Intelligent identification and alarm method for respiratory suppression event in anesthesia revival period

Abstract

The invention provides an intelligent identification and alarm method for respiratory suppression events in an anesthesia revival period, which comprises the following steps of: continuously acquiring high-frequency physiological data such as respiration, blood oxygen and electrocardio of a patient through multi-channel equipment, and establishing a dynamic causal network model fusing medical priori knowledge and clinical guidelines after standardized processing and feature extraction; a Granger causal test and a dynamic time warping algorithm are combined, a significant causal relationship among key physiological parameters is dynamically identified, a causal network structure is updated in real time, a causal analysis result is further input into a time sequence Bayesian network, calculation of a respiratory suppression event occurrence probability and reasoning path tracing are realized, and the probability of occurrence of a respiratory suppression event is calculated. According to the method and the system, the probability score is calculated, an interpretable medical logic evidence chain and thermodynamic diagram visualization are automatically generated, and if the probability score exceeds the limit, multi-mode alarm and data locking are synchronously triggered, so that the timeliness, intelligence and interpretability of respiratory suppression detection are improved, and clinical precise intervention is facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical artificial intelligence and physiological signal intelligent monitoring, and particularly relates to an intelligent identification and alarm method for respiratory depression events in the anesthesia recovery period. BACKGROUND

[0002] In the current field of anesthesia recovery period respiratory depression monitoring, clinical mainly relies on medical staff to manually interpret the values and waveforms of physiological monitoring equipment, and combines with previous experience to conduct risk assessment. In recent years, with the development of multi-parameter monitors and intelligent medical technology, data-driven respiratory depression identification systems based on multi-modal physiological signals have been gradually applied to high-risk scenarios such as post-anesthesia care units (PACU). Such systems generally rely on black box models such as neural networks, support vector machines, and ensemble learning, and through end-to-end fitting of multi-channel data such as respiratory rate, tidal volume, oxygen saturation, end-tidal carbon dioxide, and heart rate variability, automatic identification and abnormal alarm of respiratory depression events are realized. These methods can improve the event detection sensitivity and reduce manual misjudgment to some extent, but the actual clinical adoption rate is still low. The current mainstream technology route focuses on improving the identification accuracy of the model, and usually emphasizes multi-feature fusion, large data training, and deep neural network classification. Some technologies attempt to introduce post-explanation techniques (such as SHAP, LIME) or visual attention mechanisms to provide doctors with some key feature inference reasons at the model output stage. However, the black box structure inherently determines that the reasoning process of the model is difficult to map to medical logic, and doctors can only obtain the model results but cannot understand the basis for their judgments, making it difficult to trace the specific physiological chain of abnormal events and to manually correct complex or boundary cases. Since the model decision-making process is mainly based on data correlation rather than medical causality, it is easily affected by short-term noise interference or non-specific parameter variations, resulting in a high false alarm rate, which seriously restricts the clinical actual adoption limit. SUMMARY

[0003] The present application provides an intelligent identification and alarm method for respiratory depression events in the anesthesia recovery period to solve the above-mentioned technical problems.

[0004] The technical solution of the present application is as follows: the intelligent identification and alarm method for respiratory depression events in the anesthesia recovery period comprises: S1: collecting continuous time series data from a multi-channel physiological monitoring device for patients in the anesthesia recovery period, the data including respiratory rate, tidal volume, oxygen saturation, end-tidal carbon dioxide, and heart rate variability parameters, and the sampling frequency of each parameter is set according to clinical standards and labeled with a time stamp; S2: performing sliding window segmentation processing on the collected original physiological signals, the window length is set to 15-60 seconds according to the respiratory cycle characteristics, the overlap rate is not less than 50%, and a time series data block sequence containing multi-dimensional feature vectors is generated; S3: Construct a dynamic causal network template based on a medical prior knowledge base. The template includes a set of nodes and a set of directed edges, where nodes represent physiological parameters and directed edges represent known pathophysiological relationships. The weights of each edge are determined by expert consensus. S4: The Granger causality test and dynamic time warping algorithm are used to jointly analyze the current time series data block, identify the direction of significant causal relationship and the time delay parameter, and update the local subgraph structure of the dynamic causal network template according to the test results to form a real-time causal subgraph. S5: Input the real-time causal subgraph into the pre-trained temporal Bayesian network model. The model takes the breathing inhibition event as the root node, derives the distribution of downstream parameter changes layer by layer through the conditional probability table, and outputs a decision vector containing the event probability score. S6: Generate an explicit medical logic chain based on the Bayesian inference path. The logic chain consists of nodes and edges in the causal subgraph and includes a temporal causal sequence of abnormal initial parameters, changes in mediating effect parameters, and deviations in terminal parameters, forming an interpretable chain of evidence. S7: Convert the evidence chain into a natural language description and generate a corresponding heatmap visualization matrix. The horizontal axis of the matrix represents the time node, the vertical axis represents the parameter dimension, and the color intensity represents the contribution weight of each feature to the decision vector. S8: When the probability score of respiratory depression events exceeds a preset threshold, a multimodal alarm signal is triggered simultaneously and the corresponding visual evidence chain is locked. The alarm signal includes audible and visual prompts and electronic medical record markers. The lock duration is set according to the clinical treatment process.

[0005] The intelligent identification and alarm method for respiratory depression events during anesthesia recovery provided by this invention has the following beneficial effects: The intelligent identification and alarm method for respiratory depression events during anesthesia recovery provided by this invention achieves significant optimization through multi-dimensional innovation: On the one hand, it abandons the black-box model approach that simply pursues recognition accuracy, constructs a dynamic causal network template that integrates prior medical knowledge, and combines Granger causality tests and dynamic time warping algorithms to identify the true causal relationship and time delay between physiological parameters in real time, forming a traceable real-time causal subgraph. This subgraph is then input into a temporal Bayesian network to deduce the probability of respiratory depression events, fundamentally avoiding misjudgments caused by model dependence on data correlation, reducing interference from short-term noise and non-specific parameter variations, and effectively solving the problem of high false alarm rates in black-box models. On the other hand, this method does not rely on post-interpretable techniques in the model output stage, but rather generates a model that includes "abnormal initial parameters - changes in mediating effect parameters - deviations in terminal parameters". The explicit medical logic chain, combined with natural language description and heatmap visualization, clearly presents the pathophysiological chain of an event, enabling doctors to understand the basis for judgment and make manual corrections for complex or boundary cases. At the same time, the combination of dynamic causal networks and Bayesian inference significantly reduces the workload of medical staff in manual interpretation, improves the timeliness and accuracy of detection, and ultimately significantly improves the actual clinical adoption rate of the method, providing a solution for monitoring respiratory depression during anesthesia recovery that is intelligent, interpretable, and reliable. Attached Figure Description

[0006] Figure 1 This is a flowchart of the intelligent identification and alarm method for respiratory depression events during anesthesia recovery according to the present invention; Figure 2 This is a sub-flowchart of the intelligent identification and alarm method for respiratory depression events during anesthesia recovery according to the present invention; Figure 3 This is another sub-flowchart of the intelligent identification and alarm method for respiratory depression events during anesthesia recovery according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides an intelligent identification and alarm method for respiratory depression events during anesthesia recovery, specifically including: S1: Collect continuous time-series data from a multi-channel physiological monitoring device for patients emerging from anesthesia. The data includes respiratory rate, tidal volume, blood oxygen saturation, end-tidal carbon dioxide, and heart rate variability parameters. The sampling frequency of each parameter is set according to clinical standards and timestamp is marked. S2: Perform sliding window segmentation processing on the collected raw physiological signals. The window length is set to 15-60 seconds according to the respiratory cycle characteristics, and the overlap rate is not less than 50%, generating a time-series data block sequence containing multi-dimensional feature vectors. S3: Construct a dynamic causal network template based on a medical prior knowledge base. The template includes a set of nodes and a set of directed edges, where nodes represent physiological parameters and directed edges represent known pathophysiological relationships. The weights of each edge are determined by expert consensus. S4: The Granger causality test and dynamic time warping algorithm are used to jointly analyze the current time series data block, identify the direction of significant causal relationship and the time delay parameter, and update the local subgraph structure of the dynamic causal network template according to the test results to form a real-time causal subgraph. S5: Input the real-time causal subgraph into the pre-trained temporal Bayesian network model. The model takes the breathing inhibition event as the root node, derives the distribution of downstream parameter changes layer by layer through the conditional probability table, and outputs a decision vector containing the event probability score. S6: Generate an explicit medical logic chain based on the Bayesian inference path. The logic chain consists of nodes and edges in the causal subgraph and includes a temporal causal sequence of abnormal initial parameters, changes in mediating effect parameters, and deviations in terminal parameters, forming an interpretable chain of evidence. S7: Convert the evidence chain into a natural language description and generate a corresponding heatmap visualization matrix. The horizontal axis of the matrix represents the time node, the vertical axis represents the parameter dimension, and the color intensity represents the contribution weight of each feature to the decision vector. S8: When the probability score of respiratory depression events exceeds a preset threshold, a multimodal alarm signal is triggered simultaneously and the corresponding visual evidence chain is locked. The alarm signal includes audible and visual prompts and electronic medical record markers. The lock duration is set according to the clinical treatment process.

[0010] Step S1: Continuous time-series data is collected from a multi-channel physiological monitoring device for patients emerging from anesthesia. This data includes respiratory rate, tidal volume, blood oxygen saturation, end-tidal carbon dioxide, and heart rate variability. The sampling frequency for each parameter is set according to clinical standards and timestamped. Specifically, this includes: S1.1: Based on the physiological characteristics of patients in the recovery period after anesthesia, a multi-channel physiological signal acquisition device is configured. The device includes a respiratory monitoring belt, a pulse oximeter, an end-tidal carbon dioxide sensor, and an electrocardiogram module to collect key parameters including respiratory rate, tidal volume, blood oxygen saturation, end-tidal carbon dioxide, and heart rate variability, ensuring that the collected parameters cover the key physiological indicators of respiratory depression events. Based on the physiological recovery characteristics of patients during the anesthesia recovery period, a multi-channel physiological signal acquisition unit covering key indicators of respiratory depression events was selected as the equipment configuration target. A respiratory monitoring belt sensor module (parameters: tensile strain sensitivity ≥2.0 mV / με, sampling interface output impedance ≤10 Ω) is used to realize the displacement detection function of thoracic expansion and contraction; Furthermore, through the pulse oximeter sensing module (parameters: red light wavelength 660 nm, infrared wavelength 940 nm, detection gap ≤ 1 ms), the blood oxygen saturation SpO2 and pulse rate are simultaneously acquired, and photovolume pulse wave signal data are obtained. Furthermore, by using an end-tidal carbon dioxide sensor (parameters: non-dispersive infrared detection principle, resolution ≤0.1mmHg, response time ≤100 ms), real-time measurement of PETCO2 concentration is achieved, and a time series of end-tidal gas components is generated. Furthermore, through the electrocardiogram module (parameters: number of leads ≥3, analog front-end noise ≤1 μV rms, sampling resolution 16 bit), multi-lead synchronous acquisition of electrocardiogram signals is realized, and raw QRS complex data for calculating heart rate variability (HRV) is generated; Through the multi-sensor integrated bus interface, the raw signals of various physiological parameters acquired in the previous step are transformed into a data stream that can be processed uniformly, realizing the synchronous acquisition and output of multi-channel signals, and establishing parameter coverage integrity for subsequent sampling frequency setting and signal processing. For example, in a post-anesthesia recovery room setting, the aforementioned multi-channel physiological signal acquisition equipment was configured for a patient in the early stages of recovery. The strain sensitivity of the respiratory monitoring belt was set to 2.5 mV / με, the detection interval of the pulse oximeter was set to 0.5 ms, the resolution of the end-tidal carbon dioxide sensor was configured to 0.05 mmHg, and the sampling resolution of the electrocardiogram module was configured to 18 bits. Each sensor was connected to a central acquisition server via an RS-485 bus. The server's built-in time synchronization module timestamped each data stream at the sub-millisecond level. During the acquisition process, the chest displacement signal was recorded as two peak events per second, SpO2 was maintained between 97.2 and 98.5, the PETCO2 curve showed that the end-tidal concentration ranged from 38.4 to 43.1 mmHg, and the HRV synchronous calculation yielded a low-frequency power of 512 ms. 2 The high-frequency power is 136 ms2 This output demonstrates stable performance in terms of data completeness and coverage of key indicators, allowing direct entry into the next sub-step of fine-tuning the sampling frequency, thus combining clinical usability with a foundation for causal reasoning. S1.2: Set the sampling frequency for each parameter according to clinical monitoring standards, with respiratory rate sampling rate not less than 1Hz, tidal volume not less than 10Hz, SpO2 and PETCO2 not less than 20Hz, and HRV not less than 250Hz, to meet the capture requirements of dynamic changes in different physiological parameters and ensure signal temporal resolution and clinical usability. The raw data of multi-channel physiological signals output from step S1.1 is received as input. The data includes respiratory rate, tidal volume, blood oxygen saturation, end-tidal carbon dioxide and heart rate variability parameters, and all of them are continuous analog signals. The sampling frequency setting method based on clinical monitoring standards (parameter range: respiratory rate ≥1Hz, tidal volume ≥10Hz, SpO2 and PETCO2 ≥20Hz, HRV ≥250Hz) was adopted to initialize the sampling rate of each channel. Furthermore, by using a frequency segmentation matching algorithm (parameters: target sampling rate, sampling period), the sampling period of each sensor is accurately calculated, and the sampling control command for each type of physiological parameter is obtained. Furthermore, by using a sampling rate calibration algorithm (parameters: device clock deviation, target sampling rate), the dynamic correction of the device's internal sampling timer is achieved, and a sampling synchronization flag is generated; Furthermore, by using a real-time sampling quality monitoring method (parameters: sampling interval variance, packet loss rate), the integrity of the sampling process is detected, and a sampling quality assessment report is generated. By using the above sampling frequency setting and calibration methods, the results of the previous step are transformed into a clinically usable multi-channel digital physiological signal sampling configuration, thereby ensuring the temporal resolution of parameters with different dynamic change rates. For example, this sub-step is performed on a patient emerging from anesthesia in a PACU scenario, with the respiratory rate signal acquired by a respiratory monitoring band and the target sampling rate configured as follows. The tidal volume signal is acquired by a flow sensor at a frequency of Hz, with the target sampling rate configured as follows: The SpO2 signal was acquired by a pulse oximeter at a frequency of Hz, with the target sampling rate configured as follows: The PETCO2 signal was acquired by an infrared gas analyzer at a Hz sampling rate of [value missing]. The target sampling rate was configured as follows: [value missing]. The HRV signal is acquired by the ECG module at a frequency of Hz, with the target sampling rate configured as follows: Hz. For the tidal volume channel, calculate the sampling period. Seconds, and configure the device timer to every second using a frequency segmentation matching algorithm. Sampling is triggered once per second; the sampling period is calculated for the HRV channel. Seconds, and clock skew is corrected by a sampling rate calibration algorithm. milliseconds corrected to Milliseconds. In continuous sampling After seconds, the variance of the sampling interval is Within the millisecond range, the packet loss rate remained at zero, verifying that the sampled signal has high resolution and stability under different parameter dynamic rates, providing a precise time series data foundation for subsequent filtering, timestamp annotation and causal analysis; S1.3: Perform pre-filtering and analog-to-digital conversion on the acquired raw analog signals. Use a Butterworth low-pass filter to filter the SpO2 and PETCO2 signals at a 5Hz cutoff frequency, and use a band-pass filter to extract the ECG signal in the 0.5-40Hz frequency band to obtain a high-quality digital physiological signal sequence and remove environmental noise and electromyography interference. S1.4: Based on the system clock synchronization mechanism, timestamps are applied to the signals of each channel. The NTP protocol and GPS clock source are used to achieve sub-millisecond time alignment, ensuring that the multi-channel data has comparability and causal inference capability in the time sequence dimension, and providing a time reference for the subsequent construction of dynamic causal networks. S1.5: Store the time-stamped multi-channel physiological signals in a time-series database, organize them in TSDB format according to the triplet structure of channel, time stamp and value, and generate a structured data stream as the input data source for subsequent sliding window segmentation and causal relationship analysis, ensuring data integrity and traceability.

[0011] Step S2: Perform sliding window segmentation processing on the acquired raw physiological signals. The window length is set to a range of 15-60 seconds based on the respiratory cycle characteristics, with an overlap rate of no less than 50%, generating a time-series data block sequence containing multi-dimensional feature vectors. Specifically, this includes: S2.1: Based on the physiological characteristics of the respiratory cycle, calculate the window length parameter. The window length is determined according to the average respiratory cycle distribution of patients during the anesthesia recovery period, and the range is set to 15-60 seconds to ensure that the complete evolution of respiratory events is captured. S2.2: A sliding time window mechanism is used to perform segmented processing on continuous physiological signals. The moving step size of the sliding time window is calculated based on the window length and the overlap rate. The overlap rate is not less than 50% to maintain the temporal continuity between adjacent data blocks. The window length parameter calculated by S2.1 is used to divide the continuous multi-channel physiological signal into sub-sequence data blocks with fixed durations by adopting the sliding time window segmentation method (parameters: window length L, sampling frequency f). Furthermore, the moving step size of the time window is determined using the step size calculation formula, as follows:

[0012] in For the movement step size, For window length, Window overlap ratio (value not less than) This calculation ensures that there is at least half of the data overlap between adjacent windows to maintain temporal continuity; Furthermore, an index recursion algorithm (parameters: starting index i, moving step size S) is used to generate a set of start and end indices for each time window in the collected continuous time series data, and the corresponding data segments are extracted according to the channel. Furthermore, a multi-channel signal synchronous interception method is adopted (synchronization reference: timestamp alignment result) to achieve the interception of cross-channel signals within the same time window, so as to avoid feature mismatch caused by sampling phase differences; By using time window queuing storage processing, the extracted time window data blocks are arranged into a sliding window sequence according to time order, forming a set of raw segments that can be used as input for S2.3 signal preprocessing operations; For example, in a post-anesthesia recovery room scenario, if the patient's average respiratory cycle is 20 seconds, and the window length L is set to 20 seconds according to S2.1, with an overlap rate R of 0.6, then according to the formula... The movement step size S was set to 8 seconds. With a sampling frequency f of 20Hz, each window contained 400 sampling points, and adjacent windows shared 240 sampling points. Using this parameter, data segments of 20 seconds each were extracted for the respiratory rate, tidal volume, SpO2, PETCO2, and HRV channels, and timestamp alignment was used to ensure signal synchronization within the same window for each channel. The queuing-organized sliding window sequence significantly improved the accuracy and stability of capturing the dynamic changes in respiratory depression events during subsequent feature extraction. S2.3: Perform signal preprocessing operations on the original physiological signal segments captured by the sliding window. The preprocessing includes noise reduction filtering, baseline drift correction, and outlier removal to improve the stability and accuracy of subsequent feature extraction. S2.4: Based on the preprocessed physiological signal fragments, extract multidimensional feature vectors, including but not limited to: mean respiratory rate, standard deviation of tidal volume, slope of blood oxygen saturation trend, end-tidal carbon dioxide change rate, and low-frequency / high-frequency ratio of heart rate variability, to form a set of multidimensional feature vectors; S2.5: The extracted multidimensional feature vectors are organized into a time-series data block sequence in chronological order. The time-series data block sequence contains timestamp information and feature dimension identifiers, which serve as the input data structure for subsequent Granger causality analysis and dynamic time warping algorithms.

[0013] like Figure 2 As shown, step S3 involves constructing a dynamic causal network template based on a medical prior knowledge base. The template includes a set of nodes and a set of directed edges, where nodes represent physiological parameters and directed edges represent known pathophysiological relationships. The weights of each edge are determined through expert consensus. Specifically, this includes: S3.1: Based on clinical medical guidelines and pathophysiological mechanisms, define a set of key physiological parameters related to respiratory depression, including respiratory rate, tidal volume, blood oxygen saturation, end-tidal carbon dioxide, and heart rate variability, as a set of nodes in a dynamic causal network to construct an initial atlas structure driven by medical knowledge. Using the multidimensional physiological feature sequence segmented by the sliding window as input, and based on the pathophysiological feature range of patients in the anesthesia recovery period, the clinical medical guideline database interface is called (parameters: normal respiratory rate 8-20 breaths / min, normal tidal volume 6-8 mL / kg, critical value of blood oxygen saturation <95%, normal range of end-tidal carbon dioxide 35-45 mmHg, reference value of low-frequency / high-frequency heart rate variability ratio 0.5-2.0) to achieve parameter filtering function; Furthermore, through the medical knowledge graph loading module (parameter: node labels are standardized medical terms), the pathological mechanism mapping rules are invoked to convert the selected physiological parameters into node identifiers that conform to the dynamic causal network node specifications, and the node set data structure is obtained. Furthermore, a parameter correlation constraint algorithm (basic data: causal chain recommended by clinical guidelines) is used to calculate the participation index of each potential node in the causal chain of respiratory depression events, and redundant parameters with participation indices below a preset threshold are screened out to generate an optimized set of key physiological parameters. Furthermore, using the named entity standardization processing module (parameters: ICD-10 encoding and SNOMED CT mapping table), node naming standardization processing is performed, merging nodes with the same or similar medical concepts into standard nodes, forming a set of nodes with no duplication and strong consistency; By using a node definition method driven by pathophysiological mechanisms, the screening and standardization results of the previous step are transformed into a standardized set of nodes for respiratory rate, tidal volume, blood oxygen saturation, end-tidal carbon dioxide and heart rate variability, thereby realizing the medical knowledge-driven construction of the initial graph structure of the dynamic causal network. For example, in a single embodiment of respiratory depression monitoring during anesthesia recovery, the input sliding window feature data includes a respiratory rate of Hours / minute, tidal volume is Blood oxygen saturation was %, end-tidal carbon dioxide was mmHg, heart rate variability low-frequency / high-frequency ratio is By accessing a medical guideline database and comparing parameters to normal ranges, three parameters—respiratory rate, blood oxygen saturation, and end-tidal carbon dioxide—were identified as abnormal. A correlation constraint algorithm was then executed to determine the participation index of respiratory rate in the causal chain. Blood oxygen saturation was End-tidal carbon dioxide All are above the threshold. These parameters are retained. Through entity standardization, the node names are unified as three node labels: RespRate, SpO2, and PETCO2. The final output node set is {RespRate, TVol, SpO2, PETCO2, HRV}. This set covers the key physiological parameters of respiratory depression events and significantly improves medical logic consistency and interpretability during the initialization phase of the dynamic causal network. Subsequent causal edge definition steps can accurately establish causal relationships between parameters based on this set. S3.2: Based on the known pathophysiological correlation mechanism, establish a set of directed edges between nodes. The directed edges represent the direction of the causal relationship between parameters. For example, hypoventilation leads to a decrease in blood oxygen saturation and an increase in end-tidal carbon dioxide, so as to construct a causal graph structure with medical logic. S3.3: The expert consensus method is used to assign an initial weight to each directed edge. The weight is based on the multi-center clinical expert scores and the strength of literature support to quantify the strength and credibility of the causal relationship between parameters and provide a prior probability basis for subsequent Bayesian inference. Based on the set of directed edges between nodes established by S3.2, the expert consensus method (method parameters: expert group size, rating scale dimension, rating round setting) is used to perform initial weight assignment on each directed edge. Furthermore, by setting up a multi-center clinical expert group (parameters: including the heads of anesthesiology departments and respiratory monitoring specialists from no less than 5 tertiary hospitals, with a number of no less than 12 people), the pathological association strength of each directed edge was scored by experts to form an original scoring matrix and obtain the central tendency index of the expert scores. Furthermore, using a literature support strength quantification model (parameters: clinical study sample size, statistical significance level α=0.05), the corresponding literature evidence level is matched to each side of the scoring matrix to generate a weighted scoring matrix that combines expert scores and literature evidence. Furthermore, the scoring matrix is ​​input into the weight quantization formula to calculate the initial weight value for each causal edge. :

[0014] in, Normalized values ​​for expert scores Normalized values ​​for the strength of support in the literature; Furthermore, through the consistency test method (parameter: Kendall concordance coefficient threshold set to 0.7), directed edges with consistency below the threshold are filtered out and marked as requiring further clinical validation, forming an initial causal weight set that meets the quality standards; By combining expert consensus with literature-supported data processing methods, the directed edge set result from the previous step is transformed into a quantitative weight parameter matrix, thereby achieving a quantifiable initialization effect of causal relationship strength and credibility in the dynamic causal network template. For example, in a real-world respiratory monitoring scenario during anesthesia recovery, a scoring group of 15 clinical experts from 6 hospitals in Beijing, Shanghai, and Guangzhou was selected. A 5-point scale was used, where 5 represents strong correlation and 1 represents weak correlation. For the directed edge between the nodes "decreased tidal volume" and "decreased SpO2," the mean expert score was 4.2. A literature search revealed 320 randomized controlled trials supported by this edge, with a statistical significance p-value of 0.01. The corresponding normalized value of the literature support strength is... Normalized value of expert scores Substituting into the quantization formula:

[0015] Calculate the initial weight values The Kendall consistency coefficient was 0.78, which is higher than the set threshold of 0.7, therefore the directed edge was retained. This process ensures that the edge weights in the causal network template numerically reflect the real clinical mechanism associations and provides a stable data foundation for the prior probability configuration in subsequent Bayesian inference. S3.4: Based on the expert consensus scoring results, the fuzzy comprehensive evaluation method is used to normalize the causal relationship weights to eliminate the subjective bias of expert scoring and obtain a standardized causal network edge weight matrix for the initial configuration of the dynamic causal network template. Based on the expert consensus scoring results, the fuzzy comprehensive evaluation method (parameters: expert scoring matrix, indicator weight set, membership function family) is adopted to realize the fuzzy quantitative mapping of the initial weights of causal relationships; The expert scoring matrix is ​​input into the fuzzification module, and the scoring values ​​are converted into membership vectors of each evaluation level through a family of membership function families (types: triangular, trapezoidal, bell-shaped). This allows fuzzy set theory to be introduced to describe uncertainty while preserving expert judgment information. Furthermore, the membership degrees of each causal directed edge are weighted and aggregated using a fuzzy weighted average algorithm (parameters: membership vector matrix, index weight set) to obtain the comprehensive membership degree distribution of each edge and form an aggregated fuzzy evaluation vector for subsequent defuzzification processing. Furthermore, the centroid method (parameters: fuzzy evaluation vector, domain of membership function) in the defuzzification algorithm is used to calculate the sharpening weight value of each edge. This process is implemented according to the following formula:

[0016] in, The quantitative score corresponding to the evaluation level. The membership degree of this score. The number of evaluation levels; Furthermore, a weight vector is constructed based on the clarified weight values ​​of all causal relationships, and a normalization calculation (parameters: weight vector, normalization interval [0,1]) is used to eliminate differences in scale and preference among expert ratings. The normalization formula is as follows:

[0017] in, The original sharpening weight values, and These are the minimum and maximum values ​​in the weight set, respectively; By using fuzzy comprehensive evaluation and defuzzification, the expert consensus scoring results from the previous step are transformed into a standardized and normalized causal network edge weight matrix, achieving consistency and comparability in causal relationship quantification, and providing a mathematically rigorous weight configuration basis for the initialization of dynamic causal network templates. For example, in a respiratory monitoring scenario during anesthesia recovery, multicenter clinical experts rated the strength of the causal relationship "decreased tidal volume → increased end-tidal carbon dioxide" as 5, 4, 5, 3, and 4 (out of 5). A triangular membership function was used to map the score values ​​to the membership degree of the rank set {low, medium, high}, where the rank quantification scores were 0.2, 0.5, and 0.8, respectively. A weighted average aggregation of the causal relationship score data yielded a fuzzy evaluation vector {0.05, 0.20, 0.75}. Defuzzification was performed using the centroid method.

[0018] The calculated sharpening weight is 0.73. Assuming the minimum value of the weight set is 0.60 and the maximum value is 0.95, normalization is then performed.

[0019] The normalized weight value was 0.37. This value was written into the causal network weight matrix, and all edge weights of the matrix underwent the same fuzzy comprehensive evaluation and standardization process, which ultimately eliminated cross-expert scoring differences and ensured that the input weights of the subsequent Bayesian inference model had a unified numerical benchmark. S3.5: Integrate the initialized set of nodes and the weighted directed edge set into a dynamic causal network template. The template is represented in the form of an adjacency matrix and serves as the structured input basis for subsequent real-time causal subgraph updates and Bayesian inference, ensuring that the model's inference path has medical interpretability.

[0020] like Figure 3 As shown, step S4 involves jointly analyzing the current time-series data block using the Granger causality test and the dynamic time warping algorithm to identify significant causal relationship directions and time delay parameters. Based on the test results, the local subgraph structure of the dynamic causal network template is updated to form a real-time causal subgraph. Specifically, this includes: S4.1: Based on the static causal network template constructed from the medical prior knowledge base, Granger causality test is performed on the multidimensional physiological time series data blocks after sliding window segmentation to identify the significant causal relationship directions between parameters in the current data stream and obtain a preliminary causal direction matrix; For the multidimensional physiological time-series data blocks segmented by the sliding window, the Granger causality test method (parameters: significance level α=0.05, maximum lag order is set to the number of sampling points of 1~5 seconds according to the respiratory cycle) is used to determine the causal direction of each parameter pair in the current data stream. Furthermore, by using a multivariate Granger causal model (the parameter matrix contains five input variables: respiratory rate, tidal volume, SpO2, PETCO2, and HRV), a linear regression equation based on lag multinomials was fitted, and the F-statistic and p-value of each parameter pair were obtained. Furthermore, based on the F-statistic and the preset significance level, it is determined whether the causal effect is valid, and a binary matrix of causal directionality is generated. In the matrix, "1" indicates that there is a significant causal relationship, and "0" indicates that there is no significant causal relationship. Furthermore, for parameter pairs with a significant causal relationship, their Granger causality strength index is calculated using the following formula:

[0021] in, To limit the sum of squared residuals in the model, This is the sum of squared residuals of the unrestricted model, used to quantify the degree of improvement in model fit after adding lag terms; Furthermore, the binary matrix of causal directionality is combined with the causal strength index to form a preliminary causal direction matrix. This matrix not only reflects directionality but also includes strength weights, providing a screening basis for subsequent dynamic time warping delay estimation. The Granger causality test algorithm is used to transform the sliding window feature data from the previous step into a preliminary causal direction matrix that reflects the significant causal relationship and strength between the current patient's physiological indicators, thus providing the initial data basis for real-time causal subgraph updates. For example, in a respiratory monitoring scenario for a patient emerging from anesthesia, the sliding window length is 30 seconds, the overlap rate is 50%, and the sampling frequencies are respiratory rate 1Hz, tidal volume 10Hz, SpO2 and PETCO2 20Hz each, and HRV 250Hz. The characteristic sequence extracted within this window is {respiratory rate: 1 peak per second, tidal volume: mean 450mL ± 40mL, SpO2: decreasing trend -0.2% / second, PETCO2: increasing trend 0.3mmHg / second, HRV: LF / HF ratio increases by 0.5 units}. In the Granger causality test, the maximum lag order is set to 3 seconds, and significance is determined with α = 0.05. The calculated F-statistic for tidal volume → PETCO2 is 6.3, the p-value is 0.002, and the causality strength index is... The F-statistic for tidal volume → SpO2 was approximately 1.35, indicating a significant and strong causal relationship. The F-statistic for tidal volume → SpO2 was 5.1, with a p-value of 0.009 and an intensity index of approximately 1.28, also confirming the relationship. This causal direction matrix was used in subsequent dynamic time warping to calculate accurate time delay parameters, ultimately improving the physiological mechanism explanation capability of the real-time causal subgraph. Clinical validation demonstrated significantly improved accuracy in identifying respiratory depression events and transparent reasoning paths. S4.2: Apply the dynamic time warping algorithm to estimate the time delay of the identified causal direction pair, and calculate the causal response delay time based on the nonlinear alignment path of the two-parameter sequence to obtain the time delay parameter set of the causal relationship; S4.3: The causal direction matrix output by the Granger causality test is fused with the time delay parameter set obtained by dynamic time warping to generate a dynamic causal edge set with time delay annotations, which serves as a local causal subgraph under the current window. For the causal direction matrix output by the Granger causality test, the causal edge identifier matching method (parameter: direction matrix index set) is used to realize the mapping relationship between causal direction data and corresponding node numbers. Furthermore, by using the time delay parameter set (parameter: delay time series) output by the dynamic time warping algorithm, a one-to-one correspondence is achieved between the time delay data and the elements of the direction matrix between causal nodes, resulting in a preliminary combined data structure containing direction and time delay. Furthermore, a matrix fusion algorithm (parameter: weighting coefficient matrix) is used to achieve numerical fusion of the above-mentioned causal direction matrix and the time delay parameter set, and generate a direction-delay mapping table with time delay annotation. The weighting coefficient is adjusted to balance the immediate causal strength and the delay effect based on medical prior knowledge. Furthermore, using a structured recombination method, the direction-delay mapping table is transformed into a dynamic causal edge set, where each edge contains a start node, an end node, a causal confidence weight, and a delay time. Furthermore, by using the adjacency matrix update function, the dynamic causal edge set is written into the local causal subgraph data structure of the current time window, thereby realizing the synchronous update of the causal graph in terms of time delay and direction. Through the above fusion process, the causal direction and time delay data are transformed into a dynamic causal edge set with time delay annotation, realizing the expected technical effects of real-time causal subgraph input preparation and enhanced causal chain interpretability. For example, in the post-anesthesia recovery room environment, for the collected end-tidal carbon dioxide and tidal volume time-series signals, the directional matrix element output by the Granger causality test is SpO2←tidal volume, with a confidence weight of 0.82. The dynamic time warping algorithm calculates the causal response delay between the two as 15 seconds, and the delay parameter set is recorded as

[15] seconds. The matrix fusion algorithm is used to fuse the elements of the directional matrix with the delay parameter set, and the weighted coefficient matrix is ​​used. The fusion weight of tidal volume → SpO2 is set to 0.78, based on the formula:

[0022] in, The weight of a priori medicine is 0.75. The confidence level for Granger is 0.82. The normalized value for the delay time is 0.25. The delay penalty coefficient is 0.1. After calculation, the fusion weight is approximately 0.78. This fusion weight, along with the 15-second delay, is encapsulated into a dynamic causal edge {tidal volume → SpO2, w=0.78, delay=15s} and updated to the local causal subgraph. In this embodiment, the real-time causal subgraph can clearly present the physiological logic that a decrease in tidal volume leads to a decrease in SpO2 after 15 seconds, effectively enhancing the transparency of the judgment process and providing accurate structural input for path probability calculation in subsequent Bayesian inference. S4.4: Compare the generated local causal subgraph with the preset dynamic causal network template to identify causal edges that are not covered in the template but are significantly supported by the current data, and retain new causal relationships with confidence levels higher than the preset level by threshold filtering. S4.5: Update the local subgraph structure of the dynamic causal network template based on the filtered causal edge set to form a real-time causal subgraph reflecting the current physiological state of the patient, which is used as input for the subsequent temporal Bayesian inference model; The set of causal edges after threshold filtering in step S4.4 is received as input data. The set of causal edges includes node pairs, causal direction labels and corresponding time delay parameters. A template structure matching algorithm (parameters: adjacency matrix, node set, edge weight) is used to realize the node index mapping function between the causal edge set and the dynamic causal network template. Furthermore, by using the edge weight update method (parameters: initial weight, confidence coefficient), the confidence values ​​in the causal edge set are mapped to the weight increments of the corresponding template edges, and the updated edge weight matrix is ​​obtained. Furthermore, by using a subgraph replacement algorithm (parameters: local adjacency submatrix, updated edge set), the edge weights and direction labels of the corresponding sub-regions in the template are replaced with the real-time analysis results under the current window, and a local adjacency matrix containing the latest physiological state is generated. Furthermore, a time-delay embedding method (parameters: delay time set Δt, dynamic time warping alignment path) is adopted to embed the delay parameters into the edge attributes of the adjacency matrix and generate a real-time causal network subgraph with time-delay annotations. By using a structured integration process, the updated local adjacency matrix is ​​replaced in its entirety at the corresponding position in the dynamic causal network template, thereby realizing the dynamic transformation of the template into a real-time causal state. The causal subgraph generation algorithm transforms the results of the previous step into a real-time causal subgraph data structure, thereby providing the structured interpretability support required for the input of the subsequent Bayesian inference model. For example, in a respiratory monitoring scenario during anesthesia recovery, the input causal edge set contains five edges: (tidal volume → end-tidal CO2, delay 15s, confidence 0.92), (end-tidal CO2 → SpO2, delay 30s, confidence 0.88), (respiratory rate → tidal volume, delay 5s, confidence 0.95), (HRV → respiratory rate, delay 12s, confidence 0.85), and (SpO2 → HRV, delay 20s, confidence 0.81). A template structure matching algorithm is used to map these edges to template node index positions; for example, tidal volume corresponds to node ID=2, and end-tidal CO2 corresponds to node ID=4. An edge weight update method is used to map the confidence parameter to a weight increment; for example, a confidence of 0.92 corresponds to an increment of 0.15. After updating the local adjacency matrix, a subgraph replacement is performed, replacing the original edge (2→4) in the template with a delayed-labeled version, with the delay parameter embedded as attribute Δt=15. The above process is performed on all five edges, forming a real-time causal network subgraph with delayed annotations. This subgraph serves as input, allowing the subsequent Bayesian model to perform probabilistic inference based on the evidence path of decreased tidal volume leading to increased end-tidal CO2, which in turn causes a decrease in SpO2. In test cases, the interpretability was significantly improved, physician interpretation time was greatly reduced, and the clinical acceptance rate of system alarms was increased.

[0023] Step S5: Input the real-time causal subgraph into a pre-trained temporal Bayesian network model. This model uses the breathing inhibition event as the root node and derives the downstream parameter variation distribution layer by layer through a conditional probability table, outputting a decision vector containing event probability scores. Specifically, this includes: S5.1: Based on medical prior knowledge and statistical characteristics of clinical data, a Bayesian network topology model with causal structure is constructed. The topology model takes the respiratory depression event as the root node and connects parameter nodes such as respiratory rate, tidal volume, blood oxygen saturation, end-tidal carbon dioxide and heart rate variability to form a directed acyclic graph with hierarchical dependencies to support subsequent Bayesian inference calculations. Based on the real-time causal subgraph data structure output from step S4, the set of nodes related to respiratory depression events and causal relationship rules defined in the medical prior knowledge base are loaded as input conditions for the construction of the Bayesian network structure. A directed acyclic graph (DAG) construction method (parameters: node set = respiratory depression events, respiratory rate, tidal volume, blood oxygen saturation, end-tidal carbon dioxide, heart rate variability; edge set is set according to pathophysiological mechanisms) is used to generate the initial network topology. Furthermore, a network structure topology sorting algorithm (parameter: node priority based on the degree of pathological impact) is used to achieve an ordered arrangement of node hierarchical dependencies and generate a causal propagation path index table to ensure the correctness of the dependency order in inference calculations. Furthermore, the adjacency matrix representation method is adopted (parameter: the matrix size is determined by the total number of nodes) to realize the matrix encoding of causal relationships, and the edge weight values ​​from the consensus of medical experts are written into the matrix cells to obtain structured causal weight matrix data; Furthermore, the topology model is subjected to consistency verification (algorithm: loop detection + edge constraint verification) to eliminate potential loops or invalid edge structures and generate a final acyclic causal relationship graph that satisfies the inference conditions; The graph structure construction method described above transforms the real-time causal relationship results from the previous step into executable Bayesian network topology data, thereby preparing structured input for subsequent time-series inference computation. For example, in a scenario involving respiratory depression monitoring of a patient emerging from anesthesia, the input node set includes six parameter nodes: respiratory depression event node (RE), respiratory rate node (RR), tidal volume node (TV), blood oxygen saturation node (SpO2), end-tidal carbon dioxide node (PETCO2), and heart rate variability node (HRV). The edge set is defined by medical mechanisms, such as decreased RR → decreased SpO2, decreased TV → increased PETCO2, increased PETCO2 → decreased SpO2, etc. During construction, the node hierarchy dependency order generated by topological sorting is: RE → {RR, TV} → PETCO2 → SpO2 → HRV. The adjacency matrix has a dimension of 6×6, where the cell values ​​are weights after normalization of expert scores. For example, the edge weight for RR → SpO2 is 0.82, and the edge weight matrix has a weight of 0.76. The edge weight matrix generates corresponding weights by row and column. Acyclicity is detected during consistency verification, and the final output topological data is used for Bayesian conditional probability inference. In this embodiment, the network model can significantly improve the transparency of the physiological logic chain in real-time reasoning based on structured input, ensuring interpretability while maintaining the stability of the probability calculation of respiratory depression events; S5.2: Using the physiological dataset of patients in the historical anesthesia recovery period, parameters are learned for the conditional probability table of the Bayesian network model. The parameter learning is based on the maximum likelihood estimation method to model the joint probability distribution of each node, so as to obtain the state transition probability of each parameter under different respiratory states and form a generalizable reasoning basis. S5.3: Using the real-time causal subgraph generated in S4 as input, initialize the node states in the Bayesian network model. The initialization is based on the multi-dimensional feature vector extracted within the current sliding window, and assigns the prior probability distribution of the current state to each parameter node to construct the initial evidence set required for inference. Based on the real-time causal subgraph generated in step S4 as the input dataset, a node state initialization method (parameters: node ID set, causal edge weight matrix) is used to map the current physiological state to the corresponding node in the temporal Bayesian network topology. Furthermore, through a feature matching algorithm (parameters: node physiological type, sliding window feature vector index), a one-to-one correspondence is achieved between the causal subgraph nodes and the multidimensional feature vectors extracted within the sliding window, and a node-feature value binding table is obtained; Furthermore, a probability distribution fitting method (parameters: feature value sample group, historical probability model) is adopted to estimate the prior probability distribution of the current state of each parameter node and generate the node prior probability vector; Furthermore, based on the conditional independence hypothesis testing algorithm (parameters: node prior probability vector, causal edge set), the consistency of the initialization assignment is verified, abnormal assignment records that contradict medical priors are eliminated, and the prior probability set that conforms to the causal relationship is retained; By using node initialization assignment, the structural information in the real-time causal subgraph is transformed into the initial evidence set for Bayesian inference, thereby enabling the model to construct accurate inference inputs in the current time series state. For example, in a monitoring scenario in an anesthesia recovery room, the real-time causal subgraph includes a set of nodes {respiratory rate, tidal volume, SpO2, PETCO2, HRV} and a corresponding set of causal edges. Feature matching binds the current feature value of the respiratory rate node to 8 breaths per minute, the tidal volume node to 350 ml, the SpO2 node to 92%, the PETCO2 node to 50 mmHg, and the HRV node to a low-frequency / high-frequency ratio of 2.1. A prior distribution of the respiratory rate values ​​is generated using a probability distribution fitting method, as shown in the following formula:

[0024] in, The standard deviation represents the historical statistical values. A normal model was used to estimate the prior distribution of the tidal volume value of 350 ml. After testing for the conditional independence hypothesis, a prior probability set consistent with medical logic was obtained: respiratory rate node 0.82, tidal volume node 0.77, SpO2 node 0.65, PETCO2 node 0.88, and HRV node 0.59. This set was used as the initialization input for the Bayesian network. The model significantly improved the probability of inferring respiratory depression events at this time, demonstrating the enhanced accuracy and clinical interpretability of the initialization method based on real-time causal subgraphs and feature matching. S5.4: Perform temporal inference calculation on the network model based on Bayesian update rules. The inference calculation recursively updates the influence of the parent node on the child node and uses a conditional probability table to derive the posterior probability distribution of the respiratory inhibition event in the current time window to generate a decision vector containing the probability of the event occurring. S5.5: Normalize the generated decision vector to obtain a probability score for respiratory depression events. The score is standardized by the softmax function to the original posterior probability and outputs a value in the range of [0,1] for subsequent alarm threshold judgment and visualization of evidence chain generation. The decision vector generated by S5.4 is normalized using the Softmax normalization method (parameter: temperature coefficient τ is set to 1.0) to standardize the posterior probability distribution. Furthermore, through the exponential operation module (parameter: base e), an exponential transformation is performed on each element in the decision vector to obtain the exponential value sequence corresponding to each event; Furthermore, the summation calculation module performs a summation operation on the above exponential value sequence to obtain the numerical result of the normalization factor; Furthermore, by using the proportional operation module, each exponent value is divided by the normalization factor to achieve the normalization mapping of the probability values ​​and obtain a standardized probability sequence in the range [0,1]. Furthermore, the probability values ​​of the corresponding locations of respiratory depression events are extracted through the index mapping module to form event probability score data; By using Softmax normalization, the temporal Bayesian inference results from the previous step are transformed into numerical indicators that can be directly used for alarm threshold determination and evidence chain visualization, thereby achieving the stability and clinical usability of the probability score. For example, in the anesthesia recovery room scenario, the input parameters are taken from the decision vector [0.35, 0.60, 0.05] in S5.4, the temperature coefficient τ is set to 1.0, the exponential operation yields [1.419, 1.822, 1.051], the summation result is 4.292, the proportional operation yields the normalized probability sequence [0.331, 0.424, 0.245], and the index mapping yields the probability score of 0.424 for the respiratory depression event. In this scenario, the system compares the score value with the preset alarm threshold of 0.40, successfully determines that the alarm condition has been met, and provides a stable probability input for the subsequent generation of the evidence chain in S6. This example verifies the effect of the Softmax function in maintaining a stable score range and improving numerical interpretability under different input distributions.

[0025] Step S6: Generate an explicit medical logic chain based on a Bayesian inference path. This logic chain consists of nodes and edges in a causal subgraph, containing a temporal causal sequence of abnormal initial parameters, changes in mediating effect parameters, and deviations in terminal parameters, forming an interpretable chain of evidence. Specifically, this includes: S6.1: Based on the output decision vector of the temporal Bayesian network model, extract the node activation path related to the respiratory depression event. The node activation path includes the respiratory depression root node and its downstream child node paths in the conditional probability table to obtain the complete causal propagation path of model inference. S6.2: Perform causal directionality analysis on the extracted node activation path, and combine the directed edge information in the causal subgraph to identify the causal temporal relationship between each node in the path, so as to determine the logical order of the initial anomaly parameter, the mediation effect parameter and the terminal deviation parameter. S6.3: Based on causal time sequence relationships and timestamp annotations, time alignment processing is performed on the parameters of each node, and the time delay parameters between nodes are calculated using the dynamic time warping algorithm to construct a causal propagation sequence containing the time dimension; S6.4: Based on the time-aligned causal propagation sequence, generate an explicit medical logic chain consisting of 'cause → mediator → result'. The logic chain is represented in the form of a structured causal path, including the abnormal state of each node parameter and its propagation order on the time axis. For the time-aligned causal propagation sequence, a structured path reorganization algorithm (parameters: sequence event set, causal directed edge set, time delay matrix) is used to combine each abnormal node and its propagation direction into the starting segment, intermediate segment and terminal segment of the medical logic chain; Furthermore, by using state labeling and event classification methods (parameters: node parameter type, abnormal amplitude threshold), we can clearly identify the abnormality of the parameter status of each node and obtain a node information table containing physiological parameter category, abnormal attribute and temporal position. Furthermore, by using the temporal path encoding method (parameters: event category sequence, time index set), the causal node sequence is transformed into a directed path encoding structure, and a logical path matrix arranged in the pattern of "cause → mediator → result" is generated. Furthermore, a path sorting and time series mapping method (parameters: logical path matrix, timestamp mapping table) is adopted to arrange the logical paths sequentially on the time axis, ensuring that the causal chain maintains the same temporal order as the actual propagation in the structured representation; By using path structuring, the time-aligned sequence results of the previous step are transformed into structured causal path data containing anomalies in initial parameters, changes in mediating effect parameters, and deviations in terminal parameters, thereby generating explicit medical logic chains and enhancing interpretability. For example, in a scenario for determining respiratory depression during anesthesia recovery, the input causal propagation sequence includes decreased respiratory rate (time 0s), increased end-tidal carbon dioxide (time 15s), and decreased blood oxygen saturation (time 45s). A structured path reconstruction algorithm is used to reorganize the node order into... The causal chain is established. Using a state labeling method, a decrease in respiratory rate is marked as an abnormal state below the lower limit of normal, an increase in end-tidal carbon dioxide is marked as an abnormal state exceeding the upper limit of normal, and a decrease in blood oxygen saturation is marked as an abnormal state below the lower limit of normal. Path encoding is then used to encode this chain in the time series. The temporal index vector is assigned and mapped to the node event category. A time-series mapping method is used to ensure that the causal chain is arranged in the structured representation matrix according to the aforementioned temporal order, thereby generating an explicit medical logical chain. This logical chain accurately reflects the temporal characteristics and causal direction of abnormal propagation in the visualization interface, providing clinicians with a physiologically explainable basis for event judgment. S6.5: Perform semantic mapping processing on the generated explicit medical logic chain, and transform the parameter nodes and their states into clinically understandable terminology based on the medical terminology mapping table to form an interpretable evidence chain text output.

[0026] Step S7: Convert the evidence chain into a natural language description and generate a corresponding heatmap visualization matrix. The horizontal axis of the matrix represents time nodes, the vertical axis represents parameter dimensions, and the color intensity represents the contribution weight of each feature to the decision vector. Specifically, this includes: S7.1: Perform structured parsing on the causal node sequence in the Bayesian inference path. The causal nodes include abnormal initial parameters, changes in mediation effect parameters, and deviations in terminal parameters. Based on predefined medical semantic mapping rules, convert them into natural language fragments to form a preliminary logical description text. S7.2: Based on the timestamp information of the changes in each parameter in the causal path, perform time alignment processing on the natural language fragments to ensure that the order of events in the descriptive text is consistent with the actual physiological change process, and generate a complete medical logic chain description containing time stamps; S7.3: Normalize the parameter change intensity at each time point and the conditional probability value in the Bayesian network to obtain the contribution weight coefficient of each parameter in the decision vector. Based on this coefficient, construct a parameter-time two-dimensional matrix to provide a data foundation for subsequent heat map visualization. S7.4: The color mapping function is used to color-encode the values ​​in the parameter-time two-dimensional matrix, and the color intensity represents the contribution weight of the parameter to generate a heat map visualization matrix, where the horizontal axis represents the time node, the vertical axis represents the parameter dimension, and the color distribution reflects the degree of influence of each feature on the determination of respiratory depression events. Based on the parameter-time two-dimensional matrix output by step S7.3, the input data is a numerical matrix structure containing the contribution weight coefficients of parameters at each time point. The horizontal axis corresponds to the time point sequence, the vertical axis corresponds to the parameter dimension label, and the values ​​in the matrix represent the normalized contribution weight coefficients. A color mapping function (parameters: mapping interval [0,1], color gradient scheme is the standard spectrum for medical visualization) is used to map the values ​​in the matrix to the corresponding color intensity levels, so as to intuitively represent the changing trend of parameter contribution weights at different time points. Furthermore, a linear interpolation algorithm (parameter: interpolation precision ≥ 0.01) is used to achieve a continuous mapping from numerical values ​​to color codes, and a color coding matrix is ​​obtained. Each element of this matrix is ​​an RGB or RGBA color value, which is used for color rendering of the two-dimensional matrix. Furthermore, by employing matrix traversal and position indexing methods (parameters: the horizontal axis index is a timestamp sequence, and the vertical axis index is a parameter dimension sequence), the color coding matrix is ​​precisely located to the corresponding parameter-time coordinate point, and a data structure that binds position and color is generated. Furthermore, the bound data structure is visualized by using the rasterization method of the rendering engine (parameter: resolution ≥ 300 DPI) to generate a heatmap visualization matrix. The color intensity is distributed gradually according to the weight values, and the image clarity is ensured to meet medical display standards. By using color mapping and rendering, the parameter-time weight matrix from the previous step is transformed into heatmap data with both time and parameter axes, thus achieving an intuitive visualization of the contribution weights of each feature in the determination of respiratory depression events. For example, for time-series monitoring data of patients emerging from anesthesia, 60 time points are set with 1-second intervals. The parameters include five categories: respiratory rate, tidal volume, SpO2, PETCO2, and HRV. The normalized contribution weight matrix has a maximum value of 0.92 and a minimum value of 0.05. A linear color mapping function is used, assigning a weight of 0.05 to medical standard blue (0,0,255) and a weight of 0.92 to medical standard red (255,0,0). Interval values ​​are calculated using linear interpolation; for example, a weight of 0.5 corresponds to the color (128,64,64). The calculation formula is as follows:

[0027] in This is the current matrix element value. It is 0.05. It is 0.92. and This is an RGB vector. The calculated color code... Filling the matrix to 60×5, corresponding to the horizontal axis time nodes and the vertical axis parameter dimensions, the final rendered heatmap can clearly show the dynamic visualization effect of the tidal volume weight increasing significantly at t=15s and the SpO2 weight increasing sharply at t=45s on a high-resolution display, which makes it easy for doctors to quickly identify and adopt the system judgment results in the PACU scenario. S7.5: Synchronously encapsulate the natural language description text and the heatmap visualization matrix to generate an interactive visual evidence chain module, allowing clinicians to view the complete reasoning trajectory through the monitoring system interface, thereby enhancing their trust in and willingness to adopt the system's judgment.

[0028] Step S8: When the probability score of respiratory depression events exceeds a preset threshold, a multimodal alarm signal is simultaneously triggered and the corresponding visual evidence chain is locked. The alarm signal includes audible and visual prompts and electronic medical record markers. The locking duration is set according to the clinical treatment process. Specifically, it includes: S8.1: Based on the real-time output of the probability score of respiratory depression events, the score results are compared with thresholds to determine whether the alarm triggering conditions for respiratory depression events are met, and the comparison result Boolean value is obtained. S8.2: When the Boolean value of the comparison result is true, a multimodal alarm signal trigger instruction is generated. The instruction includes an audible and visual alarm control signal and an electronic medical record mark writing instruction to realize immediate warning and event recording for clinical staff. The Boolean value of the comparison result output from step S8.1 is used as the trigger condition input to the alarm signal generation module. An event-driven mechanism (parameter: trigger condition = true) is used to realize the alarm command initialization configuration. Furthermore, by using the sound and light alarm control signal generation method (parameters: sound output frequency range 300Hz-3kHz, light flashing frequency 1Hz-5Hz), the behavior parameters of the physical alarm subsystem are set, and a data structure including sound signal configuration file and light signal control sequence is obtained. Furthermore, by using the electronic medical record tag writing instruction generation method (parameters: tag type = respiratory depression event, tag timestamp = NTP synchronization time, tag priority = emergency), alarm events are registered in the electronic medical record system in real time, and a database update transaction package containing event ID, triggering conditions, and evidence chain index is generated. Furthermore, an alarm instruction encapsulation algorithm (parameters: protocol type = HTTP or MQTT, encryption method = TLS1.2, data format = JSON) is adopted to realize the multimodal combination encapsulation of audible and visual alarm control signals and electronic medical record mark writing instructions, and generate a unified alarm signal frame for monitoring equipment and information systems; By using a multimodal alarm signal synchronous triggering processing method, the Boolean triggering condition of the previous step is transformed into software instructions that can directly drive the audio-visual alarm hardware and electronic medical record data updates, thereby achieving the technical effect of immediate warning and event recording for clinical staff. For example, in a post-anesthesia recovery room scenario, the probability score for a respiratory depression event is 0.87, the preset threshold is 0.8, and after the comparison result in step S8.1 is true, the alarm signal generation module receives the trigger condition and calls the audible and visual alarm control method, setting the sound output frequency to 1000Hz for a duration of 5 seconds, and the light signal flashing frequency to 2Hz for a duration of 10 seconds. The electronic medical record tag writing instruction sets the event type to "respiratory depression event", the timestamp to 2024-04-18T10:15:30Z, the priority to "urgent", the event ID to "RDE-20240418-101530", and the evidence chain index to the corresponding real-time causal subgraph number "CG-20240418-15". The alarm command encapsulation algorithm uses the MQTT protocol and TLS 1.2 encryption. It combines the audible and visual control signals with the medical record marker transaction packet to generate a unified alarm frame with a length of 512 bytes. This frame is sent to the monitoring system via the network interface to achieve synchronous updates of hardware alarms and EMR records. Clinical personnel have confirmed that this process is completed within 2 seconds after the event occurs, significantly improving real-time performance. S8.3: Call the preset visual evidence chain locking module, based on the Bayesian inference path and causal subgraph of the current respiratory depression event, to lock the complete medical logic chain and heat map visualization matrix corresponding to the event for subsequent review and backtracking; S8.4: Based on the preset lock duration parameters in the clinical treatment process, the lock status is controlled for time. The countdown operation is performed through the timer module to ensure that the chain of evidence remains locked within the specified time and to obtain the lock status flag bit. S8.5: Feedback the audible and visual alarm status, electronic medical record marking information, and lock status flag to the system status monitoring module to achieve full lifecycle management of alarm events and real-time feedback of system operation status.

[0029] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of intelligent identification and alarm of respiratory depression events during the post-anesthesia recovery period, characterized in that, The method comprises the following steps: S1: collecting continuous time series data from a multi-channel physiological monitoring device of a patient in the anesthesia recovery period, and setting the sampling frequency of each parameter according to clinical standards and marking the time stamp; S2: performing sliding window segmentation processing on the collected original physiological signals to generate a time series data block sequence containing a multi-dimensional feature vector; S3: constructing a dynamic causal network template based on a medical prior knowledge base, wherein the dynamic causal network template contains a node set and a directed edge set, wherein the node represents a physiological parameter, and the directed edge represents a known pathophysiological correlation, and the weight of each edge is determined by an expert consensus method; S4: using Granger causality test and dynamic time warping algorithm to analyze the current time series data block to identify significant causal relationship direction and time delay parameter, updating the local subgraph structure of the dynamic causal network template according to the test result, and forming a real-time causal subgraph; S5: inputting the real-time causal subgraph into a pre-trained time series Bayesian network model, the model taking respiratory depression event as a root node, deriving the downstream parameter change distribution layer by layer through a conditional probability table, and outputting a decision vector containing an event probability score; S6: generating an explicit medical logic chain based on the Bayesian inference path, the logic chain being composed of nodes and edges in the causal subgraph to form an interpretable evidence chain; S7: converting the evidence chain into a natural language description and generating a corresponding heat map visualization matrix.

2. The method of claim 1, wherein the method further comprises: The step S7 further comprises: S8: when the probability score of the respiratory depression event exceeds a preset threshold, synchronously triggering a multi-modal alarm signal and locking the corresponding visual evidence chain.

3. The method of claim 1, wherein the method further comprises: The step S1 specifically comprises: Based on the physiological characteristics of patients in the anesthesia recovery period during the recovery period, a multi-channel physiological signal acquisition device is configured to acquire key physiological parameters; According to the clinical monitoring standard, the sampling frequency of each parameter is set; Performing pre-filtering and analog-to-digital conversion processing on the key physiological parameters to obtain a high-quality digital physiological signal sequence; Based on the system clock synchronization mechanism, timestamp the channel signals, and use NTP protocol combined with GPS clock source for time alignment; Store the multi-channel physiological signals with timestamp in the time series database, and organize them in the form of triplets of channel, timestamp and value, and generate a structured data stream.

4. The method of claim 3, wherein the method further comprises: The physiological signal acquisition device includes a respiratory monitoring belt, a pulse oximeter, an end-tidal carbon dioxide sensor and an electrocardiogram module, which acquires key physiological parameters including respiratory rate, tidal volume, oxygen saturation, end-tidal carbon dioxide and heart rate variability.

5. The method of claim 1, wherein the method further comprises: The step S2 specifically comprises: Based on the physiological characteristics of the respiratory cycle, calculate the window length parameter; Perform segmentation processing on the continuous physiological signals using a sliding time window mechanism, wherein the moving step of the sliding time window mechanism is calculated according to the window length and overlap rate; Performing signal preprocessing operation on the original physiological signal segments intercepted by the sliding window to generate preprocessed physiological signal segments; Based on the preprocessed physiological signal segments, extract a multi-dimensional feature vector to form a multi-dimensional feature vector set; Organize the multi-dimensional feature vector set into a time series data block sequence in chronological order.

6. The intelligent identification and alarm method for respiratory depression events during the anesthetic recovery period according to claim 1, characterized in that, The step S3 specifically comprises: Based on clinical medical guidelines and pathophysiological mechanisms, a set of key physiological parameters related to respiratory depression is defined as a set of nodes in a dynamic causal network, and an initial graph structure driven by medical knowledge is constructed; According to known pathophysiological correlation mechanisms, a set of directed edges between nodes is established, which represents the causal relationship direction between parameters, and a causal graph structure is constructed; Using expert consensus method, an initial weight is assigned to each directed edge; Based on the expert consensus score results, the fuzzy comprehensive evaluation method is used to normalize the causal relationship weight, and a standardized causal network edge weight matrix is obtained; The initialized node set and weighted directed edge set are integrated into a dynamic causal network template, which is represented in the form of an adjacency matrix.

7. The intelligent identification and alarm method for respiratory depression events during the anesthetic recovery period according to claim 6, characterized in that, The step S3 further comprises defining a node set based on clinical medical guidelines and expert consensus pathophysiological mechanisms, the nodes including respiratory rate, tidal volume, blood oxygen saturation, end-tidal carbon dioxide, and heart rate variability, and obtaining directed causal edge weights through expert multi-center scoring and literature support, normalizing using the fuzzy comprehensive evaluation method, using an adjacency matrix to represent the network, and performing real-time Granger causality test and causal edge and time delay fusion update of dynamic time warping.

8. The method of claim 1, wherein the method further comprises: The step S4 specifically comprises: Based on the static causal network template constructed based on the medical prior knowledge base, the Granger causality test is performed on the multi-dimensional physiological time series data blocks segmented by the sliding window, the significant causal relationship direction between the parameters in the current data stream is identified, and a preliminary causal direction matrix is obtained; The dynamic time warping algorithm is applied to the identified causal direction pair for time delay estimation, the causal response delay time is calculated based on the nonlinear alignment path of the two parameter sequences, and a time delay parameter set of the causal relationship is obtained; The preliminary causal direction matrix and the time delay parameter set are fused to generate a dynamic causal edge set as a local causal subgraph under the current window; The local causal subgraph and the preset dynamic causal network template are structurally compared to identify causal edges that are not covered in the template but are significantly supported by the current data, and the newly added causal relationships with a confidence level higher than a preset level are retained through threshold screening; The local subgraph structure of the dynamic causal network template is updated based on the screened causal edge set to form a real-time causal subgraph.

9. The method of claim 1, wherein the method further comprises: The step S5 specifically comprises: Based on the medical prior knowledge and the statistical characteristics of clinical data, a Bayesian network model with causal structure is constructed; Using the physiological data set of the historical anesthesia recovery period patients, parameter learning is performed on the conditional probability table of the Bayesian network model, and the parameter learning is based on the maximum likelihood estimation method to model the joint probability distribution of each node; The generated real-time causal subgraph is input to initialize and assign values to the node states in the Bayesian network model, and the initialization and assignment are based on the multi-dimensional feature vector extracted in the current sliding window; Based on the Bayesian update rule, the Bayesian network model is executed for time series reasoning calculation to generate a decision vector; The decision vector is normalized to obtain a probability score of the respiratory depression event.

10. The intelligent identification and alarm method for respiratory depression events during the anesthetic recovery period according to claim 9, characterized in that, The score is normalized by a softmax function on the raw posterior probabilities, outputting values in the range [0, 1].

Images