Intelligent Adjustment Method and System for ICU Ventilator Parameters
By acquiring multimodal monitoring data, a nonlinear coupling relationship between respiratory mechanics and oxygenation cycle is established, multi-objective constraint optimization is performed, and a parameter adjustment instruction set is generated. This solves the problems of lag and safety in ICU ventilator parameter adjustment and achieves precise and individualized parameter adjustment.
Patent Information
- Application Number
- CN202611285616.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing ICU ventilator parameter adjustment methods lack dynamic coupling analysis of respiratory mechanics and oxygenation cycle, resulting in parameter adjustment lag, which may lead to ventilator-associated lung injury or circulatory function deterioration, and lack of safety constraints, increasing the risk of barotrauma.
By acquiring multimodal monitoring data, respiratory mechanics and oxygenation cycle parameters are extracted, nonlinear coupling relationships are established, fusion state feature vectors are calculated, air pressure safety boundaries and human-machine synchronization constraints are introduced, multi-objective constraint optimization is performed to generate parameter adjustment instruction sets, and adaptive updates are performed.
It significantly improves the accuracy and safety of parameter adjustment, avoids barotrauma and human-ventilator aggression, and enhances the comfort and safety of ventilation support.
Smart Images

Figure CN123101792A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical equipment technology, and in particular to a method and system for intelligent adjustment of ICU ventilator parameters. Background Technology
[0002] In ICU ventilator support, parameter adjustments have long relied primarily on physicians' clinical experience and subjective judgment. Physicians typically adjust parameters such as tidal volume, respiratory rate, and inhaled oxygen concentration manually based on intermittently obtained blood gas analysis results, heart rate, blood pressure, and blood oxygen saturation—one or a few physiological indicators—combined with their own experience. This traditional method is widely used in clinical practice and constitutes the standard practice for ventilator management in intensive care units.
[0003] To improve adjustment efficiency and accuracy, some ventilators have introduced simple automatic feedback control mechanisms. However, such methods are difficult to capture the complex nonlinear interaction between the patient's respiratory and circulatory systems. Especially when patients experience pathological conditions such as acute respiratory distress syndrome or shock, single parameter adjustment often cannot take into account the overall balance of respiratory mechanics and oxygenation cycle.
[0004] Existing adjustment methods have significant drawbacks. They lack dynamic coupling analysis of respiratory mechanics parameters and oxygenation cycle parameters, relying solely on indirect indicators such as blood oxygen saturation or end-expiratory carbon dioxide. This can easily lead to parameter adjustments lagging behind the patient's actual physiological needs, and even cause ventilator-associated lung injury or circulatory deterioration. Furthermore, existing methods generally lack clear safety constraints, such as upper limits for airway pressure and patient-ventilator synchrony thresholds. This can increase the risk of barotrauma or exacerbate patient-ventilator asynchrony during adjustment, reducing treatment safety and patient comfort. These shortcomings hinder the further development of precise and individualized ventilator support therapy. Summary of the Invention
[0005] This invention provides a method and system for intelligent adjustment of ICU ventilator parameters, which can solve the problems in the prior art.
[0006] A first aspect of the present invention provides a method for intelligent adjustment of ICU ventilator parameters, comprising: Acquire multimodal monitoring data of the target patient, including respiratory flow waveform signals, airway pressure signals, blood oxygen saturation signals, and heart rate variability signals; Feature extraction is performed on the multimodal monitoring data to calculate respiratory mechanics parameters and oxygenation cycle parameters. By establishing a nonlinear coupling relationship between respiratory mechanics parameters and oxygenation cycle parameters, respiratory power consumption characteristics and oxygen supply and demand matching characteristics are calculated to generate a fusion state feature vector. Extract the current ventilator parameters and combine them with the fusion state feature vector to calculate the effectiveness score and risk level of the current respiratory support strategy; With the optimization objectives of maximizing the efficiency score and minimizing the risk level, pressure safety boundary constraints and human-machine synchronization constraints are introduced as solution boundaries. The ventilator parameter adjustment instruction set is generated through multi-objective constraint optimization. The parameter adjustment instruction set is sent to the ventilator and the adjusted physiological signal feedback data is continuously collected to adaptively update the parameters in the multi-objective constraint optimization.
[0007] Feature extraction is performed on the multimodal monitoring data to calculate respiratory mechanics parameters and oxygenation cycle parameters, including: Based on the multimodal monitoring data, the respiratory flow waveform signal and airway pressure signal are time-aligned, and a dynamic curve of flow and pressure is constructed in each respiratory cycle. The respiratory work is quantified by calculating the loop area enclosed by the dynamic curve. The slope of the major axis of the dynamic curve is extracted as a comprehensive compliance index. The curvature change points in the pressure rise phase of the dynamic curve are identified, and the pressure values corresponding to the curvature change points are used as alveolar recruitment pressure thresholds. The respiratory work, the comprehensive compliance index, and the alveolar recruitment pressure thresholds are weighted and summed to obtain respiratory mechanics parameters.
[0008] The blood oxygen saturation signal is processed by a sliding window. The time gradient of blood oxygen saturation is calculated in each time window. The proportion of time with negative time gradient is used as the risk factor for oxygenation deterioration. The heart rate variability signal is subjected to continuous wavelet transform. The energy distribution characteristics of wavelet coefficients are extracted within the target scale range and the concentration is calculated. The oxygenation cycle parameters are obtained by nonlinear mapping in combination with the risk factor for oxygenation deterioration.
[0009] By establishing a nonlinear coupling relationship between respiratory mechanics parameters and oxygenation cycle parameters, the respiratory power consumption characteristics and oxygen supply-demand matching characteristics are calculated, generating a fusion state feature vector, including: The mutual information values of the respiratory mechanics parameters and the oxygenation cycle parameters are calculated at multiple time scales. The time scale corresponding to the inflection point of the mutual information value curve is extracted as the coupling feature time, and a parameter mapping function is constructed based on the coupling feature time. The respiratory mechanics parameters are mapped to predicted values of corresponding oxygenation cycle parameters through the parameter mapping function. The predicted values are then subtracted from the synchronously acquired actual oxygenation cycle parameters to obtain a residual sequence. The decay rate is extracted from the autocorrelation function of the residual sequence. The residual sequence is decomposed into time-frequency components to obtain the dominant components. The cumulative energy of the dominant components within the sliding window is calculated. The fluctuation amplitude of the cumulative energy and the dominant frequency of the dominant components are extracted and combined to generate respiratory power consumption features. The distribution entropy value of the dominant components is calculated. The adjacent intervals of the extreme points of the mutual information value curve are statistically analyzed and the standard deviation of the interval is calculated. The distribution entropy value and the standard deviation are combined to generate oxygen supply and demand matching features. Based on the decay rate and the coupling feature time, a normalized fusion weight is calculated. The breathing power consumption feature and the oxygen supply and demand matching feature are then weighted and combined according to the fusion weight to obtain a fusion state feature vector.
[0010] Extract the current ventilator parameters and combine them with the fusion state feature vector to calculate the effectiveness score and risk level of the current respiratory support strategy, including: Extract the current inspiratory pressure setting, positive end-expiratory pressure setting, ventilation rate setting, and inspiratory-expiratory ratio setting of the ventilator, and construct a parameter state vector; The parameter state vector and the fused state feature vector are subjected to sliding matching within a time window. The Pearson correlation coefficient between each dimension component within each time window is calculated. The correlation coefficients of each dimension are combined to form a dynamic correlation matrix. The ratio of the diagonal elements to the off-diagonal elements in the dynamic correlation matrix is extracted as the coupling concentration degree. The time-varying coupling concentration degree curve is obtained by expanding it along the time axis. The cumulative synergy index is obtained by performing trapezoidal integral operation on the time-varying coupling concentration curve. The cumulative synergy index is input into the hyperbolic tangent function for nonlinear mapping to obtain the performance score. The condition number of the dynamic correlation matrix is extracted as a state stability measure. The number of components in the parameter state vector that deviate from the preset safe range is counted. The state stability measure and the number of components are weighted and combined and segmented threshold discrimination is performed. The risk level is determined based on the discrimination result.
[0011] With the optimization objectives of maximizing the efficiency score and minimizing the risk level, pressure safety boundary constraints and human-machine synchronization constraints are introduced as solution boundaries. A set of ventilator parameter adjustment instructions is generated through multi-objective constraint optimization, including: The effectiveness score and the risk level are used to construct an objective function. In the ventilator parameter space, the set of non-dominated solutions for ventilator parameters is obtained by Pareto front search. The non-dominated solution with the highest objective function value is selected as the optimization benchmark solution. A pressure safety corridor is constructed based on the peak envelope of the airway pressure signal. A time-varying upper bound for inspiratory pressure and a time-varying lower bound for positive end-expiratory pressure are set and smoothed to generate a pressure safety boundary constraint function. The respiratory flow waveform signal is segmented into respiratory cycles to identify the patient's spontaneous inhalation start time and the ventilator trigger response time and calculate the phase difference angle. The phase difference angles of multiple respiratory cycles are statistically analyzed to construct a synchronization deviation sequence and variance analysis is performed to obtain the fluctuation intensity index. The joint constraint conditions of trigger window width and flow rate threshold are set to generate human-ventilator synchronization constraint function. Using the optimized benchmark solution as the initial search point, a sequential quadratic programming method is used to iteratively search for optimization within the feasible region defined by the air pressure safety boundary constraint function and the human-machine synchronization constraint function. The iteration terminates when the gradient magnitude of the objective function is less than the convergence criterion, and the corresponding parameter combination is extracted to generate a ventilator parameter adjustment instruction set.
[0012] In the ventilator parameter space, the non-dominated solution set of ventilator parameters is obtained through Pareto front search, including: In the ventilator parameter space, the range of values for inspiratory pressure, positive end-expiratory pressure, ventilation rate and inspiratory-expiratory ratio are set and the grid is discretized. Multiple initial parameter combinations are randomly selected from the discrete grid points, and the efficacy score and risk level value are calculated for each initial parameter combination. Perform Pareto dominance judgment on all initial parameter combinations. If the effectiveness score of the first parameter combination is not lower than that of the second parameter combination and the risk level value is not higher than that of the second parameter combination, then the first parameter combination dominates the second parameter combination. Extract all parameter combinations that are not dominated to form the current non-dominated solution set. For each parameter combination in the current non-dominated solution set, identify the nearest neighbor parameter combination and calculate the local sparsity, and select the parameter combination with a local sparsity greater than the sparsity threshold as the seed solution; Around various sub-solutions, perturbations are applied to parameters of each dimension in the parameter space to generate new parameter combinations, which are then merged with the current non-dominated solution set. The Pareto dominance determination is then re-executed to obtain an expanded non-dominated solution set. The iteration terminates when the coverage of the expanded non-dominated solution set reaches the convergence threshold, and the expanded non-dominated solution set is output.
[0013] A second aspect of the present invention provides an intelligent adjustment system for ICU ventilator parameters, comprising: The data acquisition unit is used to acquire multimodal monitoring data of the target patient, including respiratory flow waveform signals, airway pressure signals, blood oxygen saturation signals, and heart rate variability signals. The fusion feature unit is used to extract features from the multimodal monitoring data, calculate respiratory mechanics parameters and oxygenation cycle parameters, and calculate respiratory power consumption characteristics and oxygen supply and demand matching characteristics by establishing a nonlinear coupling relationship between respiratory mechanics parameters and oxygenation cycle parameters, thereby generating a fusion state feature vector. The strategy evaluation unit is used to extract the current ventilator parameters and calculate the effectiveness score and risk level of the current respiratory support strategy by combining the fusion state feature vector. The instruction optimization unit is used to maximize the efficiency score and minimize the risk level as optimization objectives. It introduces air pressure safety boundary constraints and human-machine synchronization constraints as solution boundaries, and generates a set of ventilator parameter adjustment instructions through multi-objective constraint optimization. The feedback update unit is used to send the parameter adjustment instruction set to the ventilator and continuously collect the adjusted physiological signal feedback data to adaptively update the parameters in the multi-objective constraint optimization.
[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.
[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.
[0016] This method significantly improves the accuracy of characterizing the dynamic coupling relationship between respiratory mechanics and oxygenation cycle by acquiring multimodal monitoring data from patients and performing deep feature extraction. The fused state feature vector constructed based on respiratory power consumption characteristics and oxygen supply-demand matching characteristics can comprehensively reflect the patient's physiological reserves and pathological progression, providing a high-dimensional real benchmark for subsequent parameter optimization, thereby effectively avoiding one-sided decisions caused by adjusting a single indicator.
[0017] The combined quantitative assessment based on efficacy scores and risk levels gives ventilator parameter adjustment clear target guidance and risk warning capabilities. After introducing pressure safety boundary constraints and human-ventilator synchronization constraints, the optimization solution process naturally ensures that the airway pressure does not exceed the safety threshold, while maintaining the coordination between the patient's spontaneous breathing and mechanical ventilation. This significantly reduces the probability of barotrauma, volumetric injury and human-ventilator asynchrony, and improves the comfort and safety of ventilation support.
[0018] The multi-objective constraint optimization mechanism achieves a dynamic balance between maximizing the efficiency score and minimizing the risk level. The generated parameter adjustment instruction set can accurately match the current patient's pathological state, avoiding the lag and bias caused by empirical adjustments or single algorithm optimization. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the intelligent adjustment method for ICU ventilator parameters according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the method for calculating the fusion state feature vector according to an embodiment of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.
[0022] Figure 1 This is a flowchart illustrating the intelligent adjustment method for ICU ventilator parameters according to an embodiment of the present invention. The present invention provides an intelligent adjustment method for ICU ventilator parameters, comprising: Acquire multimodal monitoring data of the target patient, including respiratory flow waveform signal, airway pressure signal, blood oxygen saturation signal and heart rate variability signal; Feature extraction is performed on the multimodal monitoring data to calculate respiratory mechanics parameters and oxygenation cycle parameters. By establishing a nonlinear coupling relationship between respiratory mechanics parameters and oxygenation cycle parameters, respiratory power consumption characteristics and oxygen supply and demand matching characteristics are calculated to generate a fusion state feature vector. Extract the current ventilator parameters and combine them with the fusion state feature vector to calculate the effectiveness score and risk level of the current respiratory support strategy; With the optimization objectives of maximizing the efficiency score and minimizing the risk level, pressure safety boundary constraints and human-machine synchronization constraints are introduced as solution boundaries. The ventilator parameter adjustment instruction set is generated through multi-objective constraint optimization. The parameter adjustment instruction set is sent to the ventilator and the adjusted physiological signal feedback data is continuously collected to adaptively update the parameters in the multi-objective constraint optimization.
[0023] Feature extraction is performed on the multimodal monitoring data to calculate respiratory mechanics parameters and oxygenation cycle parameters, including: Based on the multimodal monitoring data, the respiratory flow waveform signal and airway pressure signal are time-aligned, and a dynamic curve of flow and pressure is constructed in each respiratory cycle. The respiratory work is quantified by calculating the loop area enclosed by the dynamic curve. The slope of the long axis of the dynamic curve is extracted as a comprehensive compliance index. The curvature change points in the pressure rise phase of the dynamic curve are identified, and the pressure values corresponding to the curvature change points are used as alveolar recruitment pressure thresholds. The respiratory work, the comprehensive compliance index, and the alveolar recruitment pressure thresholds are weighted and summed to obtain respiratory mechanics parameters.
[0024] The blood oxygen saturation signal is processed by a sliding window. The time gradient of blood oxygen saturation is calculated in each time window. The proportion of time with negative time gradient is used as the risk factor for oxygenation deterioration. The heart rate variability signal is subjected to continuous wavelet transform. The energy distribution characteristics of wavelet coefficients are extracted within the target scale range and the concentration is calculated. The oxygenation cycle parameters are obtained by nonlinear mapping in combination with the risk factor for oxygenation deterioration.
[0025] After obtaining the respiratory flow waveform signal, airway pressure signal, blood oxygen saturation signal and heart rate variability signal of the target patient, it is necessary to extract features from the above multimodal monitoring data and calculate respiratory mechanics parameters and oxygenation cycle parameters respectively to support the subsequent construction of fusion state feature vectors.
[0026] For the extraction of respiratory mechanics parameters, time alignment processing of the respiratory flow waveform signal and airway pressure signal is required. Since the two signals are acquired from different sensor channels, there is a difference in hardware sampling delay. Therefore, it is necessary to estimate the time offset between the two signals using a cross-correlation method, with the starting point of the respiratory cycle as the reference. Time axis shift compensation is then applied to the lagging signal to ensure a strict time correspondence between flow and pressure at each sampling point. After time alignment, within each respiratory cycle, airway pressure is plotted on the horizontal axis and respiratory flow on the vertical axis, and the sampling points are connected sequentially according to the sampling time to construct a flow-pressure dynamic curve for that respiratory cycle, i.e., a pressure-flow loop (PF loop). This dynamic curve forms different trajectory paths during the inspiration and expiration phases, and the two trajectories together form a closed loop.
[0027] Calculating the area of this closed loop allows for the quantification of the respiratory work done in a single respiratory cycle. A larger loop area indicates that, under the same tidal volume conditions, more energy is consumed by airway resistance or elastic resistance, resulting in a heavier respiratory work burden. The loop area is calculated using a discrete integration method, taking the absolute value of the difference between the area under the inspiratory and expiratory curves to obtain a quantified value of the respiratory work, denoted as [the value is missing from the original text]. .
[0028] When extracting the comprehensive compliance index, principal component analysis is performed on the dynamic curve to determine the principal direction of the data point distribution, i.e., the major axis. The slope of the major axis reflects the comprehensive trend of flow rate changing with pressure throughout the respiratory cycle. This slope is defined as the comprehensive compliance index, denoted as [insert index here]. This indicator integrates dynamic compliance information from both the inspiratory and expiratory phases, and compared to traditional static compliance measurement methods, it can more comprehensively reflect the elastic properties of lung tissue during dynamic ventilation. The higher the value, the greater the flow response of the lung tissue under a unit pressure change, and the better the lung compliance; conversely, it indicates that the lung tissue is more rigid and requires a higher driving pressure to maintain effective ventilation.
[0029] Identifying the alveolar recruitment pressure threshold requires attention to the local geometric features of the dynamic curve during the pressure rise phase. During inspiration, as airway pressure gradually increases, when a certain critical value is reached, some collapsed alveoli begin to open, causing a significant change in the slope of the flow-pressure relationship—that is, a curvature abrupt change point on the curve. By performing second-order numerical differentiation on the inspiratory curve, the curvature values at each sampling point are calculated, and the locations where significant changes in curvature occur are identified. The airway pressure value corresponding to this location is defined as the alveolar recruitment pressure threshold, denoted as [insert threshold here]. This threshold is of great clinical significance, serving as a reference for setting the lower limit of positive end-expiratory pressure (PEEP) to prevent alveolar collapse at the end of expiration due to excessively low PEEP settings, thereby reducing the risk of ventilator-associated lung injury.
[0030] In obtaining , and After considering the three fundamental characteristics, a weighted sum is performed on the three to obtain the comprehensive respiratory mechanics parameters. The calculation method is as follows ,in , , These are the weighting coefficients for the corresponding feature quantities, and the sum of the three is 1. The initial values of each weighting coefficient are set based on clinical experience and are dynamically adjusted according to feedback data in the subsequent adaptive update phase. The weighted summation method can unify and integrate three feature quantities with different dimensions into a respiratory mechanics parameter with comprehensive characterization capabilities, which is convenient for subsequent joint participation with oxygenation cycle parameters in the construction of the fusion state feature vector.
[0031] For the extraction of oxygenation cycle parameters, a sliding window processing method was applied to the blood oxygen saturation signal. A fixed-length time window was set, the window length of which was selected based on the typical range of the patient's respiratory rate, so that each window contained several complete respiratory cycles. Within each time window, the blood oxygen saturation signal was subjected to first-order time difference analysis to calculate the gradient sequence of blood oxygen saturation changing over time. The time gradient at each sampling point was denoted as... Statistics within this window The proportion of sampling points with negative values to the total number of sampling points in the window is defined as the oxygenation deterioration risk factor, denoted as [missing information]. . The closer to 1, the more pronounced the downward trend in blood oxygen saturation within that time window, and the higher the risk of hypoxemia in the patient; when When the value is close to 0, it indicates that the blood oxygen saturation remains stable or shows an upward trend within this window, and the oxygenation status is relatively good.
[0032] A continuous wavelet transform was performed on the heart rate variability signal, using the Morlet wavelet as the mother wavelet function. The signal was decomposed within the target scale range to obtain the wavelet coefficient matrix at each scale. The target scale range corresponds to the physiologically significant low-frequency band (0.04Hz to 0.15Hz) and high-frequency band (0.15Hz to 0.4Hz) in heart rate variability analysis. The former mainly reflects the joint regulatory effect of the sympathetic and parasympathetic nervous systems, while the latter mainly reflects the respiratory regulatory activity of the parasympathetic nervous system. Within the target scale range, the energy values of the wavelet coefficients at each scale were calculated to obtain the energy distribution sequence, denoted as... Based on this, the concentration index of energy distribution is calculated, denoted as . It is defined as the ratio of the maximum value in the energy distribution sequence to the total energy of the sequence. , The higher the value, the more concentrated the energy of heart rate variability is in a few specific frequency bands, indicating a relatively simple spectral structure of autonomic nervous regulation. This usually suggests impaired autonomic nervous function or aggravated stress.
[0033] Will and As input, oxygenation cycle parameters are calculated using a nonlinear mapping function. The nonlinear mapping employs a sigmoid-like function to ensure that the output value is constrained within a bounded interval, preventing the parameter values from increasing infinitely under extreme physiological conditions and causing subsequent computational instability. Specifically, ,in These are the mapping coefficients for each input feature. For bias terms, This is the Sigmoid function. The initial values of the mapping coefficients are determined through regression fitting of historical patient data and continuously corrected during the adaptive update phase using real-time feedback data. Oxygenation cycle parameters This comprehensive approach reflects the dynamic trend of blood oxygenation and the state of autonomic nervous system regulation, providing a quantitative basis for subsequent nonlinear coupling modeling from the perspective of circulatory-oxygenation coupling, and is consistent with respiratory mechanics parameters. Together they form the basic input of the fused state feature vector.
[0034] Figure 2 This is a flowchart illustrating the method for calculating the fusion state feature vector according to an embodiment of the present invention. By establishing a nonlinear coupling relationship between respiratory mechanics parameters and oxygenation cycle parameters, the respiratory power consumption characteristics and oxygen supply-demand matching characteristics are calculated to generate the fusion state feature vector, including: The mutual information values of the respiratory mechanics parameters and the oxygenation cycle parameters are calculated at multiple time scales. The time scale corresponding to the inflection point of the mutual information value curve is extracted as the coupling feature time, and a parameter mapping function is constructed based on the coupling feature time. The respiratory mechanics parameters are mapped to predicted values of corresponding oxygenation cycle parameters through the parameter mapping function. The predicted values are then subtracted from the synchronously acquired actual oxygenation cycle parameters to obtain a residual sequence. The decay rate is extracted from the autocorrelation function of the residual sequence. The residual sequence is decomposed into time-frequency components to obtain the dominant components. The cumulative energy of the dominant components within the sliding window is calculated. The fluctuation amplitude of the cumulative energy and the dominant frequency of the dominant components are extracted and combined to generate respiratory power consumption features. The distribution entropy value of the dominant components is calculated. The adjacent intervals of the extreme points of the mutual information value curve are statistically analyzed and the standard deviation of the interval is calculated. The distribution entropy value and the standard deviation are combined to generate oxygen supply and demand matching features. Based on the decay rate and the coupling feature time, a normalized fusion weight is calculated. The breathing power consumption feature and the oxygen supply and demand matching feature are then weighted and combined according to the fusion weight to obtain a fusion state feature vector.
[0035] After obtaining respiratory mechanics and oxygenation cycle parameters, it is necessary to further explore the dynamic coupling relationship between these two types of parameters to extract deeper features reflecting the patient's current respiratory state. To this end, the mutual information values between the respiratory mechanics parameter sequences and the oxygenation cycle parameter sequences are calculated at multiple discrete time scales, forming a mutual information value curve across time scales. Mutual information values can capture the degree of nonlinear dependence between the two sets of signals, rather than being limited to linear correlations, which has a stronger characterizing ability for the complex physiological coupling relationship between the respiratory and circulatory systems. On this curve, the time scale corresponding to the inflection point reflects the critical position where the coupling strength between the two types of parameters changes significantly; this time scale is defined as the coupling characteristic time. . The physical significance is that, around this timescale, changes in respiratory mechanics are most sensitive to the effects of oxygenation cycle, and the coupling effect tends to weaken when the timescale is longer or shorter than this.
[0036] Based on the determined coupling characteristic time ,by The parameter mapping function is constructed with the neighborhood of the time scale as the core. This mapping function takes respiratory mechanics parameters as input and outputs predicted values of the corresponding oxygenation cycle parameters. The mapping function is established using a locally weighted regression approach. Nearby samples are assigned higher weights to ensure the mapping result has a strong responsiveness to dynamic changes near the coupling feature time. Respiratory mechanics parameters are then substituted into... Subsequently, a predicted sequence of oxygenation cycle parameters was obtained. ,Will Synchronously acquired actual oxygenation cycle parameter sequence By subtracting the values, we obtain the residual sequence. The residual sequence contains physiological information that the mapping function cannot fully explain. This information is often closely related to the patient's spontaneous breathing effort, uneven ventilation distribution, and transient fluctuations in airway resistance.
[0037] For residual sequences Calculate its autocorrelation function ,in This represents the lag time step. The rate at which the autocorrelation function decays as the lag increases. This reflects the memory length of the residual sequence: the faster the decay, the stronger the random perturbation component in the residual, and the smaller the systematic deviation between respiratory mechanics and oxygenation cycle; the slower the decay, the more persistent the systematic deviation, suggesting that the current ventilator parameter settings may deviate significantly from the patient's physiological needs. Through the The exponential decay model is fitted within the effective lag range to estimate the value, i.e. ,in The initial amplitude, This is the extracted decay rate parameter.
[0038] For residual sequences Time-frequency decomposition is performed, using methods such as empirical mode decomposition or wavelet decomposition to decompose it into several eigenmodes or subband components. The component with the highest energy is selected as the dominant component according to the energy proportion of each component in the frequency domain. The dominant components embody the most significant periodic or quasi-periodic patterns in the residuals, and are directly related to the interaction between the ventilator ventilation rhythm and the patient's spontaneous breathing rhythm. The length is set above. The sliding window is used to calculate the cumulative energy window by window. ,in For window indexing. Extract. Fluctuation amplitude of the sequence , defined as the difference between the maximum and minimum values of the sequence in the global range; and simultaneously for Perform spectral analysis and extract the frequency corresponding to the peak power spectral density as the dominant frequency. .Will and Combining to form a breathing power consumption feature vector This eigenvector describes the dynamic burden of respiratory work from two dimensions: the intensity of energy fluctuations and the dominant oscillation frequency.
[0039] When generating oxygen supply and demand matching features, the dominant component is calculated. distribution entropy value Distribution entropy is obtained by binning the amplitude distribution of the dominant component and calculating the Shannon entropy of the probability of each interval. It reflects the uniformity of the amplitude distribution of the dominant component: the higher the entropy value, the more irregular the residual fluctuations between the oxygenation cycle and respiratory mechanics, indicating an unstable oxygen supply and demand matching state; the lower the entropy value, the more concentrated the residuals are in a specific amplitude interval, indicating obvious systematic deviation characteristics. Simultaneously, the adjacent interval sequence between all extreme points (including maxima and minima) on the statistical mutual information value curve is calculated. ,in Given the total number of adjacent intervals, calculate the standard deviation of the interval sequence. . The magnitude of the standard deviation reflects the uniformity of the extreme value distribution of the mutual information curve: the larger the standard deviation, the more irregular the change in coupling strength across different time scales, and the worse the stability of the dynamic matching of oxygen supply and demand. and Combining features to form an oxygen supply and demand matching feature vector .
[0040] In obtaining the respiratory power consumption feature vector Feature vector matching oxygen supply and demand Then, it is necessary to determine the decay rate. With coupling characteristic time Calculate the normalized fusion weights to determine the relative contribution ratio of the two types of features in the final fused feature vector. The calculation logic for the fusion weights is as follows: when Larger (rapid residual decay) and When the time is shorter, it indicates that respiratory mechanics has a stronger immediate impact on the oxygenation cycle, and the respiratory power consumption characteristic should be given a higher weight; when Smaller (longer residual duration) and Over a longer period, this indicates that the dynamic matching characteristics of the oxygenation cycle better reflect the current physiological state, and the oxygen supply and demand matching characteristics should be given higher weight. Specifically, the original weight should be defined. Then, the two are normalized to obtain the normalized fusion weight. and ,satisfy .
[0041] Finally, the two types of feature vectors are weighted and combined according to the normalized fusion weights to obtain the fused state feature vector. : , This method integrates the energy dynamics of respiratory work burden with the stability characteristics of oxygen supply and demand dynamic matching, and combines them through weights adaptively determined based on physiological coupling strength. This allows the fusion result to dynamically adjust its focus according to changes in the patient's real-time physiological state. This fused state feature vector will serve as the core input for subsequent respiratory support strategy effectiveness assessment and risk level calculation, providing a high-quality state representation foundation for multi-objective constrained optimization.
[0042] Extract the current ventilator parameters and combine them with the fusion state feature vector to calculate the effectiveness score and risk level of the current respiratory support strategy, including: Extract the current inspiratory pressure setting, positive end-expiratory pressure setting, ventilation rate setting, and inspiratory-expiratory ratio setting of the ventilator, and construct a parameter state vector; The parameter state vector and the fused state feature vector are subjected to sliding matching within a time window. The Pearson correlation coefficient between each dimension component within each time window is calculated. The correlation coefficients of each dimension are combined to form a dynamic correlation matrix. The ratio of the diagonal elements to the off-diagonal elements in the dynamic correlation matrix is extracted as the coupling concentration degree. The time-varying coupling concentration degree curve is obtained by expanding it along the time axis. The cumulative synergy index is obtained by performing trapezoidal integral operation on the time-varying coupling concentration curve. The cumulative synergy index is input into the hyperbolic tangent function for nonlinear mapping to obtain the performance score. The condition number of the dynamic correlation matrix is extracted as a state stability measure. The number of components in the parameter state vector that deviate from the preset safe range is counted. The state stability measure and the number of components are weighted and combined and segmented threshold discrimination is performed. The risk level is determined based on the discrimination result.
[0043] Extract the current inspiratory pressure setpoint, positive end-expiratory pressure setpoint, ventilation rate setpoint, and IPR ratio setpoint of the ventilator, and combine these four parameters into a parameter state vector. ,in This indicates the inhalation pressure setpoint. This indicates the positive end-expiratory pressure setting. This indicates the ventilation frequency setting value. This represents the setpoint for the inspiratory-to-expiratory ratio. Each component of the parameter state vector has undergone dimensional normalization to ensure that the values of each dimension are within a uniform scale, avoiding deviations in subsequent correlation calculations due to differences in dimensions. During actual data acquisition, the ventilator interface outputs real-time readings of the above four parameters at a fixed sampling period. Each sampling corresponds to a timestamp, thus forming a parameter state sequence over time.
[0044] parameter state vector With fusion state feature vector Sliding matching is performed within a time window. Specifically, the window slides along the time axis with a fixed step size, each window covering a consecutive number of sampling points. Within each time window, samples are taken... and Given the corresponding time period data, calculate the Pearson correlation coefficient between the components of each dimension of the two. Let the parameter state vector be... The time series of the dimensional components within the current window is The fusion state feature vector of the first The time series of the dimensional components within the current window is The corresponding Pearson correlation coefficient is: , in The number of sampling points within the window. For the first The mean of the dimensional parameter components within the window. For the first The mean of the dimensional feature components within the window. Arranging the Pearson correlation coefficients of all dimension combinations into a matrix form yields the dynamic correlation matrix. , its first Line number The column element is Dynamic correlation matrix The dimension is jointly determined by the dimensions of the parameter state vector and the fusion state feature vector, reflecting the multidimensional coupling strength distribution between ventilator parameters and patient physiological state characteristics within the current time window.
[0045] From the dynamic correlation matrix Extract the coupling concentration index. Diagonal elements reflect the correlation between the corresponding dimensions of the parameter state vector and the fused state feature vector, while off-diagonal elements reflect cross-dimensional cross-correlation. Coupling Concentration Defined as the ratio of the sum of the absolute values of the diagonal elements to the sum of the absolute values of the off-diagonal elements: , The larger the value, the more concentrated the synergistic relationship between ventilator parameters and patient physiological characteristics is in the corresponding dimension, and the stronger the structural nature of the human-machine coupling. When When the value is close to 1, it indicates that there is more cross-dimensional interference and the synchronization between humans and machines is weaker; when... A value significantly greater than 1 indicates a clear synergistic relationship across all dimensions, and a good match between the respiratory support strategy and the patient's physiological state. As the time window slides along the time axis, a value is calculated for each window. Value, corresponding to all time windows The values are arranged in chronological order to obtain the time-varying coupling concentration curve. ,in Index the time at the center of the window.
[0046] For time-varying coupling concentration curves Perform trapezoidal integral calculation to obtain the cumulative synergy index. :
[0047] in The total number of time windows. The time interval is the time between the center times of adjacent windows. The trapezoidal integral can make full use of the linear interpolation relationship between adjacent windows, reduce discretization error, and make the cumulative coordination index more accurately reflect the cumulative effect of human-machine coordination level over the entire observation period. The higher the value, the greater the synergy between the parameter status and the patient's physiological characteristics throughout the assessment period, and the stronger the overall effectiveness of the current respiratory support strategy.
[0048] Cumulative Synergy Index The performance score is obtained by performing a nonlinear mapping on the input hyperbolic tangent function. : , in This is the mapping gain coefficient, used to adjust the sensitivity range of the input value. The hyperbolic tangent function will... Mapped to The interval provides clear upper and lower bounds for the performance score, facilitating unified quantitative comparison of performance objectives in subsequent multi-objective optimization. When smaller, Approximately linear growth, with high resolution; when When it is large, The optimization process should be approached saturation to avoid extreme values interfering with the optimization process. The value is determined based on historical data statistical analysis, ensuring compatibility in most clinical scenarios. It falls within the linear sensitivity region of the hyperbolic tangent function.
[0049] The risk level is calculated from two independent dimensions. A dynamic correlation matrix is extracted. condition number As a measure of state stability, the condition number is defined as the ratio of the largest singular value to the smallest singular value of a matrix. A larger condition number indicates that the matrix is closer to singularity, and the numerical structure of the dynamic correlation matrix is more unstable. This implies greater uncertainty or rapid change in the coupling relationship between the current patient's physiological state and ventilator parameters, corresponding to higher clinical risk. The condition number is calculated using singular value decomposition. Take the maximum singular value after row decomposition With minimum singular value The ratio, i.e. ,in and The maximum and minimum singular values in the singular value decomposition results, and the values in the used parameter list. It is function-independent and belongs to an independently defined scalar parameter.
[0050] Statistical parameter state vector The number of components that deviate from the preset safety range For each dose, corresponding safety upper and lower limits are pre-defined according to clinical guidelines. For example, the safe range for the inspiratory pressure setpoint is usually [missing information]. The safe range for positive end-expiratory pressure (PEEP) settings is typically 100%. The ventilation rate is similar to the inspiratory-expiratory ratio. If the current value of a component exceeds its corresponding safe range, it is counted as a violation. The number of violations across all dimensions is accumulated to obtain... . The number of dimensions that directly reflect the potential safety hazards in the current ventilator parameter configuration is a direct basis for clinical risk assessment.
[0051] Measuring state stability With the number of violations A comprehensive risk index is obtained through weighted combination. :
[0052] in This is the dimensionless value after normalizing the condition number. and These are the weighting coefficients for the number of conditions and the number of violations, respectively, with their sum being 1. The specific values are determined based on clinical experience and historical data. Perform segmented threshold discrimination: when Below the low risk threshold When, it is determined to be a low-risk level; when In When the risk level is within a certain range, it is determined to be at a medium risk level; when Not lower than the high-risk threshold When the risk level is high, it is determined to be high-risk. The three-level risk classification provides clear risk quantification input for subsequent multi-objective constrained optimization, enabling the optimization process to dynamically adjust the parameter search boundary according to the safety of the current patient status, ensuring that the risk level is controlled within an acceptable range while improving the efficacy score.
[0053] With the optimization objectives of maximizing the efficiency score and minimizing the risk level, pressure safety boundary constraints and human-machine synchronization constraints are introduced as solution boundaries. A set of ventilator parameter adjustment instructions is generated through multi-objective constraint optimization, including: The effectiveness score and the risk level are used to construct an objective function. In the ventilator parameter space, the set of non-dominated solutions for ventilator parameters is obtained by Pareto front search. The non-dominated solution with the highest objective function value is selected as the optimization benchmark solution. A pressure safety corridor is constructed based on the peak envelope of the airway pressure signal. A time-varying upper bound for inspiratory pressure and a time-varying lower bound for positive end-expiratory pressure are set and smoothed to generate a pressure safety boundary constraint function. The respiratory flow waveform signal is segmented into respiratory cycles to identify the patient's spontaneous inhalation start time and the ventilator trigger response time and calculate the phase difference angle. The phase difference angles of multiple respiratory cycles are statistically analyzed to construct a synchronization deviation sequence and variance analysis is performed to obtain the fluctuation intensity index. The joint constraint conditions of trigger window width and flow rate threshold are set to generate human-ventilator synchronization constraint function. Using the optimized benchmark solution as the initial search point, a sequential quadratic programming method is used to iteratively search for optimization within the feasible region defined by the air pressure safety boundary constraint function and the human-machine synchronization constraint function. The iteration terminates when the gradient magnitude of the objective function is less than the convergence criterion, and the corresponding parameter combination is extracted to generate a ventilator parameter adjustment instruction set.
[0054] After calculating the efficacy score and risk level, they need to be combined to form the objective function of a multi-objective optimization problem. The optimal parameter combination that balances ventilation efficacy and safety is then searched within the ventilator parameter space. The efficacy score is negatively represented as the first objective function (transforming it into a minimization problem), and the risk level as the second objective function, together forming a bi-objective optimization framework. Within the parameter space spanned by inspiratory pressure, positive end-expiratory pressure, ventilation rate, and ISTR, a Pareto front search strategy is employed to traverse candidate solutions. For each candidate solution, its corresponding efficacy score and risk level are calculated, and it is determined whether another candidate solution is non-inferior to it in both objectives. If not, it is included in the non-dominated solution set. As the search progresses, the non-dominated solution set gradually converges to the Pareto front. After obtaining the complete non-dominated solution set, the non-dominated solution with the highest weighted sum of the objective functions is used as the optimization benchmark solution, providing a high-quality initial point for subsequent refined constraint iterations and effectively shortening the convergence path in the sequential quadratic programming stage.
[0055] The construction of airway pressure safety boundary constraints is based on the peak envelope of the airway pressure signal. In continuously acquired airway pressure time series, the pressure peak value within each respiratory cycle is extracted, and adjacent peak values are connected using cubic spline interpolation to form a peak envelope curve reflecting the dynamic upper limit of airway pressure. Based on this, a time-varying upper limit for inspiratory pressure is set. Its value is the peak envelope multiplied by the safety margin factor. (The value range is typically 0.85 to 0.95), ensuring that the actual inspiratory pressure remains below the airway damage threshold. Simultaneously, a time-varying lower bound for positive end-expiratory pressure (PEEP) is set, based on the sliding minimum of end-expiratory airway pressure, superimposed with a minimum positive pressure offset to ensure alveolar patency. This forms the lower bound of the time-varying positive end-expiratory pressure. To avoid numerical instability caused by high-frequency jitter in the constraint boundaries during optimization iterations, [the following is done]: and First-order low-pass filtering is applied for smoothing, with a filter time constant of [missing information]. The pressure is adaptively set based on the patient's respiratory cycle length, typically set to 2 to 3 respiratory cycles. After smoothing, the pressure safety boundary constraint function is expressed as the inspiratory pressure setpoint that must satisfy the following at any given time: The positive end-expiratory pressure setting must meet the following requirements. This allows for the delineation of a clear pressure safety corridor within the parameter feasible domain.
[0056] The construction of human-machine synchronization constraints relies on detailed analysis of respiratory flow waveform signals. When segmenting the respiratory flow waveform into respiratory cycles, the zero-crossing point of the flow signal (the direction from expiration to inspiration) is used as the starting marker of each respiratory cycle. An adaptive threshold detection method is then used to identify the moment when the patient's spontaneous inspiration begins. This moment corresponds to the point at which the flow signal begins to show a negative deflection (driven by negative airway pressure generated by the patient's active inspiration). Ventilator trigger response time. This is defined as the actual time point at which the ventilator begins delivering air after detecting the trigger signal, obtained by detecting the rising edge of the airway pressure signal. The difference between the two is defined as the phase difference angle. The unit is milliseconds. A positive value indicates that the ventilator response is lagging behind the patient's needs, while a negative value indicates that the ventilator is triggered too early.
[0057] In continuous Statistical phase difference angle sequence within one respiratory cycle Perform an analysis of variance on the sequence to calculate the sequence variance. This is used to quantify the fluctuation intensity of human-machine synchronization. Fluctuation Intensity Index Defined as the variance of the phase difference angle sequence and the reference variance The ratio, that is ,in It is determined by the mean variance under historical normal synchronization conditions. When Exceeding the preset threshold When the human-machine synchronization is deemed insufficient, a constraint tightening mechanism is triggered. Based on this, the trigger window width is set. (Unit: milliseconds, typically 50 to 150 milliseconds) and flow rate threshold (Unit: liters / second) together constitute the trigger sensitivity constraint: requiring that within the trigger window, the change in the flow signal must exceed [a certain value]. Only then can a valid trigger be confirmed, thus filtering out false trigger interference. The human-machine synchronization constraint function integrates the absolute value constraint of the mean phase difference angle (…). ,in (Maximum allowable average phase deviation) and fluctuation intensity constraint ( These parameters together limit the range of values for ventilator parameters within the synchronicity feasible domain.
[0058] Within the feasible region jointly defined by the pressure safety boundary constraint function and the human-machine synchronization constraint function, the optimal benchmark solution obtained from the Pareto front search is used as the initial search point to initiate a sequential quadratic programming iterative optimization process. In each iteration, the sequential quadratic programming method approximates the original nonlinear constraint optimization problem at the current point using a second-order Taylor expansion, transforming it into a quadratic programming subproblem with linear constraints. Solving this subproblem yields the search direction and step size, and the parameter combination is updated along this direction. After each parameter update, the performance score and risk level are recalculated for the new candidate parameter combination, and it is checked whether they satisfy the pressure safety boundary constraint and the human-machine synchronization constraint. If either constraint is violated, the penalty function mechanism is activated, converting the violation amount into a penalty increment in the objective function, driving the search direction back to the feasible region. The iteration termination condition is set as the magnitude of the gradient vector of the objective function. Less than convergence criterion ,in The value is set based on the numerical accuracy requirements after parameter space dimension normalization, and is usually taken as... to Order of magnitude. When the convergence condition is met, the iteration terminates and the current parameter combination is extracted. This combination is the optimal solution that simultaneously satisfies the requirements of maximizing efficiency and minimizing risk within the constrained feasible region.
[0059] After extracting the optimal parameter combination, the target values for inspiratory pressure, positive end-expiratory pressure, ventilation rate, and IPR are encapsulated into a structured ventilator parameter adjustment instruction set. Each instruction in the set includes a target parameter identifier, a target adjustment value, a suggested adjustment step size, and an execution priority field. The suggested adjustment step size is determined based on the difference between the current value and the target value, as well as the safety limit for the parameter change rate, to avoid physiological stress on the patient caused by parameter abrupt changes. The execution priority field is sorted according to risk level, with adjustment instructions for high-risk parameters (such as excessive inspiratory pressure) having higher priority than those for low-risk parameters, ensuring that safety hazards are addressed first during instruction issuance and execution. After the instruction set is generated, it is sent to the ventilator control interface, and the post-adjustment physiological signal feedback acquisition process is initiated, providing a data foundation for subsequent adaptive parameter updates.
[0060] In the ventilator parameter space, the non-dominated solution set of ventilator parameters is obtained through Pareto front search, including: In the ventilator parameter space, the range of values for inspiratory pressure, positive end-expiratory pressure, ventilation rate and inspiratory-expiratory ratio are set and the grid is discretized. Multiple initial parameter combinations are randomly selected from the discrete grid points, and the efficacy score and risk level value are calculated for each initial parameter combination. Perform Pareto dominance judgment on all initial parameter combinations. If the effectiveness score of the first parameter combination is not lower than that of the second parameter combination and the risk level value is not higher than that of the second parameter combination, then the first parameter combination dominates the second parameter combination. Extract all parameter combinations that are not dominated to form the current non-dominated solution set. For each parameter combination in the current non-dominated solution set, identify the nearest neighbor parameter combination and calculate the local sparsity, and select the parameter combination with a local sparsity greater than the sparsity threshold as the seed solution; Around various sub-solutions, perturbations are applied to parameters of each dimension in the parameter space to generate new parameter combinations, which are then merged with the current non-dominated solution set. The Pareto dominance determination is then re-executed to obtain an expanded non-dominated solution set. The iteration terminates when the coverage of the expanded non-dominated solution set reaches the convergence threshold, and the expanded non-dominated solution set is output.
[0061] When performing a Pareto front search in the ventilator parameter space, the parameter space needs to be structurally partitioned. The four dimensions—inspiratory pressure, positive end-expiratory pressure, tidal frequency, and IPR ratio—each have clinically constrained ranges: inspiratory pressure is typically between 10 and 35 cmH2O, positive end-expiratory pressure between 3 and 20 cmH2O, tidal frequency between 8 and 35 breaths / min, and IPR ratio between 1:1 and 1:4. Each dimension is discretized into a grid with uniform intervals. The step size is set according to clinical accuracy requirements; for example, the step size for inspiratory pressure is 1 cmH2O, for tidal frequency is 1 breath / min, and for IPR ratio is 0.25. After discretization, several initial parameter combinations are randomly selected from the candidate set consisting of all grid nodes. The number of selections is denoted as . The number of parameters is usually set to 80 to 200 to ensure the breadth of the parameter space coverage of the initial population.
[0062] For each initial parameter combination, the performance score calculation process and risk level assessment process are used to obtain the corresponding performance score and risk level values, respectively. A higher performance score indicates that the current parameter combination performs better in terms of maintaining oxygenation and reducing respiratory work; a lower risk level value indicates that the parameter combination deviates less from the safety boundary. There is an inherent tension between these two objectives: increasing ventilation support may improve oxygenation efficiency, but it will also increase the risk of barotrauma. Therefore, there is no single optimal solution, and a multi-objective optimization framework is needed to obtain the trade-off frontier.
[0063] The core logic of the Pareto dominance criterion is as follows: for any two combinations of parameters in the parameter space... and ,like Corresponding performance score Not less than Corresponding performance score ,and Corresponding risk level value Not higher than Corresponding risk level value If at least one of the two inequalities above is a strict inequality, then it is said that... Dominate Perform the above judgment on each pair of all initial parameter combinations, and extract all parameter combinations that are not dominated by any other parameter combination to form the current non-dominated solution set. This set represents the Pareto front approximation for the current search phase.
[0064] After establishing the non-dominated solution set, it is necessary to evaluate the uniformity of the distribution of each parameter combination within the set in the target space to identify regions with sparse frontal coverage and guide subsequent searches. For the set... For each parameter combination, calculate the Euclidean distance between it and its nearest neighbor parameter combination in the target space, denoted as the local sparsity. Let the sparsity threshold be... When the local sparsity of a certain parameter combination satisfies When this parameter combination is selected, it is chosen as the seed solution. Sparse threshold. The scale can be normalized based on the target space, for example, set to 5% to 10% of the diagonal length of the target space. Through this mechanism, the search is preferentially expanded in areas with insufficient Pareto front coverage, avoiding the waste of resources in densely covered areas.
[0065] Solve for each seed Centered on a central point, random perturbations are applied to each dimension of the parameter space to generate new candidate parameter combinations. The perturbation method employs Gaussian perturbation, applying a perturbation to the th... The dimensional parameter components are applied with a mean of 0 and a standard deviation of . random noise, where The parameter value is adaptively set based on the range of values for that dimension, typically ranging from 3% to 8% of that range. If the perturbed parameter value exceeds the preset range for that dimension, it is truncated to the boundary value to ensure that all newly generated candidate parameter combinations satisfy the parameter space constraints. Each seed solution is generated... A new parameter combination, It is usually set to 5 to 15.
[0066] Combine the new parameter combinations generated from all seed solutions with the current set of non-dominated solutions The candidates are merged to form an expanded candidate set. The Pareto dominance decision is then re-executed on the expanded candidate set, and the non-dominated parameter combinations are extracted to obtain the expanded non-dominated solution set. Compared to the initial stage, More approximate solutions were obtained in the sparse region of the Pareto front, improving the front coverage. Coverage was measured using a hypervolume index, which is the hypervolume in the target space enclosed by the set of non-dominated solutions relative to the reference point. The reference point is the corner point in the target space with the lowest efficiency score and the highest risk level.
[0067] Iteration termination condition based on coverage convergence determination: Let the first... The hypervolume index after the next iteration is: When the hypervolume increment of two adjacent iterations satisfies At that time, it was determined that the frontier search had converged, among which The convergence threshold can be set to 0.1% to 0.5% of the initial hypervolume value. Additionally, the maximum number of iterations can also be set. As a hard termination condition, it prevents iteration stagnation caused by locally flat regions. When any termination condition is satisfied, the current expanded set of non-dominated solutions is output as the final Pareto front approximation.
[0068] The output non-dominated solution set contains multiple parameter combinations that achieve different trade-offs between efficacy and risk. Subsequently, the final parameter adjustment instruction set can be selected from these combinations based on the clinician's preferences or the urgency of the patient's current condition. For example, when a patient is in the acute hypoxemia stage, the parameter combination with the highest efficacy score can be prioritized from the non-dominated solution set; when the patient is in the stable maintenance stage, the parameter combination with the lowest risk level can be selected. This Pareto front-based decision-making framework decouples the multi-objective trade-off calculation process from the final clinical decision, ensuring both the safety and effectiveness of parameter recommendations while preserving ample room for personalized clinical adjustments.
[0069] In practical implementation, the accuracy of grid discretization and the initial sampling number are related. There exists a balance between computational complexity and optimization. For real-time ICU applications, it is recommended to use a coarse-grained grid for initial discretization, followed by local refinement near the front using a seed expansion mechanism, rather than performing high-density discretization across the entire parameter space. This ensures the accuracy of the front approximation while keeping the computation time per iteration within a clinically acceptable range. The computational complexity of each iteration primarily stems from the Pareto dominance determination, whose time complexity is proportional to the square of the candidate set size. Therefore, when merging and expanding the candidate set, an upper limit constraint on the set size is necessary; solutions exceeding this limit should be randomly discarded to maintain computational efficiency.
[0070] A second aspect of the present invention provides an intelligent adjustment system for ICU ventilator parameters, comprising: The data acquisition unit is used to acquire multimodal monitoring data of the target patient, including respiratory flow waveform signals, airway pressure signals, blood oxygen saturation signals, and heart rate variability signals. The fusion feature unit is used to extract features from the multimodal monitoring data, calculate respiratory mechanics parameters and oxygenation cycle parameters, and calculate respiratory power consumption characteristics and oxygen supply and demand matching characteristics by establishing a nonlinear coupling relationship between respiratory mechanics parameters and oxygenation cycle parameters, thereby generating a fusion state feature vector. The strategy evaluation unit is used to extract the current ventilator parameters and calculate the effectiveness score and risk level of the current respiratory support strategy by combining the fusion state feature vector. The instruction optimization unit is used to maximize the efficiency score and minimize the risk level as optimization objectives. It introduces air pressure safety boundary constraints and human-machine synchronization constraints as solution boundaries, and generates a set of ventilator parameter adjustment instructions through multi-objective constraint optimization. The feedback update unit is used to send the parameter adjustment instruction set to the ventilator and continuously collect the adjusted physiological signal feedback data to adaptively update the parameters in the multi-objective constraint optimization.
[0071] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.
[0072] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.
[0073] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.
[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. An intelligent adjustment method for ICU ventilator parameters, characterized in that, include: Acquire multimodal monitoring data of the target patient, including respiratory flow waveform signals, airway pressure signals, blood oxygen saturation signals, and heart rate variability signals; Feature extraction is performed on the multimodal monitoring data to calculate respiratory mechanics parameters and oxygenation cycle parameters. By establishing a nonlinear coupling relationship between respiratory mechanics parameters and oxygenation cycle parameters, respiratory power consumption characteristics and oxygen supply and demand matching characteristics are calculated to generate a fusion state feature vector. Extract the current ventilator parameters and combine them with the fusion state feature vector to calculate the effectiveness score and risk level of the current respiratory support strategy; With the optimization objectives of maximizing the efficiency score and minimizing the risk level, pressure safety boundary constraints and human-machine synchronization constraints are introduced as solution boundaries. The ventilator parameter adjustment instruction set is generated through multi-objective constraint optimization. The parameter adjustment instruction set is sent to the ventilator and the adjusted physiological signal feedback data is continuously collected to adaptively update the parameters in the multi-objective constraint optimization.
2. The method according to claim 1, characterized in that, Feature extraction is performed on the multimodal monitoring data to calculate respiratory mechanics parameters and oxygenation cycle parameters, including: Based on the multimodal monitoring data, the respiratory flow waveform signal and airway pressure signal are time-aligned, and a dynamic curve of flow and pressure is constructed in each respiratory cycle. The respiratory work is quantified by calculating the loop area enclosed by the dynamic curve. The slope of the long axis of the dynamic curve is extracted as a comprehensive compliance index. The curvature change points in the pressure rise phase of the dynamic curve are identified, and the pressure values corresponding to the curvature change points are used as alveolar recruitment pressure thresholds. The respiratory work, the comprehensive compliance index, and the alveolar recruitment pressure thresholds are weighted and summed to obtain respiratory mechanics parameters. The blood oxygen saturation signal is processed by a sliding window. The time gradient of blood oxygen saturation is calculated in each time window. The proportion of time with negative time gradient is used as the risk factor for oxygenation deterioration. The heart rate variability signal is subjected to continuous wavelet transform. The energy distribution characteristics of wavelet coefficients are extracted within the target scale range and the concentration is calculated. The oxygenation cycle parameters are obtained by nonlinear mapping in combination with the risk factor for oxygenation deterioration.
3. The method according to claim 1, characterized in that, By establishing a nonlinear coupling relationship between respiratory mechanics parameters and oxygenation cycle parameters, the respiratory power consumption characteristics and oxygen supply-demand matching characteristics are calculated, generating a fusion state feature vector, including: The mutual information values of the respiratory mechanics parameters and the oxygenation cycle parameters are calculated at multiple time scales. The time scale corresponding to the inflection point of the mutual information value curve is extracted as the coupling feature time, and a parameter mapping function is constructed based on the coupling feature time. The respiratory mechanics parameters are mapped to predicted values of corresponding oxygenation cycle parameters through the parameter mapping function. The predicted values are then subtracted from the synchronously acquired actual oxygenation cycle parameters to obtain a residual sequence. The decay rate is extracted from the autocorrelation function of the residual sequence. The residual sequence is decomposed into time-frequency components to obtain the dominant components. The cumulative energy of the dominant components within the sliding window is calculated. The fluctuation amplitude of the cumulative energy and the dominant frequency of the dominant components are extracted and combined to generate respiratory power consumption features. The distribution entropy value of the dominant components is calculated. The adjacent intervals of the extreme points of the mutual information value curve are statistically analyzed and the standard deviation of the interval is calculated. The distribution entropy value and the standard deviation are combined to generate oxygen supply and demand matching features. Based on the decay rate and the coupling feature time, a normalized fusion weight is calculated. The breathing power consumption feature and the oxygen supply and demand matching feature are then weighted and combined according to the fusion weight to obtain a fusion state feature vector.
4. The method according to claim 1, characterized in that, Extract the current ventilator parameters and combine them with the fusion state feature vector to calculate the effectiveness score and risk level of the current respiratory support strategy, including: Extract the current inspiratory pressure setting, positive end-expiratory pressure setting, ventilation rate setting, and inspiratory-expiratory ratio setting of the ventilator, and construct a parameter state vector; The parameter state vector and the fused state feature vector are subjected to sliding matching within a time window. The Pearson correlation coefficient between each dimension component within each time window is calculated. The correlation coefficients of each dimension are combined to form a dynamic correlation matrix. The ratio of the diagonal elements to the off-diagonal elements in the dynamic correlation matrix is extracted as the coupling concentration degree. The time-varying coupling concentration degree curve is obtained by expanding it along the time axis. The cumulative synergy index is obtained by performing trapezoidal integral operation on the time-varying coupling concentration curve. The cumulative synergy index is input into the hyperbolic tangent function for nonlinear mapping to obtain the performance score. The condition number of the dynamic correlation matrix is extracted as a state stability measure. The number of components in the parameter state vector that deviate from the preset safe range is counted. The state stability measure and the number of components are weighted and combined and segmented threshold discrimination is performed. The risk level is determined based on the discrimination result.
5. The method according to claim 1, characterized in that, With the optimization objectives of maximizing the efficiency score and minimizing the risk level, pressure safety boundary constraints and human-machine synchronization constraints are introduced as solution boundaries. A set of ventilator parameter adjustment instructions is generated through multi-objective constraint optimization, including: The effectiveness score and the risk level are used to construct an objective function. In the ventilator parameter space, the set of non-dominated solutions for ventilator parameters is obtained by Pareto front search. The non-dominated solution with the highest objective function value is selected as the optimization benchmark solution. A pressure safety corridor is constructed based on the peak envelope of the airway pressure signal. A time-varying upper bound for inspiratory pressure and a time-varying lower bound for positive end-expiratory pressure are set and smoothed to generate a pressure safety boundary constraint function. The respiratory flow waveform signal is segmented into respiratory cycles to identify the patient's spontaneous inhalation start time and the ventilator trigger response time and calculate the phase difference angle. The phase difference angles of multiple respiratory cycles are statistically analyzed to construct a synchronization deviation sequence and variance analysis is performed to obtain the fluctuation intensity index. The joint constraint conditions of trigger window width and flow rate threshold are set to generate human-ventilator synchronization constraint function. Using the optimized benchmark solution as the initial search point, a sequential quadratic programming method is used to iteratively search for optimization within the feasible region defined by the air pressure safety boundary constraint function and the human-machine synchronization constraint function. The iteration terminates when the gradient magnitude of the objective function is less than the convergence criterion, and the corresponding parameter combination is extracted to generate a ventilator parameter adjustment instruction set.
6. The method according to claim 5, characterized in that, In the ventilator parameter space, the non-dominated solution set of ventilator parameters is obtained through Pareto front search, including: In the ventilator parameter space, the range of values for inspiratory pressure, positive end-expiratory pressure, ventilation rate and inspiratory-expiratory ratio are set and the grid is discretized. Multiple initial parameter combinations are randomly selected from the discrete grid points, and the efficacy score and risk level value are calculated for each initial parameter combination. Perform Pareto dominance judgment on all initial parameter combinations. If the effectiveness score of the first parameter combination is not lower than that of the second parameter combination and the risk level value is not higher than that of the second parameter combination, then the first parameter combination dominates the second parameter combination. Extract all parameter combinations that are not dominated to form the current non-dominated solution set. For each parameter combination in the current non-dominated solution set, identify the nearest neighbor parameter combination and calculate the local sparsity, and select the parameter combination with a local sparsity greater than the sparsity threshold as the seed solution; Around various sub-solutions, perturbations are applied to parameters of each dimension in the parameter space to generate new parameter combinations, which are then merged with the current non-dominated solution set. The Pareto dominance determination is then re-executed to obtain an expanded non-dominated solution set. The iteration terminates when the coverage of the expanded non-dominated solution set reaches the convergence threshold, and the expanded non-dominated solution set is output.
7. An intelligent adjustment system for ICU ventilator parameters, used to implement the method as described in any one of claims 1-6, characterized in that, include: The data acquisition unit is used to acquire multimodal monitoring data of the target patient, including respiratory flow waveform signals, airway pressure signals, blood oxygen saturation signals, and heart rate variability signals. The fusion feature unit is used to extract features from the multimodal monitoring data, calculate respiratory mechanics parameters and oxygenation cycle parameters, and calculate respiratory power consumption characteristics and oxygen supply and demand matching characteristics by establishing a nonlinear coupling relationship between respiratory mechanics parameters and oxygenation cycle parameters, thereby generating a fusion state feature vector. The strategy evaluation unit is used to extract the current ventilator parameters and calculate the effectiveness score and risk level of the current respiratory support strategy by combining the fusion state feature vector. The instruction optimization unit is used to maximize the efficiency score and minimize the risk level as optimization objectives. It introduces air pressure safety boundary constraints and human-machine synchronization constraints as solution boundaries, and generates a set of ventilator parameter adjustment instructions through multi-objective constraint optimization. The feedback update unit is used to send the parameter adjustment instruction set to the ventilator and continuously collect the adjusted physiological signal feedback data to adaptively update the parameters in the multi-objective constraint optimization.
8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.
9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.
