A multi-center respiratory parameter fusion risk assessment method for respiratory failure patients

Abstract

This invention relates to the field of data fusion technology, specifically to a multi-center respiratory parameter fusion risk assessment method for patients with respiratory failure. It includes the following steps: acquiring multi-center time-series clinical data and initializing a reliability index; performing perturbation separation processing based on the reliability index and center difference characteristics to reconstruct the true physiological estimate; dynamically updating the reliability index based on intervention response consistency; performing multi-parameter coupled state analysis to invert latent state variables, detecting logical conflicts between information from different sources and performing hierarchical adjudication to determine the final parameter set; parallel calculation of multiple risk components and dynamic adjustment of weights, outputting the risk assessment result. This method achieves dynamic decoupling between observed values ​​and true physiological signals through multi-source perturbation decomposition and reliability feedback adjustment, and provides interpretable and highly robust risk assessment results through conflict adjudication and multi-path risk fusion.

Description

Technical Field

[0001] This invention relates to the field of data fusion technology, and more specifically, to a multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure. Background Technology

[0002] Respiratory failure is one of the most common syndromes in intensive care units (ICUs). Its rapid progression and frequent treatment interventions make accurate and timely risk assessment crucial for optimizing treatment decisions and reducing patient mortality. Currently, widely used risk assessment tools include the ROX index, HACOR score, and Integrated Lung Index (IPI). In recent years, machine learning-based predictive models have also emerged. Most of these methods are based on single-center data or simply combine multi-center data to train the model.

[0003] However, existing technologies have the following problems: First, changes in observed values ​​are mixed with the actual evolution of the disease, clinical intervention response, and equipment error. Existing methods cannot dynamically decouple these factors, leading to distorted assessments. Second, each physiological parameter is processed independently, lacking a mutual verification mechanism, making it impossible to identify the source of conflict when changes in oxygenation and ventilation parameters contradict each other. Third, cross-center calibration uses static deviation compensation, which cannot adapt to dynamic changes in equipment stability, data loss, sampling delay, etc., between different centers; Fourth, the model outputs a single risk value, which cannot distinguish whether the risk is caused by disease progression, ineffective intervention, or data quality issues, resulting in poor clinical interpretability.

[0004] Therefore, there is an urgent need to provide a multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure. Summary of the Invention

[0005] The purpose of this invention is to provide a multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention aims to provide a multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure, comprising the following steps: S1. Obtain time-series clinical data of respiratory failure patients in multiple medical centers. The clinical data includes at least a set of physiological parameters, intervention operation records, and differential characteristics corresponding to each center. The differential characteristics include center baseline reliability and center stability index, and an initial reliability index is assigned to each physiological parameter. S2. Based on the differences in the centers, perform perturbation separation processing on the physiological parameter values ​​at each observation time, reconstruct the true physiological estimate, and record the perturbation separation residuals. S3. Based on the intervention operation records, determine the consistency between the actual response and the expected response of physiological parameters before and after each intervention, generate response category labels and response confidence, and thus dynamically update the confidence index of each physiological parameter. S4. Based on the updated credibility index and the reconstructed true physiological estimates, perform multi-parameter coupled state analysis to invert latent state variables, thereby detecting logical conflicts between physiological parameters, between different medical centers, and between intervention and response. When a conflict is detected, make a graded decision based on the credibility index of each parameter, the stability index of the center to which it belongs, and the confidence level of the response, and determine the final set of parameters to be adopted and the set of parameters to be suppressed. S5. Based on the final set of parameters, calculate multiple risk components in parallel, and dynamically adjust the weight of each risk component according to the current confidence distribution, intervention frequency and stability of potential state variables, and output the risk assessment result.

[0007] As a further improvement to this technical solution, the disturbance separation process in step S2 specifically includes: Separate equipment-related error components using the equipment deviation coefficient of the respective center; Estimate the time delay and separate the time delay components using historical sampling intervals; The transient impact components caused by the intervention are separated using intervention operation records and a pre-set intervention response transfer function; The true physiological estimate is obtained by subtracting each of the above components from the observed values ​​in sequence.

[0008] As a further improvement to this technical solution, in step S3, the generation of response category labels and response confidence levels are as follows: Compare the relationship between the direction of change in intervention intensity and the direction and magnitude of change in physiological parameters; When the intervention is enhanced but the expected improvement in oxygenation parameters does not change significantly, it is marked as an ineffective oxygenation response. When the intervention is enhanced but ventilation parameters show an abnormal increase, it is marked as a ventilation-oxygenation mismatch response. When the physiological response is delayed beyond the preset tolerance window, it is marked as a time delay-affected response. Otherwise, it is marked as a normal response. Calculate the response confidence level based on the stability of the response waveform.

[0009] As a further improvement to this technical solution, in step S3, the dynamic updating of the reliability index for each physiological parameter specifically includes: Calculate the coefficient of variation of this parameter value within a preset time window in the past, and use it as the fluctuation stability factor; Obtain the baseline credibility preset by the relevant center; Based on the consistency of the intervention response, the credibility score is increased when the direction of change of this parameter is consistent with the expectation; otherwise, the credibility score is decreased. Calculate the root mean square of the disturbance separation residuals within the preset time window in the past, and use it as the historical deviation factor; The updated credibility index is obtained by fusing the volatility stability factor, benchmark credibility, adjusted credibility score, and historical deviation factor through weighted summation or nonlinear mapping.

[0010] As a further improvement to this technical solution, in S4, the multi-parameter coupled state analysis is implemented through a learnable mapping network, and physiological constraint regularization terms are introduced during the training process of this network. The physiological constraints include constraints on the oxygenation curve relationships between oxygenation parameters, constraints on the ventilation balance relationships between ventilation parameters, and constraints on the mechanical equations between respiratory mechanics parameters. When the inverted potential state variables violate any of the above physiological constraints, a coupling anomaly marker is triggered and sent to conflict detection.

[0011] As a further improvement to this technical solution, the specific steps in S4 for determining the final set of parameters to be adopted and the set of parameters to be suppressed are as follows: The conflicts include logical contradictions and conflicts among detection parameters, conflicts arising from differences in state estimations among centers, and conflicts arising from the failure of the expected response to occur after intervention. For each conflict, collect the set of parameters involved, their current credibility metrics, the stability metrics of the center to which they belong, and the response confidence level; then perform the following three-level adjudication: Level 1 decision: Prioritize parameters with credibility indices higher than the preset high threshold; Second-level decision: If conflicts still exist among the high-confidence parameters, compare the stability characteristics of the centers to which each parameter belongs, and adopt the parameter with better stability; Level 3 decision: If the problem remains unresolved, then rely on the confidence level of the response and use parameters consistent with the expected direction of intervention; The final output after the ruling is the set of parameters to be used, the set of parameters to be suppressed, and the conflict reason identifier.

[0012] As a further improvement to this technical solution, in S5, multiple risk components include the risk of disease evolution trend, the risk of abnormal coupling between parameters, and the risk of center drift. The risk assessment results include a comprehensive risk index and the dominant risk type, key influencing parameters, and information on suppressed parameters.

[0013] As a further improvement to this technical solution, the risk of disease evolution trend is obtained in the following way: based on the time series of each parameter in the final set of parameters, a trend estimation algorithm is used to extract the intensity of the deterioration trend, and the magnitude of the trend risk is determined according to the intensity of the deterioration trend. The coupling anomaly risk between parameters is obtained by: calculating the coupling anomaly index based on the actual correlation strength and theoretical correlation strength between multiple preset parameter pairs using a deviation measurement method, and determining the magnitude of the coupling risk based on the level of the coupling anomaly index. The center drift risk is obtained in the following way: based on the frequency of conflict occurrence within a preset time window and the stability index of each center, a weighted fusion mechanism is used to calculate the drift risk value, and the magnitude of the drift risk is determined according to the frequency of conflict and the quality of center stability.

[0014] As a further improvement to this technical solution, in step S5, the dynamic adjustment method for the weights of multiple risk components is as follows: Calculate the average of the credibility indices of all valid parameters in the current final set of parameters, and increase the weight of trend risk when the average is higher than the preset credibility threshold. Calculate the frequency of collisions within a preset time window, and increase the weight of drift risk when the collision frequency is higher than a preset collision frequency threshold; State stability is determined based on the temporal fluctuation of potential state variables, and the weight of coupling risk is reduced when the state stability is higher than a preset stability threshold. Based on the density of intervention operations within a preset time window, the weight of trend risk is temporarily reduced when the intervention density exceeds a preset intervention density threshold.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure achieves dynamic decoupling between observed values ​​and real physiological signals through multi-source perturbation decomposition and intervention response consistency determination, eliminating interference from equipment errors, sampling delays, and intervention coupling effects. Through reliability feedback adjustment and hierarchical conflict adjudication, a mutual verification mechanism between parameters is established, realizing the reasonable resolution of contradictory information and adaptive correction of data quality, and finally outputting interpretable multipath risk assessment results. Attached Figure Description

[0016] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0017] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Example

[0018] Please see Figure 1 As shown in the figure, this embodiment provides a multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure, including the following steps: Because multicenter clinical data vary in terms of equipment, recording habits, and missing data patterns, in order to differentiate the data quality of different centers and dynamically adjust the correction intensity in subsequent processing, it is first necessary to obtain complete time-series clinical data and establish the differences between centers.

[0019] In this embodiment, the system obtains time-series clinical data of respiratory failure patients from multiple medical centers through the interface of the hospital information system or data import tools.

[0020] Clinical data should include at least: Physiological parameters: including blood oxygen saturation Oxygenation index ( ), respiratory rate (RR), tidal volume (Vt), end-tidal carbon dioxide ( ), etc., with a sampling interval of once per minute; Intervention operation records include oxygen flow rate, ventilation mode, positive end-expiratory pressure (PEEP), body position, sedation level, etc. Each record includes intervention type, start time, end time, and intensity value. The differences between the centers were obtained through offline calibration or online statistics. Specifically, for each center... Pre-calibrate the equipment offset coefficient (For example, by comparing the percentage shift in blood oxygen saturation measured with a standard gas source), calculate the average sampling delay. (Based on the average of historical timestamp intervals), statistical parameter missing probability (The percentage of missing values ​​out of total records), and equipment calibration stability. (i.e., the complement of the coefficient of variation of the calibration parameters) Simultaneously, the credibility of the central benchmark is set. This value is manually set based on the equipment's accuracy level and the quality of historical data, and its range is 0 to 1. Simultaneously for each physiological parameter Initial credibility index This indicates that the initial level of confidence in this parameter is moderate.

[0021] Because the observed physiological parameter values ​​are mixed with systematic equipment bias, sampling delay error and instantaneous intervention effects, in order to obtain pure and true physiological signals, the observed values ​​must be decomposed into multi-source perturbations.

[0022] In this embodiment, for each observation time t and each physiological parameter i, the system performs the following perturbation separation process: (1) Separation of equipment-related error components Based on the equipment deviation coefficient of center c to which this data belongs. Calculate the components of equipment error :

[0023] in This was obtained during the offline calibration phase; Physiological parameters observed at time t The value. For example, if the SpO2 equipment deviation coefficient of center A is +0.02 (i.e., the reading is 2% higher), and it is calibrated annually using a standard gas source, then... ; (2) Separation of time-delay components According to the center Average sampling delay (Unit: minutes), the time delay error at the current moment is estimated using linear interpolation or Kalman smoothing. A preferred embodiment is as follows:

[0024] in, Use the difference between the first two time points as an approximation. If minute, =-0.5% / minute, then =-0.75%; (3) Intervention affects component separation instantaneously Based on the intervention operation record, the preset intervention-response transfer function is invoked. Taking oxygen flow increase as an example, the transfer function can be modeled as a first-order inertial element:

[0025] in, This is an instantaneous effect error; For steady-state gain (e.g.) The gain on oxygen flow rate is 2% / L. This is a time constant (e.g., 3 minutes). This represents the change in oxygen flow rate (L / min). The moment the intervention begins; The system calculates the time from the start of the intervention to the current moment based on the actual intervention records. cumulative instantaneous impact .

[0026] (4) Reconstruction of true physiological estimates The true physiological estimate is obtained by subtracting each of the above components from the observed values ​​in sequence. :

[0027] Simultaneously record the disturbance separation residual. This is used for subsequent credibility updates.

[0028] Through the above decomposition, non-disease factors in the observations are effectively removed, and subsequent risk assessments are based on pure, true physiological estimates, avoiding misjudgments of disease deterioration or improvement caused by equipment drift, sampling lag, and treatment procedures.

[0029] Since clinical interventions serve as natural "probes" for verifying the rationality of physiological parameters, in order to use intervention events to reverse-evaluate the reliability of each parameter and identify treatment ineffectiveness or physiological mismatch, it is necessary to establish an intervention response consistency judgment mechanism and dynamically update the reliability indicators accordingly, as follows: The system monitors the records of each intervention operation. For a single intervention operation from time [time]... arrive Intervention events (e.g., oxygen flow rate increased from 3 L / min to 5 L / min) were used to calculate the change in intervention intensity. and changes in physiological parameters The specific judgment logic is as follows: like (Enhanced intervention) and expected improvement in oxygenation parameters (e.g.) If the increase in oxygenation response is less than a preset threshold (preferably 2%), the response category label is marked as "ineffective oxygenation response". like And ventilation parameters (such as If the blood pressure rises above a preset threshold (preferably 5 mmHg), it is marked as a "ventilation-oxygenation mismatch response". If the time delay from the start of intervention to the appearance of significant changes in physiological parameters exceeds the preset tolerance window (preferably 10 minutes), it is marked as "time delay affecting response"; Otherwise, mark it as a "normal response".

[0030] Response confidence Calculations are based on the stability of the response waveform. Specifically, the coefficient of variation (CV) of the physiological parameter waveform within the time window after intervention is calculated, and a value is taken as... ,in For adjustment coefficients (e.g.) If the waveform is stable, then... It is close to 1, but if there is a sharp fluctuation, it will be close to 0.

[0031] In order to dynamically adjust the credibility of each parameter by integrating multiple pieces of evidence, the system integrates four factors: fluctuation stability, central benchmark credibility, intervention response consistency, and historical reconstruction residuals.

[0032] For each physiological parameter The confidence index is updated after each intervention or at fixed time intervals (e.g., 5 minutes). : Volatility stability factor: Calculates the coefficient of variation of this parameter value within a preset time window (e.g., 1 hour). ,Pick The smaller the CV, the closer Stab is to 1.

[0033] Central benchmark credibility: directly obtain the pre-set reliability of the center. .

[0034] Response Consistency Adjustment: Based on the response consistency results obtained above, if the direction of change of this parameter is consistent with expectations, then... If inconsistent and the response label is "ineffective oxygenation" or "mismatch", then If it is a "time delay effect", then (No adjustment).

[0035] Historical bias factor: used to calculate the residuals of disturbance separation within a past time window. root mean square ,Pick (Assume that the RMSE has been normalized to the range [0,1].)

[0036] Fusion Update: The four factors mentioned above are weighted and summed, then mapped to [0,1] using the Sigmoid function:

[0037] in The weights can be preset to 1. .

[0038] Furthermore, when the data distribution of the same physiological parameter shows statistically significant differences across different medical centers (e.g., p < 0.05 via a two-sample t-test), the system applies a uniform punitive reduction to the reliability index of that parameter across all centers, for example, by multiplying it by 0.9.

[0039] Furthermore, since single-parameter independent processing easily overlooks the physiological constraints between variables, in order to invert potential state variables that cannot be directly observed (such as lung compliance and oxygenation reserve) and identify logical conflicts between information from different sources, it is necessary to establish a multi-parameter coupled analysis network and a hierarchical adjudication mechanism.

[0040] This embodiment employs a learnable mapping network (e.g., a three-layer fully connected neural network, with the number of input layer nodes equal to twice the number of physiological parameters, 64 hidden layers, and 3 output layer nodes corresponding to lung ventilation efficiency, oxygenation reserve, and respiratory drive state) to achieve the inversion from corrected physiological estimates to latent state variables. The network structure is as follows: Input: Corrected physiological estimate vector Intervention variables Credibility index vector ; Output: Latent state variables , This represents lung ventilation efficiency, reflecting the current quality of alveolar ventilation. A value closer to 1 indicates higher ventilation efficiency (i.e., a larger proportion of effective alveolar ventilation to minute ventilation, and smaller dead space). and (Close to normal value), the closer the value is to 0, the lower the ventilation efficiency (such as increased dead space, insufficient ventilation or hyperventilation). This represents oxygenation reserve status, reflecting lung gas exchange function and oxygenation reserve capacity. A value closer to 1 indicates sufficient oxygenation reserve (i.e., at the current oxygenation level). The lower body can maintain normal function. (Small intrapulmonary shunts) The closer the value is to 0, the poorer the oxygenation reserve (e.g., increased intrapulmonary shunts, ventilation / perfusion mismatch, requiring higher oxygenation). (to maintain oxygenation) This represents the respiratory drive intensity, reflecting the excitability of the respiratory center and the output level of the respiratory muscles. A value closer to 1 indicates extremely strong respiratory drive (often manifested as a rapid respiratory rate, increased tidal volume, or involvement of accessory respiratory muscles), while a value closer to 0 indicates respiratory drive inhibition or fatigue (such as a slow respiratory rate, shallow and slow breathing, or central inhibition). Each component ranges from 0 to 1.

[0041] Physiological constraint regularization terms are introduced during training to ensure the physiological rationality of the inversion results. Specific constraints include: Oxygenation constraint: and The relationship between them should conform to the Severinghaus oxygenation curve equation.

[0042] Ventilation constraint: The relationship between RR and Vt should conform to minute ventilation and The resulting balance.

[0043] Mechanical constraints: The relationship between Vt, PEEP, and transpulmonary pressure should conform to the respiratory mechanics equations. Where C represents compliance; The platform pressure is calculated using residuals to determine the degree of deviation.

[0044] Total regularization loss for: in, This represents the loss term due to oxygenation constraint. The weights of the loss term for oxygenation constraint; This represents the loss term due to ventilation constraints. The weight of the loss term due to ventilation constraints; This represents the loss term due to mechanical constraints. The weights are the weights of the loss term for mechanical constraints; all three weights are between 0.1 and 0.5. When the inverted latent state variable violates any of the above physiological constraints by more than a preset tolerance threshold (e.g., relative deviation > 20%), the system triggers a "coupling anomaly marker" and sends the sample to the conflict detection process.

[0045] In order to make reasonable decisions and output a reliable set of parameters, the system performs the following steps: Conflict identification: The system monitors three types of conflicts in real time; Logical contradictions and conflicts between parameters: for example An increase in the relative risk (RR) indicates improvement, but a simultaneous increase in RR (recovery risk) indicates worsening, which defies common physiological sense. This can be detected by calculating the correlation coefficient between the two parameters within a sliding window and checking if the sign is opposite to the expected value.

[0046] Inter-center state estimation discrepancy conflict: After correction, the latent state variable estimates obtained from data of the same patient at different centers differ by more than a threshold (0.3).

[0047] Conflicts where the expected response did not occur after intervention: After the response delay window following intervention, physiological parameters still did not show the expected changes, and the response confidence level was higher than 0.7.

[0048] Three-tier adjudication process (taking a specific conflict as an example): Collect the set of parameters involved in the conflict and their current credibility metrics. Stability indicators of the affiliated center Response confidence Historical trend continuity (the consistency of the direction of this parameter over the past hour).

[0049] First-level ruling: Priority shall be given to this ruling. The parameters are specified. If only one side satisfies the requirements, the parameters or data source corresponding to that side are directly adopted.

[0050] Second-level adjudication: If conflicts still exist among the high-confidence parameters (i.e.) If all values ​​are >0.7 but the conclusions are opposite, then compare the stability indices of the centers to which each parameter belongs. ,use Larger parameters.

[0051] Third-level ruling: If the issue remains unresolved (e.g.) (If they are similar), then it depends on the response confidence. Parameters consistent with the expected direction of intervention are used.

[0052] Output after adjudication: The final set of parameters adopted. The set of suppressed parameters (In subsequent integration, the weight is reset to zero), conflict reason identifier (such as " Contradicting the RR trend - First-level ruling adopted (”).

[0053] To comprehensively assess and provide interpretable results, it is necessary to compute three risk components in parallel and dynamically adjust their fusion weights. In this embodiment, three risk components are defined: disease evolution trend risk, parameter coupling anomaly risk, and center drift risk. The risk of disease progression trend is obtained in the following way: based on the time series of each parameter in the final parameter set, a trend estimation algorithm is used to extract the intensity of the deterioration trend, and the magnitude of the trend risk is determined according to the intensity of the deterioration trend. The risk of coupling anomalies between parameters is obtained in the following way: based on the actual correlation strength and theoretical correlation strength between multiple preset parameter pairs, the coupling anomaly index is calculated using the deviation measurement method, and the magnitude of the coupling risk is determined according to the level of the coupling anomaly index. Center drift risk is obtained in the following way: based on the frequency of conflict occurrence within a preset time window and the stability index of each center, a weighted fusion mechanism is used to calculate the drift risk value, and the magnitude of the drift risk is determined according to the frequency of conflict and the quality of center stability.

[0054] Because the reliability of each risk component varies across different clinical scenarios, the system dynamically calculates the weights based on the current confidence distribution, conflict frequency, state stability, and intervention density to adaptively adjust the fusion weights. The dynamic adjustment method for the weights of multiple risk components is as follows: Calculate the average confidence index of all parameters in the final parameter set used. .like (With a preset credibility threshold), the trend risk weight will be adjusted. Multiply by 1.2 (to increase the contribution of the trend path).

[0055] Calculate the preset time window Frequency of conflict occurrence within 1 hour. If (Preset conflict frequency threshold), then the drift risk weight will be applied. Multiply by 1.3.

[0056] Calculate latent state variables The degree of temporal volatility (variance over the past hour). If the variance is less than 0.1 (preset stability threshold), then risk weights will be coupled. Multiply by 0.8 (to reduce the contribution of the coupling path, since anomalies may be noise rather than a true mismatch).

[0057] Calculate the occurrence density of intervention operations within a preset time window. (Number of interventions per unit time). If This is to avoid misinterpreting physiological fluctuations caused by intervention as a worsening trend.

[0058] Finally, the three weights are normalized so that... .

[0059] Overall Risk Index :

[0060] In the formula, Risks associated with the progression of the disease; This is to mitigate the risk of abnormal coupling between parameters. Risk of center drift; Output risk assessment results: Overall Risk Index The value ranges from 0 to 1, and can be classified as low risk (<0.3), medium risk (0.3~0.6), and high risk (>0.6). Risk-dominant type: Compare the normalized values ​​of the three risk components and select the type corresponding to the largest value; Key influencing parameters: The 1-3 parameters that contribute the most to trend risk and coupling risk (e.g., through gradient or sensitivity analysis). The set of suppressed parameters (Obtained from conflict resolution).

[0061] Through dynamic weight adjustment, the system emphasizes disease trends when data quality is high, data drift when conflicts are frequent, and avoids false positives during frequent interventions, making the assessment results more reliable and interpretable. Clinicians can intuitively see whether the source of risk is disease progression, disruption of physiological relationships, or data issues, thus making accurate decisions.

[0062] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure, characterized in that, Includes the following steps: S1. Obtain time-series clinical data of respiratory failure patients in multiple medical centers. The clinical data includes at least a set of physiological parameters, intervention operation records, and differential characteristics corresponding to each center. The differential characteristics include center baseline reliability and center stability index, and an initial reliability index is assigned to each physiological parameter. S2. Based on the differences in the centers, perform perturbation separation processing on the physiological parameter values ​​at each observation time, reconstruct the true physiological estimate, and record the perturbation separation residuals. S3. Based on the intervention operation records, determine the consistency between the actual response and the expected response of physiological parameters before and after each intervention, generate response category labels and response confidence, and thus dynamically update the confidence index of each physiological parameter. S4. Based on the updated credibility index and the reconstructed true physiological estimates, perform multi-parameter coupled state analysis to invert latent state variables, thereby detecting logical conflicts between physiological parameters, between different medical centers, and between intervention and response. When a conflict is detected, make a graded decision based on the credibility index of each parameter, the stability index of the center to which it belongs, and the confidence level of the response, and determine the final set of parameters to be adopted and the set of parameters to be suppressed. S5. Based on the final set of parameters, calculate multiple risk components in parallel, and dynamically adjust the weight of each risk component according to the current confidence distribution, intervention frequency and stability of potential state variables, and output the risk assessment result.

2. The multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure according to claim 1, characterized in that: In step S2, the disturbance separation process specifically includes: Separate equipment-related error components using the equipment deviation coefficient of the respective center; Estimate the time delay and separate the time delay components using historical sampling intervals; The transient impact components caused by the intervention are separated using intervention operation records and a pre-set intervention response transfer function; The true physiological estimate is obtained by subtracting each of the above components from the observed values ​​in sequence.

3. The multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure according to claim 2, characterized in that: In step S3, the generation of response category labels and response confidence levels are as follows: Compare the relationship between the direction of change in intervention intensity and the direction and magnitude of change in physiological parameters; When the intervention is enhanced but the expected improvement in oxygenation parameters does not change significantly, it is marked as an ineffective oxygenation response. When the intervention is enhanced but ventilation parameters show an abnormal increase, it is marked as a ventilation-oxygenation mismatch response. When the physiological response is delayed beyond the preset tolerance window, it is marked as a time delay-affected response. Otherwise, it is marked as a normal response. Calculate the response confidence level based on the stability of the response waveform.

4. The multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure according to claim 3, characterized in that: In step S3, the dynamic updating of the reliability index for each physiological parameter specifically includes: Calculate the coefficient of variation of this parameter value within a preset time window in the past, and use it as the fluctuation stability factor; Obtain the baseline credibility preset by the relevant center; Based on the consistency of the intervention response, the credibility score is increased when the direction of change of this parameter is consistent with the expectation; otherwise, the credibility score is decreased. Calculate the root mean square of the disturbance separation residuals within the preset time window in the past, and use it as the historical deviation factor; The updated credibility index is obtained by fusing the volatility stability factor, benchmark credibility, adjusted credibility score, and historical deviation factor through weighted summation or nonlinear mapping.

5. The multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure according to claim 4, characterized in that: In S4, the multi-parameter coupled state analysis is implemented through a learnable mapping network, which incorporates physiological constraint regularization terms during training. The physiological constraints include constraints on the oxygenation curve relationships between oxygenation parameters, constraints on the ventilation balance relationships between ventilation parameters, and constraints on the mechanical equations between respiratory mechanics parameters. When the inverted potential state variables violate any of the above physiological constraints, a coupling anomaly marker is triggered and sent to conflict detection.

6. The multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure according to claim 5, characterized in that: In step S4, the specific steps for determining the final set of parameters to be used and the set of parameters to be suppressed are as follows: The conflicts include logical contradictions and conflicts among detection parameters, conflicts arising from differences in state estimations among centers, and conflicts arising from the failure of the expected response to occur after intervention. For each conflict, collect the set of parameters involved, their current credibility index, the stability index of the center to which they belong, and the response confidence level; The following three levels of rulings shall be executed: Level 1 decision: Prioritize parameters with credibility indices higher than the preset high threshold; Second-level decision: If conflicts still exist among the high-confidence parameters, compare the stability characteristics of the centers to which each parameter belongs, and adopt the parameter with better stability; Level 3 decision: If the problem remains unresolved, then rely on the confidence level of the response and use parameters consistent with the expected direction of intervention; The final output after the ruling is the set of parameters to be used, the set of parameters to be suppressed, and the conflict reason identifier.

7. The multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure according to claim 6, characterized in that: In S5, multiple risk components include the risk of disease evolution trend, the risk of abnormal coupling between parameters, and the risk of center drift. The risk assessment results include a comprehensive risk index and the dominant risk type, key influencing parameters, and information on suppressed parameters.

8. The multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure according to claim 7, characterized in that: The risk of disease progression trend is obtained in the following way: based on the time series of each parameter in the final parameter set, a trend estimation algorithm is used to extract the intensity of the deterioration trend, and the magnitude of the trend risk is determined according to the intensity of the deterioration trend. The coupling anomaly risk between parameters is obtained by: calculating the coupling anomaly index based on the actual correlation strength and theoretical correlation strength between multiple preset parameter pairs using a deviation measurement method, and determining the magnitude of the coupling risk based on the level of the coupling anomaly index. The center drift risk is obtained in the following way: based on the frequency of conflict occurrence within a preset time window and the stability index of each center, a weighted fusion mechanism is used to calculate the drift risk value, and the magnitude of the drift risk is determined according to the frequency of conflict and the quality of center stability.

9. The multicenter respiratory parameter fusion risk assessment method for patients with respiratory failure according to claim 8, characterized in that: In S5, the dynamic adjustment method for the weights of multiple risk components is as follows: Calculate the average of the credibility indices of all valid parameters in the current final set of parameters, and increase the weight of trend risk when the average is higher than the preset credibility threshold. Calculate the frequency of collisions within a preset time window, and increase the weight of drift risk when the collision frequency is higher than a preset collision frequency threshold; State stability is determined based on the temporal fluctuation of potential state variables, and the weight of coupling risk is reduced when the state stability is higher than a preset stability threshold. Based on the density of intervention operations within a preset time window, the weight of trend risk is temporarily reduced when the intervention density exceeds a preset intervention density threshold.

Images