieved set of biometric concepts is fused with the original text records to generate an enhanced text representation that includes the context of biometric features.

[0017] The enhanced text representation is subjected to deep semantic encoding, which transforms the unstructured text record into a structured semantic feature vector, wherein the semantic feature vector encodes the semantic relationship between the nursing event and its associated biometric features.

[0018] The semantic feature vectors are arranged in chronological order to form a semantic sequence of nursing events.

[0019] Preferably, the step of establishing and updating a dynamic vital sign baseline for the biometric signal substream includes:

[0020] The biometric signal substream is segmented by a sliding window to obtain continuous time window signal segments;

[0021] For each time window signal segment, its statistical characteristics, frequency domain characteristics, and nonlinear characteristics are calculated to form the feature vector of the time window;

[0022] The feature vector of the current time window is compared with the feature vector sequence of the historical window, and an adaptive threshold is used to determine whether the current time window is in a stable state.

[0023] The individualized baseline model of vital signs is updated by incremental learning using only the feature vectors of the time window determined to be in a stable state.

[0024] Using the updated biometric baseline model, the biometric signal sub-stream is monitored in real time, and the deviation sequence between the current signal and the dynamic baseline is calculated. The deviation sequence quantifies the instantaneous abnormality of the biometrics.

[0025] Preferably, the collaborative analysis network uses a cross-verification mechanism to uncover implicit correlation patterns between different sub-stream data, including:

[0026] Receive the semantic sequence of the nursing event, the deviation sequence, and the characteristic representation of the key waveform segment;

[0027] Within the collaborative analysis network, a temporal correlation mapping is established between the semantic sequence of nursing events and the deviation sequence, and the change pattern of biometric deviation within a preset time window after each nursing event occurs is calculated.

[0028] Establish a state-correlation mapping between deviation sequences and physiological waveform features, and analyze the distribution of physiological waveform features corresponding to different levels of biological characteristic deviations;

[0029] A bidirectional attention network is used to enable the interaction of nursing event semantics, biometric deviation and physiological waveform features, and to calculate the mutual influence weight between any two.

[0030] Based on the mutual influence weights, a three-dimensional association pattern matrix is ​​constructed. Each dimension of the association pattern matrix corresponds to the nursing event type, biometric deviation level, and physiological waveform abnormality pattern, respectively. The value of the matrix element represents the association strength of the co-occurrence of the three.

[0031] Preferably, the step of initiating a retrospective causal inference process based on the association pattern matrix includes:

[0032] Elements with association strength exceeding the significance threshold are extracted from the association pattern matrix and used as candidate strong association triples.

[0033] For each strongly associated triple, retrieve the specific time point and context data corresponding to the strongly associated triple in the original data stream;

[0034] Using the specific time point of occurrence as the endpoint, a preset inference time window is traced back, and within the inference time window, all possible preceding care events and changes in biometrics are retrieved;

[0035] A temporary causal graph is constructed, with abnormal physiological waveform patterns in the strongly correlated triples as result nodes and retrieved preceding nursing events and changes in biometrics as candidate cause nodes.

[0036] Based on the chronological order and the pre-learned association strength in the association pattern matrix, directed edges are established between candidate cause nodes and result nodes, and the causal contribution score of each edge is calculated.

[0037] The cause node with the highest causal contribution score and its path pointing to the result node are selected as an initial causal chain.

[0038] Preferably, the retrospective causal inference process further includes the verification and integration of the initial causal chain:

[0039] Merge and analyze multiple coexisting initial causal chains to identify common causal nodes or intermediate nodes in different chains;

[0040] Based on the existence of common nodes, multiple independent initial causal chains can be connected or merged to form a more complete composite causal chain that may contain multiple intermediate links.

[0041] Search the historical nursing database for case records that are similar to or opposite to the described complex causal chain;

[0042] Using the retrieved case records, the logical rationality and statistical significance of the composite causal chain are externally verified;

[0043] The verified composite causal chain, along with its causal contribution score and verification support, will be used as the final output causal chain.

[0044] Preferably, constructing a personalized nursing intervention plan that includes specific intervention points and parameter adjustment suggestions includes:

[0045] The causal chain of the final output is analyzed to identify nursing events or biometric states that are interventionable nodes in the chain;

[0046] For each interventionable node, a set of standardized nursing measures that can positively or negatively affect the interventionable node is queried from the nursing knowledge base;

[0047] Calculate the overall causal contribution score for each causal chain, and rank all causal chains according to the scores to determine the priority of intervention;

[0048] For high-priority causal chains, select the measures that are feasible and have the least expected conflict from the set of standardized nursing measures corresponding to the interventionable nodes;

[0049] The selected nursing measures are linked with specific intervention timing and execution parameters to form independent intervention sub-plans targeting the causal chain;

[0050] All independent intervention sub-plans were arranged chronologically according to the timing of intervention and conflict checks were performed to integrate them into a complete personalized nursing intervention plan document.

[0051] Preferably, the step of performing key waveform segment clipping and feature representation for the physiological waveform sub-stream includes:

[0052] Real-time monitoring of physiological waveform sub-streams; when the waveform amplitude, frequency or morphology exceeds the preset normal range, the detected moment is marked as the starting point of a critical event.

[0053] Taking the starting point of the key event as the center, a fixed-duration waveform data segment is extracted before and after it to form a key waveform segment;

[0054] Multi-scale filtering is performed on the key waveform segments to separate waveform components with different frequency components;

[0055] Calculate the time-domain statistical characteristics, frequency-domain energy distribution characteristics, and morphological descriptive characteristics of each waveform component;

[0056] The features of all waveform components are spliced ​​and normalized to form the feature representation vector of the key waveform segment.

[0057] Preferably, after constructing a personalized nursing intervention plan that includes specific intervention points and parameter adjustment suggestions, the plan also includes implementation and closed-loop update steps:

[0058] The personalized nursing intervention plan is pushed to the corresponding nursing terminal, and the execution status and result records of each measure in the plan are monitored;

[0059] After the intervention measures are implemented, the patient's original monitoring data stream continues to be collected, and the hierarchical decoupling process, dynamic baseline update, semantic anchor embedding, collaborative analysis and causal inference process are repeated.

[0060] The newly inferred causal chain is compared with the causal chain that generated the current intervention plan to evaluate the effectiveness of the intervention and detect whether any new anomalous causal patterns have emerged.

[0061] Based on the effectiveness evaluation results and newly emerging causal patterns, the personalized nursing intervention plan is dynamically adjusted and incrementally updated to form a closed-loop management system.

[0062] Preferably, the dynamic adjustment and incremental update of the personalized nursing intervention plan includes:

[0063] When the evaluation results show that the intervention is effective, strengthen the parameters of the intervention in the program or extend it to treatments with similar causal patterns;

[0064] When the assessment results show that the intervention measures are ineffective or new abnormal causal patterns emerge, the program revision process is initiated.

[0065] The revised plan process includes: analyzing the reasons for the ineffectiveness of measures or the emergence of new models, searching the nursing knowledge base and historical case database for alternative or supplementary measures, and conducting simulation and evaluation of the revised plan in a simulated environment;

[0066] The revised content, based on the deductive evaluation, will be updated to the personalized care intervention plan, and the causal context of this revision will be recorded to form a historical log of the plan's evolution.

[0067] Compared with the prior art, the beneficial effects of the present invention are:

[0068] By performing semantic anchor embedding based on biometric context on the substream of nursing event text, the time-synchronized text description is jointly encoded with the concurrent biometric time-series signals. This process generates a dynamic semantic vector representation that not only internalizes the inherent semantics of the text but also deeply integrates the patient's physiological state characteristics and trends at the corresponding moment in the text event. The nursing event is thus transformed into a high-dimensional data point carrying specific physiological context information, achieving deep fusion of unstructured text and quantitative physiological signals in a unified representation space. This lays a directly computable foundation for subsequent accurate cross-modal analysis based on the essential correlation of data.

[0069] The cross-verification mechanism built into the collaborative analysis network enables the discovery of associations among multi-source substream data. This mechanism establishes a dynamic interaction pathway at the feature level of the neural network, enabling iterative bidirectional attention calculation and feature modulation between physiological waveform features, biological signal patterns, and textual semantic vectors. This design forces the discovery of association patterns to be based on consistent support across multiple independent data sources, rather than isolated statistical features of a single data stream. The resulting association pattern matrix thus filters out spurious associations caused by noise or sporadic fluctuations in single-stream data, revealing cross-modal patterns with higher robustness and clinical interpretability.

[0070] A retrospective causal inference process is initiated based on a cross-validated association pattern matrix. This process constructs a directional causal hypothesis chain by analyzing the temporal and statistical dependencies between patterns. The personalized nursing intervention plan generated based on this causal chain is logically derived directly from a validated "event-response" causal path derived from the patient's individual historical data. The intervention goals, timing, and parameter adjustment suggestions in the plan strictly correspond to the key nodes and quantitative relationships in the causal chain, thereby realizing the transformation of nursing decision support from general empirical recommendations to individualized causal derivation. Attached Figure Description

[0071] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent management method for thoracic surgical nursing data based on biometric recognition as described in this invention.

[0072] Figure 2 This is a flowchart of semantic anchor embedding processing based on biometric context;

[0073] Figure 3 A flowchart for verifying and integrating the initial causal chain;

[0074] Figure 4 A diagram illustrating the effectiveness of conflict resolution in nursing interventions;

[0075] Figure 5 This is a simulation diagram showing the dose-physiological parameters and risk prediction of sufentanil. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Please see Figure 1This invention provides an intelligent management method for thoracic surgery nursing data based on biometric recognition. The method includes: receiving raw monitoring data streams from thoracic surgery patients from multiple bedside nursing terminals. These data streams integrate real-time physiological waveforms from multi-parameter monitors, nursing event text logs entered by nurses via mobile terminals, and biometric signals collected by wearable biosensors. The raw monitoring data streams are layered and decoupled, separated into three independent data channels based on data source and format characteristics: a biometric signal sub-stream, a nursing event text sub-stream, and a physiological waveform sub-stream. Data points in each sub-stream are timestamped and aligned using a unified time server to ensure temporal consistency in subsequent correlation analysis. For the separated biometric signal sub-stream, dynamic vital sign baselines are established and updated to obtain individualized real-time status references.

[0078] For the nursing event text sub-stream, semantic anchor embedding based on biometric context is performed to transform unstructured text into a structured sequence rich in medical semantics. For the physiological waveform sub-stream, key waveform segments are cropped and characterized to capture clinically significant waveform abnormalities. The processed data from the three sub-streams are then input into a pre-trained collaborative analysis network. This network, through a built-in cross-verification mechanism, mines implicit association patterns between different modalities and outputs a quantified association pattern matrix. Based on this association pattern matrix, the system initiates a retrospective causal inference process aimed at identifying potential causal chains between nursing events, biometric changes, and physiological waveform abnormalities. Finally, based on the structure and quantification strength of the identified causal chains, the system constructs a personalized nursing intervention plan that includes specific intervention points, timing of implementation, and parameter adjustment suggestions, and pushes it to the nursing terminal.

[0079] Example 1: See Figure 2 In the processing of the nursing event text substream, each text record in the substream undergoes word segmentation and entity recognition based on a medical dictionary to extract entities related to nursing procedures, drug names, and physical sign descriptions. From a locally or cloud-deployed thoracic surgery nursing biometric knowledge graph, using the extracted entities as query keys, a set of biometric concepts with predefined associations to each entity is retrieved. This set includes related physiological parameters, symptoms, or signs. The retrieved set of biometric concepts is fused with the original text records to generate an enhanced text representation containing biometric context. Deep semantic encoding is performed on the enhanced text representation, using a pre-trained language model to transform unstructured text records into structured semantic feature vectors. These semantic feature vectors encode the semantic relationships between nursing events and their associated biometrics. The resulting semantic feature vectors are arranged according to the timestamps of their corresponding text records to form a nursing event semantic sequence.

[0080] In the processing of biometric signal substreams, the biometric signal substreams are segmented into continuous time window signal segments by a fixed-duration sliding window. For each time window signal segment, its mean and variance are calculated, its frequency domain features are calculated using Fast Fourier Transform, and nonlinear features such as sample entropy are calculated, collectively forming the feature vector of that time window. The feature vector of the current time window is compared with the feature vector sequence of historical windows, and an adaptive threshold is used to determine whether the current time window is in a physiologically stable state. Only the feature vectors of the time windows determined to be in a stable state are used to update the baseline model of vital signs describing the normal state of the individual patient through incremental learning. Using the updated baseline model of vital signs, the subsequent input biometric signal substreams are monitored in real time, and the deviation between the current signal features and the corresponding features of the dynamic baseline is calculated, forming a deviation sequence that quantifies the instantaneous abnormality of biometrics.

[0081] In the processing of physiological waveform substreams, the substreams are monitored in real time. When waveform amplitude, frequency, or morphological indicators exceed preset individualized normal ranges, the detected moment is marked as the starting point of a critical event. A fixed-duration segment of waveform data is extracted before and after this critical event starting point to form a critical waveform segment. Multi-scale filtering is applied to the critical waveform segment to separate waveform components of different frequencies. The time-domain statistical characteristics, frequency-domain energy distribution characteristics obtained through power spectral density estimation, and morphological descriptive features based on the waveform envelope are calculated for each waveform component. All features of the waveform components are concatenated and normalized to form the characteristic representation vector of the critical waveform segment.

[0082] In a specific implementation, an example scenario is postoperative monitoring of thoracic surgery patients. The original monitoring data stream received by the nursing terminal includes physiological waveform sub-streams such as electrocardiogram waveforms and blood oxygen saturation waveforms, a nursing event text sub-stream consisting of the text record "administer furosemide 10mg intravenous bolus" entered by the nurse, and biometric signal sub-streams such as pulse wave signals collected by the fingerprint sensor. The data comparison is reflected in the fact that the text is transformed from an unstructured description to a structured vector before and after processing, the biometric signals are transformed from a raw sequence to a deviation sequence, and the physiological waveforms are transformed from a continuous stream to a characteristic representation vector. The nursing event text substream processing includes word segmentation and entity recognition based on a thoracic surgery medical dictionary for each text record in the nursing event text substream. Entities related to nursing procedures, medications, and vital signs are extracted, such as "furosemide," "intravenous injection," and "blood pressure drop." A set of biometric concepts with predefined associations to these entities is retrieved from an established thoracic surgery nursing biometric knowledge graph. For example, "furosemide" is associated with "urine output," "blood pressure," and "heart rate." The retrieved set of biometric concepts is fused with the original text record to generate an enhanced text representation containing biometric context, such as "Administer 10mg furosemide intravenously [associated biometrics: urine output, blood pressure, heart rate]." Deep semantic encoding is performed on the enhanced text representation. A pre-trained clinical language model is used to transform unstructured text records into structured semantic feature vectors. These semantic feature vectors encode the semantic relationships between nursing events and their associated biometrics. The semantic feature vectors are arranged chronologically to form a nursing event semantic sequence.

[0083] In some embodiments, biometric signal substream processing includes dividing the biometric signal substream into continuous time window signal segments by a sliding window of fixed duration of 60 seconds; calculating the mean, variance, and sample entropy for each time window signal segment; calculating the frequency domain features to form the feature vector of the time window by fast Fourier transform; comparing the feature vector of the current time window with the feature vector sequence of historical windows; determining whether the current time window is in a stable state based on an adaptive threshold; updating the individualized vital sign baseline model using only the feature vector of the time window determined to be in a stable state through incremental learning; using the updated vital sign baseline model to monitor the biometric signal substream in real time; and calculating the deviation sequence of the current signal from the dynamic baseline. In a specific implementation, when using the updated vital sign baseline model to monitor the biometric signal substream in real time, the system first divides the real-time input signal into a sliding window to obtain the current time window signal segment, and calculates the statistical features, frequency domain features, and nonlinear features of the segment to form the current feature vector. Subsequently, the system compares the current feature vector with the baseline feature vector stored in the dynamic baseline model in real time. It calculates the deviation value between the two using a predefined difference metric function. This function synthesizes a scalar result based on the differences in each dimension of the feature vector, thereby generating a deviation sequence that quantifies the instantaneous abnormality of the biometric. The deviation sequence quantifies the instantaneous abnormality of the biometric; data comparison is represented by the processed raw pulse wave signal output as a time-varying sequence of deviation values.

[0084] In some embodiments, physiological waveform substream processing includes real-time monitoring of the physiological waveform substream. When the waveform amplitude, frequency, or morphology exceeds a preset normal range, the detected moment is marked as the critical event start point. A fixed-duration waveform data segment is extracted before and after the critical event start point to form a critical waveform segment. Multi-scale filtering is performed on the critical waveform segment to separate waveform components of different frequencies. The time-domain statistical characteristics, frequency-domain energy distribution characteristics, and morphological description characteristics of each waveform component are calculated. The features of all waveform components are spliced ​​and normalized to form a characteristic representation vector of the critical waveform segment. In the example scenario, when the central electrical waveform shows ST segment elevation, it is marked as the critical event start point. The clipped critical waveform segment is separated into high-frequency and low-frequency components by multi-scale filtering. The data comparison is reflected in the transformation of the original waveform segment into a vector representation containing multi-dimensional features.

[0085] It is understandable that the semantic sequence of nursing events generated by the nursing event text substream processing, the deviation sequence generated by the biometric signal substream processing, and the feature representation vector generated by the physiological waveform substream processing will serve as inputs to the subsequent collaborative analysis network. The key difference lies in the transformation of each substream from raw heterogeneous data into structured time-series data in a unified format. It is also understandable that the adaptive threshold for determining a stable state in the biometric signal substream processing can be dynamically adjusted based on historical data distribution, and the preset normal range in the physiological waveform substream processing can be personalized based on the individual patient's baseline. These parameter adjustments do not affect the execution of core steps.

[0086] Example 2: Collaborative analysis networks uncover implicit correlation patterns between different substream data through cross-verification mechanisms. The network receives the semantic sequence of nursing events, the biometric deviation sequence, and the characteristic representations of key segments of physiological waveforms generated by preprocessing. Within the collaborative analysis network, a temporal correlation mapping is established between the semantic sequence of nursing events and the deviation sequence. The change pattern of biometric deviation within a preset time window after each nursing event is calculated, such as the peak value and duration of the deviation. A state correlation mapping is established between the deviation sequence and physiological waveform features, analyzing the distribution of physiological waveform features corresponding to different levels of biometric deviation. A bidirectional attention network enables the interaction of the semantic features of nursing events, the biometric deviation features, and the physiological waveform features, calculating the mutual influence weight between any two. Based on the calculated mutual influence weights, a three-dimensional correlation pattern matrix is ​​constructed. One dimension of this matrix corresponds to the nursing event type, another dimension corresponds to the biometric deviation level, and the third dimension corresponds to the physiological waveform abnormality pattern. The value of each element in the matrix represents the correlation strength of the co-occurrence of a specific type of nursing event, a specific level of biometric deviation, and a specific physiological waveform abnormality pattern.

[0087] In practical implementation, an example scenario involves monitoring postoperative blood pressure fluctuations and arrhythmias in thoracic surgery patients. The collaborative analysis network receives the semantic sequence of nursing events, the biometric deviation sequence, and the characteristic representations of key segments of physiological waveforms generated from previous steps. Data comparison is reflected in the network input being three heterogeneous time-series data streams, while the output is a quantified three-dimensional correlation pattern matrix. Internally, the collaborative analysis network establishes a temporal correlation mapping between the semantic sequence of nursing events and the deviation sequence, calculates the change pattern of biometric deviation within a preset time window after each nursing event, such as calculating the mean and integral area of ​​pain index deviation within 30 minutes after the "administration of analgesics" nursing event, establishes a state correlation mapping between the deviation sequence and physiological waveform features, and analyzes the distribution of physiological waveform features corresponding to different levels of biometric deviation, such as analyzing the characteristic distribution of ECG RR interval variability when pain deviation is at a high level. A bidirectional attention network is used to enable the interaction of the semantic feature vector of nursing events, the scalar sequence of biometric deviation, and the physiological waveform feature vector, and to calculate the mutual influence weight between any two of them. For example, the attention weight of the semantics of nursing events related to "pain management" on the biometric deviation of "blood pressure" is calculated.

[0088] In some embodiments, the calculation of temporal association mapping involves defining a temporal alignment operation within the collaborative analysis network to align each event in the nursing event semantic sequence with a segment of the deviation sequence within a preset time window. The calculation of state association mapping involves discretizing and binning the deviation sequence and statistically analyzing the aggregation of various physiological waveform feature vectors at each deviation level. The bidirectional attention network uses a multi-head attention mechanism to calculate the cross-attention scores between the three sets of features: nursing event semantics, biometric deviation, and physiological waveform features. Based on the calculated mutual influence weights, a three-dimensional association pattern matrix is ​​constructed. The first dimension index of the association pattern matrix corresponds to the nursing event type, the second dimension index corresponds to the biometric deviation level, and the third dimension index corresponds to the physiological waveform abnormality pattern. The value of each element in the matrix represents the association strength of the co-occurrence of a specific type of nursing event, a specific level of biometric deviation, and a specific physiological waveform abnormality pattern. In the example scenario, the matrix element M[“analgesic drugs”, “pain index_high”, “heart rate variability_low”] has a high value, indicating a strong association among the three.

[0089] In some embodiments, the length of the preset time window can be set differently according to the type of nursing event. For example, the time window corresponding to drug intervention events is longer while the time window corresponding to position change events is shorter. The binning strategy for deviation level can adopt an adaptive method based on percentiles. The training of the bidirectional attention network depends on the historical nursing dataset with labeled association information.

[0090] Optionally, the association strength of elements in the association pattern matrix can be calculated using an aggregation formula, such as the elements of the final association strength matrix. The calculation formula is:

[0091]

[0092] in: Represents the elements of the final association strength matrix. This represents the attention weight of nursing event type i to biometric deviation level j. This represents the attention weight of the biometric deviation level j to the abnormal physiological waveform pattern k. The symbol represents the attention weight of nursing event type i to physiological waveform abnormality pattern k. This represents an operation function that aggregates three attention weights, such as a weighted average or a maximum value function. This formula is used to illustrate that the matrix elements are determined by a combination of multiple attention weights.

[0093] Understandably, the three-dimensional association pattern matrix serves as the foundational input for subsequent retrospective causal inference. Data comparison is reflected in the network's fusion of temporal point event sequences, continuous deviation sequences, and segment waveform feature vectors to generate a structured association representation with a clear clinical interpretive dimension. Understandably, the bidirectional attention network allows the model to dynamically focus on the most relevant parts between different data substreams. For example, when faced with an abnormal waveform of "hypotension," the network assigns higher mutual influence weights to the nursing event of "diuretic use" and the biometric deviation of "blood volume."

[0094] Example 3: See Figure 3 The retrospective causal inference process, initiated based on the association pattern matrix, first extracts elements from the association pattern matrix whose association strength exceeds a preset significance threshold; these elements are considered candidate strong association triples. For each strong association triple, the system retrieves the specific occurrence time point and contextual data corresponding to that triple from the original data stream. Using this specific occurrence time point as the endpoint, a preset inference time window is traced backward. Within this inference time window, all preceding nursing event records and biometric change data are retrieved. The system constructs a temporary causal graph, using abnormal physiological waveform patterns in the strong association triples as result nodes and the retrieved preceding nursing events and biometric changes as candidate cause nodes. Based on the chronological order and the pre-learned association strength in the association pattern matrix, directed edges are established between candidate cause nodes and result nodes, and the causal contribution score of each edge is calculated. The cause node with the highest causal contribution score and its path pointing to the result node are selected as an initial causal chain.

[0095] Multiple coexisting initial causal chains are merged and analyzed to identify common causal nodes or intermediate nodes in different chains. Based on the existence of common nodes, multiple independent initial causal chains are connected or merged at the common nodes to form a more complete composite causal chain containing multiple intermediate links. Historical nursing databases are searched for case records similar to or opposite to the composite causal chain. Using the retrieved case records, the logical rationality and statistical significance of the composite causal chain are externally validated. The validated composite causal chain, along with its causal contribution score and validation support, is used as the final output causal chain.

[0096] In practical implementation, an example scenario involves inferring the potential cause of a patient's sudden onset of premature ventricular contractions (PVCs) after surgery. The input to the process is a three-dimensional association pattern matrix output by a collaborative analysis network. Data comparison is reflected in the input being a statistical matrix describing co-occurrence associations, while the output is a causal chain with directionality and temporal order. The retrospective causal inference process extracts elements from the association pattern matrix whose association strength exceeds a preset significance threshold as candidate strong association triples. For example, elements with an association strength exceeding the threshold of 0.85 are extracted from the association pattern matrix, and the corresponding triples are ["additional administration of intravenous analgesia pump", "pain index_severe deviation", "ECG waveform_PVC pattern"]. For each strong association triple, the specific occurrence time and contextual data corresponding to the strong association triple are retrieved from the original data stream. For example, it is found that the abnormal physiological waveform corresponding to the above triple actually occurred on October 27, 2023, at 14:25:30. Using the specific time point as the endpoint, a preset inference time window is traced back. Within the inference time window, all possible preceding nursing events and changes in biometrics are retrieved. For example, within the window 2 hours before 14:25:30, the following were retrieved: "Closed chest drainage tube negative pressure", "Chief complaint of incision pain score of 8 points", and "Blood pressure showed a downward trend".

[0097] In some embodiments, a temporary causal graph is constructed, using abnormal physiological waveform patterns in strongly correlated triples as result nodes and retrieved preceding nursing events and changes in biometrics as candidate cause nodes. Directed edges are established between candidate cause nodes and result nodes based on chronological order and pre-learned correlation strength in the correlation pattern matrix, and the causal contribution score of each edge is calculated. The cause node with the highest causal contribution score and its path pointing to the result node are selected as an initial causal chain. In the example, the initial causal chain is: "Chief complaint of incision pain score 8 points" -> "Pain index_severe deviation" -> "ECG waveform_ventricular premature beat pattern".

[0098] In its implementation, the retrospective causal inference process includes verifying and integrating the initial causal chain, merging and analyzing multiple coexisting initial causal chains, and identifying common causal nodes or intermediate nodes in different chains. Based on the existence of common nodes, multiple independent initial causal chains are connected or merged to form a more complete composite causal chain. For example, another initial chain is "rapid intravenous fluid resuscitation 500ml" -> "central venous pressure - moderately elevated". If "pain index - severe deviation" and "central venous pressure - moderately elevated" show a strong correlation in the matrix, then the two chains are merged. The historical nursing database is searched for case records similar to or opposite to the composite causal chain. The retrieved case records are used to externally verify the logical rationality and statistical significance of the composite causal chain. The verified composite causal chain, along with its causal contribution score and verification support, is used as the final output causal chain.

[0099] Optionally, the causal contribution score can be calculated based on specific rules, such as directed edges. Causal contribution score The calculation formula is:

[0100]

[0101] in: Represents a directed edge The causal contribution score, This represents the decay factor of the time interval from the occurrence of event m at node m to the occurrence of event n at node n. This indicates the frequency and influence of node m as a cause in historical data. The symbol represents the association strength between node m and node n derived from the association pattern matrix. This represents a fusion function, a formula used to quantify the causal power from cause to effect.

[0102] It is understandable that the length of the inference time window is a configurable parameter, requiring a balance between the comprehensiveness of causal discovery and computational efficiency. An excessively long time window introduces irrelevant preceding events, while a window that is too short misses the true cause. It is also understandable that "similar or opposite case records" in external validation refer to causal records in the historical database that have the same or opposite node connections. Validation support can be determined by both the quantity and quality of matching cases. In some embodiments, the construction of a temporary causal graph allows for multiple edges pointing to the same result node. The calculation of the causal contribution score needs to comprehensively consider temporal proximity, correlation strength, and the specificity of the node itself. The rule for fusing multiple initial causal chains can be based on pathfinding and node overlap analysis algorithms in graph theory. Optionally, the significance threshold can be dynamically calculated based on the distribution characteristics of all element values ​​in the association pattern matrix. For example, the threshold can be set to more than twice the standard deviation of the mean of all element values. Retrieval of the historical nursing database can employ a similarity matching algorithm based on medical ontology terminology to improve recall.

[0103] Example 4: When constructing a personalized nursing intervention plan, the system analyzes the final output causal chain and identifies nursing events or biometric states that serve as interventionable nodes in the chain. For each interventionable node, a set of standardized nursing measures that can positively or negatively influence that node is queried from a structured nursing knowledge base. The overall causal contribution score of each causal chain is calculated, and all causal chains are ranked according to the score to determine the intervention priority. For high-priority causal chains, the nursing measures that are executable in that state and are expected to have the least conflict with other existing treatment measures are selected from the set of standardized nursing measures corresponding to their interventionable nodes, combined with the patient's current real-time physiological state data. The selected nursing measures are bound with specific intervention timing and execution parameters to form an independent intervention sub-plan for that causal chain. All independent intervention sub-plans are arranged chronologically according to their suggested intervention timing, and conflict checks are performed between measures. Finally, a complete and chronologically clear personalized nursing intervention plan document is generated.

[0104] In practical implementation, an example scenario involves generating a nursing plan containing specific interventions for post-thoracic surgery patients. The input is the set of causal chains output from a retrospective causal inference process. The data comparison is reflected in the input being a logical chain describing causal relationships, while the output is an executable, structured nursing plan document. Constructing a personalized nursing intervention plan involves parsing the final output causal chain and identifying nursing events or biometric states that serve as interventional nodes within the chain, such as identifying "high pain score" and "decreasing blood pressure" as interventional nodes. For each interventional node, a set of standardized nursing measures that can positively or negatively influence the interventional node is queried from a structured nursing knowledge base. For example, for the "high pain score" node, queries might yield "addition of analgesics," "adjustment of analgesia pump parameters," and "non-pharmacological analgesia intervention."

[0105] The overall causal contribution score of each causal chain is calculated, and all causal chains are ranked according to the score to determine the priority of intervention. For high-priority causal chains, from the set of standardized nursing measures corresponding to their interventionable nodes, measures that are feasible in the current state and are expected to have the least conflict with other existing treatment measures are selected, taking into account the patient's current real-time physiological state data. For example, for the chain of "severe pain leading to decreased blood pressure and premature ventricular contractions," "additional sufentanil 5μg intravenous injection" is selected from the measure set instead of "adjusting the background dose of the analgesic pump" because the patient's current blood pressure is low. The selected nursing measures are bound with specific intervention timing and execution parameters to form independent intervention sub-plans for the causal chain. For example, "additional sufentanil 5μg intravenous injection" is bound with the intervention timing of "pain score > 7 points" and the parameter "5μg, slow intravenous push". All independent intervention sub-plans are arranged chronologically according to intervention timing and conflict checks are performed to generate a complete personalized nursing intervention plan document.

[0106] In some embodiments, the calculation of the overall causal contribution score depends not only on the initial causal contribution score of the causal chain, but also comprehensively considers the validation support and time freshness factor of the causal chain. The formula for calculating the overall causal contribution score of each causal chain is as follows:

[0107]

[0108] in: This represents the overall causal contribution score of the causal chain. This represents the initial causal contribution score of the causal chain. This indicates the degree of validation support for the causal chain in historical cases. The decay factor representing the time elapsed from the occurrence of the resulting event in a causal chain to the present time is represented by the symbol […]. This represents an aggregation function, such as a weighted summation function. Nursing intervention conflict checking relies on a predefined drug interaction library and a nursing intervention compatibility matrix to ensure that selected interventions do not create contraindications with other treatments the patient is currently receiving. See Table 1 for the priority ranking results of the causal chain and the matching status of the corresponding interventions.

[0109] Table 1: Prioritization of Causal Chains and Matching of Intervention Measures

[0110] Description of causal chain Overall causal contribution score Priority Selected nursing measures Intervention timing Execution parameters Severe pain leads to a drop in blood pressure and premature ventricular contractions. 0.92 high Administer an additional 5 μg of sufentanil intravenously. When the pain score is >7 points 5μg, slowly injected via intravenous push Excessive intravenous fluid administration can lead to increased central venous pressure and arrhythmias. 0.78 middle Adjust the infusion rate to 40 ml / h Execute immediately Target speed 40ml / h Irritation from the drainage tube caused pain and tachycardia. 0.65 Low Check and adjust the position of the drainage tube. During the next inspection Assess the location and fixation of the drainage tube

[0111] In some embodiments, the process of integrating and generating personalized nursing intervention protocol documents includes an automated time-series scheduling algorithm. This algorithm arranges independent intervention sub-protocols with specific intervention timings in chronological order and performs conflict detection and analysis on measures with overlapping execution times or resource competition. The final output document format is a structured electronic nursing checklist. It is understood that the judgment logic for minimizing expected conflicts will access all current treatment orders and vital sign statuses of the patient in real time. For example, it will determine whether there is a pharmacokinetic interaction between the newly added analgesic and the current anticoagulant, or whether the postural adjustment measures conflict with the patient's current activity contraindications. Optionally, the binding of intervention timing can be based on absolute time points or relative event or physiological state thresholds, such as "pain score > 7 points" or "systolic blood pressure below 90 mmHg." The ranking of overall causal contribution scores can use a dynamic threshold method, where only chains with scores higher than the threshold are included in the high-priority group.

[0112] See Figure 4 In the nursing intervention conflict analysis phase, this chart integrates multi-dimensional data on the number of detected conflicts, the number of successfully resolved conflicts, and the corresponding resolution rates for different conflict types. Specifically, a bar chart presents the number of detected conflicts (red bars) and the number of successfully resolved conflicts (cyan bars) for four types of conflicts: "drug interactions," "operational resource competition," "time overlap," and "vital sign contraindications." A line chart (purple) displays the resolution rate (%) for each type of conflict. The data distribution shows that: "Time overlap" conflicts both had 12 detected and 12 successfully resolved cases, corresponding to a resolution rate of 100%, making it the conflict type with the best resolution effect at this stage; "operational resource competition" conflicts had 8 detected cases, with 7 successfully resolved, a resolution rate of approximately 87.5%; "drug interactions" conflicts had 6 detected cases, with 5 successfully resolved, a resolution rate of 100%; and "vital sign contraindications" conflicts had 5 detected cases, with only 4 successfully resolved, a resolution rate dropping to 80%, making it the type with the relatively weakest resolution effect among the four conflict types.

[0113] Example 5: After the personalized nursing intervention plan is generated, the system proceeds to the execution and closed-loop update steps. The system pushes the personalized nursing intervention plan to the corresponding nursing terminals and monitors the execution status and results of each measure within the plan; these records are entered into the nursing event text log. After the intervention measures are executed, the system continues to collect the patient's original monitoring data stream and repeatedly performs hierarchical decoupling processing, dynamic baseline updates, semantic anchor embedding, collaborative analysis, and causal inference processes. The system compares the newly inferred causal chain with the causal chain upon which the current intervention plan was based, assesses the effectiveness of the implemented intervention measures, and detects whether any new abnormal causal patterns have emerged. Based on the effectiveness assessment results and the newly emerging causal patterns, the personalized nursing intervention plan is dynamically adjusted and incrementally updated, forming a closed-loop management system.

[0114] When dynamically adjusting and incrementally updating personalized nursing intervention plans, if the evaluation results show that the intervention is effective, the parameters of the intervention in the plan are strengthened or it is extended to the treatment of similar causal patterns. If the evaluation results show that the intervention is ineffective or a new abnormal causal pattern emerges, the plan revision process is initiated. The plan revision process includes analyzing the reasons for the ineffectiveness of the measures or the emergence of the new pattern, searching the nursing knowledge base and historical case database for alternative or supplementary measures, and conducting a simulation evaluation of the revised plan in a patient physiological model-based environment. The revised content that has passed the simulation evaluation is updated into the personalized nursing intervention plan, and the causal context of this revision is recorded to form a historical log of the plan's evolution.

[0115] In practical implementation, an example scenario involves executing and dynamically adjusting the nursing plan for a patient who has undergone lobectomy. The input is a pre-generated personalized nursing intervention plan document, and the data comparison reflects the differences between the patient's physiological data flow before and after the plan's execution and the system's newly inferred causal chain. The execution and closed-loop update steps of the plan include pushing the personalized nursing intervention plan to the corresponding nursing terminal and monitoring the execution status and result recording of each measure in the plan. For example, after receiving the intervention sub-plan "When the pain score is >7, administer an additional 5μg of sufentanil intravenously," the nurse terminal executes it and records "Executed; the pain score dropped to 4 points 30 minutes after injection." After the intervention is executed, the system continues to collect the patient's original monitoring data flow and repeatedly executes the hierarchical decoupling processing, dynamic baseline update, semantic anchor embedding, collaborative analysis, and causal inference processes. The newly inferred causal chain is compared with the causal chain upon which the current intervention plan is based to assess the effectiveness of the intervention and detect any new abnormal causal patterns. For example, the newly inferred causal chain shows a strong correlation between "decreased pain score" and "increased blood pressure," while the "ventricular premature beat pattern" no longer appears, indicating that the intervention is effective. A new pattern of "decreased blood oxygen saturation after activity" is also detected. Based on the effectiveness assessment results and the newly emerging causal patterns, the personalized nursing intervention plan is dynamically adjusted and incrementally updated to form a closed-loop management system.

[0116] In some embodiments, when dynamically adjusting and incrementally updating personalized care intervention protocols, if the evaluation results show that the intervention is effective, the parameters of the intervention in the protocol are strengthened or it is extended to the treatment of similar causal patterns. For example, the additional dose parameter of sufentanil is adjusted from "5 μg" to "5-7 μg" to cover more severe pain situations. When the evaluation results show that the intervention is ineffective or a new abnormal causal pattern emerges, a protocol revision process is initiated. The protocol revision process includes analyzing the reasons for the ineffectiveness of the measures or the emergence of the new pattern, searching the nursing knowledge base and historical case database for alternative or supplementary measures, and performing a simulation evaluation of the revised protocol in a simulated environment. The revised content that has passed the simulation evaluation is updated to the personalized care intervention protocol, and the causal context of this revision is recorded to form a historical log of protocol evolution.

[0117] In practice, the logic for evaluating the effectiveness of interventions relies on the comparison between old and new causal chains. For example, if the old chain "A causes B" no longer appears or its intensity is significantly reduced in the new data, the intervention is considered effective. The comparison between old and new causal chains can be based on a quantitative difference calculation. The formula is:

[0118]

[0119] in: Indicates the degree of difference. This represents the set of causal contribution scores for the causal chain upon which the original intervention plan was based. This represents the set of causal contribution scores for newly inferred causal chains, denoted by [symbol]. This represents a comparison function, such as calculating the norm of the difference in causal contribution scores between two sets at common nodes. The detection of novel anomalous causal patterns is achieved by examining the association pattern matrix for significant, strongly correlated triples not covered by existing schemes.

[0120] It is understandable that the simulation environment assessment is based on the patient's personalized physiological model and pharmacokinetics model to predict changes in vital signs and potential risks brought about by the revised nursing measures. The historical log recording format includes the revision time, reason for revision, content of revision, and the data snapshot on which it is based. In the simulation and evaluation of the environment, the specific implementation of the personalized physiological model is based on individualized data extracted from the patient's biometric signal substream. Time window signal segments are obtained through sliding window segmentation, and the statistical, frequency domain, and nonlinear characteristics of each segment are calculated to form a feature vector. The feature vector of the time window determined to be in a stable state is used to update the baseline model of vital signs through incremental learning. This model dynamically reflects the normal fluctuation range of the patient's physiological parameters and is used to predict changes in vital signs caused by revised nursing interventions. The specific implementation of the pharmacokinetic model relies on a structured nursing knowledge base. This knowledge base stores the metabolic parameters, mechanisms of action, and individual variability adjustment factors of standardized drugs. The model combines the patient's real-time physiological status, such as liver and kidney function indicators, weight, and age data, to simulate the absorption, distribution, metabolism, and excretion of drugs in the body, predicting the trend of blood drug concentration changes and potential adverse reaction risks under specific nursing interventions, thus providing a quantitative basis for protocol revision. It is understood that there are safety boundaries for the operation of enhanced intervention parameters; any adjustment of any parameter will not exceed the safe dosage range and execution specifications of the drug or operation defined in the nursing knowledge base.

[0121] See Figure 5In the simulation environment assessment, based on the patient's personalized physiological model and pharmacokinetic model, the physiological indicators and risk prediction results under different sufentanil doses were displayed. The figure uses sufentanil dose (μg) as the horizontal axis, simultaneously presenting the trends of four dimensions: predicted pain score, predicted systolic blood pressure (unit: mmHg), predicted blood drug concentration (unit: ng / ml), and adverse reaction risk (%), and marking two reference lines: the upper limit of the safe dose and the adverse reaction risk threshold. Specifically, as the sufentanil dose increased from 3μg to 8μg, the predicted systolic blood pressure showed a continuous upward trend, the predicted blood drug concentration showed a linear increase, and the adverse reaction risk rapidly climbed and exceeded the threshold after the dose exceeded 7μg; while the predicted pain score remained at a low level after the dose reached 4μg. The core logic of this diagram is to quantify changes in blood drug concentration, vital sign response, and adverse reaction risk at different doses by using personalized physiological models (reflecting normal fluctuations in patients' physiological parameters) and pharmacokinetic models (simulating drug metabolism by combining data such as patients' liver and kidney function and weight). The upper limit of the safe dose (7μg) corresponds to the point where blood drug concentration and adverse reaction risk rise rapidly, providing a safety boundary and quantitative basis based on simulation for adjusting the sufentanil dose (such as increasing it from 5μg to 5-7μg).

[0122] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0123] 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. A method for intelligent management of thoracic surgery nursing data based on biometric recognition, characterized in that, The method includes the following steps: Raw monitoring data streams of thoracic surgery patients are received from multiple nursing terminals. These raw monitoring data streams integrate the patients' real-time physiological waveforms, nursing event text logs, and biometric signals collected by biosensors. The original monitoring data stream is subjected to layered decoupling processing to separate the biometric signal sub-stream, the nursing event text sub-stream, and the physiological waveform sub-stream, and timestamp synchronization and alignment are performed on each sub-stream; For the biometric signal substream, dynamic vital sign baselines are established and updated; for the nursing event text substream, semantic anchor embedding based on biometric context is performed; and for the physiological waveform substream, key waveform segments are clipped and characterized. The processed biometric signal sub-stream, nursing event text sub-stream, and physiological waveform sub-stream are input into a collaborative analysis network. The collaborative analysis network uses a cross-verification mechanism to mine the implicit correlation patterns between different sub-stream data and outputs a correlation pattern matrix. Based on the aforementioned association pattern matrix, a retrospective causal inference process is initiated, which is used to identify the causal chain between nursing events, changes in biometrics, and abnormal physiological waveforms. Based on the structure and strength of the causal chain, a personalized nursing intervention plan is constructed, which includes specific intervention points and parameter adjustment suggestions.

2. The intelligent management method for thoracic surgery nursing data based on biometric recognition according to claim 1, characterized in that, The semantic anchor embedding process based on biometric context is performed on the sub-stream of nursing event text, including: Each text record in the nursing event text substream is segmented and entity recognition is performed to extract entities related to nursing procedures, medications, and vital sign descriptions. From the established biometric knowledge graph of thoracic surgery nursing, retrieve the set of biometric concepts that have a predefined association with the entity; The retrieved set of biometric concepts is fused with the original text records to generate an enhanced text representation that includes the context of biometric features. The enhanced text representation is subjected to deep semantic encoding, which transforms the unstructured text record into a structured semantic feature vector, wherein the semantic feature vector encodes the semantic relationship between the nursing event and its associated biometric features. The semantic feature vectors are arranged in chronological order to form a semantic sequence of nursing events.

3. The intelligent management method for thoracic surgery nursing data based on biometric recognition according to claim 1, characterized in that, The establishment and updating of dynamic vital sign baselines for the biometric signal substream includes: The biometric signal substream is segmented by a sliding window to obtain continuous time window signal segments; For each time window signal segment, its statistical characteristics, frequency domain characteristics, and nonlinear characteristics are calculated to form the feature vector of the time window; The feature vector of the current time window is compared with the feature vector sequence of the historical window, and an adaptive threshold is used to determine whether the current time window is in a stable state. The individualized baseline model of vital signs is updated by incremental learning using only the feature vectors of the time window determined to be in a stable state. Using the updated biometric baseline model, the biometric signal sub-stream is monitored in real time, and the deviation sequence between the current signal and the dynamic baseline is calculated. The deviation sequence quantifies the instantaneous abnormality of the biometrics.

4. The intelligent management method for thoracic surgery nursing data based on biometric recognition according to claims 2 and 3, characterized in that, The collaborative analysis network uses a cross-verification mechanism to uncover implicit correlation patterns between different sub-stream data, including: Receive the semantic sequence of the nursing event, the deviation sequence, and the characteristic representation of the key waveform segment; Within the collaborative analysis network, a temporal correlation mapping is established between the semantic sequence of nursing events and the deviation sequence, and the change pattern of biometric deviation within a preset time window after each nursing event occurs is calculated. Establish a state-correlation mapping between deviation sequences and physiological waveform features, and analyze the distribution of physiological waveform features corresponding to different levels of biological characteristic deviations; A bidirectional attention network is used to enable the interaction of nursing event semantics, biometric deviation and physiological waveform features, and to calculate the mutual influence weight between any two. Based on the mutual influence weights, a three-dimensional association pattern matrix is ​​constructed. Each dimension of the association pattern matrix corresponds to the nursing event type, biometric deviation level, and physiological waveform abnormality pattern, respectively. The value of the matrix element represents the association strength of the co-occurrence of the three.

5. The intelligent management method for thoracic surgery nursing data based on biometric recognition according to claim 4, characterized in that, Based on the association pattern matrix, a retrospective causal inference process is initiated, including: Elements with association strength exceeding the significance threshold are extracted from the association pattern matrix and used as candidate strong association triples. For each strongly associated triple, retrieve the specific time point and context data corresponding to the strongly associated triple in the original data stream; Using the specific time point of occurrence as the endpoint, a preset inference time window is traced back, and within the inference time window, all possible preceding care events and changes in biometrics are retrieved; A temporary causal graph is constructed, with abnormal physiological waveform patterns in the strongly correlated triples as result nodes and retrieved preceding nursing events and changes in biometrics as candidate cause nodes. Based on the chronological order and the pre-learned association strength in the association pattern matrix, directed edges are established between candidate cause nodes and result nodes, and the causal contribution score of each edge is calculated. The cause node with the highest causal contribution score and its path pointing to the result node are selected as an initial causal chain.

6. The intelligent management method for thoracic surgery nursing data based on biometric recognition according to claim 5, characterized in that, The retrospective causal inference process also includes the verification and integration of the initial causal chain: Merge and analyze multiple coexisting initial causal chains to identify common causal nodes or intermediate nodes in different chains; Based on the existence of common nodes, multiple independent initial causal chains can be connected or merged to form a more complete composite causal chain that may contain multiple intermediate links. Search the historical nursing database for case records that are similar to or opposite to the stated complex causal chain; Using the retrieved case records, the logical rationality and statistical significance of the composite causal chain are externally verified; The verified composite causal chain, along with its causal contribution score and verification support, will be used as the final output causal chain.

7. The intelligent management method for thoracic surgery nursing data based on biometric recognition according to claim 6, characterized in that, The construction of a personalized nursing intervention plan that includes specific intervention points and parameter adjustment suggestions includes: The causal chain of the final output is analyzed to identify nursing events or biometric states that are interventionable nodes in the chain; For each interventionable node, a set of standardized nursing measures that can positively or negatively affect the interventionable node is queried from the nursing knowledge base; Calculate the overall causal contribution score for each causal chain, and rank all causal chains according to the scores to determine the priority of intervention; For high-priority causal chains, select the measures that are feasible and have the least expected conflict from the set of standardized nursing measures corresponding to the interventionable nodes; The selected nursing measures are linked with specific intervention timing and execution parameters to form independent intervention sub-plans targeting the causal chain; All independent intervention sub-plans were arranged chronologically according to the timing of intervention and conflict checks were performed to integrate them into a complete personalized nursing intervention plan document.

8. The intelligent management method for thoracic surgery nursing data based on biometric recognition according to claim 1, characterized in that, The process of pruning and characterizing key waveform segments for the physiological waveform sub-stream includes: Real-time monitoring of physiological waveform sub-streams; when the waveform amplitude, frequency or morphology exceeds the preset normal range, the detected moment is marked as the starting point of a critical event. Taking the starting point of the key event as the center, a fixed-duration waveform data segment is extracted before and after it to form a key waveform segment; Multi-scale filtering is performed on the key waveform segments to separate waveform components with different frequency components; Calculate the time-domain statistical characteristics, frequency-domain energy distribution characteristics, and morphological descriptive characteristics of each waveform component; The features of all waveform components are spliced ​​and normalized to form the feature representation vector of the key waveform segment.

9. The intelligent management method for thoracic surgery nursing data based on biometric recognition according to claim 1, characterized in that, After constructing a personalized nursing intervention plan that includes specific intervention points and parameter adjustment suggestions, the plan also includes implementation and closed-loop update steps: The personalized nursing intervention plan is pushed to the corresponding nursing terminal, and the execution status and result records of each measure in the plan are monitored; After the intervention measures are implemented, the patient's original monitoring data stream continues to be collected, and the hierarchical decoupling process, dynamic baseline update, semantic anchor embedding, collaborative analysis and causal inference process are repeated. The newly inferred causal chain is compared with the causal chain that generated the current intervention plan to evaluate the effectiveness of the intervention and detect whether any new anomalous causal patterns have emerged. Based on the effectiveness evaluation results and newly emerging causal patterns, the personalized nursing intervention plan is dynamically adjusted and incrementally updated to form a closed-loop management system.

10. The intelligent management method for thoracic surgery nursing data based on biometric recognition according to claim 9, characterized in that, The dynamic adjustment and incremental update of the personalized nursing intervention plan includes: When the evaluation results show that the intervention is effective, strengthen the parameters of the intervention in the program or extend it to treatments with similar causal patterns; When the assessment results show that the intervention measures are ineffective or new abnormal causal patterns emerge, the program revision process is initiated. The revised plan process includes: analyzing the reasons for the ineffectiveness of measures or the emergence of new models, searching the nursing knowledge base and historical case database for alternative or supplementary measures, and conducting simulation and evaluation of the revised plan in a simulated environment; The revised content, based on the deductive evaluation, will be updated to the personalized care intervention plan, and the causal context of this revision will be recorded to form a historical log of the plan's evolution.

Images