Postanesthesia reviving delay risk assessment method based on adaptive fuzzy reasoning

Abstract

The invention discloses an anesthesia post-operation awakening delay risk assessment method based on adaptive fuzzy reasoning. The method comprises the following steps: S1, obtaining a normalized multi-modal physiological information data stream with consistent time sequence; s2, dividing the standardized multi-modal physiological information data stream into a high-frequency physiological information data substream and a low-frequency physiological information data substream according to a sampling frequency; s3, generating a high-frequency fuzzy reasoning result based on the initial high-frequency fuzzy rule base and an adaptive membership function; s4, generating a low-frequency fuzzy reasoning result based on the initial low-frequency fuzzy rule base and an adaptive membership function; s5, outputting a multi-scale fusion risk scoring curve; and S6, generating a wake-up delay risk alarm signal and recording a risk driving factor list. According to the method, the risk identification sensitivity and robustness of high-risk patients, complex operations and multi-complication cases are remarkably improved, and missing report and false report caused by a single-dimension or static threshold value are avoided.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a method for assessing the risk of delayed awakening after anesthesia based on adaptive fuzzy reasoning. Background Art

[0002] With the rapid development of digital medicine and intelligent monitoring technology, anesthesia management has gradually evolved towards intelligence, personalization and high-risk prevention and control. The risk assessment of delayed awakening after anesthesia has become one of the core functions of the perioperative intelligent monitoring system. At this stage, hospitals generally adopt risk warning methods based on single physiological indicator thresholds or multi-indicator static statistical models.

[0003] Specifically, the current mainstream risk assessment systems mostly make static judgments based on empirical thresholds, ignoring the changes in anesthetic sensitivity caused by individual differences in patients' age, weight, and liver and kidney function, resulting in high false alarm rates for high-risk patients and missed alarm rates for low-risk patients. At the same time, traditional methods are usually unable to integrate the complex dynamic relationships between multimodal data during surgery in real time, making it difficult to accurately reflect the risk of delayed awakening.

[0004] On the other hand, with the increasing digitalization and data interconnection of operating rooms, the amount of clinical monitoring data continues to grow. The traditional approach of relying on manual observation and static rule judgment can no longer meet the needs of efficient and personalized risk warning in real-time high-concurrency environments. Existing technologies have limitations in the effective fusion of multi-scale signals, the identification of abnormal short-term fluctuations, and the dynamic modeling of trend risks. They are insufficient in risk identification of groups with a high incidence of delayed awakening and lack a systematic adaptive optimization capability and dynamic evolution mechanism.

[0005] Therefore, there is an urgent need to propose a new risk assessment method that can integrate multiple scales, dynamically adapt, have a transparent mechanism, and facilitate clinical decision support, so as to better meet the needs of intelligent and precise monitoring in modern operating rooms. Summary of the Invention

[0006] One purpose of the present invention is to propose a method for assessing the risk of delayed postoperative awakening after anesthesia based on adaptive fuzzy reasoning. The present invention significantly improves the sensitivity and robustness of risk identification for high-risk patients, complex surgeries and cases with multiple complications, and avoids missed reports and false alarms caused by single dimensions or static thresholds.

[0007] According to an embodiment of the present invention, a method for assessing the risk of delayed recovery after anesthesia based on adaptive fuzzy reasoning includes the following steps:

[0008] S1. Collect and preprocess the patient's multimodal physiological information data stream to obtain a normalized multimodal physiological information data stream with consistent time sequence;

[0009] S2. Dividing the normalized multimodal physiological information data stream into a high-frequency physiological information data sub-stream and a low-frequency physiological information data sub-stream according to the sampling frequency;

[0010] S3. Establish a high-frequency fuzzy inference channel and input a high-frequency physiological information data substream into the high-frequency fuzzy inference channel to generate a high-frequency fuzzy inference result based on the initial high-frequency fuzzy rule base and the adaptive membership function;

[0011] S4. Establish a low-frequency fuzzy inference channel and input a low-frequency physiological information data substream into the low-frequency fuzzy inference channel to generate a low-frequency fuzzy inference result based on the initial low-frequency fuzzy rule base and the adaptive membership function;

[0012] S5. Input the high-frequency fuzzy inference results and the low-frequency fuzzy inference results into the fuzzy fusion module to fuse the high-frequency fuzzy inference results and the low-frequency fuzzy inference results, and output a multi-scale fusion risk score curve;

[0013] S6. Compare the multi-scale fusion risk score curve with the risk safety threshold in real time. If the multi-scale fusion risk score curve exceeds the risk safety threshold at any time, a delayed awakening risk alarm signal is generated and a list of risk driving factors is recorded. The multi-scale fusion risk score curve, delayed awakening risk alarm signal and risk driving factor list are pushed to the anesthesia information interaction terminal for reference and intervention by clinical personnel.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Collecting a multimodal physiological information data stream of the patient, including a time series of hemodynamic parameters, a time series of bispectral index, a time series of end-tidal concentration of a gaseous anesthetic agent, a time series of intraoperative drug infusion parameters, and a static basic feature vector of the patient;

[0016] S12. Perform synchronous alignment processing on the collected time series physiological information data, set a unified time step, resample all time series to the same set of time points, and obtain the original aligned data matrix under the unified time axis;

[0017] S13. Perform missing value compensation on each item in the original aligned data matrix;

[0018] S14. Perform outlier detection and elimination on the data matrix after missing value compensation to generate a cleaned data matrix;

[0019] S15. Performing normalization processing on the cleaned data matrix to obtain a normalized data matrix;

[0020] S16. Concatenate the normalized data matrix with the patient's static basic feature vector to obtain the final feature input matrix.

[0021] Optionally, the hemodynamic parameter time series is the systolic and diastolic blood pressure values ​​at each moment, the bispectral index time series is the bispectral index at each moment, the end-tidal concentration time series of the gas anesthetic is the minimum alveolar anesthetic concentration at each moment, the intraoperative drug infusion parameter time series is the drug infusion rate at each moment, including the dosage of propofol and remifentanil, and the patient's static basic feature vector includes immutable individual attributes and preoperative evaluation data including age, body mass index, glomerular filtration rate, alanine aminotransferase and surgery type.

[0022] Optionally, S2 includes the following steps:

[0023] S21. Set the time sampling window set W = {W H ,W L}, where W H is the high-frequency sampling window used to extract second-level physiological fluctuations, W L A low-frequency sampling window used to capture minute-level anesthetic metabolic trends;

[0024] S22. Based on the uniform sampling step Δt, the final feature input matrix Perform sliding window segmentation processing to extract high-frequency time period sequence blocks and low-frequency time period sequence blocks respectively, and obtain two independent data stream sets:

[0025] X H ={X final [t:t+W H -1]|t=1,1+Δt,…,TW H +1};

[0026] X L ={X final [t:t+W L -1]|t=1,1+Δt,…,TW L +1};

[0027] Among them, X H represents the set of high-frequency physiological information data substreams, X L represents the set of low-frequency physiological information data substreams, P represents the patient's static basic feature vector, T is the total number of sampling time steps, and t is the time index variable, which represents each data sampling time point in the entire monitoring period, with a value range from 1 to T;

[0028] S23. High-frequency physiological information data sub-stream set X H Each submatrix in Extracting volatility eigenvectors The volatility eigenvector is composed of the range, variance and first-order difference series mean of each mode in the time period;

[0029] S24. Low-frequency physiological information data sub-stream set X L Each submatrix in Extract trend feature vector The trend feature vector consists of the linear regression slope, mean, and moving average of each mode in the time period;

[0030] S25. Arrange the volatility feature vectors of all high-frequency physiological information data substreams in chronological order as a high-frequency reasoning input sequence Arrange the trend feature vectors of all low-frequency physiological information data substreams in chronological order as a low-frequency reasoning input sequence

[0031] Optionally, S3 includes the following steps:

[0032] S31. Constructing a high-frequency fuzzy inference channel, the high-frequency fuzzy inference channel is used to receive a high-frequency inference input sequence, each volatility feature vector in the high-frequency inference input sequence includes multiple modal features;

[0033] S32. Aiming at the individual differences in anesthesia sensitivity of patients during delayed awakening after anesthesia, an adaptive membership function adjustment mechanism of the high-frequency fuzzy inference channel is designed. The adaptive membership function in the adaptive membership function adjustment mechanism Used to measure the fuzzy membership of each input feature to risk reasoning in the current time window:

[0034]

[0035] in, is the mean center of the membership function dynamically corrected based on the previous postoperative awakening delay feedback, is the fuzzy expansion coefficient updated online based on the risk prediction error of the wake-up feedback, represents the adaptive parameter of the i-th feature in the k-th time window;

[0036] S33. Build a high-frequency fuzzy rule base, which consists of multiple fuzzy rules. Each fuzzy rule contains multiple preconditions and a fuzzy output result. The condition is that each modal feature in the current time window belongs to a fuzzy set. The output result is the awakening delay risk level corresponding to the fuzzy rule.

[0037] S34. For each fuzzy rule, calculating the activation strength value of the fuzzy rule under the current volatility feature vector;

[0038] S35. Perform sensitivity screening on all activation intensity values, and only retain fuzzy rules whose activation intensity values ​​exceed the dynamic threshold for inference output in the current time window;

[0039] S36. Perform fuzzy weighted fusion on all fuzzy rules that pass the screening, and calculate the high-frequency fuzzy inference results in the current time window.

[0040] Optionally, the S4 includes the following steps:

[0041] S41. Construct a low-frequency fuzzy inference channel, where each trend feature vector in the low-frequency inference input sequence contains multiple modal features;

[0042] S42. Design an adaptive membership function adjustment mechanism in the low-frequency fuzzy inference channel. The low-frequency adaptive membership function is used to measure the fuzzy membership degree of each trend input feature to the risk of delayed awakening within the current time window, which is used to adapt to the differences in anesthetic drug metabolism capacity among different patients.

[0043] S43. Construct a low-frequency fuzzy rule base, which consists of multiple fuzzy rules for drug metabolism trend modeling. Each fuzzy rule contains multiple fuzzy preconditions and a fuzzy output result. The fuzzy output result is the awakening delay risk level corresponding to the fuzzy rule;

[0044] S44. Calculating the activation strength value of each low-frequency fuzzy rule under the current trend feature vector;

[0045] S45. Introduce a trend sensitivity weighting mechanism to perform weighted adjustment on the activation strength value of each fuzzy rule. Based on the consistency between the anesthetic drug concentration trend and the EEG index trend in the current time window, dynamically increase the sensitivity weight of the pharmacokinetic-related modality and recalculate the weighted activation strength value of the rule under the current input.

[0046] S46. Perform weighted averaging on the weighted activation strength values ​​of all fuzzy rules in each time window and their corresponding risk level output values ​​to obtain a low-frequency fuzzy inference result in the time window;

[0047] S47. Arrange the low-frequency fuzzy inference results of each time window in chronological order to obtain a complete low-frequency fuzzy inference result sequence.

[0048] Optionally, the S5 includes the following steps:

[0049] S51. Aligning the high-frequency fuzzy inference result sequence with the low-frequency fuzzy inference result sequence according to the time index to obtain a set of corresponding values;

[0050] S52. For each time index, calculate the high-frequency credibility coefficient and the low-frequency credibility coefficient of the high-frequency risk score value and the low-frequency risk score value respectively;

[0051] S53. Normalize the high-frequency credibility coefficient and the low-frequency credibility coefficient, and weight the high-frequency risk score and the low-frequency risk score based on the normalized high-frequency weight coefficient and the low-frequency weight coefficient to obtain a fused score.

[0052] S54. Dynamically smooth the fusion score value to obtain a smoothed score value. The smoothed score value is used to suppress local spike noise caused by short-term signal fluctuations. The smoothed score values ​​of all time points are connected and arranged in chronological order to obtain a curve. The curve is the multi-scale fusion risk score curve.

[0053] Optionally, the S6 includes the following steps:

[0054] S61. Set a risk safety threshold, which is determined based on the statistical distribution of historical large-scale postoperative recovery data, different surgical types, and clinical factors of patient characteristics;

[0055] S62. Each score value of the multi-scale fusion risk score curve is compared with the risk safety threshold in real time. If the multi-scale fusion risk score value at any time point exceeds the risk safety threshold, a wake-up delay risk warning signal is automatically generated;

[0056] S63. For the time point at which the wake-up delay risk warning signal is generated, the inference engine is called to record the risk driving factor list of the multi-scale fusion risk score curve at that moment;

[0057] S64. Push the current multi-scale fusion risk score curve, awakening delay risk warning signal and risk driving factor list to the anesthesia information interaction terminal in real time through the data interface. The anesthesia information interaction terminal is used to synchronously display the score curve, warning signal and driving factor details on the terminal platform for real-time reference by clinical staff.

[0058] Optionally, the risk safety threshold classification rules are as follows:

[0059] When the multi-scale fusion risk score curve is continuously greater than or equal to 0.8 at any time point, or the high-frequency risk score value is continuously greater than or equal to 0.85, or the low-frequency risk score value is continuously greater than or equal to 0.75, the system automatically outputs a high-risk warning;

[0060] When the multi-scale fusion risk score curve is between 0.5 and 0.8 at any time point, or the high-frequency risk score value is between 0.6 and 0.85, or the low-frequency risk score value is between 0.5 and 0.75, the system outputs a medium risk warning;

[0061] When the multi-scale fusion risk score curve is lower than 0.5 at any time point, and both the high-frequency risk score value and the low-frequency risk score value are lower than 0.5, the system determines it as low risk and maintains routine monitoring.

[0062] The beneficial effects of the present invention are:

[0063] (1) The present invention constructs a high-frequency fuzzy inference channel and a low-frequency fuzzy inference channel, and designs a multi-scale risk scoring fusion mechanism, which can collaboratively model second-level physiological signals and minute-level pharmacokinetic trends, thereby realizing full-cycle, continuous, and dynamic perception of the risk of delayed awakening after anesthesia. The fusion module dynamically allocates high- and low-frequency inference results based on adaptive weights according to the current risk contribution, and combines a dynamic smoothing mechanism to effectively suppress misjudgments caused by short-term fluctuations, significantly improving the risk identification sensitivity and robustness for high-risk patients, complex surgeries, and cases with multiple complications, and avoiding omissions and false alarms caused by single dimensions or static thresholds.

[0064] (2) The present invention introduces an adaptive membership function updating method based on individual patient characteristics, actual postoperative feedback, and model prediction errors. It can perform membership self-learning and real-time fine-tuning of rule base parameters for each input variable in the fuzzy reasoning channel according to the multi-dimensional static and dynamic characteristics of the patient's age, weight, liver and kidney function, and surgery type. The high-frequency and low-frequency channels adopt sensitivity weighting and trend weighting mechanisms, respectively, to achieve individualized risk level output in the high-risk group of delayed awakening.

[0065] (3) The present invention supports the real-time display of risk score curves, short-term fluctuations, and key indicators of trend changes through explicit management of fuzzy rules throughout the entire process and visual output of factor contribution weights, providing clinical anesthesiologists with a clear risk evolution trajectory and a basis for ranking influencing factors. The system has a variety of clinical interaction modes, including hierarchical display of high-frequency and low-frequency information, intelligent triggering of risk warning thresholds, and intelligent push of intervention suggestions, which can provide early warning of high-risk situations of delayed postoperative awakening. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0067] Figure 1 This is a flow chart of a method for risk assessment of delayed awakening after anesthesia based on adaptive fuzzy reasoning proposed by the present invention. DETAILED DESCRIPTION

[0068] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0069] refer to Figure 1 A method for assessing the risk of delayed recovery after anesthesia based on adaptive fuzzy reasoning includes the following steps:

[0070] S1. Collect and preprocess the patient's multimodal physiological information data stream to obtain a normalized multimodal physiological information data stream with consistent time sequence;

[0071] S2. Divide the normalized multimodal physiological information data stream into a high-frequency physiological information data substream and a low-frequency physiological information data substream based on the sampling frequency. The high-frequency physiological information data substream is used to describe second-level physiological fluctuations, and the low-frequency physiological information data substream is used to describe minute-level anesthetic metabolic trends.

[0072] S3 establishes a high-frequency fuzzy inference channel and inputs a high-frequency physiological information data substream into the high-frequency fuzzy inference channel. The high-frequency fuzzy inference channel generates a high-frequency fuzzy inference result based on the initial high-frequency fuzzy rule base and the adaptive membership function;

[0073] S4. Establish a low-frequency fuzzy inference channel and input a low-frequency physiological information data substream into the low-frequency fuzzy inference channel. The low-frequency fuzzy inference channel generates a low-frequency fuzzy inference result based on the initial low-frequency fuzzy rule base and the adaptive membership function;

[0074] S5. Input the high-frequency fuzzy inference results and the low-frequency fuzzy inference results into the fuzzy fusion module, which fuses the high-frequency fuzzy inference results and the low-frequency fuzzy inference results according to the multi-scale weight coefficient and outputs a multi-scale fusion risk score curve;

[0075] S6. Compare the multi-scale fusion risk score curve with the risk safety threshold in real time. If the multi-scale fusion risk score curve exceeds the risk safety threshold at any time, a delayed awakening risk alarm signal is generated and a list of risk driving factors is recorded. The multi-scale fusion risk score curve, delayed awakening risk alarm signal and risk driving factor list are pushed to the anesthesia information interaction terminal for reference and intervention by clinical personnel.

[0076] In this embodiment, S1 includes the following steps:

[0077] S11. Collect multimodal physiological information data streams from the patient, including hemodynamic parameter time series, bispectral index time series, end-tidal concentration time series of gaseous anesthetics, intraoperative drug infusion parameter time series, and the patient's static basic feature vector. Each data type is collected once at the same time step, with a total number of collection steps of T and a number of modal types of M.

[0078] S12. Perform synchronous alignment processing on the collected time series physiological information data, set a unified time step, resample all time series to the same set of time points, and obtain the original aligned data matrix under the unified time axis. The number of rows in the original aligned data matrix is ​​equal to the total number of sampling time steps, and the number of columns is equal to the number of modal types;

[0079] S13. Perform missing value compensation on each item in the original aligned data matrix. If a modality is missing at a certain time point, a compensation function is constructed using forward filling and sample mean interpolation methods. The compensation function is used to assign values ​​to the missing items. After compensation, each item in the data matrix has a value. Non-missing items retain their original values, while missing items are assigned values ​​using the compensation function.

[0080] S14. Perform outlier detection and elimination on the data matrix after missing value compensation. For each mode, calculate the mean and standard deviation of the mode at all time points. Determine the statistical interval based on the mean and standard deviation. Determine all data items outside the statistical interval as outliers. Perform local averaging correction on outliers using the valid values ​​of adjacent time points. After correction, generate a cleaned data matrix.

[0081] S15. Perform normalization on the cleaned data matrix. For each mode, take the minimum and maximum values ​​of the mode at all time points as the interval endpoints, linearly transform each item to the interval [0, 1], and obtain a normalized data matrix after normalization.

[0082] S16. Concatenate the normalized data matrix with the patient's static basic feature vector. The concatenation method is to append the patient's static basic feature vector to the data at each time point to obtain the final feature input matrix.

[0083] In this embodiment, the hemodynamic parameter time series is the systolic and diastolic blood pressure values ​​at each moment, the bispectral index time series is the bispectral index at each moment, the end-tidal concentration time series of the gas anesthetic is the minimum alveolar anesthetic concentration at each moment, and the intraoperative drug infusion parameter time series is the drug infusion rate at each moment, including the dosage of propofol and remifentanil. The patient's static basic feature vector includes immutable individual attributes and preoperative evaluation data including age, body mass index, glomerular filtration rate, alanine aminotransferase, and surgery type.

[0084] In this embodiment, S2 includes the following steps:

[0085] S21. Set the time sampling window set W = {W H ,W L}, where W H is the high-frequency sampling window used to extract second-level physiological fluctuations, W Lis a low-frequency sampling window used to capture minute-level anesthetic metabolic trends, satisfying W H <W L ;

[0086] S22. Based on the uniform sampling step Δt, the final feature input matrix Perform sliding window segmentation processing to extract high-frequency time period sequence blocks and low-frequency time period sequence blocks respectively, and obtain two independent data stream sets:

[0087] X H ={X final [t:t+W H -1]|t=1,1+Δt,…,TW H +1};

[0088] X L ={X final [t:t+W L -1]|t=1,1+Δt,…,TW L +1};

[0089] Among them, X H represents the set of high-frequency physiological information data substreams, X L represents the set of low-frequency physiological information data substreams, reflecting the time period submatrix of the sedative drug metabolism process and the slow-changing trend of anesthetic concentration. P represents the patient's static basic feature vector, T is the total number of sampling time steps, and t is the time index variable, representing each data sampling time point in the entire monitoring period, with a value range from 1 to T.

[0090] S23. High-frequency physiological information data sub-stream set X H Each submatrix in Extracting volatility eigenvectors The volatility eigenvector is composed of the range, variance and first-order difference series mean of each mode in the time period;

[0091] S24. Low-frequency physiological information data sub-stream set X L Each submatrix in Extract trend feature vector The trend feature vector consists of the linear regression slope, mean, and moving average of each mode in the time period;

[0092] S25. Arrange the volatility feature vectors of all high-frequency physiological information data substreams in chronological order as a high-frequency reasoning input sequence Arrange the trend feature vectors of all low-frequency physiological information data substreams in chronological order as a low-frequency reasoning input sequence

[0093] In this embodiment, S3 includes the following steps:

[0094] S31. Construct a high-frequency fuzzy inference channel, which is used to receive a high-frequency inference input sequence. Each volatility feature vector in the high-frequency inference input sequence contains multiple modal features, each modal feature is used to describe the patient's physiological fluctuation state within the current time window, and the dimension of the volatility feature vector is three times the total number of modal features in each time window.

[0095] S32. Aiming at the differences in individual patients’ sensitivity to anesthesia during delayed awakening after anesthesia, an adaptive membership function adjustment mechanism for high-frequency fuzzy inference channel is designed. Used to measure the fuzzy membership of each input feature to risk reasoning in the current time window:

[0096]

[0097] in, is the mean center of the membership function dynamically corrected based on the previous postoperative awakening delay feedback, is the fuzzy expansion coefficient updated online based on the risk prediction error of the wake-up feedback, represents the adaptive parameter of the i-th feature in the k-th time window;

[0098] The core principle of the adaptive membership function in step S32 is that each high-frequency inference input feature not only relies on static, empirically set fuzzy divisions, but also introduces online adjustable mean center parameters and fuzzy expansion parameters. Both parameters are derived from the model's dynamic correction of the previous postoperative true awakening feedback and risk score errors, so that the membership function can be adjusted in real time for different patients, different anesthesia schemes and dynamic changes during surgery, so that the model can perceive the risk response characteristics of individual patients under high-frequency physiological fluctuations, and transform the static mapping of the standard fuzzy system into a self-learning mapping, ensuring that the inference mechanism has the ability to migrate and adapt across populations and pathological types.

[0099] In terms of construction, the adaptive membership function expression adopted by the present invention has the following parameters: and The risk concentration center and physiological fluctuation tolerance range of the characteristics in the current patient history are respectively depicted in real time. By continuously introducing feedback on the patient's actual awakening time and the model output error, each step is calibrating the membership function's ability to express the risk of sensitive groups, taking into account the high dynamics of physiological signals and the medical scenario's demand for personalized high-precision risk judgment.

[0100] S33. Build a high-frequency fuzzy rule base, which consists of multiple fuzzy rules. Each fuzzy rule contains multiple preconditions and a fuzzy output result. The condition is that each modal feature in the current time window belongs to a fuzzy set. The output result is the awakening delay risk level corresponding to the fuzzy rule.

[0101] S34. For each fuzzy rule, calculate the activation strength value of the fuzzy rule under the current volatility feature vector. The activation strength value is determined by multiplying the adaptive membership function values ​​of all modal input features under their corresponding fuzzy sets. The membership value of each input feature introduces a feature sensitivity weight, which is dynamically set based on whether abnormal physiological fluctuations occur during the operation.

[0102] S35. Perform sensitivity screening on all activation intensity values, retaining only fuzzy rules whose activation intensity exceeds a dynamic threshold for inference output in the current time window. The dynamic threshold is calculated based on the mean and standard deviation of the activation intensity of all fuzzy rules. The standard deviation is used to quantify the degree of dispersion of the current activation intensity distribution. The dynamic threshold is used to screen out fuzzy rules that are insensitive to the current patient status.

[0103] S36. Perform fuzzy weighted fusion on all fuzzy rules that pass the screening, and calculate the high-frequency fuzzy inference result in the current time window. The high-frequency fuzzy inference result is the weighted average of the product of the awakening delay risk level corresponding to each fuzzy rule, its activation strength, and the fuzzy rule sensitivity weight. The weighted average is used to represent the overall awakening delay risk score in the time window.

[0104] In this embodiment, S4 includes the following steps:

[0105] S41. Construct a low-frequency fuzzy inference channel, the low-frequency fuzzy inference channel being configured to receive a low-frequency inference input sequence, wherein each trend feature vector in the low-frequency inference input sequence comprises multiple modal features, each modal feature being configured to describe the pharmacokinetic trend state of the patient within the current time window, and the dimension of the trend feature vector being equal to three times the total number of modal features within each time window;

[0106] S42. Design an adaptive membership function adjustment mechanism in the low-frequency fuzzy inference channel. The low-frequency adaptive membership function is used to measure the fuzzy membership degree of each trend input feature to the risk of delayed awakening within the current time window. It is used to adapt to the differences in anesthetic drug metabolism capacity among different patients. The formula is the same as S32.

[0107] S43. Construct a low-frequency fuzzy rule base, which consists of multiple fuzzy rules for drug metabolism trend modeling. Each fuzzy rule contains multiple fuzzy preconditions and a fuzzy output result, where the fuzzy precondition is that each modal feature in the current trend feature vector belongs to a specific fuzzy set, and the fuzzy output result is the awakening delay risk level corresponding to the fuzzy rule;

[0108] S44. Calculate the activation strength value of each low-frequency fuzzy rule under the current trend feature vector, where the activation strength value is obtained by continuously multiplying the membership values ​​of all modal features in the trend feature vector under their corresponding fuzzy sets;

[0109] S45. A trend sensitivity weighting mechanism is introduced to perform weighted adjustment on the activation strength value of each fuzzy rule. Based on the consistency between the anesthetic drug concentration trend and the EEG index trend in the current time window, the sensitivity weight of the pharmacokinetic-related modality is dynamically increased. The sensitivity weight is used to adjust the participation of the modality in fuzzy reasoning. Ultimately, the membership values ​​of all modalities are weighted according to their respective sensitivity weights, and the weighted activation strength value of the rule under the current input is recalculated.

[0110] S46. Perform a weighted average fusion of the weighted activation strength values ​​of all fuzzy rules in each time window and their corresponding risk level output values ​​to obtain a low-frequency fuzzy inference result for that time window. The low-frequency fuzzy inference result represents the awakening delay trend risk score for the current time period. The awakening delay trend risk score is the sum of the product of the risk level outputs of all rules and their weighted activation strengths, divided by the sum of the weighted activation strengths of all rules.

[0111] S47. Arrange the low-frequency fuzzy inference results of each time window in chronological order to obtain a complete low-frequency fuzzy inference result sequence.

[0112] In this embodiment, S5 includes the following steps:

[0113] S51. Align the high-frequency fuzzy inference result sequence and the low-frequency fuzzy inference result sequence according to the time index to obtain a set of corresponding values, where each pair of corresponding values ​​represents a high-frequency risk score value and a low-frequency risk score value in the same time window;

[0114] S52. For each time index, calculate the high-frequency credibility coefficient and low-frequency credibility coefficient of the high-frequency risk score and the low-frequency risk score, respectively. The high-frequency credibility coefficient is used to measure the reliability of the high-frequency score in expressing the overall risk at that time point, and the low-frequency credibility coefficient is used to measure the ability of the low-frequency score to explain the risk trend at that time point.

[0115] S53. Normalize the high-frequency credibility coefficient and the low-frequency credibility coefficient. Based on the normalized high-frequency weight coefficient and low-frequency weight coefficient, weight the high-frequency risk score and the low-frequency risk score respectively to obtain a fused score. The fused score is used to express the multi-scale comprehensive risk score at the current time point. The fusion process adopts a linear weighting strategy, and the weight distribution is determined by the credibility of the high-frequency and low-frequency scores at the time point.

[0116] S54. Dynamically smooth the fusion score value to obtain a smoothed score value. The smoothed score value is used to suppress local spike noise caused by short-term signal fluctuations. The smoothed score values ​​of all time points are connected and arranged in chronological order to obtain a curve. The curve is the multi-scale fusion risk score curve. The multi-scale fusion risk score curve is used to display the patient's comprehensive risk change trend in real time throughout the postoperative recovery process.

[0117] In this embodiment, S6 includes the following steps:

[0118] S61. Set a risk safety threshold. The risk safety threshold is determined based on the statistical distribution of historical large-scale postoperative recovery data, different surgical types, and clinical factors of patient characteristics. The risk safety threshold is used to distinguish between the safe zone and the warning zone in the multi-scale fusion risk score curve;

[0119] S62. Each score value of the multi-scale fusion risk score curve is compared with the risk safety threshold in real time. If the multi-scale fusion risk score value at any time point exceeds the risk safety threshold, a wake-up delay risk warning signal is automatically generated. The wake-up delay risk warning signal includes the current time point, the current score value, and the corresponding threshold-exceeding amplitude;

[0120] S63. At the time point when the awakening delay risk warning signal is generated, the inference engine is called to record the risk driving factor list of the multi-scale fusion risk score curve at that moment. The risk driving factor list is generated based on the marginal contribution of each modal physiological feature to the total score at that moment. The marginal contribution is calculated based on the sensitivity weight of each input feature under the activated core fuzzy rule and the change rate of the inference result during the fuzzy inference process;

[0121] S64. Push the current multi-scale fusion risk score curve, awakening delay risk warning signal and risk driving factor list to the anesthesia information interaction terminal in real time through the data interface. The anesthesia information interaction terminal is used to synchronously display the score curve, warning signal and driving factor details on the terminal platform for real-time reference by clinical staff.

[0122] In this implementation, the risk safety threshold classification rules are as follows:

[0123] When the multi-scale fusion risk score curve is continuously greater than or equal to 0.8 at any time point, or the high-frequency risk score value is continuously greater than or equal to 0.85, or the low-frequency risk score value is continuously greater than or equal to 0.75, the system automatically outputs a high-risk warning;

[0124] When the multi-scale fusion risk score curve is between 0.5 and 0.8 at any time point, or the high-frequency risk score value is between 0.6 and 0.85, or the low-frequency risk score value is between 0.5 and 0.75, the system outputs a medium risk warning;

[0125] When the multi-scale fusion risk score curve is lower than 0.5 at any time point, and both the high-frequency risk score value and the low-frequency risk score value are lower than 0.5, the system determines it as low risk and maintains routine monitoring.

[0126] Example 1: In October 2024, the Anesthesiology Department of the First People's Hospital mainly admitted elderly patients undergoing hip fracture surgery, complex abdominal tumor resection, and chronic liver and kidney diseases combined with general anesthesia. The hospital purchased and deployed the present invention, and conducted a one-month control trial with the traditional anesthesia risk assessment system based on the fixed threshold BIS method.

[0127] The trial subjects included 135 patients undergoing general anesthesia surgery, with an age distribution of 48-87 years (average 67.4 years), including 36 patients with chronic liver disease, 28 patients with chronic kidney disease, and 23 patients with BMI ≥ 30 (obese). The trial covered major departments of orthopedics, general surgery, and urology. Each patient was connected to the in-hospital monitoring system throughout the operation and recovery period, and all monitoring data were synchronously connected to the intelligent recovery risk management system.

[0128] Before surgery, the patient's basic information, including gender, age, body mass index, American Society of Anesthesiologists (ASA) score, liver and kidney function indicators, comorbidities, and surgery type, was entered into the electronic medical record. In Example 1, a female patient numbered "20241014-037," 67 years old, with a BMI of 31.4, ASA grade 3, and comorbidities with chronic hepatitis B and type 2 diabetes, was scheduled to undergo right femoral fracture resection under general anesthesia.

[0129] After anesthesia induction, the system automatically collects the patient's multimodal physiological data: including real-time blood pressure (once per minute), continuous BIS index (once every 10 seconds), end-tidal isoflurane concentration (once per minute), remifentanil and propofol pump speed (once per minute), and routine indicators of pulse, blood oxygen, and heart rate. The data is accessed in real time, and the system synchronously aligns, cleans, fills in missing data, and normalizes them to form a standardized data stream for subsequent analysis.

[0130] During the anesthesia maintenance and awakening period, the system performs multi-scale diversion processing, splitting the data into two types of data sub-streams: high-frequency (BIS, heart rate, blood pressure fluctuations, window 30 seconds, sliding step 10 seconds) and low-frequency (anesthetic concentration, metabolic indicators, window 5 minutes, sliding step 1 minute).

[0131] The high-frequency inference channel automatically selects the characteristics of drastic blood pressure fluctuations and continuous abnormal BIS fluctuations, and applies an adaptive membership function for online correction (when the BIS window mean fluctuates from 60.3 to 42.1 and the propofol rate decreases from 0.15 mg / kg / h to 0.05 mg / kg / h, the system sensitively identifies an increase in its high-risk weight); the low-frequency inference channel performs trend analysis on the changing trend of end-tidal isoflurane concentration and liver function metabolic rate (the liver enzyme ALT in patients with chronic liver disease during the recovery period is higher than 80 U / L, and the end-tidal isoflurane degradation curve slows down).

[0132] Subsequently, the fusion module automatically matches the two types of inference results and gives adaptive weighted fusion to high-frequency abnormalities (BIS abnormal fluctuation greater than 20, blood pressure fluctuation amplitude greater than 30 mmHg) and low-frequency slow metabolism (isoflurane clearance half-life extended by 30%).

[0133] For a patient 15 minutes after surgery, the high-frequency risk score was 0.91, the low-frequency risk score was 0.78, and the fusion weights were 0.55 and 0.45, respectively. The multi-scale fusion risk score was 0.85. At this time, the system detection score continued to be higher than the safety threshold (0.8), and a high-risk warning for delayed awakening was automatically pushed on the large screen and mobile terminal, and it was suggested to reduce the infusion of sedative drugs and strengthen language awakening stimulation.

[0134] The system simultaneously records the actual awakening time (from drug discontinuation to BIS>80 and eye opening) and compares it with the predicted risk score curve. If the patient finally wakes up fully 42 minutes after drug discontinuation, the system will feedback to the learning module and automatically update the center and expansion coefficients of the high-frequency and low-frequency membership functions for this type of patient (elderly women + chronic liver disease + high BMI + high-dose isoflurane), optimizing the adaptive ability of the model for subsequent similar cases.

[0135] In the above 135 surgical cases, the risk assessment results of the system based on the present invention and the traditional BIS threshold method were manually compared. The specific data are shown in Table 1 below:

[0136] Table 1 Risk assessment results of the system of the present invention and the traditional BIS threshold method

[0137]

[0138]

[0139] Training sample, number "20241009-022": 68-year-old male, chronic kidney disease, BIS average value 50.2, isoflurane pump rate 0.22mg / kg / h, actual recovery time after surgery 58 minutes; the system of the present invention predicted a high-risk score peak of 0.93, issued an early warning 22 minutes in advance, and the final recovery time was shortened by 14 minutes after reducing isoflurane during surgery; the traditional BIS method did not provide an early warning and ultimately missed the report.

[0140] Case number "20241023-111": A 74-year-old female with a BMI of 34.6 and abnormal liver function. Her BIS was repeatedly below 60 after surgery and she had slow end-tidal isoflurane degradation. The system of the present invention integrated high- and low-frequency reasoning, scored a persistently high risk, and intervened 16 minutes in advance. Ultimately, no serious delay occurred, and the traditional method mistakenly reported it as normal.

[0141] Table 2 Multiple sets of real training data

[0142]

[0143]

[0144] After one month of clinical comparison, it was found that the method of the present invention can accurately identify high-risk patients in complex cases and drug-sensitive groups, with the false alarm rate reduced by 76%, the missed alarm rate reduced by 78%, the high-risk intervention response rate increased by 73%, and the average awakening time after intervention shortened by more than 2 times.

[0145] The present invention constructs high-frequency fuzzy inference channels and low-frequency fuzzy inference channels and designs a multi-scale risk scoring fusion mechanism. It can collaboratively model second-level physiological signals and minute-level pharmacokinetic trends, thereby realizing full-cycle, continuous and dynamic perception of the risk of delayed awakening after anesthesia. The fusion module is based on adaptive weights and dynamically allocates high- and low-frequency inference results according to the current risk contribution. Combined with a dynamic smoothing mechanism, it effectively suppresses misjudgments caused by short-term fluctuations, significantly improving the risk identification sensitivity and robustness for high-risk patients, complex surgeries and cases with multiple complications, and avoiding omissions and false alarms caused by single dimensions or static thresholds.

[0146] The present invention introduces an adaptive membership function updating method based on individual patient characteristics, real postoperative feedback, and model prediction errors. It can perform membership self-learning and real-time fine-tuning of rule base parameters for each input variable in the fuzzy reasoning channel according to the multi-dimensional static and dynamic characteristics of the patient's age, weight, liver and kidney function, and surgical type. The high-frequency and low-frequency channels adopt sensitivity weighting and trend weighting mechanisms, respectively, to achieve individualized risk level output in the high-risk group for delayed awakening.

[0147] The present invention uses explicit management of fuzzy rules throughout the entire process and visualized output of factor contribution weights to support real-time display of risk scoring curves, short-term fluctuations, and key trend change indicators, providing clinical anesthesiologists with a clear risk evolution trajectory and a basis for ranking influencing factors. The system features multiple clinical interaction modes, including hierarchical display of high-frequency and low-frequency information, intelligent triggering of risk warning thresholds, and intelligent push of intervention suggestions, enabling early warning of high-risk situations of delayed postoperative recovery.

[0148] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for assessing the risk of delayed recovery after anesthesia based on adaptive fuzzy reasoning, characterized in that: The steps include: S1. Collect and preprocess the patient's multimodal physiological information data stream to obtain a normalized multimodal physiological information data stream with consistent time sequence; S2. Dividing the normalized multimodal physiological information data stream into a high-frequency physiological information data sub-stream and a low-frequency physiological information data sub-stream according to the sampling frequency; S3. Establish a high-frequency fuzzy inference channel and input a high-frequency physiological information data substream into the high-frequency fuzzy inference channel to generate a high-frequency fuzzy inference result based on the initial high-frequency fuzzy rule base and the adaptive membership function; S4. Establish a low-frequency fuzzy inference channel and input a low-frequency physiological information data substream into the low-frequency fuzzy inference channel to generate a low-frequency fuzzy inference result based on the initial low-frequency fuzzy rule base and the adaptive membership function; S5. Input the high-frequency fuzzy inference results and the low-frequency fuzzy inference results into the fuzzy fusion module to fuse the high-frequency fuzzy inference results and the low-frequency fuzzy inference results, and output a multi-scale fusion risk score curve; S6. Compare the multi-scale fusion risk score curve with the risk safety threshold in real time. If the multi-scale fusion risk score curve exceeds the risk safety threshold at any time, a delayed awakening risk alarm signal is generated and a list of risk driving factors is recorded. The multi-scale fusion risk score curve, delayed awakening risk alarm signal and risk driving factor list are pushed to the anesthesia information interaction terminal for reference and intervention by clinical personnel.

2. The method for assessing the risk of delayed recovery after anesthesia based on adaptive fuzzy reasoning according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Collecting a multimodal physiological information data stream of the patient, including a time series of hemodynamic parameters, a time series of bispectral index, a time series of end-tidal concentration of a gaseous anesthetic agent, a time series of intraoperative drug infusion parameters, and a static basic feature vector of the patient; S12. Perform synchronous alignment processing on the collected time series physiological information data, set a unified time step, resample all time series to the same set of time points, and obtain the original aligned data matrix under the unified time axis; S13. Perform missing value compensation on each item in the original aligned data matrix; S14. Perform outlier detection and elimination on the data matrix after missing value compensation to generate a cleaned data matrix; S15. Performing normalization processing on the cleaned data matrix to obtain a normalized data matrix; S16. Concatenate the normalized data matrix with the patient's static basic feature vector to obtain the final feature input matrix.

3. The method for risk assessment of delayed recovery after anesthesia based on adaptive fuzzy reasoning according to claim 2, characterized in that: The hemodynamic parameter time series is the systolic and diastolic blood pressure values ​​at each moment, the bispectral index time series is the bispectral index at each moment, the end-tidal concentration of the gas anesthetic is the minimum alveolar anesthetic concentration at each moment, the intraoperative drug infusion parameter time series is the drug infusion rate at each moment, including the dosage of propofol and remifentanil, and the patient's static basic feature vector includes immutable individual attributes and preoperative assessment data including age, body mass index, glomerular filtration rate, alanine aminotransferase and surgery type.

4. The method for risk assessment of delayed recovery after anesthesia based on adaptive fuzzy reasoning according to claim 2, characterized in that: The S2 comprises the following steps: S21. Set the time sampling window set W = {W H ,W L }, where W H is the high-frequency sampling window used to extract second-level physiological fluctuations, W L A low-frequency sampling window used to capture minute-level anesthetic metabolic trends; S22. Based on the uniform sampling step Δt, the final feature input matrix Perform sliding window segmentation processing to extract high-frequency time period sequence blocks and low-frequency time period sequence blocks respectively, and obtain two independent data stream sets: X H =X final [t:t+W H -1]∣t=1,1+Δt,…,T-W H +1; X L =X final [t:t+W L -1]∣t=1,1+Δt,…,T-W L +1; Among them, X H represents the set of high-frequency physiological information data substreams, X L represents the set of low-frequency physiological information data substreams, P represents the patient's static basic feature vector, T is the total number of sampling time steps, and t is the time index variable, which represents each data sampling time point in the entire monitoring period, with a value range from 1 to T; S23. High-frequency physiological information data sub-stream set X H Each submatrix in Extracting volatility eigenvectors The volatility eigenvector is composed of the range, variance and first-order difference series mean of each mode in the time period; S24. Low-frequency physiological information data sub-stream set X L Each submatrix in Extract trend feature vector The trend feature vector consists of the linear regression slope, mean, and moving average of each mode in the time period; S25. Arrange the volatility feature vectors of all high-frequency physiological information data substreams in chronological order as a high-frequency reasoning input sequence Arrange the trend feature vectors of all low-frequency physiological information data substreams in chronological order as a low-frequency reasoning input sequence 5. The method for risk assessment of delayed recovery after anesthesia based on adaptive fuzzy reasoning according to claim 4, characterized in that: The S3 includes the following steps: S31. Constructing a high-frequency fuzzy inference channel, the high-frequency fuzzy inference channel is used to receive a high-frequency inference input sequence, each volatility feature vector in the high-frequency inference input sequence includes multiple modal features; S32. Aiming at the individual differences in anesthesia sensitivity of patients during delayed awakening after anesthesia, an adaptive membership function adjustment mechanism of the high-frequency fuzzy inference channel is designed. The adaptive membership function in the adaptive membership function adjustment mechanism Used to measure the fuzzy membership of each input feature to risk reasoning in the current time window: in, is the mean center of the membership function dynamically corrected based on the previous postoperative awakening delay feedback, is the fuzzy expansion coefficient updated online based on the risk prediction error of the wake-up feedback, represents the adaptive parameter of the i-th feature in the k-th time window, f k,i represents the i-th high-frequency modal input eigenvalue in the k-th high-frequency time window; S33. Build a high-frequency fuzzy rule base, which consists of multiple fuzzy rules. Each fuzzy rule contains multiple preconditions and a fuzzy output result. The condition is that each modal feature in the current time window belongs to a fuzzy set. The output result is the awakening delay risk level corresponding to the fuzzy rule. S34. For each fuzzy rule, calculating the activation strength value of the fuzzy rule under the current volatility feature vector; S35. Perform sensitivity screening on all activation intensity values, and only retain fuzzy rules whose activation intensity values ​​exceed the dynamic threshold for inference output in the current time window; S36. Perform fuzzy weighted fusion on all fuzzy rules that pass the screening, and calculate the high-frequency fuzzy inference results in the current time window.

6. The method for risk assessment of delayed recovery after anesthesia based on adaptive fuzzy reasoning according to claim 5, characterized in that: The S4 comprises the following steps: S41. Construct a low-frequency fuzzy inference channel, where each trend feature vector in the low-frequency inference input sequence contains multiple modal features; S42. Design an adaptive membership function adjustment mechanism in the low-frequency fuzzy inference channel. The low-frequency adaptive membership function is used to measure the fuzzy membership degree of each trend input feature to the risk of delayed awakening within the current time window, which is used to adapt to the differences in anesthetic drug metabolism capacity among different patients. S43. Construct a low-frequency fuzzy rule base, which consists of multiple fuzzy rules for drug metabolism trend modeling. Each fuzzy rule contains multiple fuzzy preconditions and a fuzzy output result. The fuzzy output result is the awakening delay risk level corresponding to the fuzzy rule; S44. Calculating the activation strength value of each low-frequency fuzzy rule under the current trend feature vector; S45. Introduce a trend sensitivity weighting mechanism to perform weighted adjustment on the activation strength value of each fuzzy rule. Based on the consistency between the anesthetic drug concentration trend and the EEG index trend in the current time window, dynamically increase the sensitivity weight of the pharmacokinetic-related modality and recalculate the weighted activation strength value of the rule under the current input. S46. Perform weighted averaging on the weighted activation strength values ​​of all fuzzy rules in each time window and their corresponding risk level output values ​​to obtain a low-frequency fuzzy inference result in the time window; S47. Arrange the low-frequency fuzzy inference results of each time window in chronological order to obtain a complete low-frequency fuzzy inference result sequence.

7. The method for risk assessment of delayed recovery after anesthesia based on adaptive fuzzy reasoning according to claim 6, characterized in that: The S5 comprises the following steps: S51. Aligning the high-frequency fuzzy inference result sequence with the low-frequency fuzzy inference result sequence according to the time index to obtain a set of corresponding values; S52. For each time index, calculate the high-frequency credibility coefficient and the low-frequency credibility coefficient of the high-frequency risk score value and the low-frequency risk score value respectively; S53. Normalize the high-frequency credibility coefficient and the low-frequency credibility coefficient, and weight the high-frequency risk score and the low-frequency risk score based on the normalized high-frequency weight coefficient and the low-frequency weight coefficient to obtain a fused score. S54. Dynamically smooth the fusion score value to obtain a smoothed score value. The smoothed score value is used to suppress local spike noise caused by short-term signal fluctuations. The smoothed score values ​​of all time points are connected and arranged in chronological order to obtain a curve. The curve is the multi-scale fusion risk score curve.

8. The method for assessing the risk of delayed recovery after anesthesia based on adaptive fuzzy reasoning according to claim 7, characterized in that: The S6 comprises the following steps: S61. Set a risk safety threshold, which is determined based on the statistical distribution of historical large-scale postoperative recovery data, different surgical types, and clinical factors of patient characteristics; S62. Each score value of the multi-scale fusion risk score curve is compared with the risk safety threshold in real time. If the multi-scale fusion risk score value at any time point exceeds the risk safety threshold, a wake-up delay risk warning signal is automatically generated; S63. For the time point at which the wake-up delay risk warning signal is generated, the inference engine is called to record the risk driving factor list of the multi-scale fusion risk score curve at that moment; S64. The current multi-scale fusion risk score curve, awakening delay risk alarm signal and risk driving factor list are pushed to the anesthesia information interaction terminal in real time through the data interface. The anesthesia information interaction terminal is used to synchronously display the score curve, alarm signal and driving factor details on the terminal platform for real-time reference by clinical personnel.

9. The method for assessing the risk of delayed recovery after anesthesia based on adaptive fuzzy reasoning according to claim 8, characterized in that: The classification rules for the risk safety threshold are as follows: When the multi-scale fusion risk score curve is continuously greater than or equal to 0.8 at any time point, or the high-frequency risk score value is continuously greater than or equal to 0.85, or the low-frequency risk score value is continuously greater than or equal to 0.75, the system automatically outputs a high-risk warning; When the multi-scale fusion risk score curve is between 0.5 and 0.8 at any time point, or the high-frequency risk score value is between 0.6 and 0.85, or the low-frequency risk score value is between 0.5 and 0.75, the system outputs a medium risk warning; When the multi-scale fusion risk score curve is lower than 0.5 at any time point, and both the high-frequency risk score value and the low-frequency risk score value are lower than 0.5, the system determines it as low risk and maintains routine monitoring.

Images