Machine Learning-Based Thoracic Anesthesia Control Method

Abstract

This invention discloses a machine learning-based method for controlling thoracic anesthesia, comprising the following steps: collecting a multi-source raw time-series data set for thoracic anesthesia and performing preprocessing; performing sliding segmentation on the standardized time-series data set for thoracic anesthesia; extracting a set of short-timescale perturbation feature vectors and performing short-timescale learning; extracting a set of medium-timescale evolution feature vectors and performing medium-timescale learning; extracting a set of long-timescale cumulative feature vectors and performing long-timescale learning; constructing a time-scale competition judgment set and determining a set of control weight allocation vectors; performing constrained fusion to generate a rhythmic anesthesia control strategy, distributing it to the execution unit, and cyclically updating the standardized time-series data set for thoracic anesthesia until the anesthesia for thoracic surgery ends. This machine learning-based multi-timescale anesthesia control method achieves coordinated and constrained regulation of anesthesia infusion and ventilation rhythms, possessing advantages such as high stability, controllable conflicts, and strong continuity.

Description

Technical Field

[0001] This invention relates to the fields of medical engineering and intelligent control, and in particular to a machine learning-based method for controlling thoracic anesthesia. Background Technology

[0002] During thoracic surgery anesthesia control, clinicians typically rely on anesthesiologists to make empirical adjustments based on monitoring indicators such as anesthetic drug infusion rate, end-tidal concentration of inhaled anesthetic, invasive arterial blood pressure, heart rate, blood oxygen saturation, end-tidal carbon dioxide partial pressure, peak airway pressure, respiratory rate, and depth of anesthesia index. Some existing technologies introduce regular thresholds, statistical features of single time windows, or machine learning models to output suggestions for adjusting anesthesia infusion or ventilation parameters, and form a closed-loop control through periodic data collection and updates.

[0003] Existing technologies generally lack scale-aligned modeling mechanisms for short-timescale disturbance responses, medium-timescale evolution processes, and long-timescale cumulative effects. Control proposals and control rhythms are prone to directional conflicts within the same time interval, and there is a lack of calculable conflict identification and weight allocation processes. Stability constraints often adopt static amplitude limiting or fixed thresholds, making it difficult to form a constrained fusion strategy driven by stability dissipation assessment and residual stability margin. As a result, the temporal continuity and amplitude-limited characteristics of rhythmic control strategies cannot be satisfied simultaneously. Summary of the Invention

[0004] One objective of this invention is to propose a machine learning-based method for controlling thoracic anesthesia. This invention provides a multi-timescale anesthesia control method based on machine learning, which achieves coordinated and constrained regulation of anesthesia infusion and ventilation rhythms, and has the advantages of high stability, controllable conflicts, and strong continuity.

[0005] The machine learning-based thoracic anesthesia control method according to an embodiment of the present invention includes the following steps: During thoracic surgery anesthesia, a multi-source raw time-series data set of thoracic anesthesia was collected and preprocessed to generate a standardized time-series data set of thoracic anesthesia. Sliding segmentation was performed on the standardized time-series data set of thoracic anesthesia to generate short-timescale anesthesia response fragment sets, medium-timescale anesthesia response fragment sets, and long-timescale anesthesia response fragment sets, respectively; Extract a set of short-timescale perturbation feature vectors from the set of short-timescale anesthesia response fragments, perform short-timescale learning, and generate a set of short-timescale control suggestions; Extract the set of mid-timescale evolution feature vectors from the mid-timescale anesthesia response fragment set, and perform mid-timescale learning to generate a set of mid-timescale control rhythms; Extract a set of long-timescale cumulative feature vectors from the set of long-timescale anesthesia response fragments, and perform long-timescale learning to generate a set of stability dissipation assessment results and a set of remaining stability margins; Based on the short-timescale control suggestion set, the medium-timescale control rhythm set, the stability dissipation assessment result set, and the remaining stability margin set, a time-scale competition judgment set is constructed. Control layer conflict identification is performed, a time-scale competition score sequence is generated, and a control weight allocation vector set is determined. Based on the control weight allocation vector set, the stability dissipation assessment result set, and the remaining stability margin set, constrained fusion is performed to generate a rhythmic anesthesia control strategy, which is then distributed to the anesthesia infusion execution unit and the ventilator control execution unit. The standardized time-series data set for thoracic anesthesia is updated cyclically until the anesthesia for thoracic surgery ends.

[0006] Optionally, the multi-source raw time-series data for thoracic anesthesia includes anesthetic drug infusion rate, end-tidal concentration of inhaled anesthetic, respiratory rate, peak airway pressure, end-tidal carbon dioxide partial pressure, blood oxygen saturation, heart rate, invasive arterial blood pressure, and anesthesia depth index.

[0007] Optionally, the preprocessing includes time alignment, outlier removal, missing segment imputation, and normalization.

[0008] Optionally, the execution of sliding segments specifically includes: An anesthesia physiological response decomposition configuration set is constructed based on the standardized time-series data set of thoracic anesthesia. The anesthesia physiological response decomposition configuration set includes short time-scaled response window length, medium time-scaled response window length and long time-scaled response window length. Based on the short timescale response window length, the standardized time series data set of thoracic anesthesia is subjected to short timescale sliding segmentation along the time axis to generate a set of short timescale anesthesia response fragments arranged in chronological order; Based on the length of the medium-timescale response window, a medium-timescale sliding segmentation operation is performed on the standardized time-series data set of thoracic anesthesia to form a set of medium-timescale anesthesia response segments covering multiple short-timescale windows; Based on the length of the long timescale response window, a long timescale sliding segmentation operation is performed on the standardized time-series data set of thoracic anesthesia to generate a set of long timescale anesthesia response fragments. Time index mapping relationships are established for the short-timescale anesthesia response fragment set, the medium-timescale anesthesia response fragment set, and the long-timescale anesthesia response fragment set, respectively, to form a scale-aligned anesthesia physiological response fragment structure set.

[0009] Optionally, the generation of the short timescale control suggestion set includes: Obtain a set of short-timescale anesthesia response fragments. For each short-timescale anesthesia response fragment, extract the anesthetic drug infusion rate, airway peak pressure, end-tidal carbon dioxide partial pressure, blood oxygen saturation, heart rate, and invasive arterial blood pressure within the corresponding time interval to construct the short-timescale original response sequence. Based on the short-timescale raw response sequence, and combined with the direction of change of anesthetic drug infusion rate within the short-timescale response window, the transient rate of change of invasive arterial blood pressure and the transient rate of change of heart rate were calculated. Perform short-term trend analysis on blood oxygen saturation and calculate the characteristics of the slope of blood oxygen saturation decline; Identify abnormal changes in airway peak pressure within a short timescale response window and extract the amplitude characteristics of sudden changes in airway peak pressure. Transient shift analysis was performed on the end-tidal carbon dioxide partial pressure to calculate the transient shift characteristics of end-tidal carbon dioxide. The transient rate of change of invasive arterial blood pressure, transient rate of change of heart rate, slope of decrease of blood oxygen saturation, amplitude of sudden change of airway peak pressure, and transient shift of carbon dioxide at end-tidal are concatenated and normalized to obtain a set of short-timescale perturbation feature vectors. Short-timescale learning is performed based on the set of short-timescaled perturbation feature vectors to generate a set of short-timescaled control proposals.

[0010] Optionally, the generation of the time-scaled control rhythm set includes: Obtain a set of mid-timescale anesthesia response fragments. For each mid-timescale anesthesia response fragment, extract the anesthesia depth index, end-tidal concentration of inhaled anesthetic, invasive arterial blood pressure, heart rate and respiratory rate within the corresponding time interval to construct the mid-timescale original evolutionary response sequence. Based on the original evolutionary response sequence of the mid-timescale, the anesthesia depth index was subjected to time-time trend analysis, and the slope characteristics and trend stability characteristics of the anesthesia depth index were calculated. Fluctuation analysis was performed on invasive arterial blood pressure and heart rate separately to calculate the characteristics of blood pressure fluctuation amplitude, blood pressure mean drift, heart rate fluctuation amplitude, and heart rate mean drift. Perform rhythm analysis on respiratory rate to calculate the trend characteristics and rhythm stability characteristics of respiratory rate changes; Smoothing and trend analysis were performed on the end-tidal concentration of inhaled anesthetics, and the slope characteristics of the end-tidal concentration change were calculated. The characteristics of the slope of the change in the depth of anesthesia index, the trend stability of the depth of anesthesia index, the amplitude of blood pressure fluctuation, the drift of the mean blood pressure, the amplitude of heart rate fluctuation, the drift of the mean heart rate, the trend of the change in respiratory rate, the rhythm stability of respiratory rate, and the slope of the change in end-expiratory concentration are concatenated and normalized to obtain a set of mid-timescale evolution feature vectors. Based on the set of evolutionary feature vectors of the medium time scale, medium time scale learning is performed to generate a set of medium time scale control rhythms.

[0011] Optionally, the generation of the stability dissipation assessment result set and the residual stability margin set specifically includes: Obtain a set of long-timescale anesthesia response fragments. For each long-timescale anesthesia response fragment, extract the anesthesia depth index, invasive arterial blood pressure, heart rate and respiratory rate within the corresponding time interval to construct the long-timescale original cumulative response sequence. For the anesthesia depth index, the cumulative change value of the anesthesia depth index is obtained; Calculate the mean drift and cumulative fluctuation values ​​of invasive arterial blood pressure and heart rate over long time scale intervals, respectively. Calculate the cumulative value of rhythm deviation within a long time-scaled interval for respiratory rate; The cumulative change values ​​of the anesthesia depth index, the mean blood pressure drift value, the cumulative blood pressure fluctuation value, the mean heart rate drift value, the cumulative heart rate fluctuation value, and the cumulative rhythm deviation value are concatenated and normalized to obtain a long-timescale cumulative feature vector set. Long-timescale learning is performed based on the long-timescale cumulative feature vector set, and a set of stability dissipation assessment results is generated for each long-timescale anesthesia response segment. Based on the difference between the stability dissipation assessment result set and the preset stability benchmark value, the corresponding residual stability margin value is calculated to form the residual stability margin set.

[0012] Optionally, the generation of the control weight allocation vector set specifically includes: Based on the time index mapping relationship between the short-timescale anesthesia response fragment set, the medium-timescale anesthesia response fragment set, and the long-timescale anesthesia response fragment set, a corresponding medium-timescale control rhythm entry, stability dissipation assessment result value, and remaining stability margin value are matched for each short-timescale anesthesia response fragment. Extract short-timescale control suggestion entries within the corresponding time interval from the short-timescale control suggestion set, and obtain the control adjustment direction value and control adjustment amplitude value of the short-timescale control suggestion entries; Extract the mid-timescale control rhythm entries within the corresponding time interval from the mid-timescale control rhythm set, and obtain the anesthetic drug infusion rhythm direction value, anesthetic drug infusion rhythm step size value, respiratory rate regulation rhythm direction value, and respiratory rate regulation rhythm step size value of the mid-timescale control rhythm entries. Under the dimension of anesthetic drug infusion control, the consistency of the control adjustment direction value of short-timescale control suggestion items and the anesthetic drug infusion rhythm direction value of medium-timescale control rhythm items is determined, and the consistency determination value of anesthetic drug infusion direction is generated. Under the respiratory rate regulation control dimension, the consistency of the control regulation direction value of the short time-scale control suggestion item and the respiratory rate regulation rhythm direction value of the medium time-scale control rhythm item is determined, and a respiratory rate regulation direction consistency determination value is generated. The consistency judgment value of anesthetic drug infusion direction is combined with the consistency judgment value of respiratory rate regulation direction to generate control layer conflict identification value; Based on the control layer conflict identification value, the stability dissipation assessment result value and the remaining stability margin value within the corresponding time interval, the corresponding time scale competition score is generated and arranged in chronological order to form a time scale competition score sequence. Based on the time-scale competition score sequence, a weight allocation determination is performed on the short-timescale control suggestion set and the medium-timescale control rhythm set to generate a control weight allocation vector set.

[0013] Optionally, the control weight allocation vector includes a short timescale weight value and a medium timescale weight value, and the sum of the short timescale weight value and the medium timescale weight value is a fixed value of 1.

[0014] Optionally, the execution of constrained fusion specifically includes: Obtain the control weight allocation vector, stability dissipation assessment result, and remaining stability margin value within the corresponding time interval; Extract short-timescale control suggestion entries within the corresponding time interval from the short-timescale control suggestion set, and obtain the control adjustment amplitude value; Extract the mid-timescale control rhythm entries within the corresponding time interval from the mid-timescale control rhythm set, and obtain the step size values ​​of anesthetic drug infusion rhythm and respiratory rate regulation rhythm. Based on the short-timescale weight value and the medium-timescale weight value in the control weight allocation vector, the control adjustment amplitude value of the short-timescale control suggestion item and the anesthetic drug infusion rhythm step value of the medium-timescale control rhythm item are weighted and summed to generate the fused anesthetic drug infusion adjustment value. The control adjustment amplitude value of the short-timescale control suggestion item and the respiratory rate adjustment rhythm step value of the medium-timescale control rhythm item are weighted and summed to generate the fused respiratory rate adjustment value. Based on the stability dissipation assessment results and the remaining stability margin, constraint judgments are performed on the fused anesthetic drug infusion regulation value and respiratory rate regulation value to generate constrained regulation values. By linearly mapping the constrained regulation values, the anesthetic drug infusion rhythm parameter values ​​and the respiratory rate regulation rhythm parameter values ​​are obtained, and a rhythmic anesthesia control strategy is generated. The rhythmic anesthesia control strategy is distributed to the anesthesia infusion execution unit and the ventilator control execution unit and the control operation is executed. After completing the control operations, the original time-series data of thoracic anesthesia from multiple sources were re-acquired and preprocessed to update the standardized time-series data set of thoracic anesthesia. The above steps for generating and executing the rhythmic anesthesia control strategy are repeated for subsequent time intervals until the anesthesia for thoracic surgery ends.

[0015] The beneficial effects of this invention are: This invention introduces a multi-timescale collaborative modeling mechanism with short, medium, and long timescales during thoracic surgery anesthesia control. This mechanism moves beyond transient regulation or empirical threshold judgment within a single time window, enabling scale-aligned analysis of rapid perturbation behavior, continuous evolution trends, and long-term cumulative effects of anesthetic physiological responses on a unified time axis. By performing sliding segmentation on the standardized time-series data set of thoracic anesthesia and establishing a time index mapping relationship, anesthesia response fragments generated at different timescales can maintain strict correspondence. This provides a data structure foundation for collaborative decision-making of subsequent control recommendations, control rhythms, and stability assessments within the same time interval, effectively avoiding the regulatory lag or over-correction problems caused by time mismatch of multi-source control information in traditional methods.

[0016] At the control strategy generation level, this invention achieves the division of labor and cooperation between immediate response regulation, rhythm maintenance regulation, and long-term stability constraints by separately constructing a short-timescale control suggestion set, a medium-timescale control rhythm set, a stability dissipation assessment result set, and a residual stability margin set. Furthermore, a timescale competition judgment set is introduced to identify conflicts in the directional consistency of short-timescale and medium-timescale controls within the same time interval. This, combined with the stability dissipation assessment results and residual stability margin, generates a timescale competition score sequence, making the control weight allocation process computationally based rather than empirically set. This allows for adaptive adjustment of the dominance of short-timescale and medium-timescale controls in the overall decision-making process at different anesthesia stages.

[0017] At the control execution level, this invention fuses short-timescale control suggestions and medium-timescale control rhythm execution based on a set of control weight allocation vectors. It then dynamically limits and constrains the fused adjustment amount by combining stability dissipation assessment results and remaining stability margin, ensuring that the generated rhythmic anesthesia control strategy simultaneously possesses temporal continuity, limited adjustment amplitude, and controllable stability. By synchronously distributing the rhythmic anesthesia control strategy to the anesthesia infusion execution unit and the ventilator control execution unit and cyclically updating the standardized time-series data set for thoracic anesthesia, this invention can reduce the frequency of manual intervention while ensuring overall stability of anesthesia depth, circulatory status, and ventilation status. This reduces the risk of short-term fluctuations accumulating into long-term instability, thereby improving the safety, consistency, and controllability of the thoracic anesthesia control process. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the machine learning-based thoracic anesthesia control method proposed in this invention; Figure 2This is a schematic diagram illustrating the process of identifying conflicts in the execution control layer and generating a set of control weight allocation vectors for the machine learning-based thoracic anesthesia control method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 A machine learning-based method for controlling thoracic anesthesia includes the following steps: During thoracic surgery anesthesia, a multi-source raw time-series data set of thoracic anesthesia was collected and preprocessed to generate a standardized time-series data set of thoracic anesthesia. Sliding segmentation was performed on the standardized time-series data set of thoracic anesthesia to generate short-timescale anesthesia response fragment sets, medium-timescale anesthesia response fragment sets, and long-timescale anesthesia response fragment sets, respectively; Extract a set of short-timescale perturbation feature vectors from the set of short-timescale anesthesia response fragments, perform short-timescale learning, and generate a set of short-timescale control suggestions; Extract the set of mid-timescale evolution feature vectors from the mid-timescale anesthesia response fragment set, and perform mid-timescale learning to generate a set of mid-timescale control rhythms; Extract a set of long-timescale cumulative feature vectors from the set of long-timescale anesthesia response fragments, and perform long-timescale learning to generate a set of stability dissipation assessment results and a set of remaining stability margins; Based on the short-timescale control suggestion set, the medium-timescale control rhythm set, the stability dissipation assessment result set, and the remaining stability margin set, a time-scale competition judgment set is constructed. Control layer conflict identification is performed, a time-scale competition score sequence is generated, and a control weight allocation vector set is determined. Based on the control weight allocation vector set, the stability dissipation assessment result set, and the remaining stability margin set, constrained fusion is performed to generate a rhythmic anesthesia control strategy, which is then distributed to the anesthesia infusion execution unit and the ventilator control execution unit. The standardized time-series data set for thoracic anesthesia is updated cyclically until the anesthesia for thoracic surgery ends.

[0021] In this embodiment, the multi-source raw time-series data for thoracic anesthesia include anesthetic drug infusion rate, end-tidal concentration of inhaled anesthetic, respiratory rate, peak airway pressure, end-tidal carbon dioxide partial pressure, blood oxygen saturation, heart rate, invasive arterial blood pressure, and anesthesia depth index.

[0022] In this embodiment, preprocessing includes time alignment, outlier removal, missing segment imputation, and normalization.

[0023] In this embodiment, performing sliding segmentation specifically includes: An anesthesia physiological response decomposition configuration set was constructed based on the standardized time series data set of thoracic anesthesia. The anesthesia physiological response decomposition configuration set includes short time scale response window length, medium time scale response window length and long time scale response window length. Each response window length is used to limit the analysis interval of anesthesia physiological response at different time scales. Based on the short timescale response window length, the standardized time series data set of thoracic anesthesia is subjected to short timescale sliding segmentation along the time axis. A preset overlap ratio is maintained between adjacent time windows to generate a set of short timescale anesthesia response segments arranged in chronological order, which is used to characterize the transient change behavior of rapid perturbation-type physiological responses during anesthesia. Based on the length of the medium timescale response window, a medium timescale sliding segmentation operation is performed on the standardized time series data set of thoracic anesthesia to form a set of medium timescale anesthesia response segments covering multiple short timescale windows, which are used to describe the evolution of anesthesia depth, analgesia level and circulatory status in a continuous time interval. Based on the length of the long timescale response window, a long timescale sliding segmentation operation is performed on the standardized time series data set of thoracic anesthesia to generate a set of long timescale anesthesia response segments that span the entire anesthesia stage or its continuous sub-stages, which are used to reflect the long-term changes in the cumulative effect of anesthetic drugs and physiological baseline drift. Time index mapping relationships are established for the short-timescale anesthesia response fragment set, the medium-timescale anesthesia response fragment set, and the long-timescale anesthesia response fragment set, respectively, to ensure that the correspondence of anesthesia response fragments at different time scales remains consistent on the same time axis, forming a scale-aligned anesthesia physiological response fragment structure set.

[0024] In this embodiment, the generation of the short timescale control suggestion set includes: Obtain a set of short-timescale anesthesia response fragments. For each short-timescale anesthesia response fragment, extract the anesthetic drug infusion rate, airway peak pressure, end-tidal carbon dioxide partial pressure, blood oxygen saturation, heart rate, and invasive arterial blood pressure within the corresponding time interval to construct the short-timescale original response sequence. Based on the short-timescale original response sequence, combined with the direction of change of anesthetic drug infusion rate within the short-timescale response window, the transient change rate characteristics of invasive arterial blood pressure and the transient change rate characteristics of heart rate are calculated to characterize the instantaneous response intensity of the circulatory system under the condition of anesthetic drug administration changes. Within a short timescale response window, the infusion rate of anesthetic drugs is extracted, and the direction of change of the infusion rate within that short timescale response window is calculated to form a drug administration direction of change identifier. Within the same short timescale response window, the blood pressure difference sequence between adjacent sampling points is calculated in chronological order for invasive arterial blood pressure. Based on the blood pressure difference sequence and the drug administration direction of change identifier, intervals with the same direction of change are selected, and the mean blood pressure change rate within the intervals with the same direction of change is calculated to obtain the transient rate of change of invasive arterial blood pressure. Within the same short timescale response window, the heart rate difference sequence between adjacent sampling points is calculated in chronological order, and intervals with the same direction of change of heart rate are selected based on the drug administration direction of change identifier. The mean heart rate change rate within the intervals with the same direction of change is calculated to obtain the transient rate of change of heart rate. Short-term trend analysis of blood oxygen saturation was performed to calculate the slope characteristics of the decrease in blood oxygen saturation, which was used to characterize the degree of transient deterioration of oxygenation under the current ventilation parameters. Blood oxygen saturation is extracted within a short time-scaled response window. The blood oxygen saturation difference sequence between adjacent sampling points is calculated in chronological order. Continuous negative difference segments are extracted from the blood oxygen difference sequence, and the cumulative decrease in blood oxygen saturation is calculated for each negative difference segment, yielding the cumulative decrease in blood oxygen saturation. The number of consecutive sampling points within the negative difference segment is counted, and the duration of the decrease is calculated using a uniform sampling step size, yielding the duration of the decrease. Based on the cumulative decrease in blood oxygen saturation and the duration of the decrease, a blood oxygen saturation decrease slope feature is generated. Identify abnormal changes in airway peak pressure within a short timescale response window and extract the amplitude characteristics of airway peak pressure mutations to characterize rapid perturbation behavior caused by changes in ventilation load and lung compliance. Transient offset analysis was performed on the end-tidal carbon dioxide partial pressure to calculate the transient offset characteristics of end-tidal carbon dioxide, which are used to reflect the short-term deviation of the gas exchange state under the current ventilation rhythm. The transient rate of change of invasive arterial blood pressure, transient rate of change of heart rate, slope of decrease of blood oxygen saturation, amplitude of sudden change of airway peak pressure, and transient shift of carbon dioxide at end-tidal are concatenated and normalized to obtain a set of short-timescale perturbation feature vectors. Short-timescale learning is performed based on the set of short-timescale perturbation feature vectors to generate a set of short-timescale control proposals, which are used to characterize the immediate control adjustment direction and adjustment amplitude constraints under the current short timescale. During short-timescale learning, short-timescale disturbance feature vector sets are recorded within multiple consecutive short-timescale response windows to form a historical sample set. For each record in the historical sample set, a control response effect value is calculated. The control response effect value is the decrease in the sum of the absolute values ​​of the blood pressure transient rate of change feature value and the heart rate transient rate of change feature value in the next short-timescale response window after the control response is executed, relative to the corresponding value in the current window. The effect value is then divided into a fixed number of effect level values ​​according to preset segment intervals. For the current short-timescale disturbance feature vector, the sum of the absolute values ​​of the five feature differences between it and each short-timescale disturbance feature vector in the historical sample set is calculated to obtain a similarity value sequence. A fixed number of historical samples with the smallest similarity value are selected, and the level with the most frequent occurrence of its effect level value is counted to obtain the predicted level value. Based on the predicted level value, corresponding control adjustment direction values ​​and control adjustment amplitude values ​​are generated, resulting in short-timescale control suggestion entries. All short-timescale control suggestion entries are summarized to obtain a short-timescale control suggestion set.

[0025] In this embodiment, the generation of the time-scaled control rhythm set includes: Obtain a set of mid-timescale anesthesia response fragments. For each mid-timescale anesthesia response fragment, extract the anesthesia depth index, end-tidal concentration of inhaled anesthetic, invasive arterial blood pressure, heart rate and respiratory rate within the corresponding time interval to construct the mid-timescale original evolutionary response sequence. Based on the original evolution response sequence of the mid-timescale, a time trend analysis was performed on the anesthesia depth index to calculate the slope characteristics and trend stability characteristics of the anesthesia depth index, which are used to characterize the evolution direction and stability of the anesthesia depth within the mid-timescale time interval. Fluctuation analysis was performed on invasive arterial blood pressure and heart rate to calculate the characteristics of blood pressure fluctuation amplitude, blood pressure mean drift, heart rate fluctuation amplitude, and heart rate mean drift, which were used to characterize the stability evolution of the circulatory system over a continuous time interval. Rhythm analysis was performed on respiratory rate to calculate the trend characteristics of respiratory rate changes and rhythm stability characteristics, which were used to reflect the adjustment trend of ventilation control at the mid-time scale. Within the time interval corresponding to each mid-timescale anesthesia response segment, a respiratory rate sequence is acquired as the original evolution sequence of the mid-timescale respiratory rate. The respiratory rate difference between adjacent sampling points is calculated in chronological order to form a respiratory rate variation difference sequence. Sign statistics are performed on the respiratory rate variation difference sequence in chronological order to calculate the number of occurrences of positive and negative differences. Based on the relationship between the number of positive and negative differences, the direction of the respiratory rate change trend is determined, yielding the respiratory rate change trend characteristics. The absolute value of the respiratory rate variation difference sequence is taken to form a respiratory rate change amplitude sequence. The mean amplitude of the respiratory rate change amplitude sequence is calculated within the mid-timescale response window to obtain the average respiratory rate change amplitude value. The variance of the respiratory rate change amplitude sequence is calculated to obtain the respiratory rate change dispersion value. The average respiratory rate change amplitude value and the respiratory rate change dispersion value are weighted to generate a respiratory rate rhythm stability feature, used to characterize the rhythm stability of the respiratory rate at the mid-timescale time scale. Smoothing and trend analysis were performed on the end-tidal concentration of inhaled anesthetics, and the slope characteristics of the end-tidal concentration change were calculated to characterize the continuous change behavior of the effect of inhaled anesthetics in the mid-timescale interval. The characteristics of the slope of the change in the depth of anesthesia index, the trend stability of the depth of anesthesia index, the amplitude of blood pressure fluctuation, the drift of the mean blood pressure, the amplitude of heart rate fluctuation, the drift of the mean heart rate, the trend of the change in respiratory rate, the rhythm stability of respiratory rate, and the slope of the change in end-expiratory concentration are concatenated and normalized to obtain a set of mid-timescale evolution feature vectors. Midtimescale learning is performed based on the midtimescale evolution feature vector set to generate a midtimescale control rhythm set, which is used to characterize the control constraints of the anesthesia depth maintenance rhythm and ventilation rhythm adjustment within a continuous time interval. During mid-timescale learning, a set of mid-timescale evolutionary feature vectors arranged chronologically is obtained. For each mid-timescale evolutionary feature vector, a mid-timescale control effect value is calculated. The mid-timescale control effect value is the decrease in the sum of the deviations between the anesthesia depth index and the invasive arterial blood pressure at the end of the corresponding mid-timescale anesthesia response segment, relative to the sum of the corresponding deviations at the end of the next mid-timescale anesthesia response segment. The mid-timescale control effect value is quantified into a mid-timescale effect level value according to a preset segmented interval. The current mid-timescale evolutionary feature vector and the historical mid-timescale features in the mid-timescale evolutionary feature vector set are compared. The sum of the absolute values ​​of the feature differences of the evolutionary feature vectors is calculated to obtain the mid-timescale similarity value sequence. A fixed number of historical samples with the smallest values ​​in the mid-timescale similarity value sequence are selected, and the mid-timescale effect level value that appears most frequently is counted to obtain the mid-timescale prediction level value. The direction value of the anesthetic drug infusion rhythm and the direction value of the respiratory rate regulation rhythm are determined based on the mid-timescale prediction level value, and the corresponding rhythm step size value is determined respectively to generate mid-timescale control rhythm entries. All mid-timescale control rhythm entries in the current time interval are summarized to obtain the mid-timescale control rhythm set.

[0026] In this embodiment, the generation of the stability dissipation assessment result set and the remaining stability margin set specifically includes: Obtain a set of long-timescale anesthesia response fragments. For each long-timescale anesthesia response fragment, extract the anesthesia depth index, invasive arterial blood pressure, heart rate and respiratory rate within the corresponding time interval to construct the long-timescale original cumulative response sequence. The difference in the change between adjacent sampling points in the original cumulative response sequence of the long timescale is calculated in chronological order. The absolute value of the difference is then summed to obtain the cumulative change value of the anesthesia depth index, which is used to characterize the cumulative fluctuation of the anesthesia depth in the long timescale interval. The mean drift and cumulative fluctuation values ​​of invasive arterial blood pressure and heart rate were calculated over long time intervals to characterize the continuous shift and stability loss of the circulatory system baseline state over long time scales. Within the time interval corresponding to each long-timescale anesthesia response segment, invasive arterial blood pressure and heart rate sequences were acquired as long-timescale raw circulatory response sequences. The arithmetic mean of these long-timescale raw circulatory response sequences was calculated within the corresponding time interval to obtain the long-timescale average blood pressure and heart rate values. Blood pressure and heart rate values ​​at the start of each long-timescale anesthesia response segment were acquired as blood pressure and heart rate baseline values, respectively. The absolute value of the difference between the long-timescale average blood pressure and the blood pressure baseline values ​​was calculated to obtain the blood pressure mean drift value. The absolute value of the difference between the long-timescale average heart rate and the heart rate baseline values ​​was calculated to obtain the heart rate mean drift value. For the invasive arterial blood pressure sequence, the blood pressure difference between adjacent sampling points was calculated in chronological order, and the absolute value of the blood pressure difference was accumulated within the long-timescale time interval to obtain the cumulative blood pressure fluctuation value. For the heart rate sequence, the heart rate difference between adjacent sampling points was calculated in chronological order, and the absolute value of the heart rate difference was accumulated within the long-timescale time interval to obtain the cumulative heart rate fluctuation value. The cumulative value of rhythm deviation within a long time-scale interval is calculated for respiratory rate to reflect the continuous adjustment intensity of respiratory control rhythm within the long time-scale interval; Within the time interval corresponding to each long-timescale anesthesia response segment, a respiratory rate sequence is obtained as the original evolution sequence of the long-timescale respiratory rate. The respiratory rate value corresponding to the start time of the long-timescale anesthesia response segment is obtained as the baseline value of the respiratory rate. The original evolution sequence of the long-timescale respiratory rate is traversed chronologically for each sampling point. The difference between the respiratory rate value of the current sampling point and the baseline value of the respiratory rate is calculated, and the absolute value of the difference is taken to obtain the respiratory rate offset value of the current sampling point. The respiratory rate offset values ​​of each sampling point are accumulated within the long-timescale time interval to obtain the cumulative value of the respiratory rate rhythm offset, and the cumulative value of the respiratory rate rhythm offset is used as the cumulative value of the rhythm offset. The cumulative change values ​​of the anesthesia depth index, the mean blood pressure drift value, the cumulative blood pressure fluctuation value, the mean heart rate drift value, the cumulative heart rate fluctuation value, and the cumulative rhythm deviation value are concatenated and normalized to obtain a long-timescale cumulative feature vector set. Long-timescale learning is performed based on the long-timescale cumulative feature vector set. For each long-timescale anesthesia response segment, a set of stability dissipation assessment results is generated to characterize the degree to which the stability of the anesthesia system is continuously consumed within the corresponding long-timescale time interval. During long-timescale learning, a set of long-timescale cumulative feature vectors arranged chronologically is obtained. Based on the time index mapping relationship of the long-timescale anesthesia response segment set, the corresponding next long-timescale anesthesia response segment is determined for each long-timescale cumulative feature vector. For the next long-timescale anesthesia response segment, the sum of the cumulative change value of the anesthesia depth index, the mean blood pressure drift value, the cumulative blood pressure fluctuation value, the mean heart rate drift value, the cumulative heart rate fluctuation value, and the cumulative respiratory rate rhythm deviation value is calculated to obtain the next long-timescale comprehensive consumption value. The same comprehensive consumption value is calculated for the long-timescale anesthesia response segment corresponding to the current long-timescale cumulative feature vector to obtain the current long-timescale comprehensive consumption value. The next long-timescale comprehensive consumption value is then calculated and... The difference between the current long-term comprehensive consumption values ​​yields the long-term stability change value; the historical long-term stability change values ​​are divided into levels according to preset segment intervals to form a long-term stability level set; the sum of the absolute values ​​of the feature differences between the current long-term cumulative feature vector and the historical long-term cumulative feature vectors is calculated to obtain a long-term similarity value sequence; a fixed number of historical samples with the smallest long-term similarity values ​​are selected, and the level corresponding to the most frequent occurrence of the long-term stability level is counted to obtain the current long-term stability prediction level value; based on the current long-term stability prediction level value, the corresponding stability dissipation assessment result value is generated, and the above process is performed on all long-term anesthesia response segments to form a stability dissipation assessment result set; Based on the difference between the stability dissipation assessment result set and the preset stability benchmark value, the corresponding residual stability margin value is calculated to form a residual stability margin set, which is used to characterize the stability adjustment space that the system can still withstand during subsequent anesthesia control.

[0027] In this embodiment, the generation of the weight allocation vector set specifically includes: Based on the time index mapping relationship between the short-timescale anesthesia response fragment set, the medium-timescale anesthesia response fragment set, and the long-timescale anesthesia response fragment set, a corresponding medium-timescale control rhythm entry, stability dissipation assessment result value, and remaining stability margin value are matched for each short-timescale anesthesia response fragment. Extract short-timescale control suggestion entries within the corresponding time interval from the short-timescale control suggestion set, and obtain the control adjustment direction value and control adjustment amplitude value of the short-timescale control suggestion entries; Extract the mid-timescale control rhythm entries within the corresponding time interval from the mid-timescale control rhythm set, and obtain the anesthetic drug infusion rhythm direction value, anesthetic drug infusion rhythm step size value, respiratory rate regulation rhythm direction value, and respiratory rate regulation rhythm step size value of the mid-timescale control rhythm entries. Under the dimension of anesthetic drug infusion control, the consistency of the control adjustment direction value of short-timescale control suggestion items and the anesthetic drug infusion rhythm direction value of medium-timescale control rhythm items is determined, and the consistency determination value of anesthetic drug infusion direction is generated. Obtain short-timescale control suggestion entries within the corresponding time interval and read the control adjustment direction value; obtain mid-timescale control rhythm entries matching the short-timescale anesthesia response segment and read the anesthetic drug infusion rhythm direction value; perform sign unification processing on the control adjustment direction value and the anesthetic drug infusion rhythm direction value, mapping the increasing direction to a direction sign value of 1, the decreasing direction to a direction sign value of -1, and the holding direction to a direction sign value of 0, obtaining the short-timescale direction sign value and the mid-timescale direction sign value; calculate the absolute value of the difference between the short-timescale direction sign value and the mid-timescale direction sign value to obtain the direction difference value; generate an anesthetic drug infusion direction consistency judgment value based on the direction difference value, outputting a consistency judgment value of 1 when the direction difference value is 0, and outputting a consistency judgment value of 0 when the direction difference value is not 0; generate an anesthetic drug infusion direction consistency judgment value. Under the respiratory rate regulation control dimension, the consistency of the control regulation direction value of the short time-scale control suggestion item and the respiratory rate regulation rhythm direction value of the medium time-scale control rhythm item is determined, and a respiratory rate regulation direction consistency determination value is generated. Obtain short-timescale control suggestion entries within the corresponding time interval and read the control regulation direction value; obtain mid-timescale control rhythm entries matching the short-timescale anesthesia response fragment and read the respiratory rate regulation rhythm direction value; perform sign unification processing on the control regulation direction value and the respiratory rate regulation rhythm direction value, mapping the increment direction to a direction sign value of 1, the decrement direction to a direction sign value of -1, and the hold direction to a direction sign value of 0, thus obtaining the short-timescale direction sign value and the mid-timescale direction sign value; calculate the absolute value of the difference between the short-timescale direction sign value and the mid-timescale direction sign value to obtain the direction difference value; generate a respiratory rate regulation direction consistency judgment value based on the direction difference value, outputting a consistency judgment value of 1 when the direction difference value is 0, and outputting a consistency judgment value of 0 when the direction difference value is not 0; generate a respiratory rate regulation direction consistency judgment value. The consistency judgment value of anesthetic drug infusion direction is combined with the consistency judgment value of respiratory rate regulation direction to generate control layer conflict identification value, which is used to characterize whether there is a control direction conflict between the short time scale control suggestion set and the medium time scale control rhythm set in the same time interval; Obtain the consistency judgment values ​​for the anesthetic drug infusion direction and the respiratory rate regulation direction within the corresponding time interval; perform binary validity checks on the consistency judgment values ​​for the anesthetic drug infusion direction and the respiratory rate regulation direction respectively to obtain the valid values ​​for anesthetic drug infusion consistency and respiratory rate regulation consistency; sum the valid values ​​for anesthetic drug infusion consistency and respiratory rate regulation consistency to obtain the consistency count value; generate a control layer conflict identification value based on the consistency count value. When the consistency count value is 2, output the control layer conflict identification value 0; when the consistency count value is less than 2, output the control layer conflict identification value 1. Based on the control layer conflict identification value, the stability dissipation assessment result value and the remaining stability margin value within the corresponding time interval, the corresponding time scale competition score is generated and arranged in chronological order to form a time scale competition score sequence. The process involves: acquiring control layer conflict identification values, stability dissipation assessment results, and remaining stability margin values ​​within the corresponding time interval; calculating the ratio between the stability dissipation assessment results and the remaining stability margin values ​​to obtain a dissipation margin ratio, which characterizes the degree of stability consumption relative to the occupancy of the tolerable stability space; performing interval quantization based on the dissipation margin ratio to generate dissipation margin level values; summing the control layer conflict identification values ​​and the dissipation margin level values ​​to obtain a timescale competition score; repeating the acquisition, ratio calculation, interval quantization, and summation steps for each short-timescale anesthesia response segment to obtain a timescale competition score set arranged in chronological order, and outputting the timescale competition score set as a timescale competition score sequence. Based on the time-scale competition score sequence, a weight allocation determination is performed on the short-timescale control suggestion set and the medium-timescale control rhythm set to generate a control weight allocation vector set; The system acquires a timescale competition score sequence arranged chronologically and determines corresponding short-timescale control suggestion entries and medium-timescale control rhythm entries for each timescale competition score. It pre-sets initial values ​​for short-timescale and medium-timescale control weights for the system, using these initial values ​​as the weight allocation benchmark. It performs interval determination on the timescale competition scores corresponding to the current time interval, generating a timescale competition level value to characterize the competition intensity level between short-timescale and medium-timescale controls within the current time interval. Based on the timescale competition level value, it determines the weight adjustment amount. During system initialization, it pre-constructs a correspondence set between timescale competition level values ​​and weight adjustment values, configuring a unique short-timescale weight adjustment value and a unique medium-timescale weight adjustment value for each timescale competition level value in the correspondence set. This correspondence set is then stored as a weight adjustment mapping set. The weight adjustment is determined by experience or historical statistics. Based on the timescale competition level value of the current time interval, a matching mapping entry is searched in the weight adjustment mapping set, and the corresponding short-timescale weight adjustment value and medium-timescale weight adjustment value are read. The read short-timescale weight adjustment value and medium-timescale weight adjustment value are output as the weight adjustment value determined in the current time interval. The short-timescale weight adjustment value is applied to the initial value of the short-timescale control weight to obtain the short-timescale weight value in the current time interval, and the medium-timescale weight adjustment value is applied to the initial value of the medium-timescale control weight to obtain the medium-timescale weight value in the current time interval. The short-timescale weight value and the medium-timescale weight value are normalized to generate a control weight allocation vector. The above weight determination and update steps are repeated for each time interval in the timescale competition score sequence to form a set of control weight allocation vectors arranged in chronological order.

[0028] In this embodiment, the control weight allocation vector includes a short timescale weight value and a medium timescale weight value, and the sum of the short timescale weight value and the medium timescale weight value is a fixed value of 1.

[0029] In this embodiment, performing constrained fusion specifically includes: Obtain the control weight allocation vector, stability dissipation assessment result, and remaining stability margin value within the corresponding time interval; Extract short-timescale control suggestion entries within the corresponding time interval from the short-timescale control suggestion set, and obtain the control adjustment amplitude value; Extract the mid-timescale control rhythm entries within the corresponding time interval from the mid-timescale control rhythm set, and obtain the step size values ​​of anesthetic drug infusion rhythm and respiratory rate regulation rhythm. Based on the short-timescale weight value and the medium-timescale weight value in the control weight allocation vector, the control adjustment amplitude value of the short-timescale control suggestion item and the anesthetic drug infusion rhythm step value of the medium-timescale control rhythm item are weighted and summed to generate the fused anesthetic drug infusion adjustment value. The control adjustment amplitude value of the short-timescale control suggestion item and the respiratory rate adjustment rhythm step value of the medium-timescale control rhythm item are weighted and summed to generate the fused respiratory rate adjustment value. Based on the stability dissipation assessment results and the remaining stability margin, constraint judgments are performed on the fused anesthetic drug infusion regulation value and respiratory rate regulation value to generate constrained regulation values. The system acquires the stability dissipation assessment result, remaining stability margin, fused anesthetic drug infusion adjustment value, and fused respiratory rate adjustment value corresponding to the current time interval. Based on the stability dissipation assessment result, it performs margin correction on the remaining stability margin value to generate an effective remaining stability margin value. Based on the effective remaining stability margin value, it determines the upper limits of allowable adjustment for anesthetic drug infusion and respiratory rate adjustment, respectively. It takes the absolute value of the fused anesthetic drug infusion adjustment value and compares it with the upper limits of allowable adjustment for anesthetic drug infusion to generate an anesthetic drug infusion exceedance judgment value. Based on the anesthetic drug infusion exceedance judgment value, it generates a constrained adjustment value for anesthetic drug infusion. When the anesthetic drug infusion exceedance judgment value indicates that the limit has not been exceeded, the output value is equal to the fused anesthetic drug infusion adjustment value. When the anesthetic drug infusion exceedance judgment value indicates an exceedance, the output value is equal to the upper limit of the allowable adjustment of the anesthetic drug infusion and the amplitude limit result of the fused anesthetic drug infusion adjustment value in the same direction; the absolute value of the fused respiratory rate adjustment value is taken and compared with the upper limit of the allowable adjustment of the respiratory rate to generate a respiratory rate exceedance judgment value; a restricted respiratory rate adjustment value is generated based on the respiratory rate exceedance judgment value. When the respiratory rate exceedance judgment value indicates no exceedance, the output value is equal to the fused respiratory rate adjustment value; when the respiratory rate exceedance judgment value indicates exceedance, the output value is equal to the upper limit of the allowable adjustment of the respiratory rate and the amplitude limit result of the fused respiratory rate adjustment value in the same direction; the restricted anesthetic drug infusion adjustment value and the restricted respiratory rate adjustment value are output as the restricted adjustment value; By linearly mapping the constrained regulation values, the anesthetic drug infusion rhythm parameter values ​​and the respiratory rate regulation rhythm parameter values ​​are obtained, and a rhythmic anesthesia control strategy is generated. The rhythmic anesthesia control strategy, after integrating short-timescale control suggestions, medium-timescale control rhythms and stability constraints, forms a set of execution parameters with time continuity and amplitude limitation characteristics for simultaneously regulating the anesthetic drug infusion rate and respiratory rate. The rhythmic anesthesia control strategy is distributed to the anesthesia infusion execution unit and the ventilator control execution unit and the control operation is executed. After completing the control operations, the original time-series data of thoracic anesthesia from multiple sources were re-acquired and preprocessed to update the standardized time-series data set of thoracic anesthesia. The above steps for generating and executing the rhythmic anesthesia control strategy are repeated for subsequent time intervals until the anesthesia for thoracic surgery ends.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to the actual application scenario of thoracic surgery anesthesia in a tertiary general hospital. During thoracic surgeries such as thoracoscopic lobectomy, the hospital commonly experienced frequent manual adjustments to anesthetic drug infusion and ventilation parameters during the anesthesia maintenance phase. This was especially true under conditions of surgical traction, one-lung ventilation switching, and significant individual patient differences. Short-term fluctuations, lags in regulation, and repeated corrections were common in the depth of anesthesia, circulatory status, and respiratory status, increasing the workload of anesthesiologists and placing higher demands on the stability of the anesthesia process.

[0031] In this scenario, after the surgery begins, the anesthesia monitoring system continuously collects multi-source raw time-series data on thoracic anesthesia, including anesthetic drug infusion rate, end-tidal concentration of inhaled anesthetic, respiratory rate, peak airway pressure, end-tidal carbon dioxide partial pressure, blood oxygen saturation, heart rate, invasive arterial blood pressure, and anesthesia depth index, through the ventilator interface. Real-time time alignment, outlier removal, missing segment imputation, and normalization are performed to form a standardized time-series data set for thoracic anesthesia. This standardized data set is continuously updated in the background control system and divided into short-timescale, medium-timescale, and long-timescale anesthesia response segments according to preset sliding segmentation rules. Anesthesia response segments at different time scales are maintained on the same time axis through a time index mapping relationship.

[0032] At short timescales, the system extracts a set of perturbation feature vectors reflecting transient changes in circulation, ventilation load, and oxygenation status from short-timescale anesthesia response segments. This set is then combined with historical response samples for short-timescale learning to generate a set of short-timescale control proposals describing the constraints on immediate adjustment direction and amplitude. Within continuous time intervals, the system simultaneously performs evolutionary feature extraction and mid-timescale learning on mid-timescale anesthesia response segments, generating a set of mid-timescale control rhythms reflecting the maintenance rhythm of anesthesia depth and the rhythm of ventilation regulation. Simultaneously, cumulative feature analysis is performed on anesthesia response segments covering a longer time range to continuously assess the stability consumption of anesthesia depth, circulatory status, and respiratory rhythm, forming a set of stability dissipation assessment results. Based on this, a set of remaining stability margins is calculated to characterize the stability space that the system can withstand during subsequent adjustments.

[0033] During the control decision-making process, the system constructs a time-scale competition judgment set based on the short-timescale control suggestion set, the medium-timescale control rhythm set, and stability-related assessment results. It performs consistency analysis and conflict identification on control directions at different time scales and generates a time-scale competition score sequence based on the degree of stability dissipation. On this basis, the system dynamically determines the weight allocation relationship between short-timescale and medium-timescale control within the current time interval and performs constrained fusion of the two types of control information under stability constraints to generate a rhythmic anesthesia control strategy. This rhythmic anesthesia control strategy is distributed in real-time to the anesthesia infusion execution unit and the ventilator control execution unit to coordinately adjust the anesthetic drug infusion rate and respiratory rate parameters. After execution, it continuously updates the standardized time-series data set for thoracic anesthesia, forming a closed-loop control process.

[0034] To verify the performance of this invention in practice, it was compared with traditional methods, and the results are shown in Table 1. Table 1: Performance Comparison of the Method of this Invention and Traditional Anesthesia Control Methods in Thoracic Surgery.

[0035] As shown in Table 1, in terms of the fluctuation range of the anesthesia depth index and the transient fluctuation range of invasive arterial blood pressure, traditional anesthesia control methods have limited ability to suppress external stimuli and operational disturbances during thoracic surgery. The anesthesia depth and circulatory status are prone to significant fluctuations in a short period of time, especially during traction, body position changes, and ventilation adjustment phases, where fluctuations tend to superimpose. The method of this invention shows advantages in both of these indicators, indicating that by learning short-timescale disturbance characteristics and working synergistically with medium-timescale control rhythm, it is possible to suppress the rapid amplification of physiological responses at different time scales and prevent short-term disturbances from evolving into a continuous unstable state.

[0036] At the level of control behavior, traditional methods show a high number of adjustments to both the anesthetic drug infusion rate and respiratory rate, reflecting that the control strategy mainly relies on human experience or fixed rules for frequent corrections. The regulatory behavior exhibits discrete and iterative characteristics. This invention constructs a time-scale competitive decision set and assigns weights to short-timescale control suggestions and medium-timescale control rhythms, enabling control decisions to form a clear division of labor at different time scales, reducing unnecessary iterative adjustments, and making anesthetic infusion and ventilation regulation smoother.

[0037] The comparison of the proportion of control conflicts further demonstrates the advantages of this invention in terms of control consistency. In traditional methods, different control logics are prone to generating inconsistent adjustment instructions for anesthetic drug infusion and respiratory rate within the same time interval, which increases the risk of system oscillation. This invention introduces a control layer conflict identification mechanism and uses directional consistency as an important constraint condition for control decision-making, which significantly reduces the proportion of control conflicts and fundamentally improves the stability of multi-control source collaborative work.

[0038] Regarding long-term regulatory capacity, the improvement in the average level of remaining stability margin indicates that the present invention can effectively control the rate of stability dissipation throughout the anesthesia process, avoiding premature consumption of the system's regulatory space. This effect stems from the participation of the stability dissipation assessment result set and the remaining stability margin set in the control decision, ensuring that the system always retains sufficient regulatory margin when executing control.

[0039] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based method for controlling thoracic anesthesia, characterized in that, Includes the following steps: During thoracic surgery anesthesia, a multi-source raw time-series data set of thoracic anesthesia was collected and preprocessed to generate a standardized time-series data set of thoracic anesthesia. Sliding segmentation was performed on the standardized time-series data set of thoracic anesthesia to generate short-timescale anesthesia response fragment sets, medium-timescale anesthesia response fragment sets, and long-timescale anesthesia response fragment sets, respectively; Extract a set of short-timescale perturbation feature vectors from the set of short-timescale anesthesia response fragments, perform short-timescale learning, and generate a set of short-timescale control suggestions; Extract the set of mid-timescale evolution feature vectors from the mid-timescale anesthesia response fragment set, and perform mid-timescale learning to generate a set of mid-timescale control rhythms; Extract a set of long-timescale cumulative feature vectors from the set of long-timescale anesthesia response fragments, and perform long-timescale learning to generate a set of stability dissipation assessment results and a set of remaining stability margins; Based on the short-timescale control suggestion set, the medium-timescale control rhythm set, the stability dissipation assessment result set, and the remaining stability margin set, a time-scale competition judgment set is constructed. Control layer conflict identification is performed, a time-scale competition score sequence is generated, and a control weight allocation vector set is determined. Based on the control weight allocation vector set, the stability dissipation assessment result set, and the remaining stability margin set, constrained fusion is performed to generate a rhythmic anesthesia control strategy, which is then distributed to the anesthesia infusion execution unit and the ventilator control execution unit. The standardized time-series data set for thoracic anesthesia is updated cyclically until the anesthesia for thoracic surgery ends.

2. The machine learning-based thoracic anesthesia control method according to claim 1, characterized in that, The multi-source raw time-series data for thoracic anesthesia include anesthetic drug infusion rate, end-tidal concentration of inhaled anesthetic drugs, respiratory rate, peak airway pressure, end-tidal carbon dioxide partial pressure, blood oxygen saturation, heart rate, invasive arterial blood pressure, and anesthesia depth index.

3. The machine learning-based thoracic anesthesia control method according to claim 1, characterized in that, The preprocessing includes time alignment, outlier removal, missing segment imputation, and normalization.

4. The machine learning-based thoracic anesthesia control method according to claim 1, characterized in that, The execution of sliding segmentation specifically includes: An anesthesia physiological response decomposition configuration set is constructed based on the standardized time-series data set of thoracic anesthesia. The anesthesia physiological response decomposition configuration set includes short time-scaled response window length, medium time-scaled response window length and long time-scaled response window length. Based on the short timescale response window length, the standardized time series data set of thoracic anesthesia is subjected to short timescale sliding segmentation along the time axis to generate a set of short timescale anesthesia response fragments arranged in chronological order; Based on the length of the medium-timescale response window, a medium-timescale sliding segmentation operation is performed on the standardized time-series data set of thoracic anesthesia to form a set of medium-timescale anesthesia response segments covering multiple short-timescale windows; Based on the length of the long timescale response window, a long timescale sliding segmentation operation is performed on the standardized time-series data set of thoracic anesthesia to generate a set of long timescale anesthesia response fragments. Time index mapping relationships are established for the short-timescale anesthesia response fragment set, the medium-timescale anesthesia response fragment set, and the long-timescale anesthesia response fragment set, respectively, to form a scale-aligned anesthesia physiological response fragment structure set.

5. The machine learning-based thoracic anesthesia control method according to claim 1, characterized in that, The generation of the short-timescale control suggestion set includes: Obtain a set of short-timescale anesthesia response fragments. For each short-timescale anesthesia response fragment, extract the anesthetic drug infusion rate, airway peak pressure, end-tidal carbon dioxide partial pressure, blood oxygen saturation, heart rate, and invasive arterial blood pressure within the corresponding time interval to construct the short-timescale original response sequence. Based on the short-timescale raw response sequence, and combined with the direction of change of anesthetic drug infusion rate within the short-timescale response window, the transient rate of change of invasive arterial blood pressure and the transient rate of change of heart rate were calculated. Perform short-term trend analysis on blood oxygen saturation and calculate the characteristics of the slope of blood oxygen saturation decline; Identify abnormal changes in airway peak pressure within a short timescale response window and extract the amplitude characteristics of sudden changes in airway peak pressure. Transient shift analysis was performed on the end-tidal carbon dioxide partial pressure to calculate the transient shift characteristics of end-tidal carbon dioxide. The transient rate of change of invasive arterial blood pressure, transient rate of change of heart rate, slope of decrease of blood oxygen saturation, amplitude of sudden change of airway peak pressure, and transient shift of carbon dioxide at end-tidal are concatenated and normalized to obtain a set of short-timescale perturbation feature vectors. Short-timescale learning is performed based on the set of short-timescaled perturbation feature vectors to generate a set of short-timescaled control proposals.

6. The machine learning-based thoracic anesthesia control method according to claim 1, characterized in that, The generation of the time-scaled control rhythm set includes: Obtain a set of mid-timescale anesthesia response fragments. For each mid-timescale anesthesia response fragment, extract the anesthesia depth index, end-tidal concentration of inhaled anesthetic, invasive arterial blood pressure, heart rate and respiratory rate within the corresponding time interval to construct the mid-timescale original evolutionary response sequence. Based on the original evolutionary response sequence of the mid-timescale, the anesthesia depth index was subjected to time-time trend analysis, and the slope characteristics and trend stability characteristics of the anesthesia depth index were calculated. Fluctuation analysis was performed on invasive arterial blood pressure and heart rate separately to calculate the characteristics of blood pressure fluctuation amplitude, blood pressure mean drift, heart rate fluctuation amplitude, and heart rate mean drift. Perform rhythm analysis on respiratory rate to calculate the trend characteristics and rhythm stability characteristics of respiratory rate changes; Smoothing and trend analysis were performed on the end-tidal concentration of inhaled anesthetics, and the slope characteristics of the end-tidal concentration change were calculated. The characteristics of the slope of the change in the depth of anesthesia index, the trend stability of the depth of anesthesia index, the amplitude of blood pressure fluctuation, the drift of the mean blood pressure, the amplitude of heart rate fluctuation, the drift of the mean heart rate, the trend of the change in respiratory rate, the rhythm stability of respiratory rate, and the slope of the change in end-expiratory concentration are concatenated and normalized to obtain a set of mid-timescale evolution feature vectors. Based on the set of evolutionary feature vectors of the medium time scale, medium time scale learning is performed to generate a set of medium time scale control rhythms.

7. The machine learning-based thoracic anesthesia control method according to claim 1, characterized in that, The generation of the stability dissipation assessment result set and the residual stability margin set specifically includes: Obtain a set of long-timescale anesthesia response fragments. For each long-timescale anesthesia response fragment, extract the anesthesia depth index, invasive arterial blood pressure, heart rate and respiratory rate within the corresponding time interval to construct the long-timescale original cumulative response sequence. For the anesthesia depth index, the cumulative change value of the anesthesia depth index is obtained; Calculate the mean drift and cumulative fluctuation values ​​of invasive arterial blood pressure and heart rate over long time scale intervals, respectively. Calculate the cumulative value of rhythm deviation within a long time-scaled interval for respiratory rate; The cumulative change values ​​of the anesthesia depth index, the mean blood pressure drift value, the cumulative blood pressure fluctuation value, the mean heart rate drift value, the cumulative heart rate fluctuation value, and the cumulative rhythm deviation value are concatenated and normalized to obtain a long-timescale cumulative feature vector set. Long-timescale learning is performed based on the long-timescale cumulative feature vector set, and a set of stability dissipation assessment results is generated for each long-timescale anesthesia response segment. Based on the difference between the stability dissipation assessment result set and the preset stability benchmark value, the corresponding residual stability margin value is calculated to form the residual stability margin set.

8. The machine learning-based thoracic anesthesia control method according to claim 1, characterized in that, The generation of the control weight allocation vector set specifically includes: Based on the time index mapping relationship between the short-timescale anesthesia response fragment set, the medium-timescale anesthesia response fragment set, and the long-timescale anesthesia response fragment set, a corresponding medium-timescale control rhythm entry, stability dissipation assessment result value, and remaining stability margin value are matched for each short-timescale anesthesia response fragment. Extract short-timescale control suggestion entries within the corresponding time interval from the short-timescale control suggestion set, and obtain the control adjustment direction value and control adjustment amplitude value of the short-timescale control suggestion entries; Extract the mid-timescale control rhythm entries within the corresponding time interval from the mid-timescale control rhythm set, and obtain the anesthetic drug infusion rhythm direction value, anesthetic drug infusion rhythm step size value, respiratory rate regulation rhythm direction value, and respiratory rate regulation rhythm step size value of the mid-timescale control rhythm entries. Under the dimension of anesthetic drug infusion control, the consistency of the control adjustment direction value of short-timescale control suggestion items and the anesthetic drug infusion rhythm direction value of medium-timescale control rhythm items is determined, and the consistency determination value of anesthetic drug infusion direction is generated. Under the respiratory rate regulation control dimension, the consistency of the control regulation direction value of the short time-scale control suggestion item and the respiratory rate regulation rhythm direction value of the medium time-scale control rhythm item is determined, and a respiratory rate regulation direction consistency determination value is generated. The consistency judgment value of anesthetic drug infusion direction is combined with the consistency judgment value of respiratory rate regulation direction to generate control layer conflict identification value; Based on the control layer conflict identification value, the stability dissipation assessment result value and the remaining stability margin value within the corresponding time interval, the corresponding time scale competition score is generated and arranged in chronological order to form a time scale competition score sequence. Based on the time-scale competition score sequence, a weight allocation determination is performed on the short-timescale control suggestion set and the medium-timescale control rhythm set to generate a control weight allocation vector set.

9. The machine learning-based thoracic anesthesia control method according to claim 8, characterized in that, The control weight allocation vector includes a short timescale weight value and a medium timescale weight value, and the sum of the short timescale weight value and the medium timescale weight value is a fixed value of 1.

10. The machine learning-based thoracic anesthesia control method according to claim 1, characterized in that, The execution of constrained fusion specifically includes: Obtain the control weight allocation vector, stability dissipation assessment result, and remaining stability margin value within the corresponding time interval; Extract short-timescale control suggestion entries within the corresponding time interval from the short-timescale control suggestion set, and obtain the control adjustment amplitude value; Extract the mid-timescale control rhythm entries within the corresponding time interval from the mid-timescale control rhythm set, and obtain the step size values ​​of anesthetic drug infusion rhythm and respiratory rate regulation rhythm. Based on the short-timescale weight value and the medium-timescale weight value in the control weight allocation vector, the control adjustment amplitude value of the short-timescale control suggestion item and the anesthetic drug infusion rhythm step value of the medium-timescale control rhythm item are weighted and summed to generate the fused anesthetic drug infusion adjustment value. The control adjustment amplitude value of the short-timescale control suggestion item and the respiratory rate adjustment rhythm step value of the medium-timescale control rhythm item are weighted and summed to generate the fused respiratory rate adjustment value. Based on the stability dissipation assessment results and the remaining stability margin, constraint judgments are performed on the fused anesthetic drug infusion regulation value and respiratory rate regulation value to generate constrained regulation values. By linearly mapping the constrained regulation values, the anesthetic drug infusion rhythm parameter values ​​and the respiratory rate regulation rhythm parameter values ​​are obtained, and a rhythmic anesthesia control strategy is generated. The rhythmic anesthesia control strategy is distributed to the anesthesia infusion execution unit and the ventilator control execution unit and the control operation is executed. After completing the control operations, the original time-series data of thoracic anesthesia from multiple sources were re-acquired and preprocessed to update the standardized time-series data set of thoracic anesthesia. The above steps for generating and executing the rhythmic anesthesia control strategy are repeated for subsequent time intervals until the anesthesia for thoracic surgery ends.

Images