Breathing machine intelligent off-line evaluation system based on multiple parameters

Abstract

The invention relates to the technical field of breathing machine off-line evaluation, in particular to a breathing machine intelligent off-line evaluation system based on multiple parameters, which is characterized in that a data acquisition unit acquires breathing and physiological parameters of a patient in real time; the data analysis unit extracts a respiratory rate instantaneous value and a blood oxygen saturation continuous measurement value through preprocessing, inputs the respiratory rate instantaneous value and the blood oxygen saturation continuous measurement value into an intelligent analysis model composed of a deep residual network and a time convolution network, fuses a physiological hysteresis compensation coefficient to generate an offline index, and outputs a preliminary offline risk index; the data processing unit is embedded into a clinical rule base through a decision tree algorithm, generates a risk correction factor through a taboo check factor, allocates differentiation weights according to severity levels, corrects initial risk deviation and generates a comprehensive offline risk index, and the clinical rule base is continuously optimized through online learning, so that the offline risk assessment accuracy is improved; and a reliable basis is provided for clinical decision-making.

Description

Technical Field

[0001] This invention relates to the field of ventilator weaning assessment technology, and more specifically, to a multi-parameter-based intelligent ventilator weaning assessment system. Background Technology

[0002] Ventilator weaning assessment is an important technology, specifically applied to the risk assessment of ventilator-supported patients. Its core principle is to improve the accuracy of weaning risk assessment by integrating respiratory physiological parameters with clinical contraindications, thus meeting the core clinical needs for weaning safety and effectiveness. In ventilator weaning assessments, respiratory rate is prone to sudden abnormal fluctuations, blood oxygen saturation changes exhibit physiological lag characteristics, and individual patients often have diverse clinical contraindications. Because these factors intertwine and affect the patient's weaning tolerance, single-parameter or simple quantitative assessments are insufficient to comprehensively capture the true risk, potentially leading to misjudgments of weaning timing and resulting in weaning failure or safety hazards. To address this technical problem, we provide a multi-parameter-based intelligent ventilator weaning assessment system. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-parameter-based intelligent ventilator weaning assessment system to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, a multi-parameter-based intelligent ventilator weaning assessment system is provided, comprising: The data acquisition unit is used to collect the patient's respiratory parameters in real time through the ventilator's built-in sensors, as well as physiological parameters collected through physiological monitoring equipment; The data analysis unit receives the respiratory parameters and physiological parameters, performs real-time preprocessing and feature extraction on the respiratory parameters and physiological parameters, obtains respiratory rate and blood oxygen saturation, and inputs the respiratory rate and blood oxygen saturation into a preset intelligent analysis model. The intelligent analysis model performs comprehensive quantitative analysis of respiratory rate and blood oxygen saturation based on machine learning pattern recognition algorithm, obtains the weaning index, and outputs a preliminary weaning risk index based on the weaning index. The data processing unit applies a decision tree algorithm to the preliminary offline risk index and integrates the data with a preset clinical rule base to correct risk bias and generate a comprehensive offline risk index.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention synchronously acquires respiratory and physiological parameters through a data acquisition unit. After adaptive noise suppression and motion artifact removal preprocessing, it accurately extracts instantaneous respiratory rate and continuous blood oxygen saturation measurements, laying a high-quality data foundation for assessment. The intelligent analysis model employs a dual-channel architecture: a deep residual network captures abrupt changes in respiratory rate, and a temporal convolutional network tracks continuous trends in blood oxygen. It integrates physiological hysteresis compensation coefficients to generate a weaning index, and combines this with a three-dimensional risk surface to output a preliminary weaning risk index, effectively addressing assessment biases caused by parameter fluctuations and hysteresis characteristics. The data processing unit embeds a decision tree algorithm into a clinical rule base, generating risk correction factors through contraindication verification factors. Differential weights are assigned according to severity levels to correct preliminary risk biases and generate a comprehensive weaning risk index. Furthermore, the clinical rule base is continuously optimized through online learning, achieving synergistic integration of respiratory physiological parameters and clinical contraindications. This improves the accuracy and personalization of weaning risk assessment, providing reliable quantitative evidence for clinical weaning decisions and reducing the risk of weaning failure. Attached Figure Description

[0006] Figure 1 This is an overall block diagram of the present invention.

[0007] The meanings of the labels in the diagram are as follows: 1. Data acquisition unit; 2. Data analysis unit; 3. Data processing unit. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] This invention provides a multi-parameter-based intelligent ventilator weaning assessment system. Please refer to [link / reference]. Figure 1 As shown, it includes: The data acquisition unit 1 is used to collect the patient's respiratory parameters in real time through the built-in sensors of the ventilator and the physiological parameters collected through the physiological monitoring equipment; Data analysis unit 2 receives respiratory parameters and physiological parameters, performs real-time preprocessing and feature extraction on the respiratory parameters and physiological parameters, obtains respiratory rate and blood oxygen saturation, and inputs the respiratory rate and blood oxygen saturation into a preset intelligent analysis model. The intelligent analysis model performs comprehensive quantitative analysis of respiratory rate and blood oxygen saturation based on machine learning pattern recognition algorithm, obtains the weaning index, and outputs a preliminary weaning risk index based on the weaning index. Data processing unit 3 applies a decision tree algorithm to the preliminary offline risk index and integrates the data with a pre-set clinical rule base to correct risk bias and generate a comprehensive offline risk index.

[0010] Real-time preprocessing and feature extraction of respiratory and physiological parameters are performed, specifically including: After filtering respiratory parameters using adaptive noise suppression technology, the peak intervals of the respiratory flow curve are extracted. The instantaneous respiratory frequency value of the peak intervals of the respiratory flow curve per unit time is calculated by time-series integration. Physiological parameters are processed simultaneously. Based on the elimination of motion artifacts of physiological parameters, the rate of change of absorption intensity of characteristic wavelengths of arterial blood oxygen saturation is identified. Continuous blood oxygen saturation measurement values ​​are output through dynamic calibration algorithm.

[0011] The construction process of the pre-defined intelligent analysis model specifically includes: The respiratory rate time series matrix and blood oxygen saturation decay trajectory of historical ventilator-associated cases were collected as training sets. A deep residual network was used to extract the abrupt change pattern features of respiratory rate, and a temporal convolutional network was used to capture the continuous trend features of blood oxygen saturation. A dual-channel input ventilator-associated risk prediction architecture was established through a feature fusion module, and the intelligent analysis model parameters were solidified after adversarial training optimization.

[0012] The working logic of the pre-set intelligent analysis model for respiratory rate and blood oxygen saturation inputs specifically includes: Instantaneous respiratory rate values ​​are constructed as a sliding time window spectrogram, and continuous blood oxygen saturation measurements are converted into a dynamic decay slope sequence. The high-frequency abnormal oscillation components in the sliding time window spectrogram are analyzed through the respiratory rate branch of the offline risk prediction architecture, and the step-down segments in the dynamic decay slope sequence are analyzed through the blood oxygen saturation branch of the offline risk prediction architecture. The outputs of the two branches are weighted and fused through a cross-modal attention mechanism.

[0013] The pre-set intelligent analysis model uses mutation-sensitive convolutional kernels to detect sudden increases in high-frequency abnormal oscillation components in the analysis of respiratory rate. In the analysis of blood oxygen saturation, the pre-set intelligent analysis model deploys trend memory units to track the cumulative duration of step-down segments. When a sudden increase event is detected and the cumulative duration exceeds the coordination threshold, a high-risk state marker is triggered.

[0014] The specific steps of comprehensive quantitative analysis include: The intensity value of the sudden growth event output by the mutation-sensitive convolution kernel is mapped to the respiratory drive hyperactivity index, and the cumulative duration output by the trend memory unit is quantified as the oxygenation deterioration index. The respiratory drive hyperactivity index and the oxygenation deterioration index are fused into a continuous offline index through a nonlinear transformation function. The fusion process introduces the physiological hysteresis compensation coefficient of blood oxygen saturation on respiratory rate.

[0015] For instantaneous respiratory rate values, accelerated exhaustion patterns with variability exceeding the physiological baseline in adjacent respiratory cycles are identified by mutation-sensitive convolutional kernels. For continuous blood oxygen saturation measurements, the compensatory decline inflection point that occurs later than the respiratory rate is captured by trend memory units.

[0016] The specific steps involved in generating a preliminary offline risk index based on the offline index include: The continuous weaning index is input into the risk stratification module, and spatial positioning is performed based on the preset three-dimensional risk surface. The dimensions are the respiratory drive hyperactivity index, oxygenation deterioration index, and physiological lag compensation coefficient. When the continuous weaning index falls into the high-risk area of ​​the three-dimensional risk surface, a high preliminary weaning risk index is output, and the medium and low risk areas correspond to the graded risk index.

[0017] Using the initial offline risk index as the root node and the contraindication verification factor embedded in the clinical rule base as the decision branch, a dynamic decision path is generated through recursive splitting. When the contraindication verification factor triggers the preset rule, a risk correction factor is generated.

[0018] The specific steps for correcting risk bias and generating a comprehensive offline risk index include: The risk correction factor and the initial offline risk index are input into the weighted engine. Based on the severity level of the contraindication verification factor in the dynamic decision path, differentiated weights are assigned to output a comprehensive offline risk index. The clinical rule base updates the weight allocation strategy of the contraindication verification factor through an online learning mechanism.

[0019] It needs further explanation that after the data acquisition unit 1 acquires the patient's respiratory and physiological parameters through the ventilator's built-in sensors and physiological monitoring equipment, the data analysis unit 2 needs to perform real-time preprocessing on these raw data to remove interference, and then extract key feature parameters to provide high-quality input for the subsequent intelligent analysis model. The specific implementation method is as follows: The core logic of real-time preprocessing and feature extraction of respiratory and physiological parameters is to first reduce noise and remove interference, and then accurately extract the core indicators related to weaning assessment. Specifically, this includes first processing the respiratory parameters, filtering them using adaptive noise suppression technology, and then extracting the peak interval of the respiratory flow curve. Respiratory parameters are core data related to patient breathing collected by the ventilator's built-in sensors, including respiratory flow rate, tidal volume, and airway pressure. Respiratory flow rate data is the core basis for calculating respiratory rate. Adaptive noise suppression technology is a noise reduction technique that can adjust filtering parameters in real time according to the dynamic characteristics of the respiratory signal. It can accurately distinguish between effective respiratory signals and irrelevant signals such as equipment operating noise and environmental interference, avoiding signal distortion or incomplete noise reduction caused by traditional fixed filtering parameters. The respiratory flow curve is a continuous waveform curve formed by the change of respiratory flow rate over time. Each peak corresponds to the peak inspiratory time during a single breath. The regularity of the peaks directly reflects the stability of breathing. The peak interval is the time difference between two adjacent respiratory flow peaks, a key intermediate parameter for deriving respiratory rate. The workflow for this part is as follows: First, the raw respiratory flow rate data from the collected respiratory parameters is imported into the adaptive noise suppression module. This module identifies the dominant frequency range of the respiratory flow rate signal through rapid spectrum analysis. The dominant frequency of a normal adult respiratory signal is 0.1-0.5 Hz. Then, an adaptive filtering template matching this dominant frequency is dynamically generated to perform segmented filtering on the raw signal. The effective signal within the dominant frequency range is retained, while noise signals outside the dominant frequency are attenuated and suppressed. After filtering, the peak detection algorithm is started, and a peak identification threshold is set. This threshold is 30% of the maximum amplitude of the filtered signal. It is calibrated through a large amount of clinical data to ensure accurate identification of the true respiratory peaks and avoid misjudging noise peaks. The processed respiratory flow rate curve is traversed, and all peaks that meet the criteria of amplitude exceeding the threshold, amplitude increase at the previous sampling point, and amplitude increase at the next sampling point are marked. The peak points of the decreasing value are identified, and the time coordinates corresponding to each peak point are extracted. The difference between the time coordinates of two adjacent peak points is calculated, which is the peak interval of the respiratory flow curve. Each interval corresponds to a complete respiratory cycle, providing a basis for subsequent respiratory rate calculation. The instantaneous respiratory rate value of the peak interval of the respiratory flow curve per unit time is calculated by time series integration. Time series integration is a processing method that accumulates continuous peak interval data according to the time series. The core is to derive the respiratory rate from the number of respiratory cycles per unit time. The instantaneous respiratory rate value is the number of breaths per unit time corresponding to a specific moment. It is different from the average respiratory rate over a fixed time period and can more accurately capture the dynamic changes in the patient's respiratory status, providing real-time data support for weaning risk assessment. The workflow of this part is as follows: First, a calculation time window is set, ideally 10 seconds based on clinical needs, balancing real-time performance and data stability. This avoids excessive numerical fluctuations due to a window that is too short or response delays due to a window that is too long. Extracted continuous peak interval data are sequentially included in this window. All peak interval data within the window are summed to obtain the total respiratory cycle duration. Then, using time-series integration logic, the instantaneous respiratory rate (NRFR) is obtained by dividing the time window length by the total respiratory cycle duration. For example, if the total respiratory cycle duration within 10 seconds is 8 seconds, the corresponding NRFR is 75 breaths / minute. As time progresses, the time window slides forward in 1-second increments. Each slide removes the earliest peak interval data and incorporates the most recently acquired peak interval data, recalculating the NRFR. This achieves real-time dynamic updates of the respiratory rate and synchronously processes physiological parameters. Based on the removal of motion artifacts in physiological parameters, the absorption of characteristic wavelengths of arterial oxygen saturation is identified. The intensity change rate, synchronous processing refers to processing in parallel with respiratory parameters to ensure the timestamps of the two parameters are aligned, providing time-consistent data for subsequent comprehensive analysis. Physiological parameters are the patient's physiological data collected by physiological monitoring equipment. The core here is the optical signal data used to calculate blood oxygen saturation (obtained by irradiating arterial blood vessels with red and infrared light). Motion artifacts are distortions in the optical signal caused by patient limb movements, changes in body position, or involuntary shaking, manifesting as sudden jumps and irregular fluctuations in signal amplitude, which seriously affect the accuracy of blood oxygen saturation calculation. The characteristic wavelength of arterial blood oxygen saturation is the optical wavelength that specifically reflects the concentration of oxyhemoglobin in the blood. 660 nm red light and 940 nm infrared light are typically selected. The comparison of the absorption intensity of the two wavelengths is the core principle for calculating blood oxygen saturation. The absorption intensity change rate is the proportion of change in the absorption intensity of the characteristic wavelength's optical signal over time. This change rate allows for dynamic tracking of the fluctuation trend of blood oxygen saturation. The implementation process of this part of the technical features is as follows: First, motion artifact removal is performed on the raw optical signal data in the physiological parameters. An adaptive thresholding method combined with trend judgment is used. The baseline amplitude (signal mean) of the optical signal within the past 30 seconds is calculated, and the artifact detection threshold is set to ±50% of the baseline amplitude. When the signal amplitude exceeds this threshold at a certain moment, and the duration is less than 2 seconds with no continuous trend in the signal before and after, it is determined to be a motion artifact. For the identified artifact period, linear interpolation is used to fit and fill it with normal signal data from the three sampling points before and after the artifact, ensuring the continuity and integrity of the signal. After artifact removal, the optical signal data corresponding to 660 nm red light and 940 nm infrared light are extracted. The absorption intensity difference between consecutive sampling points for each characteristic wavelength is calculated, and then this difference is divided by the sampling interval time. The sampling interval of the physiological monitoring device is fixed at 0.1 seconds to obtain the absorption intensity change rate of each characteristic wavelength. Simultaneously, by calculating the ratio of the absorption intensity change rates of the two wavelengths, interference factors such as individual skin thickness and vascular distribution differences are eliminated, ensuring that the change rate can truly reflect the dynamic changes in blood oxygen saturation. The calibration algorithm outputs continuous blood oxygen saturation measurements. The dynamic calibration algorithm is one that adjusts calibration parameters in real time based on the patient's physiological state. It pre-stores correlation models between the characteristic wavelength absorption intensity change rate and blood oxygen saturation for patients of different ages, weights, and underlying diseases, avoiding measurement deviations caused by fixed calibration parameters. The continuous blood oxygen saturation measurements are real-time, continuous values, unlike discrete measurements, and can comprehensively capture subtle changes in the patient's oxygenation status. In practice, the processed absorption intensity change rates of the two characteristic wavelengths are substituted into the dynamic calibration algorithm. The algorithm first selects an appropriate correlation model based on the patient's basic information synchronously collected by the physiological monitoring equipment, then substitutes the ratio of absorption intensity change rates into the model to calculate the instantaneous blood oxygen saturation measurement. Simultaneously, the algorithm monitors the fluctuation range of five consecutive instantaneous measurements in real time. If the fluctuation exceeds 2%, the calibration coefficient is automatically adjusted and recalculated to ensure the accuracy and stability of the output continuous blood oxygen saturation measurements. These, along with the instantaneous respiratory rate, serve as core input parameters for subsequent intelligent analysis models, providing a reliable basis for weaning risk assessment.

[0020] After data analysis unit 2 extracts the instantaneous respiratory rate and continuous blood oxygen saturation measurements, a pre-defined intelligent analysis model needs to be constructed to achieve accurate quantitative analysis of weaning risk. This model is built based on historical clinical data and is gradually improved through feature extraction, architecture construction, and adversarial optimization. The specific implementation method is as follows: The construction process of the pre-set intelligent analysis model specifically includes first collecting the respiratory rate time series matrix and blood oxygen saturation decay trajectory of historical weaning cases as training sets. The respiratory rate time series matrix is ​​a two-dimensional data matrix formed by arranging the continuous instantaneous respiratory rate values ​​of a single historical weaning case in chronological order. The row dimension corresponds to the time step and the column dimension corresponds to the respiratory rate value, which can intuitively present the dynamic changes of respiratory rate over time. The blood oxygen saturation decay trajectory is a curve of the continuous blood oxygen saturation measurement value of a single case changing over time, focusing on the change pattern of blood oxygen saturation in the stage of decline. It is the core data reflecting the deterioration of oxygenation function. The training set is a collection of historical case data containing a large number of labeled weaning results (success or failure), which is used to train the model to learn the correlation between weaning risk and characteristic parameters. The acquisition process for both is as follows: First, sufficient clinical historical data on cases that have been weaned off ventilator are collected. Each case must include a complete sequence of instantaneous respiratory rate values ​​(sampling interval consistent with real-time monitoring, i.e., 1 second / value, with a data duration of no less than 2 hours per case), a sequence of continuous blood oxygen saturation measurements, and the final weaning outcome and key clinical information of the case. The instantaneous respiratory rate values ​​for each case are organized in chronological order, with every 100 consecutive data segments forming a row vector. Multiple row vectors are arranged sequentially to form the respiratory rate time series matrix for that case, ensuring that the dimensions of all case matrices are consistent. Insufficient data is filled with the last value, and excess data is extracted from the latest 100 data segments. For the continuous blood oxygen saturation measurements, a sliding window method (window size set to 30 seconds) is used to calculate the average decline slope within each window. If the slope is negative, it is determined to be a decay segment. All decay segments are connected in chronological order. Trajectories without decay are represented by horizontal straight lines, ultimately forming... The oxygen saturation decay trajectory of this case was analyzed. The respiratory rate time-series matrix, oxygen saturation decay trajectory, and corresponding weaning result labels of all cases were integrated. Cases with a data missing rate exceeding 5% or significant measurement errors were removed, forming a diverse training set covering different ages, underlying diseases, and weaning scenarios. This ensures the model's generalization ability. A deep residual network was used to extract abrupt change pattern features of respiratory rate. Deep residual networks are deep learning models that solve the gradient vanishing problem in deep networks by introducing residual connections. They can effectively mine complex features in the data and adapt to the nonlinear variation of instantaneous respiratory rate values. Abrupt change pattern features refer to sudden and irregular changes in respiratory rate within a short period, such as a sudden increase of more than 20% in respiratory rate or a significant deviation in the amplitude of adjacent respiratory cycles exceeding the normal range. These features often indicate abnormal respiratory drive in patients and are important early warning signals of increased weaning risk. The extraction process is as follows: First, the respiratory rate time series matrix in the training set is normalized, mapping all values ​​to the 0-1 interval to eliminate the influence of differences in the magnitude of data from different cases. The normalized matrix is ​​then input into a pre-defined deep residual network containing 10 residual blocks. Each residual block consists of two convolutional layers and one residual connection. The kernel size is set to 3×3, the stride is 1, and the padding method is set to "same" to ensure that no feature dimensions are lost. The network performs layer-by-layer feature mapping on the time series matrix through forward propagation. The shallow convolutional layers capture the basic trend of respiratory rate changes, while the deep convolutional layers focus on abrupt changes within local time windows. An attention mechanism is introduced in the last three residual blocks to focus on regions where the rate of change of values ​​exceeds a pre-defined threshold (3 times the normal variation amplitude). Weighted operations are used to strengthen the feature representation of abrupt change regions. Finally, the deep features are compressed into a fixed-dimensional feature vector through a global average pooling layer. This vector represents the mutation pattern feature of respiratory rate, which can accurately quantify key information such as the intensity, frequency, and duration of mutations. Combined with a temporal convolutional network to capture the persistent trend feature of blood oxygen saturation, the temporal convolutional network is a deep learning model built on causal convolution and dilated convolution. It is good at capturing long-term dependencies in time series and can effectively extract the long-term trend of blood oxygen saturation changes, avoiding interference from short-term fluctuations. The persistent trend feature refers to the stable change trend of continuous blood oxygen saturation measurements over a long period of time (usually more than 1 minute), such as a slow and continuous decrease, a step-like decrease followed by a low level, etc., which is different from random short-term fluctuations and can truly reflect the continuous state of the patient's oxygenation function. The capture process is as follows: First, the blood oxygen saturation decay trajectory is smoothed and filtered (using a 5-second sliding window moving average method) to eliminate short-term noise caused by slight limb movements. The preprocessed trajectory is then converted into a one-dimensional time series and input into a temporal convolutional network. This network contains 8 causal convolutional layers with a kernel size of 5 and an expansion rate increasing from 1 in multiples of 2 (1, 2, 4, 8...). By expanding the receptive field through convolution, the network captures long-term trends. The network extracts blood oxygen saturation change features at different time scales through layer-by-layer convolution operations. Shallow layers capture short-term decay details, while deep layers integrate long-term trend information. Targeting the decay characteristics of blood oxygen saturation, the network enhances features in regions with negative slopes by calculating continuous time steps. The long slope accumulation value quantifies the duration and decay rate of the trend. Finally, the extracted features are converted into fixed-dimensional feature vectors through a fully connected layer, which are the persistent trend features of blood oxygen saturation. This can comprehensively reflect the degree and speed of deterioration of oxygenation function. A dual-channel input offline risk prediction architecture is established through a feature fusion module. The feature fusion module is a functional module used to integrate feature vectors from different sources. It can explore the synergistic relationship between respiratory rate mutation features and blood oxygen saturation trend features, thereby improving prediction accuracy. The dual-channel input offline risk prediction architecture refers to two independent processing paths that process respiratory rate features and blood oxygen saturation features separately. The output model architecture is then integrated through the fusion module, which can simultaneously take into account the core information of both parameters. The establishment process is as follows: First, two parallel feature processing channels are constructed. The first channel (respiratory rate branch) receives the mutation pattern feature vector output by the deep residual network. It then performs feature dimension adjustment and nonlinear transformation through two fully connected layers to adapt the feature vector to the fusion requirements. The second channel (blood oxygen saturation branch) receives the persistent trend feature vector output by the temporal convolutional network. It uses fully connected layers with the same structure for feature optimization to ensure that the output feature dimensions of the two channels are completely consistent. The output feature vectors from both channels are then input into the feature fusion module. A concatenation + attention-weighted fusion method is used: first, the two feature vectors are concatenated into a higher-dimensional joint feature vector; then, an attention mechanism is used to calculate the weight coefficients of the features from the two channels (in cases of failed weaning, the weight of mutation features is increased if their contribution is high; in cases of deteriorating oxygenation, the weight of trend features is increased). The joint feature vector is then weighted. The enhanced and fused feature vectors are input into the prediction layer, which contains a fully connected layer and a sigmoid activation function. The output is the probability value of offline risk, forming a complete dual-channel input offline risk prediction architecture. This architecture can simultaneously capture sudden abnormalities in respiratory rate and continuous deterioration of blood oxygen saturation, achieving collaborative prediction of multi-dimensional features. The intelligent analysis model parameters are then solidified after adversarial training optimization. Adversarial training is a training method that improves the robustness of the model through mutual adversarial interaction between the generator and the discriminator. The generator is the constructed dual-channel prediction architecture, and the discriminator is an additional binary classification network used to distinguish the model's predicted results from the true offline labels. The intelligent analysis model parameters include the weights and biases of the deep residual network, temporal convolutional network, feature fusion module, and prediction layer. Solidifying the parameters means determining and storing the optimal parameter combination for real-time evaluation. The optimization process is as follows: The dual-channel prediction architecture is used as the generator, while the discriminator adopts a three-layer fully connected layer structure with a softmax activation function. The discriminator's input is the offline risk probability value output by the generator and the true case label. The output is the judgment result of the true feature or the generated feature. During training, the training set features are first input into the generator to obtain the prediction result. The discriminator calculates the judgment loss based on the true label and the prediction result. The parameters of the generator and the discriminator are updated simultaneously through backpropagation. The generator's goal is to minimize the prediction error and the discrimination loss (to make the prediction result closer to the true value and able to fool the discriminator). The discriminator's goal is to maximize the judgment accuracy. After 50 rounds of alternating training, when the generator's training set prediction accuracy exceeds 95%, the validation set accuracy exceeds 90%, and the discriminator's judgment accuracy stabilizes at around 50% (the generator and discriminator reach equilibrium), training is stopped. All parameters of the generator at this time are extracted and stored in the system database for solidification, ultimately forming a preset intelligent analysis model that can accurately process real-time data.

[0021] After data analysis unit 2 extracts the instantaneous values ​​of respiratory rate and continuous measurements of blood oxygen saturation, these two types of real-time parameters need to be input into a preset intelligent analysis model for in-depth analysis. The model achieves accurate feature mining through dual-channel parallel processing and cross-modal fusion. The specific working logic is as follows: The working logic of the preset intelligent analysis model for inputting respiratory rate and blood oxygen saturation specifically includes: firstly, constructing a sliding time window spectrogram from the instantaneous respiratory rate value. The sliding time window is a data analysis window of fixed length that moves along the time axis according to a preset step size, used to extract continuous respiratory rate data segments, balancing time resolution and frequency resolution. The spectrogram is a two-dimensional graph formed by converting the time-domain respiratory rate data into frequency-domain data through Fourier transform, with the horizontal axis representing frequency and the vertical axis representing amplitude, which can intuitively present the energy distribution of different frequency components. The instantaneous respiratory rate value is the number of breaths per second, reflecting the patient's real-time respiratory status. The construction process is as follows: First, the sliding time window length is set to 20 seconds (clinically validated, this length captures short-term mutations while avoiding data fragmentation), with a step size of 5 seconds (ensuring data overlap and avoiding the omission of key features). Continuous instantaneous respiratory rate values ​​are sequentially included in the window, with each window containing four consecutive instantaneous values ​​(since instantaneous values ​​update every second, 20 seconds correspond to 20 data points). The respiratory rate data within each window is preprocessed, using a Hanning window function to reduce spectral leakage, followed by a Fast Fourier Transform to convert the time-domain data to the frequency-domain data, obtaining the frequency-amplitude distribution within that window. The frequency-domain data from all windows are then arranged chronologically. The graph is plotted as time (window start time) and frequency, with amplitude on the vertical axis, forming a two-dimensional sliding time window spectrum. This clearly shows the energy changes of respiratory rate at different times and frequency segments, providing a basis for subsequent high-frequency abnormality detection. Continuous blood oxygen saturation measurements are converted into a dynamic decay slope sequence. The continuous blood oxygen saturation measurements are updated every second, continuously reflecting the patient's oxygenation status. The dynamic decay slope sequence is a set of average decrease slopes of blood oxygen saturation within each time period, arranged in chronological order. A negative slope indicates a decrease (decline) in blood oxygen saturation, while a positive or zero slope indicates no decay. This accurately captures the dynamic trend of blood oxygen saturation changes. The conversion process is as follows: A 30-second sliding window (adapting to the physiological delay characteristics of changes in blood oxygen saturation) with a 10-second step size was used to segment continuous blood oxygen saturation measurements. Linear fitting was performed on all measurements within each window, and the slope of the fitted line was calculated (slope = (end value - initial value) / window length). If the slope ≥ 0, it indicates no decrease in blood oxygen saturation within that window, and the slope value is recorded as 0. If the slope < 0, it indicates decrease, and the original slope value is retained. All calculated slope values ​​were arranged sequentially according to the window sliding order to form a dynamic decay slope sequence. Continuous negative slope segments in the sequence correspond to the continuous decay process of blood oxygen saturation, serving as the basis for subsequent stepwise descent. The analysis provides data support, analyzing high-frequency abnormal oscillation components in the sliding time window spectrogram through the respiratory rate branch of the offline risk prediction architecture. The respiratory rate branch of the offline risk prediction architecture is an independent channel in the model specifically for processing respiratory rate features, composed of convolutional layers, pooling layers, and attention layers, adapting to the feature extraction needs of frequency domain data. High-frequency abnormal oscillation components refer to oscillating signals in the spectrogram with frequencies higher than 1 Hz (exceeding the frequency range of normal respiratory rate; the dominant frequency corresponding to normal adult respiratory rate is usually 0.1-0.5 Hz) and amplitudes significantly higher than background noise. These components are often caused by respiratory muscle fatigue, hyperactive respiratory drive, etc., and are important signals of abnormal respiratory function. The analysis process is as follows: The sliding time window spectrogram is input into the respiratory frequency branch. The first convolutional layer of this branch uses a 1×3 kernel with a stride of 1 to capture local high-frequency features along the frequency axis. After enhancing the nonlinear expression with a ReLU activation function, it is input into a max pooling layer (2×2 kernel with a stride of 2) to compress the data dimensionality while retaining key features. Subsequent layers of stacked convolution-pooling modules progressively deepen feature extraction. Shallow layers focus on basic frequency components, while deeper layers emphasize the feature expression of high-frequency regions. A channel attention mechanism is introduced after the last convolutional layer to calculate the feature weights of each frequency band. Regions with frequencies above 1 Hz and amplitudes exceeding three times the background noise are given higher weights. High-frequency anomalous signals are highlighted through weighted operations. Finally, global average pooling is used to... Multidimensional frequency domain features are compressed into fixed-dimensional feature vectors, enabling accurate extraction and quantification of high-frequency abnormal oscillation components. The blood oxygen saturation branch of the offline risk prediction architecture analyzes the step-decreasing segments in the dynamic decay slope sequence. This branch, an independent channel for processing blood oxygen saturation features, consists of temporal convolutional layers and gating mechanisms, excelling at capturing long-term dependencies and segmented features in time series. The step-decreasing segment refers to a segmented decay pattern in the dynamic decay slope sequence where the slope of multiple consecutive windows is negative (decreasing), and the absolute value change of the slope between adjacent windows does not exceed 50%, exhibiting a decreasing-stabilizing-decreasing pattern. These segments reflect a continuous deterioration of blood oxygen saturation without compensatory recovery, a key indicator of impaired oxygenation function. The analysis process is as follows: The dynamic decay slope sequence is input into the blood oxygen saturation branch. The first temporal convolutional layer of the branch uses a causal convolution with a kernel size of 5 and an expansion rate of 1 to ensure that feature extraction does not depend on future data. A gated linear unit (GLU) filters out invalid information, retaining only decay-related features. Subsequent temporal convolutional layers with increasing expansion rates (2, 4, 8) expand the receptive field to capture long-term slope change patterns. A sequence attention mechanism is introduced to focus on windows with consecutive negative slopes. Parameters such as the mean slope and the number of sustained windows within the segment are calculated. If the number of consecutive negative slope windows is ≥3 (corresponding to continuous decay for more than 30 seconds) and the absolute change in slope between adjacent windows is ≤50%, it is determined to be a step-decreasing segment. The feature parameters of the segment are then processed through a fully connected layer. The data is converted into fixed-dimensional feature vectors to quantify the step-down segment. The outputs of the two branches are weighted and fused using a cross-modal attention mechanism. The outputs of the two branches are the high-frequency abnormal oscillation feature vector from the respiratory rate branch and the step-down segment feature vector from the blood oxygen saturation branch, reflecting the weaning risk from the two dimensions of respiratory drive and oxygenation function, respectively. The cross-modal attention mechanism is a fusion technique that can uncover the correlation between different modalities of data (respiratory rate, blood oxygen saturation). By calculating the correlation between the two features and assigning weights, the contribution of key features is strengthened, avoiding misjudgments caused by single-modal bias. The weighted fusion is to superimpose the two feature vectors according to the weight ratio to form a comprehensive feature vector that takes into account the core information of both parameters. The fusion process is as follows: First, the feature vectors output from the two branches are normalized, mapping the values ​​to the 0-1 range to eliminate the difference in data magnitude between modalities. A cross-modal attention weight calculation module is then constructed. The two feature vectors are input into this module, and the similarity between the vectors is calculated through dot product operation. The higher the similarity, the stronger the correlation between the two features. Weights are assigned according to the similarity. When the similarity is high, the weights of both features are increased, while when the similarity is low, the modality with the higher contribution is highlighted. After the weights are assigned, the two feature vectors are summed element-wise according to their corresponding weights to generate a comprehensive feature vector with unified dimensions. This vector fully integrates the key features of respiratory drive abnormalities and oxygenation function deterioration, providing high-quality feature input for subsequent calculation of the offline index.

[0022] After the intelligent analysis model completes the feature extraction of high-frequency abnormal oscillation components and step-down segments through a dual-branch approach, it is still necessary to further accurately detect key risk events and track core indicators to provide targeted basis for subsequent comprehensive quantitative analysis. The specific implementation method is as follows: The pre-defined intelligent analysis model uses a mutation-sensitive convolutional kernel to detect sudden increases in high-frequency abnormal oscillation components during respiratory rate analysis. This mutation-sensitive convolutional kernel is optimized for mutation signals, exhibiting an asymmetrical weight distribution with high weights in the center and low weights on both sides. It strongly responds to sudden changes in the signal while suppressing invalid signals in stable regions, thus meeting the detection requirements for mutation features in high-frequency abnormal oscillation components. High-frequency abnormal oscillation components are oscillating signals with frequencies higher than 1 Hz and amplitudes significantly exceeding background noise in the sliding time window spectrum, reflecting irregular fluctuations in respiratory rate. A sudden increase in the amplitude of the high-frequency abnormal oscillation component refers to a rapid increase in amplitude within a short period (usually no more than 3 seconds), specifically an increase of more than 50% from the baseline level, lasting for at least 2 seconds. These events are often caused by respiratory muscle fatigue, a sudden increase in airway resistance, etc., indicating abnormal hyperactivity of the patient's respiratory drive and serving as a direct warning signal of increased weaning risk. The detection process follows a closed-loop workflow of preprocessing, convolution enhancement, threshold determination, and event confirmation. The specific workflow is as follows: First, the high-frequency abnormal oscillation component feature vector output from the respiratory rate branch is standardized to map the values ​​to a uniform interval, eliminating the influence of individual differences in data from different patients. Then, the standardized feature vector is input into a detection layer equipped with a mutation-sensitive convolutional kernel. This kernel is set to a size of 3×3, with core weights concentrated in the center and adjacent regions. Higher convolutional weights are assigned to mutation segments with high numerical change rates in the feature vector. Through convolution, the feature amplitude of the mutation region is amplified while suppressing the feature response of the stable region, thus enhancing the extraction of the mutation signal. After convolution, a ReLU activation function is used to filter negative response values, retaining only valid mutation features. Finally, dual judgment thresholds are set, including an amplitude increase threshold and a duration threshold. The amplitude increase threshold is set at 50% of the baseline amplitude of the patient's high-frequency abnormal oscillation component (the baseline amplitude is calculated from the average of the data from the previous 5 minutes), and the duration threshold is set to 2 seconds. The system monitors the activated feature vector in real time. When the feature amplitude at a certain moment exceeds the amplitude increase threshold, and the duration of this state reaches the duration threshold, a sudden increase event is detected. The system records the start time, peak amplitude, and duration of the event to provide detailed parameters for subsequent index calculations, ensuring the accuracy and traceability of the detection results. The preset intelligent analysis model deploys trend memory units to track the cumulative duration of stepwise decline segments in the analysis of blood oxygen saturation. The trend memory unit is a network designed based on a gated recurrent mechanism. The unit effectively stores trend information in time-series data, filters short-term fluctuations, and accurately accumulates the duration of target segments. A step-down segment is a series of consecutive negative slope segments in a dynamic decay slope sequence, characterized by a continuous decline in blood oxygen saturation without significant recovery. Accumulated duration refers to the total duration of a single step-down segment from start to finish, reflecting the degree of continued deterioration in blood oxygen saturation. The longer the accumulated duration, the more severe the impairment of oxygenation function. During tracking, the trend memory unit first receives the feature vector of the step-down segment output from the blood oxygen saturation branch. It then uses an update gate to determine whether the current feature belongs to a continuous decay state. If the slope value in the feature vector is continuously negative and the change is within a reasonable range, the update gate opens, and the current feature is updated. The time step (10 seconds) is added to the cumulative duration. If the slope value becomes non-negative or the change exceeds the judgment range of the step-decline segment, the update gate closes, the cumulative duration stops increasing and the current value is temporarily stored. At the same time, the tracking state is reset to prepare to capture the next step-decline segment. Throughout the tracking process, the unit filters out small changes in the slope caused by short-term fluctuations through the forget gate, ensuring that the cumulative duration only counts the true continuous decay time and avoids false accumulation. When a sudden growth event is detected and the cumulative duration exceeds the coordination threshold, a high-risk state is triggered. The coordination threshold is a preset dual threshold that comprehensively considers the correlation between respiratory drive abnormalities and oxygenation function deterioration, including the amplitude threshold of the sudden growth event and the cumulative duration threshold of the step-decline segment.Both conditions must be met simultaneously or exceed their respective thresholds to be considered high-risk. These thresholds are calibrated using extensive historical data on weaning failure cases to ensure the specificity and sensitivity of the warning. The high-risk status marker is a risk warning indicator output by the system, used to alert clinicians that the patient is at high risk of weaning failure and requires postponement of weaning or enhanced monitoring. Specifically, the system first retrieves preset synergy thresholds, comparing the peak amplitude of the detected sudden increase event with the event amplitude threshold, and simultaneously comparing the tracked cumulative duration with the duration threshold. If the peak amplitude of the sudden increase event exceeds the event amplitude threshold, and the cumulative duration of the step-down segment exceeds the duration threshold, or if either significantly exceeds the threshold, the synergy threshold condition is met. The system immediately triggers the high-risk status marker, associating and storing the marker information with the corresponding event parameters (duration of the sudden increase event and the specific value of the cumulative duration). This provides crucial risk information for subsequent weaning index calculation and risk assessment, ensuring that clinicians can quickly detect the synergistic deterioration signal of the patient's respiratory and oxygenation functions and take timely intervention measures.

[0023] After the intelligent analysis model detects a sudden increase in high-frequency abnormal oscillation components and tracks the cumulative duration of a stepwise decrease in blood oxygen saturation, these discrete risk signals need to be transformed into unified quantitative indicators through comprehensive quantitative analysis. This provides the core basis for the output of the preliminary offline risk index. The specific implementation method is as follows: The comprehensive quantitative analysis specifically involves mapping the intensity value of the sudden increase event output by the mutation-sensitive convolutional kernel to a respiratory drive hyperactivity index. The intensity value of the sudden increase event is the characteristic amplitude output by the mutation-sensitive convolutional kernel after detecting the sudden increase event. This value is positively correlated with the severity of the event; the larger the amplitude, the more significant the change in respiratory rate. The respiratory drive hyperactivity index is the core parameter for quantifying the degree of abnormal increase in the intensity of respiratory muscle work in patients. The value ranges from 0 to 10. The higher the score, the more hyperactive the respiratory drive, reflecting that the patient consumes more energy to maintain breathing and the higher the risk of respiratory muscle fatigue. Its core function is to transform the sudden increase signal of respiratory rate into an intuitive and quantifiable risk indicator, providing a basis for respiratory function assessment for weaning assessment. The mapping process is as follows: First, a baseline range for intensity values ​​is established. This range is calibrated using historical offline case data, specifically the intensity range of high-frequency abnormal oscillation components under normal breathing conditions (denoted as the baseline intensity range). The intensity value output by the mutation-sensitive convolution kernel is compared with the upper limit of the baseline intensity range, and the excess ratio is calculated (excess ratio = (actual intensity value - baseline upper limit value) / baseline upper limit value). A segmented mapping rule is used: if the excess ratio ≤ 0, it indicates no sudden increase event or the intensity is not exceeded, and the mapping score is 0; if 0 < excess ratio ≤ 50%, the mapping score is 3; if 50% < excess ratio ≤ 100%, the mapping score is 6; if the excess ratio > 100%, the mapping score is 10. Additionally, if two or more sudden increase events are detected consecutively within the same minute... An additional 2 points (maximum 10 points) are added to the current mapped score to ensure the cumulative risk effect of sustained hyperactivity is reflected. The final score is the respiratory drive hyperactivity index. The cumulative duration output by the trend memory unit is quantified as the oxygenation deterioration index. The cumulative duration is the total duration of the stepwise decline segment tracked by the trend memory unit, directly reflecting the duration of continuous deterioration of blood oxygen saturation. The oxygenation deterioration index is the core parameter for quantifying the degree and duration of impaired oxygenation function in patients. The value range is 0-10 points. The higher the score, the more severe the decline in blood oxygen saturation, the longer the duration, and the worse the oxygenation function. Its function is to convert the trend signal of blood oxygen saturation into a quantitative indicator, supplementing the weaning risk assessment dimension from the perspective of oxygenation function. The quantification process is as follows: First, a baseline threshold for cumulative duration is set. Based on clinical data, 30 seconds is defined as the critical duration for oxygenation deterioration (cumulative duration ≥ 30 seconds is considered clinically significant oxygenation deterioration). A segmented quantification rule is used: if cumulative duration < 30 seconds, it is quantified as 0 points; if 30 seconds ≤ cumulative duration < 60 seconds, it is quantified as 2 points; if 60 seconds ≤ cumulative duration < 120 seconds, it is quantified as 5 points; if 120 seconds ≤ cumulative duration < 180 seconds, it is quantified as 8 points; and if cumulative duration ≥ 180 seconds, it is quantified as 10 points. Simultaneously, considering the average decline slope in the dynamic decay slope sequence, if the absolute value of the average slope is ≥ 0.5% / second (indicating rapid decay), an additional 1 point is added to the current quantified score (maximum 10 points). By reinforcing the risk weight of rapid deterioration, the final score is the oxygenation deterioration index. A nonlinear transformation function is used to fuse the respiratory drive hyperactivity index and the oxygenation deterioration index into a continuous weaning index. This nonlinear transformation function is a specially designed mathematical function for fusing two-dimensional indices, effectively handling the nonlinear correlation between respiratory drive and oxygenation function, and avoiding risk misjudgment caused by simple linear superposition. The continuous weaning index is a unified quantitative indicator that comprehensively reflects the patient's weaning risk, with a value range of 0-100. A higher score indicates a higher weaning risk. Its core advantage is overcoming the limitations of discrete indicators, accurately capturing subtle changes in risk, and providing continuous and accurate data support for subsequent risk stratification. The fusion process is as follows: First, the respiratory drive hyperactivity index and the oxygenation deterioration index were normalized, mapping both to the 0-1 range (normalized value = original score / 10). Then, a physiological lag compensation coefficient for blood oxygen saturation on respiratory rate was introduced. This coefficient corrects for the lag between changes in blood oxygen saturation and changes in respiratory rate caused by physiological mechanisms. Since the deterioration of oxygenation function after respiratory dysfunction often has a 1-3 minute delay, direct fusion could lead to biased risk assessment. The coefficient ranged from 0.8 to 1.2. Based on historical data, when the occurrence time of a respiratory drive hyperactivity event was ≤1 minute, the coefficient was 0.8 (weakening the weight of the oxygenation index to offset the delay); when the occurrence time was >1 minute and ≤3 minutes, the coefficient was 1.0 (weighting balanced); and when the occurrence time was >3 minutes, the coefficient was 1. 2. (Strengthening the weighting of the oxygenation index to reflect the true deterioration after delay): During fusion, the normalized value of the respiratory drive hyperactivity index is first multiplied by 0.6 (the basic weight of respiratory drive), and the normalized value of the oxygenation deterioration index is multiplied by the physiological lag compensation coefficient and then by 0.4 (the dynamic weight of oxygenation function) to obtain two weighted values. The two weighted values ​​are added together and then input into a nonlinear transformation function. This function uses the characteristics of an S-curve to enhance the sensitivity of the middle score segment (normalized sum of 0.3-0.7) and appropriately compress the score segments at both ends (<0.3 or >0.7) to avoid an excessively high proportion of extreme values. Finally, the function output is multiplied by 100 to convert it into a continuous weaning index of 0-100 points, realizing a comprehensive quantification of the dual risks of respiration and oxygenation, and ensuring that the index can comprehensively and accurately reflect the patient's current weaning tolerance.

[0024] While the intelligent analysis model completes comprehensive quantitative analysis and obtains continuous disconnection index, in order to further accurately capture the deterioration trend of respiratory and oxygenation functions and avoid missing key risk signals due to a single quantitative indicator, it is also necessary to conduct in-depth feature mining on the instantaneous value of respiratory rate and the continuous measurement value of blood oxygen saturation. The specific implementation method is as follows: For instantaneous respiratory rate values, a mutation-sensitive convolutional kernel is used to identify accelerated exhaustion patterns where the variability between adjacent respiratory cycles exceeds the physiological baseline. The variability between adjacent respiratory cycles refers to the ratio of the duration difference between two consecutive respiratory cycles to the duration of the previous respiratory cycle. The duration of a respiratory cycle directly corresponds to the peak interval of the respiratory flow curve (each peak interval is the duration of a complete respiratory cycle). This variability can intuitively reflect the stability of the respiratory rhythm; the larger the variability, the more irregular the breathing. The specific process for identifying this variability is as follows: First, extract the peak interval duration of each respiratory cycle from the pre-processed respiratory flow curve, and record them sequentially as cycle 1, cycle 2, cycle 3... cycle n. Then, calculate the duration difference between adjacent cycles (the duration of the later cycle minus the duration of the previous cycle). Divide this difference by the duration of the previous cycle to obtain the variation amplitude of each adjacent combination (if the difference is negative, the variation amplitude is positive, reflecting only the magnitude of the difference, not the direction of increase or decrease), forming a continuous variation amplitude sequence. The physiological baseline refers to the value obtained by adding twice the standard deviation to the mean of the variation amplitude of all adjacent respiratory cycles within this time period, when the patient's respiratory status is stable (no sudden increase event, respiratory rate within the normal range), 5 minutes before the start of the weaning assessment. The numerical value serves as a critical standard for judging whether the amplitude of variation is abnormal. Exceeding this threshold indicates that the respiratory rhythm deviates from a stable state. The accelerated failure pattern refers to a characteristic pattern of rapid deterioration of respiratory function, specifically defined as the amplitude of variation exceeding the physiological baseline for three or more consecutive adjacent respiratory cycles, with the amplitude of variation showing a progressively increasing trend (each subsequent amplitude being larger than the previous one). This pattern indicates that respiratory muscle fatigue is continuously worsening, the ability to control the respiratory rhythm is rapidly declining, and the risk of weaning from ventilator is drastically increasing. During the identification process, the mutation-sensitive convolutional kernel, based on the previous detection of sudden growth events, further performs convolution operations on the amplitude of variation sequence, and strengthens the feature expression of continuous abnormal variations through weight allocation. When three or more consecutive amplitudes exceeding the baseline are detected and showing an increasing trend, the pattern is identified. When a trend is identified, it indicates an accelerated failure pattern. The start time and cumulative duration of this pattern are recorded, providing crucial information on the rate of respiratory function deterioration for weaning risk assessment. For continuous blood oxygen saturation measurements, a trend memory unit captures the compensatory decline inflection point, which occurs later than the respiratory rate. This compensatory decline inflection point refers to the point at which blood oxygen saturation, after a 1-3 minute physiological compensation period (the body maintains oxygenation stability through regulation) following an abnormal respiratory rate (such as an accelerated failure pattern or a sudden increase in respiratory rate), suddenly shifts from a relatively stable state to a sustained decline. This inflection point signifies the failure of the compensatory mechanism for oxygenation function, officially entering the deterioration stage, and is the core early warning signal for deteriorating oxygenation function, occurring later than the respiratory rate. This is determined by physiological mechanisms. Abnormal respiratory function first affects ventilation efficiency, and then gradually leads to a decrease in blood oxygen saturation. This delay time is usually 1-3 minutes. The trend memory unit dynamically adjusts the monitoring window according to the time when the respiratory rate abnormality occurs. During the capture process, the trend memory unit first associates the onset time of the respiratory rate abnormality (such as the onset time of accelerated failure mode), and sets a 1-3 minute delay monitoring window based on this. Within this window, it continuously tracks the dynamic decay slope sequence of continuous blood oxygen saturation measurements (the slope of the previously constructed 30-second sliding window). Within the delay monitoring window, if the slope is always non-negative or occasionally negative but not continuous, it indicates that the oxygenation function is still within the compensatory range. When it detects that the slope of a certain sliding window changes from non-negative to negative,Furthermore, if the slope of the subsequent two consecutive sliding windows remains negative and the absolute value of the slope gradually increases (i.e., the rate of decline accelerates), then the starting moment of the first window with a negative slope is the compensatory decline inflection point. The trend memory unit locks the time coordinate and slope value corresponding to this inflection point through a gating mechanism, simultaneously stopping the current monitoring window and instead tracking the downward trend after the inflection point. This ensures accurate capture of the critical turning point from compensation to decompensation of oxygenation function, providing core data on the starting point of oxygenation function deterioration for the generation of the comprehensive weaning risk index.

[0025] After obtaining the continuous offline index through comprehensive quantitative analysis, data analysis unit 2 needs to transform this continuous quantitative indicator into a clinically directly referable graded risk result. Therefore, precise risk positioning is achieved through the risk stratification module, and the specific implementation method is as follows: The process of generating a preliminary weaning risk index based on the weaning index involves first inputting the continuous weaning index into the risk stratification module. The continuous weaning index is a 0-100 point continuous quantitative indicator derived from the combined respiratory drive hyperactivity index, oxygenation deterioration index, and physiological lag compensation coefficient. A higher score indicates a higher weaning risk, but its range is broad, requiring further stratification to clarify the risk level. The risk stratification module is specifically designed to convert the continuous index into a graded risk level. It has a built-in preset three-dimensional risk surface model and region determination logic. Its core function is to output an intuitive and standardized preliminary weaning risk index through multi-dimensional collaborative positioning, facilitating rapid clinical assessment. During the input process, the risk stratification module simultaneously receives the continuous weaning index and its three corresponding core dimension parameters (respiratory drive hyperactivity index, oxygenation deterioration index, and physiological lag compensation coefficient). The hysteresis compensation coefficient ensures that the positioning process considers both comprehensive indices and single-dimensional features, avoiding one-sided judgments. Spatial positioning is performed based on a preset three-dimensional risk surface. This three-dimensional risk surface is a three-dimensional spatial model constructed based on historical weaning case data. The respiratory drive hyperactivity index (0-10 points), oxygenation deterioration index (0-10 points), and physiological hysteresis compensation coefficient (0.8-1.2) are used as the X, Y, and Z axes, respectively. It is a continuous surface formed by fitting a large number of case data with labeled weaning results. Each spatial point on the surface corresponds to a clear weaning risk level. The rise and fall of the surface intuitively reflects the risk differences under different parameter combinations. Spatial positioning is the process of finding the corresponding spatial position of the current patient in the three-dimensional risk surface using the three-dimensional parameters as coordinate values, and then determining the risk area to which the patient belongs. The specific positioning process is as follows: First, the three-dimensional parameters are normalized. The respiratory drive hyperactivity index and oxygenation deterioration index (0-10 points) are mapped to the 0-1 interval, and the physiological hysteresis compensation coefficient (0.8-1.2) is also linearly mapped to the 0-1 interval to eliminate the impact of differences in the magnitude of different-dimensional parameters on the positioning accuracy. Then, the normalized X, Y, and Z coordinate values ​​are input into the three-dimensional risk surface model. The model quickly locates the surface position corresponding to the coordinates through an interpolation algorithm, and at the same time, it retrieves the preset risk area label (high risk, medium risk, low risk) for that position. During the positioning process, a continuous offline index is also introduced for cross-validation. If the risk area located by the coordinates matches the risk interval corresponding to the continuous offline index (e.g., 70 points or above corresponds to high risk, 40-70 points corresponds to medium risk), the system will determine the risk level. If the scores (below 40 points correspond to low risk) are consistent, the location result is confirmed as valid. If a deviation occurs (e.g., the coordinate location is medium risk but the continuous index reaches 80 points), the location result is corrected based on the risk range corresponding to the continuous weaning index to ensure the accuracy of risk assessment. The dimensions are the respiratory drive hyperactivity index, the oxygenation deterioration index, and the physiological lag compensation coefficient. These three dimensions comprehensively characterize weaning risk from different perspectives. The respiratory drive hyperactivity index reflects the immediate degree of abnormality in respiratory function, the oxygenation deterioration index reflects the state of impaired oxygenation function, and the physiological lag compensation coefficient corrects parameter delay deviations caused by physiological mechanisms. The three dimensions work together to form a complete dimensional system for risk assessment, ensuring that the three-dimensional risk surface can comprehensively cover the risk characteristics under different weaning scenarios and avoid single-dimensional biases. To mitigate omissions or misjudgments caused by one-dimensional assessment, a high initial weaning risk index is output when the continuous weaning index falls into the high-risk area of ​​the three-dimensional risk surface. The medium- and low-risk areas correspond to graded risk indices. The high-risk area is defined as the spatial region on the three-dimensional risk surface with a respiratory drive hyperactivity index ≥7, an oxygenation deterioration index ≥6, a physiological lag compensation coefficient ≥1.1, and a continuous weaning index ≥70. Cases in this area have a weaning failure rate exceeding 80%, and the output high initial weaning risk index is set at 8-10 (10 being the highest risk). The medium-risk area is defined as the region with a respiratory drive hyperactivity index of 3-6, an oxygenation deterioration index of 2-5, a physiological lag compensation coefficient of 0.9-1.0, and a continuous weaning index of 40-69, corresponding to a weaning failure rate of 3... For the 0%-80% risk range, a preliminary weaning risk index of 4-7 is output. The low-risk area is defined as an area with a respiratory drive hyperactivity index ≤2, an oxygenation deterioration index ≤1, a physiological lag compensation coefficient ≤0.89, and a continuous weaning index ≤39, with a weaning failure rate of less than 30%. In this area, a low preliminary weaning risk index of 0-3 is output. Specifically, the risk stratification module retrieves the corresponding graded score range based on the risk area confirmed by spatial positioning, and then refines the score by combining the specific value of the continuous weaning index within that area (e.g., in a high-risk area, a continuous index of 90 points outputs 10 points, and 75 points outputs 8 points). This results in a precise preliminary weaning risk index, providing standardized and quantifiable basic data for the risk deviation correction of the subsequent data processing unit 3.

[0026] After the data analysis unit 2 outputs the preliminary weaning risk index, this index is generated only based on the quantitative analysis of respiratory rate and blood oxygen saturation, without fully considering key influencing factors such as the patient's clinical contraindications. Therefore, the data processing unit 3 needs to integrate the data using a decision tree algorithm combined with a preset clinical rule base to correct potential risk biases. The specific implementation method is as follows: Using the preliminary weaning risk index as the root node, which is a graded risk index from 0 to 10, reflecting the weaning risk at the respiratory and oxygenation function level, the root node is the starting node of the decision tree algorithm, serving as the starting point of the entire decision logic. All subsequent decision branches revolve around this node. Its core function is to first establish the basic risk level, and then make secondary corrections based on clinical contraindication factors. In specific implementation, data processing unit 3 first receives the preliminary weaning risk index and divides it into three basic branches according to the score: low risk (0-3 points), intermediate risk (4-7 points), and high risk (8-10 points). This lays the logical foundation for embedding contraindication verification factors. Contraindication verification factors embedded in the clinical rule base serve as decision branches. Contraindication verification factors refer to various clinical factors that have been identified in clinical practice and may lead to weaning failure or increase the risk of weaning. Specifically, these include the patient's underlying diseases, abnormal immediate physiological indicators, and acute complications. Each factor includes specific judgment criteria and severity levels (mild, moderate, severe). The acquisition process is as follows: systematic clinical... The rule base is built upon the latest clinical guidelines, expert consensus, and massive historical data on weaning failures. Through retrospective analysis of common factors in weaning failure cases, 20 core contraindications significantly impacting weaning outcomes were identified. These factors, after clinical expert validation, were quantified into verifiable factors. Each factor is associated with specific detection indicators and pre-stored in the clinical rule base, allowing for automatic verification based on real-time patient clinical data. Decision branches are the judgment paths extending from the root node in the decision tree. Each contraindication verification factor corresponds to one or more branches, used to superimpose the impact of clinical contraindications on the baseline risk level. Dynamic decision paths are generated through recursive splitting, the core logic of the decision tree algorithm. Starting from the root node, each contraindication verification factor is used as a judgment condition, progressively subdividing the current risk group to form continuous decision nodes. Dynamic decision paths are personalized decision logic chains generated based on specific patient clinical data. Different patients will form different paths due to different contraindication verification results, ensuring targeted decision-making. The generation process is as follows: First, the initial weaning risk index is divided into three basic branches (low, medium, and high) as primary branches. For each primary branch, the highest-priority contraindication verification factor is embedded as a secondary branch. If the patient's verification result is positive, it is further split into tertiary branches according to the severity of the factor. Then, in each tertiary branch, the next-highest priority contraindication verification factor is embedded, and it is again split into quaternary branches according to the verification result and severity. This process continues recursively, splitting the contraindication verification factors layer by layer according to their clinical priority, until all core contraindication verification factors are embedded. This ultimately forms a dynamic decision-making path with multiple nodes covering various clinical scenarios. When a contraindication verification factor triggers a preset rule, a risk correction factor is generated. The preset rule is a trigger condition set for each contraindication verification factor in the clinical rule base. It is formulated by clinical experts based on evidence-based medicine and clarifies which verification results will affect the weaning risk and the extent of the impact. For example, the preset rule for acute pulmonary infection is a white blood cell count > 10 × 10⁻⁶. 9 If a chest X-ray shows newly appearing inflammatory infiltrates, it is considered a trigger. The risk correction is +3 points for 'severe', +1.5 points for 'moderate', and +0.5 points for 'mild'. Each contraindication verification factor corresponds to a specific set of triggering criteria and correction rules. The risk correction factor is a quantitative value used to adjust the initial weaning risk index; it can be positive (increased risk) or 0 (no impact). Its core function is to convert the impact of clinical contraindications into a quantifiable correction value, correcting the deviation of the initial risk index. The specific generation process is as follows: In each branch of the dynamic decision-making path, the system automatically retrieves the patient's real-time clinical data and compares it with the preset rules for contraindication verification factors of the current branch. If the patient's data meets the triggering conditions of the preset rules, the corresponding correction amount is retrieved from the clinical rule base according to the severity of the factor to generate a risk correction factor. If the triggering conditions are not met, the risk correction factor is 0. For example, if the patient's initial weaning risk index is 3 (low risk), and the preset rule for severe acute pulmonary infection is triggered in the decision-making path, the corresponding risk correction factor is +3. If the moderate risk index is triggered simultaneously, the risk correction factor is 0. The preset rule for dissociative disorders corresponds to a risk correction factor of +1.5 points. Therefore, the patient's cumulative risk correction factor is +4.5 points, providing a key basis for the subsequent generation of a comprehensive weaning risk index. The entire process deeply integrates the initial weaning risk index with clinical contraindication verification factors through the recursive splitting logic of the decision tree, dynamically generating a decision path that fits the individual patient's situation. This ensures the accuracy of the risk correction factor, so that the subsequent comprehensive weaning risk index not only covers the risks at the physiological function level, but also fully takes into account the actual clinical contraindications, which is more in line with the actual needs of clinical weaning assessment.

[0027] After data processing unit 3 generates risk correction factors through dynamic decision-making paths, the initial weaning risk index still needs to be differentiated based on the severity of clinical contraindications in order to accurately reflect the patient's actual weaning risk. Therefore, a weighted engine is needed to complete the final integration and generate a comprehensive weaning risk index. The specific implementation method is as follows: The process of correcting risk bias and generating a comprehensive weaning risk index involves first inputting risk correction factors and the preliminary weaning risk index into a weighted engine. Risk correction factors are the quantitative correction values ​​corresponding to each contraindication verification factor that triggers a preset rule; multiple factors are possible, and their values ​​are positively correlated with the degree of impact of the contraindication on weaning risk. The preliminary weaning risk index is a basic risk grading indicator ranging from 0 to 10, reflecting the risk level at the respiratory and oxygenation function level. The weighted engine is a functional module specifically designed to fuse the basic index and correction factors by weight. It has built-in weight allocation logic and data fusion algorithms, and can dynamically adjust the contribution of correction factors according to the severity of contraindications, avoiding risk misjudgments caused by a single correction amount. During the input process, the weighted engine first simultaneously receives the specific value of the preliminary weaning risk index, the values ​​of all risk correction factors, and the values ​​of each correction factor. The severity level information of contraindication verification factors corresponding to positive factors is used to establish a correlation data table of basic index, modified factor, and severity level, providing structured input for subsequent weight allocation and fusion calculation. Differential weights are assigned based on the severity level of contraindication verification factors in the dynamic decision-making path. The severity level of contraindication verification factors is a preset grading standard in the clinical rule base, divided into three levels: mild, moderate, and severe, according to the degree of impact of contraindications on weaning outcomes. Mild indicates that the contraindication has a small impact on weaning risk; moderate indicates that the contraindication may increase the difficulty of weaning; and severe indicates that the contraindication is highly likely to cause weaning failure. Differential weights are weight coefficients assigned to different severity levels to adjust the contribution ratio of the corresponding risk modification factor in the comprehensive index. The higher the severity level, the larger the weight coefficient, ensuring that key high-risk factors are given sufficient attention. The specific allocation process is as follows: The weighted engine retrieves a pre-defined weight coefficient table from the clinical rule base (0.3 for mild cases, 0.7 for moderate cases, and 1.0 for severe cases; these coefficients are calibrated using extensive clinical data to balance sensitivity and specificity). Then, based on the severity level of each risk correction factor in the dynamic decision-making path, it matches the corresponding weight coefficient from the table. For example, a weight coefficient of 1.0 is matched for the correction factor corresponding to severe acute pulmonary infection, and a weight coefficient of 0.7 is matched for the correction factor corresponding to moderate electrolyte disturbance. If a patient has multiple risk correction factors, a corresponding weight coefficient is assigned to each correction factor, forming a personalized weight set. This outputs a comprehensive weaning risk index, which is the final quantitative indicator reflecting the patient's weaning risk. The index ranges from 0 to 10, with higher scores indicating higher weaning risk. Its core advantage is that it balances the individualized impact of basic physiological function risks and clinical contraindications, better meeting actual clinical assessment needs. The specific calculation process is as follows: First, calculate the weighted value of each risk correction factor, which is the correction factor value multiplied by the corresponding weight coefficient. For example, when the correction factor is +3 (severe level), the weighted value is 3 × 1.0 = 3 points; when the correction factor is +1.5 (medium level), the weighted value is 1.5 × 0.7 = 1.05 points. Then, sum the weighted values ​​of all correction factors to obtain the cumulative weighted correction amount. Next, add the initial offline risk index to the cumulative weighted correction amount to obtain the initial composite index after fusion. Finally, apply boundary constraints to the initial composite index: if the result exceeds 10 points, take 10 points (highest risk); if it is below 0 points, take 0 points (lowest risk), ensuring that the index is within a reasonable range. For example, if the initial offline risk index is 3 points (low risk), the cumulative weighted correction amount is 4.05 points. The initial comprehensive index was 7.05 points, and the final output was a comprehensive weaning risk index of 7.1 points (rounded to one decimal place to improve accuracy). This index comprehensively integrates the dual risks of physiological function and clinical contraindications, providing clinicians with direct and reliable quantitative basis for making weaning decisions. The clinical rule base updates the weight allocation strategy of contraindication verification factors through an online learning mechanism. This online learning mechanism refers to the system continuously collecting new weaning case data (including patient contraindication information, preliminary risk index, comprehensive risk index, and final weaning results) during clinical application. The dynamic adjustment mechanism of weight coefficients is optimized through algorithm iteration to ensure that the weight allocation strategy always aligns with the latest clinical practice. The weight allocation strategy refers to the rules for setting and adjusting the weight coefficients corresponding to different severity levels of contraindications. The update process is as follows: The system periodically initiates an online learning process (every 1000 valid offline cases are accumulated). It first selects cases with complete data and clear offline results, groups them according to the severity of contraindications, and analyzes the correlation between the weight coefficient of the correction factor and the offline failure rate in each group. If the accuracy of the comprehensive risk index corresponding to the weight coefficient of a certain severity level in predicting the offline failure rate is found to be lower than 85% (preset threshold), the weight coefficient of that level is adjusted through the gradient descent algorithm. At the same time, if a new clinical guideline clarifies that the risk level of a certain type of contraindication has changed, the online learning mechanism will receive manually labeled update instructions and synchronously adjust the severity level classification and corresponding weight coefficient of the contraindication. The updated weight allocation strategy will automatically cover the old strategy in the clinical rule base and be applied to the subsequent calculation of the comprehensive offline risk index, so as to achieve continuous optimization and clinical adaptation of the system.

[0028] In this invention, data acquisition unit 1 collects patients' respiratory and physiological parameters in real time; data analysis unit 2 extracts instantaneous respiratory rate and continuous blood oxygen saturation measurements after preprocessing, inputs them into an intelligent analysis model composed of a deep residual network and a temporal convolutional network, integrates physiological hysteresis compensation coefficients to generate a weaning index, and outputs a preliminary weaning risk index; data processing unit 3 embeds a decision tree algorithm into the clinical rule base, generates risk correction factors through contraindication verification factors, assigns differentiated weights according to severity level, corrects preliminary risk biases to generate a comprehensive weaning risk index, and continuously optimizes the clinical rule base through online learning to improve the accuracy of weaning risk assessment and provide a reliable basis for clinical decision-making.

[0029] 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 present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-parameter-based intelligent ventilator weaning assessment system, characterized in that, include: The data acquisition unit (1) is used to collect the patient's respiratory parameters and physiological parameters collected by the physiological monitoring equipment in real time through the built-in sensors of the ventilator; The data analysis unit (2) receives the respiratory parameters and physiological parameters, performs real-time preprocessing and feature extraction on the respiratory parameters and physiological parameters, obtains respiratory rate and blood oxygen saturation, and inputs the respiratory rate and blood oxygen saturation into a preset intelligent analysis model. The intelligent analysis model performs comprehensive quantitative analysis on respiratory rate and blood oxygen saturation based on machine learning pattern recognition algorithm, obtains the off-board index, and outputs a preliminary off-board risk index based on the off-board index. The data processing unit (3) applies a decision tree algorithm to the preliminary offline risk index and integrates the data with a preset clinical rule base to correct risk bias and generate a comprehensive offline risk index.

2. The intelligent ventilator weaning assessment system based on multiple parameters according to claim 1, characterized in that: Real-time preprocessing and feature extraction of the respiratory and physiological parameters are performed, specifically including: Based on the respiratory parameters filtered by adaptive noise suppression technology, the peak interval of the respiratory flow curve is extracted. The instantaneous value of respiratory frequency per unit time is calculated by time-series integration of the peak interval of the respiratory flow curve. Physiological parameters are processed simultaneously. Based on the elimination of motion artifacts of physiological parameters, the rate of change of absorption intensity of characteristic wavelengths of arterial blood oxygen saturation is identified. Continuous blood oxygen saturation measurement values ​​are output through dynamic calibration algorithm.

3. The intelligent ventilator weaning assessment system based on multiple parameters according to claim 1, characterized in that: The construction process of the preset intelligent analysis model specifically includes: The respiratory rate time series matrix and blood oxygen saturation decay trajectory of historical ventilator-associated cases were collected as training sets. A deep residual network was used to extract the abrupt change pattern features of respiratory rate, and a temporal convolutional network was used to capture the continuous trend features of blood oxygen saturation. A dual-channel input ventilator-associated risk prediction architecture was established through a feature fusion module, and the intelligent analysis model parameters were solidified after adversarial training optimization.

4. The intelligent ventilator weaning assessment system based on multiple parameters according to claim 3, characterized in that: The working logic of the preset intelligent analysis model for inputting respiratory rate and blood oxygen saturation specifically includes: Instantaneous respiratory rate values ​​are constructed as a sliding time window spectrogram, and continuous blood oxygen saturation measurements are converted into a dynamic decay slope sequence. The high-frequency abnormal oscillation components in the sliding time window spectrogram are analyzed through the respiratory rate branch of the offline risk prediction architecture, and the step-down segments in the dynamic decay slope sequence are analyzed through the blood oxygen saturation branch of the offline risk prediction architecture. The outputs of the two branches are weighted and fused through a cross-modal attention mechanism.

5. The intelligent ventilator weaning assessment system based on multiple parameters according to claim 4, characterized in that: The preset intelligent analysis model uses mutation-sensitive convolutional kernels to detect sudden growth events of the high-frequency abnormal oscillation components in the analysis of respiratory rate. In the analysis of blood oxygen saturation, the preset intelligent analysis model deploys trend memory units to track the cumulative duration of step-down segments. When a sudden growth event is detected and the cumulative duration exceeds the coordination threshold, a high-risk state marker is triggered.

6. The intelligent ventilator weaning assessment system based on multiple parameters according to claim 5, characterized in that: The specific operations of the comprehensive quantitative analysis include: The intensity value of the sudden growth event output by the mutation-sensitive convolution kernel is mapped to the respiratory drive hyperactivity index, and the cumulative duration output by the trend memory unit is quantified as the oxygenation deterioration index. The respiratory drive hyperactivity index and the oxygenation deterioration index are fused into a continuous offline index through a nonlinear transformation function. The fusion process introduces the physiological hysteresis compensation coefficient of blood oxygen saturation on respiratory rate.

7. The intelligent ventilator weaning assessment system based on multiple parameters according to claim 6, characterized in that: For instantaneous respiratory rate values, the mutation-sensitive convolutional kernel identifies accelerated exhaustion patterns where the variation amplitude of adjacent respiratory cycles exceeds the physiological baseline. For continuous blood oxygen saturation measurements, the trend memory unit captures the compensatory decline inflection point that appears later than the respiratory rate.

8. The intelligent ventilator weaning assessment system based on multiple parameters according to claim 6, characterized in that: The specific steps for outputting a preliminary offline risk index based on the offline index include: The continuous weaning index is input into the risk stratification module, and spatial positioning is performed based on the preset three-dimensional risk surface. The dimensions are the respiratory drive hyperactivity index, oxygenation deterioration index, and physiological lag compensation coefficient. When the continuous weaning index falls into the high-risk area of ​​the three-dimensional risk surface, a high preliminary weaning risk index is output, and the medium and low risk areas correspond to the graded risk index.

9. The intelligent ventilator weaning assessment system based on multiple parameters according to claim 8, characterized in that: Using the initial offline risk index as the root node and the contraindication verification factor embedded in the clinical rule base as the decision branch, a dynamic decision path is generated through recursive splitting. When the contraindication verification factor triggers the preset rule, a risk correction factor is generated.

10. The intelligent ventilator weaning assessment system based on multiple parameters according to claim 9, characterized in that: The specific steps for correcting risk bias and generating a comprehensive offline risk index include: The risk correction factor and the preliminary offline risk index are input into the weighted engine. Based on the severity level of the contraindication verification factor in the dynamic decision path, differentiated weights are assigned to output a comprehensive offline risk index. The clinical rule base updates the weight allocation strategy of the contraindication verification factor through an online learning mechanism.

Images