Intelligent severe case early warning analysis system based on multi-source clinical data fusion

Abstract

The invention discloses an intelligent critical case early warning analysis system based on multi-source clinical data fusion, belongs to the technical field of image processing, and can more accurately carry out respiration monitoring and early warning. Comprising a respiratory movement analysis module used for generating a respiratory movement amplitude change curve according to thoracic and abdominal movement video data; the breathing machine waveform analysis module is used for generating waveform characteristic data containing abnormal time slices according to the breathing machine waveform data; the blood oxygen signal processing module is used for generating a blood oxygen change trend vector according to the blood oxygen saturation degree signal data; the multi-modal fusion module is used for generating a multi-dimensional physiological feature vector under a synchronous timestamp according to the breathing movement amplitude change curve, the waveform feature data and the blood oxygen change trend vector; the risk assessment module is used for generating a breathing compensation state score according to the multi-dimensional physiological feature vector; and the alarm triggering module is used for triggering a grading alarm signal according to the deviation degree of the breathing compensation state score and the baseline range.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and specifically to a critical care intelligent early warning analysis system based on multi-source clinical data fusion. Background Art

[0002] In modern medicine, the monitoring and care of critically ill respiratory patients have always been an important and complex challenge. Critically ill respiratory patients usually include those in the intensive care unit (ICU), who face various serious respiratory problems and must be closely monitored and cared for in order to take effective treatment measures in a timely manner. Due to the suddenness and complexity of respiratory problems, medical staff need to use efficient and accurate monitoring means to promptly detect the abnormal conditions of patients in order to avoid potential complications and reduce medical costs.

[0003] Currently, the monitoring of critically ill respiratory patients mostly relies on traditional visual monitoring and image processing techniques. Conventional methods usually analyze the image data of patients and use the gray-scale change features in the images to judge the respiratory conditions of patients. However, the clarity of the images is often insufficient, making it difficult to effectively identify the abnormal changes of patients in actual monitoring. For example, when performing respiratory monitoring and early warning, it is often necessary to preprocess the images by means of Gaussian filtering technology to improve the readability of the images. However, traditional Gaussian filtering uses a fixed window size for processing, and this approach has significant limitations.

[0004] Gaussian filtering with a fixed window may cause the images to be overly smoothed, thus masking the detailed features in the images. This situation makes the capture of subtle changes in the patient's body less accurate, reducing the reliability and accuracy of respiratory monitoring and early warning.

[0005] The disclosure of the above background art content is only for assisting in understanding the concept and technical solution of the present invention, and it does not necessarily belong to the prior art of this patent application. Without clear evidence indicating that the above content was publicly available on the filing date of this patent application, the above background art should not be used to evaluate the novelty and inventiveness of this application. Summary of the Invention

[0006] This application provides a critical care intelligent early warning analysis system based on multi-source clinical data fusion, which can perform respiratory monitoring and early warning more accurately.

[0007] To achieve the above object, the embodiments of this application disclose the following technical solutions:

[0008] A critical care intelligent early warning analysis system based on multi-source clinical data fusion, comprising:

[0009] A data acquisition module, configured to continuously acquire chest and abdomen motion video data, ventilator waveform data, and blood oxygen saturation signal data of critically ill respiratory patients;

[0010] A respiratory motion analysis module for performing respiratory motion region segmentation processing on chest and abdomen motion video data to generate a respiratory motion amplitude change curve;

[0011] A ventilator waveform analysis module for performing dynamic threshold analysis processing on ventilator waveform data to generate waveform feature data including abnormal time segments;

[0012] An oxygen saturation signal processing module for performing adaptive window processing on oxygen saturation signal data to generate an oxygen saturation change trend vector;

[0013] A multi-modal fusion module for performing multi-modal feature fusion processing on the respiratory motion amplitude change curve, waveform feature data, and oxygen saturation change trend vector to generate a multi-dimensional physiological feature vector at a synchronous timestamp;

[0014] A risk assessment module for inputting the multi-dimensional physiological feature vector into a risk assessment model and outputting a respiratory compensation status score;

[0015] An alarm trigger module for triggering a hierarchical alarm signal according to the deviation degree of the respiratory compensation status score from the baseline range.

[0016] In the embodiment of the present application, by synchronously collecting chest and abdomen motion video data, ventilator waveform data, and oxygen saturation signal data, a multi-source information complementary mechanism is constructed. The chest and abdomen motion video data provides dynamic information of the respiratory pattern, the ventilator waveform data reflects the mechanical ventilation parameters, and the oxygen saturation signal characterizes the oxygenation state. The three cross-verify the patient's state from different physiological dimensions. The multi-modal fusion module combines physiological feature vectors from different sources into a multi-dimensional feature vector according to the synchronous timestamp, eliminating the noise or local failure interference of a single data source. The comprehensive evaluation model after data fusion can capture abnormal features missed by traditional single-image analysis, such as abnormal correlations between the oxygen saturation trend and ventilator pressure fluctuations. The system dynamically analyzes the respiratory compensation state through multi-dimensional features, avoiding the problem of excessive image smoothing caused by traditional fixed filtering windows and reducing the risk of misjudgment caused by image quality fluctuations. In this way, on the basis of retaining key physiological details, the monitoring sensitivity and reliability are improved, and accurate early warning of abnormal changes in critically ill respiratory patients is realized.

[0017] In some possible implementation manners, the respiratory motion analysis module is specifically used for:

[0018] Using a pre-trained U-Net network to perform respiratory motion region segmentation on the original video frame sequence {F t} to generate a respiratory region mask {M t}, where F t represents the t-th frame image, and M t is a binary mask;

[0019] Calculate the pixel displacement Δd of the region according to the difference between adjacent frame masks t :

[0020]

[0021] In the formula, ⊕ represents the exclusive OR operation, and (x, y) is the pixel coordinate; M t (x, y) is the pixel value of the t-th frame mask at the coordinate (x, y);

[0022] For the pixel displacement Δd of the region t Perform Butterworth low-pass filtering on the sequence to generate the respiratory motion amplitude change curve A(t):

[0023]

[0024] In the formula, N is the filter window length, and w k is the k-th coefficient of the Butterworth filter. In this way, the dependence of the traditional gray-scale analysis method on the image quality can be overcome through the image segmentation ability of the U-Net network. The respiratory motion region can be extracted by generating an accurate binary mask, avoiding feature loss caused by image blurring or occlusion. The calculation of the pixel displacement of the region is based on the segmentation result, and the exclusive OR operation is used to quantify the dynamic change of the respiratory region between adjacent frames, converting the visual motion into an analyzable numerical sequence, and improving the detection sensitivity of the respiratory amplitude change. The Butterworth low-pass filtering process performs dynamic noise suppression on the displacement sequence, balancing the signal fidelity and smoothness by adjusting the window length and filter coefficients, ensuring that the respiratory curve eliminates high-frequency interference while retaining the key fluctuation features. In this way, a high-fidelity conversion from the original video to the quantified respiratory features can be achieved, enhancing the robustness and monitoring accuracy of the system in a complex clinical environment.

[0025] In some possible implementation manners, the ventilator waveform analysis module is specifically configured to:[[]]

[0026] Calculate the standard deviation of the sliding window for the ventilator waveform data {P i}, and output the pressure fluctuation sequence σ w :[[]]

[0027]

[0028] In the formula, m is the window length; P j is the j-th pressure sampling value within the window; is the average pressure within the window;

[0029] Based on the historical normal cycle segment of the ventilator waveform data {P i}, calculate the dynamic threshold θ:[[]]

[0030] θ = μhist +α·σ hist

[0031] Wherein, μ hist is the mean value of the historical normal cycle pressure; σ hist is the standard deviation of the historical normal cycle pressure; α is the sensitivity coefficient;

[0032] According to the real-time comparison result between the pressure fluctuation sequence σ w and the dynamic threshold θ, waveform feature data including abnormal time segments {T k} and their duration {L k} is generated; wherein,

[0033]

[0034] Wherein, and are the start and end times of the k-th abnormal segment. Among them, the calculation of the standard deviation of the sliding window quantifies the real-time fluctuation of the ventilator pressure waveform to generate a pressure fluctuation sequence, providing basic data support for the adaptive adjustment of the dynamic threshold. The dynamic threshold is constructed based on the mean value and standard deviation of the pressure in the historical normal cycle segment, and an individualized alarm boundary is generated by superimposing the sensitivity coefficient, which not only retains the patient's basic physiological characteristics but also adapts to the normal fluctuation range in different disease stages. During the comparison process between the real-time pressure fluctuation sequence and the dynamic threshold, the dual quantification mechanism of the abnormal time segment and its duration distinguishes instantaneous abnormal fluctuations from persistent abnormal states. The duration parameter is calculated by the time difference between the start and end times, avoiding false alarms caused by only relying on the single-point threshold trigger in the traditional method.

[0035] The calculation of the standard deviation captures the local variation of the waveform, the dynamic threshold provides an individualized alarm benchmark, and the duration screening selects clinically significant abnormal events. With the synergistic effect of the three, the system can not only identify short-term abnormalities such as sudden pressure surges but also capture long-term trends such as slow increase in airway resistance, while filtering out instantaneous interference signals caused by coughing or body position changes. In this way, the waveform analysis of the ventilator can have both sensitivity and specificity, providing a hierarchical early warning ability for the abnormal state of mechanically ventilated patients.

[0036] In some possible implementation manners, the blood oxygen signal processing module is specifically configured to:

[0037] Calculate the variance of the blood oxygen saturation signal {S i}, and output the real-time variance

[0038]

[0039] Wherein: n is the initial window length; is the mean value of blood oxygen in the window;

[0040] Based on real-time variance Dynamically adjust the window length and output the adaptive window parameter n ′ :

[0041]

[0042] In the formula: is the preset maximum variance threshold; ∈ is a very small constant to prevent division by zero;

[0043] Perform window normalization on the blood oxygen saturation signal {S i} and output the standardized signal

[0044]

[0045] where: μ n′ is the mean of the current window; σ n′ is the standard deviation of the current window;

[0046] Perform linear regression analysis on the standardized signal to generate a blood oxygen change trend vector V containing the trend slope β, where,

[0047]

[0048] In the formula: t i is the i-th timestamp; is the mean of the timestamps within the fitting window; is within the fitting window the mean of; m is the length of the fitting window;

[0049] V = [β, σ n′ , μ n′ T ;

[0050] In the formula: σ n′ : the standard deviation of the blood oxygen saturation within the adjusted window; μ n′ ​: The mean value of blood oxygen saturation within the adjusted window. Among them, the real-time variance calculation of the blood oxygen saturation signal directly drives the dynamic adjustment of the adaptive window length, and the window length is adjusted inversely according to the variance size: when the signal fluctuates violently, the window is shortened to capture rapid changes, and when the fluctuation is gentle, the window is extended to smooth the noise. Window normalization eliminates the influence of signal amplitude differences on trend analysis, making the blood oxygen change rates of different patients comparable. Linear regression analysis is applied to the normalized signal, and the extracted trend slope quantifies the speed of blood oxygen decline or recovery. The standard deviation and mean value of the adjusted window respectively reflect the current blood oxygen stability and baseline level. The multi-dimensional combination of the trend slope, standard deviation, and mean value constitutes the blood oxygen change trend vector, which fuses the instantaneous fluctuation characteristics and long-term evolution trends into a unified indicator. The dynamic adjustment of the adaptive window ensures that the regression analysis can focus on short-term trends during rapid blood oxygen changes and evaluate long-term trends during stable phases; window normalization avoids misjudgment of trends caused by individual blood oxygen baseline differences. The closed-loop linkage of variance calculation, window adjustment, and regression analysis enables the system to not only promptly capture sudden drops in blood oxygen but also identify slow and hidden hypoxia, while filtering out instantaneous measurement noise through the standard deviation. In this way, it breaks through the lag of traditional threshold alarms, provides a comprehensive description of the blood oxygen status including rate, stability, and baseline offset for the multi-modal fusion module, and ultimately realizes early warning of respiratory compensation imbalance.

[0051] In some possible implementation manners, the multi-modal fusion module is specifically configured to:

[0052] Perform downsampling processing on the ventilator waveform data {P i} and the blood oxygen signal {S j}, and output the downsampled sequences {P′ k} and {S′ k}:

[0053]

[0054] Where: r: The downsampling rate, satisfying 1 / r is equal to the frame rate f of the chest and abdomen movement video data v ; k: The serial number of the downsampling window, an integer index; i: The original sequence index of the ventilator waveform data {P i}; j: The original sequence index of the blood oxygen signal {S j};

[0055] Align the respiratory movement amplitude change curve A(t), the downsampled sequences {P′ k} and {S′ k} based on the dynamic time warping algorithm, and output the synchronization timestamp τ m :

[0056]

[0057] where m is the index of the synchronization timestamp (1 ≤ m ≤ M); M is the total number of timestamps to be aligned;

[0058] τ m : the m-th synchronization timestamp; P′ m : the mean value of the ventilator waveform data in the m-th time window after downsampling; S′ m : the mean value of the blood oxygen signal in the m-th time window after downsampling;

[0059] According to the synchronization timestamp τ m extract the set of local extreme points {E m} of A(t), and generate the synchronization timestamp feature:

[0060]

[0061] where: E m : the phase marker point of the synchronization timestamp, used for time alignment of the multi-dimensional physiological feature vector; δ: the tolerance range of time alignment. Among them, the downsampling process can unify the high-frequency sampling rates of the ventilator waveform data and the blood oxygen signal to the frame rate of the chest and abdomen video, eliminating the time resolution conflict caused by the sampling rate difference of multiple devices. After the ventilator waveform data is downsampled by the sliding window mean, the pressure fluctuation trend is retained, and after the blood oxygen signal is downsampled by the mean, the oxygenation change profile is retained. The forced synchronization of the two with the video frame rate lays the foundation for subsequent time alignment. The dynamic time warping algorithm takes the respiratory amplitude change curve as the benchmark, and generates the synchronization timestamp by minimizing the combined difference between the waveform and the downsampled sequence of blood oxygen, so that different modality data are strictly aligned within the same respiratory cycle. The local extreme points of the respiratory amplitude curve mark the turning moments of the respiratory phase, and the tolerance range constraint of the synchronization timestamp ensures the precise correlation of cross-modal data on key physiological events.

[0062] The abnormal segments of the ventilator pressure, the change of the blood oxygen trend slope, and the fluctuation of the respiratory movement amplitude form cross-verification in the synchronous time dimension. For example, the synchronous occurrence of a sudden increase in pressure and a sudden decrease in respiratory amplitude can clearly indicate the risk of airway obstruction. The spatio-temporal consistency of multi-modal data eliminates the one-sidedness of single-source data analysis. The high-frequency details of the ventilator waveform, the metabolic information of the blood oxygen signal, and the spatial characteristics of the video movement jointly construct a complete portrait of the respiratory compensation state. In this way, the bidirectional improvement of the abnormal detection sensitivity and the false alarm suppression ability can be achieved.

[0063] In some possible implementation manners, the steps of generating the risk assessment model include:

[0064] screening the multi-dimensional physiological feature vectors corresponding to the stable respiratory cycles from the historical patient data as the training data set;

[0065] Use a long short - term memory network to perform time - series modeling training on the training data set to generate an initial risk assessment model;

[0066] Optimize the hyperparameters of the initial risk assessment model through cross - validation to generate a pre - trained long short - term memory network risk assessment model.

[0067] In some possible implementation manners, the steps of determining the baseline range include:

[0068] Obtain the sequence of respiratory compensation scores {Y q} during the initial monitoring period of the patient;

[0069] According to the sequence of respiratory compensation scores {Y q}, calculate the mean value μ b and the standard deviation σ b :

[0070]

[0071] where Y q : the q - th respiratory compensation score; Q: the total number of scores during the initial monitoring period;

[0072] According to the mean value μ b and the standard deviation σ b , determine the baseline range, and the baseline range is [μ b - 3σ b , μ b +3σ b .

[0073] In this way, it is possible to provide statistical support for the baseline range by calculating the mean value and standard deviation of the sequence of respiratory compensation scores during the initial monitoring period of the patient. Setting the baseline range as the interval of the mean value plus or minus three times the standard deviation can fully cover the normal physiological fluctuation range. The personalized baseline construction process integrates the time - series characteristics and individual differences of the patient's historical data, enabling the system to distinguish between the group - general threshold and the patient - specific threshold, and avoiding the neglect of individual differences by the traditional fixed threshold. In addition, setting the three - standard - deviation interval can also balance the sensitivity and specificity of anomaly detection, allowing physiological changes within the normal fluctuation range while capturing abnormal deviations beyond the reasonable range.

[0074] In some possible implementation manners, the alarm trigger module is specifically used for:

[0075] Calculate the absolute value of the difference between the real - time respiratory compensation status score and the median value of the baseline range;

[0076] When the absolute value of the difference exceeds the first threshold but is less than the second threshold, trigger a first - level alarm signal;

[0077] When the absolute value of the difference reaches the second threshold but is less than the third threshold, trigger a second - level alarm signal;

[0078] When the absolute value of the difference exceeds the third threshold, a third-level alarm signal is triggered. In this way, the degree of deviation can be divided by multiple preset threshold gradients. The first-level alarm prompts potential risks, the second-level alarm triggers moderate intervention, and the third-level alarm requires emergency treatment, forming a clinical priority with a step-by-step response. The dynamic comparison between the real-time score and the baseline median reduces false alarms caused by individual differences or physiological fluctuations, and the multi-level threshold design avoids the response lag of a single alarm threshold to abnormalities of different degrees.

[0079] In some possible implementation manners, the severe critical intelligent early warning analysis system based on multi-source clinical data fusion further includes a treatment suggestion generation module, which is used to generate a corresponding clinical treatment suggestion plan according to the level of the classification alarm signal. In this way, it is beneficial to improve the scheduling efficiency of medical resources and the success rate of treating critical cases.

[0080] In some possible implementation manners, the treatment suggestion generation module is specifically used for:

[0081] Query the pre-stored abnormal type-treatment plan mapping table according to the level of the classification alarm signal to obtain a candidate treatment plan set; wherein, the abnormal type-treatment plan mapping table contains treatment plans and priorities corresponding to different alarm levels;

[0082] Select the highest-priority plan from the candidate treatment plan set according to the preset priority rule to generate a clinical treatment suggestion plan. Brief Description of the Drawings

[0083] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those of ordinary skill in the art, without creative efforts, other implementation drawings can be obtained according to the provided drawings.

[0084] Figure 1 It is a schematic structural diagram of a severe critical intelligent early warning analysis system based on multi-source clinical data fusion provided by some embodiments of the present application. Detailed Embodiments

[0085] Now the specific implementation manners of the present invention will be described in detail. Although the present invention is described in conjunction with these specific implementation manners, it should be understood that it is not intended to limit the present invention to these specific implementation manners. On the contrary, these implementation manners are intended to cover alternative, modified, or equivalent implementation manners that may be included within the spirit and scope of the invention defined by the claims. In the following description, a large number of specific details are set forth in order to provide a comprehensive understanding of the present invention. The present invention can be implemented without some or all of these specific details.

[0086] When used in conjunction with the terms "comprising," "a method comprises," or similar language in this specification and the appended claims, the singular forms "a," "an," and "the" include plural references unless the context clearly indicates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0087] Please refer to Figure 1 , a critical care intelligent early warning analysis system based on multi-source clinical data fusion, comprising:

[0088] A data acquisition module 1 for continuously acquiring chest and abdominal motion video data, ventilator waveform data, and blood oxygen saturation signal data of critically ill patients with respiratory diseases;

[0089] A respiratory motion analysis module 2 for performing respiratory motion region segmentation processing on the chest and abdominal motion video data to generate a respiratory motion amplitude change curve;

[0090] A ventilator waveform analysis module 3 for performing dynamic threshold analysis processing on the ventilator waveform data to generate waveform feature data containing abnormal time segments;

[0091] A blood oxygen signal processing module 4 for performing adaptive window processing on the blood oxygen saturation signal data to generate a blood oxygen change trend vector;

[0092] A multi-modal fusion module 5 for performing multi-modal feature fusion processing on the respiratory motion amplitude change curve, waveform feature data, and blood oxygen change trend vector to generate a multi-dimensional physiological feature vector at the same synchronous timestamp;

[0093] A risk assessment module 6 for inputting the multi-dimensional physiological feature vector into a risk assessment model and outputting a respiratory compensation status score;

[0094] An alarm trigger module 7 for triggering a graded alarm signal according to the deviation degree of the respiratory compensation status score from the baseline range.

[0095] The system constructs a multi-source information complementary mechanism by synchronously collecting chest and abdominal motion video data, ventilator waveform data, and blood oxygen saturation signal data. The chest and abdominal motion video data provides dynamic information on the breathing pattern, the ventilator waveform data reflects the mechanical ventilation parameters, and the blood oxygen saturation signal characterizes the oxygenation status. The three cross-verify the patient's status from different physiological dimensions. The multi-modal fusion module combines physiological feature vectors from different sources into a multi-dimensional feature vector according to the synchronous timestamp, eliminating the noise or local failure interference of a single data source. The comprehensive evaluation model after data fusion can capture abnormal features missed by traditional single-image analysis, such as abnormal correlations between blood oxygen trends and ventilator pressure fluctuations. The system dynamically analyzes the respiratory compensation status through multi-dimensional features, avoiding the problem of excessive image smoothing caused by traditional fixed filtering windows and reducing the risk of misjudgment caused by image quality fluctuations. In this way, the monitoring sensitivity and reliability are improved while retaining key physiological details, achieving accurate early warning of abnormal changes in critically ill respiratory patients.

[0096] In some embodiments, the respiratory motion analysis module 2 is specifically configured to:

[0097] In the first step, use a pre-trained U-Net network to segment the respiratory motion area of the original video frame sequence {F t} to generate a respiratory area mask {M t}, where F t represents the t-th frame image, and M t is a binary mask;

[0098] In the second step, calculate the regional pixel displacement Δd t according to the difference between adjacent frame masks:

[0099]

[0100] In the formula, ⊕ represents the exclusive OR operation, and (x, y) is the pixel coordinate; M t (x, y) is the pixel value of the t-th frame mask at the coordinate (x, y);

[0101] In the third step, perform Butterworth low-pass filtering on the regional pixel displacement Δd t sequence to generate a respiratory motion amplitude change curve A(t):

[0102]

[0103] In the formula, N is the filtering window length, w kis the coefficient of the nth Butterworth filter. In this way, the dependence of traditional gray-scale analysis methods on image quality can be overcome through the image segmentation ability of the U-Net network. By generating an accurate binary mask, the respiratory motion area can be extracted, avoiding feature loss caused by image blurring or occlusion. The calculation of the regional pixel displacement is based on the segmentation result. The exclusive OR operation is used to quantify the dynamic changes in the respiratory area of adjacent frames, converting visual motion into an analyzable numerical sequence and improving the detection sensitivity of respiratory amplitude changes. Butterworth low-pass filtering is performed on the displacement sequence for dynamic noise suppression. By adjusting the window length and filter coefficients, the signal fidelity and smoothness are balanced to ensure that the respiratory curve eliminates high-frequency interference while retaining key fluctuation features. In this way, a high-fidelity conversion from the original video to quantified respiratory features can be achieved, enhancing the robustness and monitoring accuracy of the system in complex clinical environments.

[0104] In some embodiments, the ventilator waveform analysis module 3 is specifically configured to:

[0105] In the first step, calculate the standard deviation of the sliding window for the ventilator waveform data {P i}, and output the pressure fluctuation sequence σ w :

[0106]

[0107] where m is the window length; P j is the jth pressure sampling value within the window; is the average pressure within the window;

[0108] In the second step, based on the historical normal cycle segment of the ventilator waveform data {P i}, calculate the dynamic threshold θ:

[0109] θ = μ hist + α · σ hist

[0110] where μ hist is the average pressure of the historical normal cycle; σ hist is the standard deviation of the pressure in the historical normal cycle; α is the sensitivity coefficient;

[0111] In the third step, according to the real-time comparison result of the pressure fluctuation sequence σ w and the dynamic threshold θ, generate waveform feature data including abnormal time segments {T k} and their durations {L k}; where,

[0112]

[0113] where, and is the start and end time of the k-th abnormal segment. Among them, the calculation of the sliding window standard deviation quantifies the real-time fluctuations of the ventilator pressure waveform, generates a pressure fluctuation sequence, and provides basic data support for the adaptive adjustment of the dynamic threshold. The dynamic threshold is constructed based on the pressure mean and standard deviation of the historical normal cycle segments, and generates an individualized alarm boundary by superimposing a sensitivity coefficient, which not only retains the patient's basic physiological characteristics but also adapts to the normal fluctuation range at different disease stages. During the comparison process between the real-time pressure fluctuation sequence and the dynamic threshold, the dual quantification mechanism of the abnormal time segment and its duration distinguishes instantaneous abnormal fluctuations from persistent abnormal states. The duration parameter is calculated by the time difference between the start and end times, avoiding false alarms caused by relying solely on single-point threshold triggering in traditional methods.

[0114] The standard deviation calculation captures local waveform variations, the dynamic threshold provides an individualized alarm benchmark, and the duration screening selects clinically significant abnormal events. With the synergistic effect of the three, the system can not only identify short-term abnormalities such as sudden pressure surges but also capture long-term trends such as slow increases in airway resistance, while filtering out instantaneous interference signals caused by coughing or body position changes. In this way, the ventilator waveform analysis can have both sensitivity and specificity, providing a hierarchical early warning ability for the abnormal states of mechanically ventilated patients.

[0115] In some embodiments, the blood oxygen signal processing module 4 is specifically configured to:

[0116] In the first step, calculate the variance of the blood oxygen saturation signal {S i}, and output the real-time variance

[0117]

[0118] where: n is the initial window length; is the blood oxygen mean within the window;

[0119] In the second step, dynamically adjust the window length based on the real-time variance and output the adaptive window parameter n':

[0120]

[0121] where: is the preset maximum variance threshold; ∈ is a very small constant to prevent division by zero;

[0122] In the third step, perform window normalization processing on the blood oxygen saturation signal {S i}, and output the standardized signal

[0123]

[0124] where: μ n′is the mean of the current window; σ n′ is the standard deviation of the current window;

[0125] In the fourth step, for the standardized signal perform a linear regression analysis to generate a blood oxygen change trend vector V containing the trend slope β, where

[0126]

[0127] In the formula: t i is the i-th timestamp; is the mean of the timestamps within the fitting window; is within the fitting window mean; m is the length of the fitting window;

[0128] V = [β, σ n′ , μ n′ T ;

[0129] In the formula: σ n′ : the standard deviation of the blood oxygen saturation within the adjusted window; μ n′ : the mean of the blood oxygen saturation within the adjusted window. Among them, the real-time variance calculation of the blood oxygen saturation signal directly drives the dynamic adjustment of the adaptive window length, and the window length is adjusted inversely according to the variance size: when the signal fluctuates violently, the window is shortened to capture rapid changes, and when the fluctuation is gentle, the window is extended to smooth the noise. The window normalization process eliminates the influence of signal amplitude differences on trend analysis, making the blood oxygen change rates of different patients comparable. The linear regression analysis acts on the normalized signal, and the extracted trend slope quantifies the speed of blood oxygen decline or rise, while the standard deviation and mean of the adjusted window respectively reflect the current blood oxygen stability and baseline level. The multi-dimensional combination of the trend slope, standard deviation, and mean constitutes the blood oxygen change trend vector, integrating the instantaneous fluctuation characteristics and long-term evolution trends into a unified index. The dynamic adjustment of the adaptive window ensures that the regression analysis can focus on short-term trends when the blood oxygen changes sharply and evaluate long-term trends during stable phases; window normalization avoids misjudgment of trends caused by individual blood oxygen baseline differences. The closed-loop linkage of variance calculation, window adjustment, and regression analysis enables the system to not only capture sudden blood oxygen drops in a timely manner but also identify slow-developing hidden hypoxia, while filtering out instantaneous measurement noise through the standard deviation. In this way, breaking through the lag of traditional threshold alarms, it provides a comprehensive description of the blood oxygen status including rate, stability, and baseline offset for the multi-modal fusion module, ultimately achieving early warning of respiratory compensation imbalance.

[0130] In some embodiments, the multi-modal fusion module 5 is specifically configured to:

[0131] In the first step, for the ventilator waveform data {P i} and the blood oxygen signal {S​j}Perform downsampling processing and output the downsampled sequence {P' k} and {S' k}:

[0132]

[0133] Where: r: The downsampling rate, satisfying 1 / r is equal to the frame rate f of the chest and abdomen movement video data v ; k: The serial number of the downsampling window, an integer index; i: The original sequence index of the ventilator waveform data {P i}; j: The original sequence index of the blood oxygen signal {S j};

[0134] In the second step, based on the dynamic time warping algorithm, align the respiratory movement amplitude change curve A(t), the downsampled sequence {P' k} and {S' k}, and output the synchronization timestamp τ m :

[0135]

[0136] Where, m: The index of the synchronization timestamp (1 ≤ m ≤ M); M: The total number of timestamps to be aligned; τ m : The m-th synchronization timestamp; P' m : The mean value of the ventilator waveform data in the m-th time window after downsampling; S' m : The mean value of the blood oxygen signal in the m-th time window after downsampling;

[0137] In the third step, according to the synchronization timestamp τ m Extract the local extreme point set {E m} of A(t) and generate the synchronization timestamp feature:

[0138]

[0139] Where: E m: Phase marker points for synchronizing timestamps, used for time alignment of multi-dimensional physiological feature vectors; δ: Tolerance range for time alignment. Among them, downsampling can unify the high-frequency sampling rates of ventilator waveform data and blood oxygen signals to the frame rate of chest and abdomen videos, eliminating the time resolution conflict caused by the sampling rate differences of multiple devices. After downsampling the ventilator waveform data by moving window mean, the pressure fluctuation trend is retained, and after downsampling the blood oxygen signal by mean, the oxygenation change profile is retained. The forced synchronization of the two with the video frame rate lays the foundation for subsequent time alignment. The dynamic time warping algorithm takes the respiratory amplitude change curve as a reference, generates synchronized timestamps by minimizing the combined difference between the waveform and the downsampled blood oxygen sequence, so that different modality data are strictly aligned within the same respiratory cycle. The local extreme points of the respiratory amplitude curve mark the turning moments of the respiratory phase, and the tolerance range constraint of the synchronized timestamps ensures the precise correlation of cross-modal data on key physiological events.

[0140] The abnormal segments of ventilator pressure, the change in the slope of the blood oxygen trend, and the fluctuation of respiratory movement amplitude form cross-verification in the synchronized time dimension. For example, the synchronous occurrence of a sudden increase in pressure and a sudden decrease in respiratory amplitude can clearly indicate the risk of airway obstruction. The spatio-temporal consistency of multi-modal data eliminates the one-sidedness of single-source data analysis. The high-frequency details of the ventilator waveform, the metabolic information of the blood oxygen signal, and the spatial characteristics of video movement jointly construct a complete portrait of the respiratory compensation state through synergistic effects. In this way, the bidirectional improvement of the abnormal detection sensitivity and false alarm suppression ability can be achieved.

[0141] In some embodiments, the steps of generating a risk assessment model include:

[0142] The first step is to screen the multi-dimensional physiological feature vectors corresponding to stable respiratory cycles from historical patient data as the training data set;

[0143] The second step is to use a long short-term memory network to perform time series modeling training on the training data set to generate an initial risk assessment model;

[0144] The third step is to optimize the hyperparameters of the initial risk assessment model through cross-validation to generate a pre-trained long short-term memory network risk assessment model.

[0145] In some embodiments, the steps of determining the baseline range include:

[0146] The first step is to obtain the sequence of respiratory compensation scores {Y q} during the initial monitoring period of the patient;

[0147] The second step is to calculate the mean μ q and the standard deviation σ b according to the sequence of respiratory compensation scores {Y b}:

[0148]

[0149] Wherein, Y q : the q-th respiratory compensation score; Q: the total number of scores during the initial monitoring period;

[0150] Thirdly, according to the mean value μ b and the standard deviation σ b , determine the baseline range, and the baseline range is [μ b - 3σ b , μ b + 3σ b .

[0151] In this way, the mean value and standard deviation of the respiratory compensation score sequence during the patient's initial monitoring period can be calculated to provide statistical support for the baseline range. Setting the baseline range as the interval of the mean value plus or minus three times the standard deviation can fully cover the normal physiological fluctuation range. The personalized baseline construction process integrates the temporal characteristics and individual differences of the patient's historical data, enabling the system to distinguish between the group general threshold and the patient-specific threshold, and avoiding the neglect of individual differences by the traditional fixed threshold. In addition, the three-fold standard deviation interval setting can also balance the sensitivity and specificity of anomaly detection, allowing physiological changes within the normal fluctuation range while capturing abnormal deviations beyond the reasonable range.

[0152] In some embodiments, the alarm trigger module 7 is specifically configured to:

[0153] First, calculate the absolute value of the difference between the real-time respiratory compensation status score and the median of the baseline range;

[0154] Second, when the absolute value of the difference exceeds the first threshold but is less than the second threshold, trigger a first-level alarm signal;

[0155] Third, when the absolute value of the difference reaches the second threshold but is less than the third threshold, trigger a second-level alarm signal;

[0156] Fourth, when the absolute value of the difference exceeds the third threshold, trigger a third-level alarm signal. In this way, the deviation degree can be divided by a preset multiple threshold gradient. The first-level alarm prompts potential risks, the second-level alarm triggers medium intervention, and the third-level alarm requires emergency treatment, forming a clinical priority with a gradual response. The dynamic comparison between the real-time score and the baseline median reduces false alarms caused by individual differences or physiological fluctuations, and the multi-level threshold design avoids the response lag of a single alarm threshold to different degrees of anomalies.

[0157] In some embodiments, the severe critical intelligent early warning analysis system based on multi-source clinical data fusion further includes a treatment suggestion generation module 8, which is used to generate corresponding clinical treatment suggestion schemes according to the levels of the hierarchical alarm signals. In this way, it is beneficial to improve the scheduling efficiency of medical resources and the success rate of treating critical cases.

[0158] Based on the above embodiments, the disposal suggestion generation module 8 is specifically configured to:

[0159] First, query the pre-stored abnormal type-disposal plan mapping table according to the level of the hierarchical alarm signal to obtain a candidate disposal plan set; wherein, the abnormal type-disposal plan mapping table contains disposal plans and priorities corresponding to different alarm levels;

[0160] Second, select the highest-priority plan from the candidate disposal plan set according to the preset priority rule to generate a clinical disposal suggestion plan.

[0161] Although the present invention has been described in detail with general descriptions and specific embodiments above, on the basis of the present invention, some modifications or improvements can be made, which are obvious to those skilled in the art. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.

Claims

1. An intelligent early warning analysis system for critical illness based on multi-source clinical data fusion, characterized in that, Including: A data acquisition module for continuously acquiring video data of the chest and abdomen movements, ventilator waveform data, and blood oxygen saturation signal data of critically ill patients with respiratory diseases; A respiratory movement analysis module for performing respiratory movement area segmentation processing on the chest and abdomen movement video data to generate a respiratory movement amplitude change curve; A ventilator waveform analysis module for performing dynamic threshold analysis processing on the ventilator waveform data to generate waveform feature data including abnormal time segments; A blood oxygen signal processing module for performing adaptive window processing on the blood oxygen saturation signal data to generate a blood oxygen change trend vector; A multi-modal fusion module for performing multi-modal feature fusion processing on the respiratory movement amplitude change curve, waveform feature data, and blood oxygen change trend vector to generate a multi-dimensional physiological feature vector at the same synchronization timestamp; A risk assessment module for inputting the multi-dimensional physiological feature vector into a risk assessment model and outputting a respiratory compensation status score; An alarm trigger module for triggering a graded alarm signal according to the deviation degree of the respiratory compensation status score from the baseline range.

2. The critical care intelligent early warning analysis system based on multi-source clinical data fusion according to claim 1, wherein The respiratory movement analysis module is specifically used for: Use the pre-trained U-Net network to segment the respiratory motion region from the original video frame sequence {F t}, generating a respiratory region mask {M t}, where F t represents the t-th frame image, and M t is a binary mask; Calculate the pixel displacement Δd of the region based on the difference between adjacent frame masks t : In the formula, represents the exclusive OR operation, and (x, y) are pixel coordinates; M t (x, y) is the pixel value of the t-th frame mask at the coordinate (x, y); For the regional pixel displacement Δd t Perform Butterworth low-pass filtering on the sequence to generate the respiratory motion amplitude change curve A(t): where N is the length of the filtering window, and w k is the k-th coefficient of the Butterworth filter.

3. The intelligent early warning analysis system for critical illness based on multi-source clinical data fusion according to claim 2, characterized in that, The ventilator waveform analysis module is specifically used for: Perform a sliding window standard deviation calculation on the ventilator waveform data {P i}, and output the pressure fluctuation sequence σ w : Where m is the window length; P j is the j-th pressure sampling value within the window; is the average pressure within the window; Based on the historical normal cycle segments of the ventilator waveform data {P i}, calculate the dynamic threshold θ: θ = μ hist + α·σ hist where μ hist is the mean value of historical normal cycle pressure; σ hist is the standard deviation of historical normal cycle pressure; α is the sensitivity coefficient; According to the real-time comparison result between the pressure fluctuation sequence σ w and the dynamic threshold θ, generate waveform feature data including abnormal time segments {T k} and their durations {L k}; where In the formula, and are the start and end times of the k-th abnormal segment.

4. The intelligent early warning analysis system for critical illness based on multi-source clinical data fusion according to claim 3, wherein The blood oxygen signal processing module is specifically used for: Calculate the variance of the blood oxygen saturation signal {S i}, and output the real-time variance Where: n is the initial window length; is the average blood oxygen level within the window; Based on real-time variance Dynamically adjust the window length and output the adaptive window parameter n′: Wherein: is a preset maximum variance threshold; ∈ is a very small constant to prevent division by zero; Perform window normalization on the blood oxygen saturation signal {S i}, and output the standardized signal where: μ n′ is the mean of the current window; σ n′ is the standard deviation of the current window; Perform linear regression analysis on the standardized signal to generate a blood oxygen change trend vector V containing the trend slope β, where where: t i is the i-th timestamp; is the mean of timestamps within the fitting window; is within the fitting window mean value; m is the length of the fitting window; V = [β, σ n′ , μ n′ T ;​ where: σ n′ : the standard deviation of the blood oxygen saturation within the adjusted window; μ n′ : the mean value of the blood oxygen saturation within the adjusted window.

5. The intelligent early warning analysis system for critical illness based on multi-source clinical data fusion according to claim 4, wherein The multi-modal fusion module is specifically used for: Downsample the ventilator waveform data {P i} and the blood oxygen signal {S j}, and output the downsampled sequences {P' k} and {S' k}: where: r: downsampling rate, satisfying that 1 / r is equal to the frame rate f of the chest and abdomen motion video data v ; k: serial number of the downsampling window, integer index; i: The original sequence index of the ventilator waveform data {P i}; j: The original sequence index of the blood oxygen signal {S j}; Align the respiratory movement amplitude change curve A(t), the downsampled sequence {P′ k} and {S′ k} based on the dynamic time warping algorithm, and output the synchronization timestamp τ m : Where m is the index of the synchronization timestamp (1 ≤ m ≤ M); M : The total number of timestamps to be aligned; τ m : the m-th synchronization timestamp; P' m : the mean value of the ventilator waveform data in the m-th time window after downsampling; S' m : the mean value of the blood oxygen signal in the m-th time window after downsampling; According to the synchronization timestamp τ m Extract the set of local extreme points {E m} of A(t) to generate the synchronization timestamp feature: Where: E m : The phase marker point of the synchronous timestamp, used for time alignment of the multi-dimensional physiological feature vector; δ: The tolerance range for time alignment.

6. The intelligent early warning analysis system for critical illness based on multi-source clinical data fusion according to claim 5, characterized in that The steps for generating the risk assessment model include: Screening multi-dimensional physiological feature vectors corresponding to stable respiratory cycles from historical patient data as a training data set; Performing time series modeling training on the training data set using a long short-term memory network to generate an initial risk assessment model; Optimizing the hyperparameters of the initial risk assessment model through cross-validation to generate the pre-trained long short-term memory network risk assessment model.

7. The intelligent early warning analysis system for critical illness based on multi-source clinical data fusion according to claim 6, characterized in that, The steps for determining the baseline range include: Obtain the respiratory compensation score sequence {Y of the patient during the initial monitoring period q}; According to the respiratory compensation score sequence {Y q}, calculate the mean value μ b and the standard deviation σ b : where Y q : the q-th respiratory compensation score; Q: the total number of scores during the initial monitoring period; According to the mean μ b and the standard deviation σ b , determine the baseline range, and the baseline range is [μ b - 3σ b , μ b + 3σ b .

8. The critical care intelligent early warning analysis system based on multi-source clinical data fusion according to claim 7, wherein The alarm trigger module is specifically used for: Calculating the absolute value of the difference between the real-time respiratory compensation status score and the median of the baseline range; When the absolute value of the difference exceeds the first threshold but is less than the second threshold, triggering a first-level alarm signal; When the absolute value of the difference reaches the second threshold but is less than the third threshold, triggering a second-level alarm signal; When the absolute value of the difference exceeds the third threshold, triggering a third-level alarm signal.

9. The intelligent early warning analysis system for critical illness based on multi-source clinical data fusion according to claim 1, characterized in that, It further includes a treatment recommendation generation module for generating a corresponding clinical treatment recommendation plan according to the level of the graded alarm signal.

10. The intelligent early warning analysis system for critical illness based on multi-source clinical data fusion according to claim 9, wherein The treatment recommendation generation module is specifically used for: Querying a pre-stored abnormal type-treatment plan mapping table according to the level of the graded alarm signal to obtain a candidate treatment plan set; wherein, the abnormal type-treatment plan mapping table contains treatment plans and priorities corresponding to different alarm levels; Selecting the highest-priority plan from the candidate treatment plan set according to a preset priority rule to generate the clinical treatment recommendation plan.

Images