Physiological parameter monitoring and intelligent early warning system for critical patient

Abstract

The invention discloses a critical patient physiological parameter monitoring and intelligent early warning system, which relates to the technical field of intelligent medical treatment, and is provided with a multi-modal data collection module for collecting multi-modal physiological parameter data of a critical patient through a sensor array of flexible electronic skin, and a signal alignment module for aligning the multi-modal physiological parameter data of the critical patient. Aligning signal sequences with different sampling rates in the multi-modal physiological parameter data, setting a physiological theoretical range updating module, constructing a transfer learning model according to historical medical records of a patient, updating a physiological theoretical range of a heart rate variation coefficient in real time, setting a grading early warning module, and triggering a Bayesian network composite event detection model. Generating a graded early warning signal, setting a decision support module, dynamically loading a clinical knowledge graph through a micro-service architecture, and pushing decision support information containing rescue priority labels to a monitoring terminal; and the precision and efficiency of intensive care are obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent healthcare, and specifically to a physiological parameter monitoring and intelligent early warning system for critically ill patients. Background Art

[0002] The physiological states of critically ill patients are highly complex and dynamically variable. Their conditions may deteriorate rapidly within a short period, requiring continuous, accurate monitoring and rapid intervention. Traditional monitoring methods mostly rely on single-parameter threshold alarms such as heart rate and blood oxygen. However, single parameters are easily interfered by noise and cannot comprehensively reflect the overall state of patients. In addition, fixed thresholds are difficult to adapt to individual differences such as underlying diseases, age, and physical constitution, often resulting in false alarms or missed alarms. An intelligent monitoring and grading early warning system can integrate multimodal physiological data such as electrocardiogram, blood oxygen, and respiratory impedance, combine dynamic threshold adjustment and composite event analysis, significantly improving the accuracy and timeliness of early warning, and providing a scientific basis for clinical decision-making.

[0003] In existing monitoring and early warning technologies, there are problems that traditional devices only monitor single parameters (such as electrocardiogram or blood oxygen), lack the collaborative analysis of multimodal data, cannot capture abnormal signals of multi-system linkage, or use fixed threshold alarms and cannot dynamically adjust according to the individual historical data of patients, resulting in insufficient sensitivity to high-risk patients.

[0004] Therefore, the present invention proposes a physiological parameter monitoring and intelligent early warning system for critically ill patients. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art. For this purpose, the present invention proposes a physiological parameter monitoring and intelligent early warning system for critically ill patients, significantly improving the accuracy and efficiency of intensive care.

[0006] To achieve the above object, a physiological parameter monitoring and intelligent early warning system for critically ill patients is proposed, including a multimodal data collection module, a signal alignment module, a physiological theoretical range update module, a grading early warning module, and a decision support module; wherein, each module is electrically connected; The multimodal data collection module collects multimodal physiological parameter data of critically ill patients through a sensor array of flexible electronic skin, including electrocardiogram waveforms, blood oxygen saturation, and respiratory impedance signals, and sends the multimodal physiological parameter data to the signal alignment module; The signal alignment module aligns the signal sequences with different sampling rates in the multimodal physiological parameter data and sends the signal sequences to the grading early warning module; A physiological theory range update module constructs a transfer learning model based on the patient's historical medical records. The transfer learning model automatically generates an individualized sign fluctuation threshold interval, updates the physiological theory range of the heart rate variability coefficient in real time, and sends the physiological theory range to the hierarchical warning module; A hierarchical warning module triggers a Bayesian network composite event detection model when at least two physiological parameters in the signal sequence simultaneously touch outside the physiological theory range. The Bayesian network composite event detection model calculates the multi-parameter joint abnormal probability, generates a hierarchical warning signal, and sends the hierarchical warning signal to the decision support module; A decision support module dynamically loads a clinical knowledge graph through a microservice architecture. The clinical knowledge graph associates the hierarchical warning signal with a preset disposal plan, and pushes decision support information with rescue priority markings to the monitoring terminal; The steps of collecting multi-modal physiological parameter data of critically ill patients include the following: Step 11: Design and deploy a sensor array of flexible electronic skin. The sensor array includes multiple signal acquisition channels for respectively collecting electrocardiogram waveforms, blood oxygen saturation, and respiratory impedance signals; Step 12: Calibrate the sensitivity parameters of the flexible electronic skin sensor array. The sensitivity parameters include the amplitude threshold of the electrocardiogram waveform, the photoelectric signal intensity of blood oxygen saturation, and the frequency response characteristics of respiratory impedance; Step 13: Synchronously collect multi-modal physiological parameter data on the body surface of critically ill patients. The multi-modal physiological parameter data includes electrocardiogram waveforms, blood oxygen saturation, and respiratory impedance signals; Step 14: Perform preliminary processing on the multi-modal physiological parameter data. The preliminary processing includes signal denoising and normalization processing, and outputs a standardized multi-modal physiological parameter data set; The method of aligning signal sequences with different sampling rates in the multi-modal physiological parameter data using the dynamic time warping algorithm is as follows: Align the signal sequences with different sampling rates in the synchronized multi-modal physiological parameter data using the dynamic time warping algorithm; The steps of constructing a transfer learning model based on the patient's historical medical records include the following: Step 31: Extract multi-modal physiological parameter feature data from the patient's historical medical records, and use feature extraction technology to generate a standardized feature vector sequence; The historical medical records are stored in a relational database in a time series form, and each record contains a sequence of electrocardiogram waveform sampling point voltage values; The historical medical records also contain blood oxygen saturation photoplethysmogram signals; Step 32: Construct a feature distribution model for the patient population. Input the standardized feature vector sequence, and use the clustering analysis method to identify the feature distribution rules of different patient populations, and output the patient population feature distribution model.

[0007] Step 33: Construct a transfer learning model. Input the patient population feature distribution model and the individual feature vector sequence, and use the transfer learning algorithm to map the common feature distribution to the individual feature distribution, and output the individualized physical sign fluctuation threshold interval. Step 34: Update the normal range of the heart rate variability coefficient in real time. Input the individualized physical sign fluctuation threshold interval and the physiological parameter data monitored in real time, and use the dynamic adjustment algorithm to update the normal range of the heart rate variability coefficient.

[0008] The method for triggering the Bayesian network composite event detection model is as follows: The Bayesian network composite event detection model calculates the multi-parameter joint anomaly probability based on the at least two physiological parameter data and outputs the multi-parameter joint anomaly probability.

[0009] The method for generating the hierarchical warning signal is as follows: Generate a hierarchical warning signal according to the multi-parameter joint anomaly probability and the preset hierarchical rules. The hierarchical warning signal includes warning information at different levels.

[0010] The step of pushing the decision support information including the rescue priority annotation to the monitoring terminal includes the following steps: Step 51: Dynamically load the clinical knowledge graph through the microservice architecture, and retrieve the treatment plan associated with the warning signal from the clinical knowledge graph. Step 52: Associate the warning signal with the retrieved treatment plan to generate a rescue priority annotation. Step 53: Push the decision support information including the rescue priority annotation to the monitoring terminal.

[0011] Compared with the prior art, the beneficial effects of the present invention are: First, this solution uses a flexible electronic skin sensor array to synchronously collect multi-modal physiological signals such as electrocardiogram, blood oxygen, and respiratory impedance. The dynamic time warping algorithm is used to align data streams with different sampling rates, eliminating the problem of time-frequency misalignment and ensuring the temporal consistency of the signals. Then, an individualized physical sign fluctuation threshold is constructed based on a transfer learning model. By analyzing the patient's historical medical records and the distribution of group characteristics, the normal ranges of key parameters such as the heart rate variability coefficient are dynamically adjusted to achieve personalized adaptation of the warning threshold. Subsequently, a Bayesian network compound event detection model is used to calculate the joint abnormal probability of multiple parameters. When at least two physiological parameters exceed the threshold simultaneously, a hierarchical warning mechanism is triggered to reduce the false alarm risk caused by the noise interference of a single parameter. Finally, combined with a clinical knowledge graph and a microservice architecture, the system automatically associates the warning signal with a standardized treatment plan, generates a rescue priority label, and pushes it to the monitoring terminal in real time, forming a closed-loop support from monitoring to decision-making, thus significantly improving the accuracy and efficiency of intensive care. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 FIG. is a diagram showing the module connection relationship of a physiological parameter monitoring and intelligent warning system for critically ill patients in Embodiment 1 of the present invention; Figure 2 FIG. is a flowchart for collecting multi-modal physiological parameter data of critically ill patients in Embodiment 1 of the present invention; Figure 3 FIG. is a flowchart for constructing a transfer learning model based on the historical medical records of patients in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0013] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0014] Embodiment 1

[0015] As Figure 1 shown, a physiological parameter monitoring and intelligent warning system for critically ill patients includes a multi-modal data collection module, a signal alignment module, a physiological theory range update module, a hierarchical warning module, and a decision support module; among them, each module is connected electrically; The multi-modal data collection module collects multi-modal physiological parameter data of critically ill patients through a sensor array of flexible electronic skin, including electrocardiogram waveforms, blood oxygen saturation, and respiratory impedance signals, and sends the multi-modal physiological parameter data to the signal alignment module; A signal alignment module that aligns signal sequences with different sampling rates in the multimodal physiological parameter data and sends the signal sequences to a hierarchical warning module; A physiological theory range update module that constructs a transfer learning model based on the patient's historical medical records. The transfer learning model automatically generates an individualized sign fluctuation threshold range, updates the physiological theory range of the heart rate variability coefficient in real time, and sends the physiological theory range to the hierarchical warning module; A hierarchical warning module that triggers a Bayesian network composite event detection model when at least two physiological parameters in the signal sequence simultaneously touch outside the physiological theory range. The Bayesian network composite event detection model calculates the joint abnormal probability of multiple parameters, generates a hierarchical warning signal, and sends the hierarchical warning signal to the decision support module; A decision support module that dynamically loads a clinical knowledge graph through a microservice architecture. The clinical knowledge graph associates the hierarchical warning signal with a preset disposal plan and pushes decision support information with rescue priority markings to the monitoring terminal; In an embodiment of the present invention, as Figure 2 shown, the steps of collecting multimodal physiological parameter data of critically ill patients include the following: Step 11: Design and deploy a sensor array of flexible electronic skin. The sensor array includes multiple signal acquisition channels for respectively collecting electrocardiogram waveforms, blood oxygen saturation, and respiratory impedance signals; Specifically, the sensor array of the flexible electronic skin adopts a multi-layer heterogeneous material composite structure, and an electrocardiogram signal acquisition module, a blood oxygen optical sensing unit, and a respiratory impedance detection circuit are integrated on a flexible polyimide substrate.

[0016] The flexible polyimide substrate refers to using a polyimide film with a coefficient of thermal expansion of 3.1×10 -5 / K as the base material. Its dielectric constant of 4.2@1MHz can effectively isolate the parasitic capacitance between electrodes.

[0017] The electrocardiogram signal acquisition module is composed of 3 groups of interdigital silver nanowire electrodes. The distance between each group of electrodes is 2.5±0.1mm. "Interdigital" means that the electrodes are designed with an interdigital comb-like structure, which increases the effective contact area by 120%. The setting of the electrode distance of 2.5mm is based on the attenuation characteristics of the human epidermal electric potential, and can capture effective signals above 0.5mV in the QRS complex, and is connected to the signal conditioning circuit through silver paste wires.

[0018] The blood oxygen optical sensing unit includes dual-light source LEDs with wavelengths of 660 nm and 880 nm and a photodiode receiver. Its photoelectric conversion sensitivity is adjusted to 5 mV / μW through a built-in transimpedance amplifier. Among them, the 660-nm wavelength corresponds to the absorption peak of deoxyhemoglobin, and the 880 nm corresponds to the isosbestic point of oxyhemoglobin, etc. This dual-wavelength design realizes the calculation of blood oxygen saturation through the Beer-Lambert law. The setting of the photoelectric conversion sensitivity of 5 mV / μW is based on the quantum efficiency curve of the photodiode, which can ensure a signal-to-noise ratio > 60 dB under typical skin transmittance.

[0019] The respiratory impedance detection circuit uses the four-electrode method for measurement. Drive electrodes and detection electrodes with a spacing of 10 mm are arranged on a silicone rubber substrate, and a 50-kHz sinusoidal excitation signal is applied.

[0020] During deployment, the sensor array is fixed to the region from the manubrium sterni to the xiphoid process of the patient through a medical-grade acrylic adhesive patch, ensuring that the contact impedance between each sensing unit and the skin is less than 2 kΩ.

[0021] Step 12: Calibrate the sensitivity parameters of the flexible electronic skin sensor array. The sensitivity parameters include the amplitude threshold of the electrocardiogram waveform, the photoelectric signal intensity of blood oxygen saturation, and the frequency response characteristics of respiratory impedance. It can be understood that since the sensor array needs to adapt to the physiological characteristics of different patients, it is necessary to calibrate the sensitivity parameters to ensure the measurement accuracy and stability of the sensor array. Specifically, during the calibration process, a NIBP simulator is used to generate a standard electrocardiogram waveform. The NIBP simulator refers to a non-invasive blood pressure simulation device, and its standard electrocardiogram waveform generation is based on the Lown-Canong-Levine electrocardiogram model. By adjusting the gain of the preamplifier of the silver nanowire electrode, the R-wave amplitude reaches a threshold of 1.2 ± 0.05 V. The 1.2-V threshold of the R-wave amplitude corresponds to the input range of lead II of a clinical ECG device, and this voltage value can be adapted to the range of a 12-bit ADC after gain adjustment.

[0022] The calibration of blood oxygen saturation uses a simulated solution containing 35% reduced hemoglobin. The 35% concentration in the reduced hemoglobin simulated solution is set to simulate the blood oxygen dynamic equilibrium state when the arterial blood oxygen saturation of the human body is 95%. Adjust the bias voltage of the photodiode so that the light intensity ratio of 880 nm / 660 nm reaches 0.48 ± 0.02 when SpO2 is 95%. The calibration point of the light intensity ratio of 0.48 comes from the linear interval of the oxygen dissociation curve, and the conversion relationship between this ratio and SpO2 follows the Nelder-Mead nonlinear fitting algorithm.

[0023] Respiratory impedance calibration applies a sinusoidal displacement load of 0.5 - 2.0 Hz through a robotic arm, optimizes the parameters of the lock-in amplifier to achieve a power frequency interference rejection ratio of -80 dB at 50 Hz, with a linearity error of the frequency response less than 1.5%. The requirement of a -80 dB power frequency interference rejection ratio at 50 Hz is derived from the YY 0505 standard for medical electrical equipment, and this indicator is achieved by adjusting the quality factor Q value of the lock-in amplifier.

[0024] Finally, the gain coefficients of each channel are stored in the EEPROM to form a calibration matrix containing 32 sets of compensation parameters.

[0025] Step 13: Synchronously collect multi-modal physiological parameter data on the body surface of critically ill patients. The multi-modal physiological parameter data includes electrocardiogram waveforms, blood oxygen saturation, and respiratory impedance signals. It can be understood that since the multi-modal physiological parameter data needs to be consistent in time, synchronous collection is required to ensure the relevance and accuracy of the data. Specifically, when synchronously collecting, a 10 MHz main clock signal is generated by the FPGA. Distributed to the ADC sampling units of each sensing channel, the electrocardiogram signal is sampled at a rate of 1 kHz, and the right leg drive technology is used to eliminate common-mode interference. The "right leg" refers to the improved circuit of the Wilson central terminal, which reduces the common-mode voltage by 40 dB through negative feedback. The blood oxygen signal is collected as a photoplethysmogram at a rate of 125 Hz, and crosstalk is eliminated through time-division multiplexing. The 125 Hz sampling rate of the photoplethysmogram is set to meet the sampling requirements of the pulse rate signal (usually <5 Hz) according to the Nyquist theorem. The time-division multiplexing interval of 8 ms can avoid crosstalk of the LED light source. The respiratory impedance is measured as a complex impedance at a rate of 50 Hz, and the real part is used to calculate the change in thoracic volume. In the complex impedance measurement of respiratory impedance, the real part reflects tissue impedance, and the imaginary part corresponds to capacitive reactance. The real part is selected because the change in thoracic volume mainly causes a resistance change rate of 0.15 Ω / cm³.

[0026] The three-way signals are packed into a data frame aligned with the time stamp after CRC check and transmitted to the edge node through Bluetooth 5.0, with the delay jitter controlled within ±2 ms.

[0027] Step 14: Perform preliminary processing on the multi-modal physiological parameter data. The preliminary processing includes signal denoising and normalization processing, and outputs a standardized multi-modal physiological parameter data set. It can be understood that since the multi-modal physiological parameter data may contain noise and different dimensions, preliminary processing is required to improve the data quality and comparability.

[0028] Specifically, for signal denoising, an improved wavelet threshold algorithm is adopted to perform 6-layer db4 wavelet decomposition on the electrocardiogram (ECG) signal. The db4 wavelet refers to the Daubechies 4th-order wavelet, and its support length of 8 is suitable for capturing the characteristics of the QRS complex.

[0029] The soft threshold function is set to σ√(2lnN) (where σ is the noise variance and N = 1024). The noise variance σ is estimated through the median of the detail coefficients at Level 1. In the formula σ√(2lnN), N = 1024 corresponds to the number of sampling points in 1 second of ECG data.

[0030] In the normalization process, the ECG signal is baseline corrected and calibrated at 1 mV / unit according to the standard of limb lead II; the blood oxygen signal is converted to the optical density ratio ODR = (AC660 / DC660) / (AC880 / DC880). The AC component in the optical density ratio ODR refers to the alternating current component (corresponding to pulsatile blood flow), and DC is the direct current component (reflecting tissue absorption). The construction of this ratio makes the calculation error of blood oxygen saturation < 1.5%; the respiratory impedance value is mapped to a relative thoracic volume of 0 - 100%.

[0031] Finally, the matrix dimension of the standardized data set is 3000×12, including time-domain waveforms, frequency-domain features, and 12-dimensional statistics.

[0032] Furthermore, the method for aligning the signal sequences with different sampling rates in the multimodal physiological parameter data using the dynamic time warping algorithm is as follows: The dynamic time warping algorithm is used to align the signal sequences with different sampling rates for the synchronized multimodal physiological parameter data. It can be understood that since the multimodal physiological parameter data may come from different sensors and their sampling rates may vary, directly performing data processing may lead to signal misalignment or analysis errors. Therefore, it is necessary to use the dynamic time warping algorithm to align the signal sequences with different sampling rates to ensure data consistency and accuracy.

[0033] Specifically, in the signal sequence alignment process, the dynamic time warping algorithm adopts an improved local path constraint strategy to define the cost matrix as:

[0034] where is the cumulative cost matrix, representing the minimum cumulative difference path from the signal start point to point (i, j). i is the time step of the high-frequency signal, such as the 250 Hz ECG, and j is the time step of the low-frequency signal, such as the 50 Hz respiratory impedance, reflecting the global optimal solution for the time-domain alignment of the two signals , and That is, it represents the i-th sampling point of the ECG signal and the j-th sampling point of the SpO2 or respiratory signal respectively. It represents the morphological difference metric between the ECG signal and the blood oxygen signal. The constraint weight parameters (α, β, γ) are set to (0.3, 1, 0.3) according to clinical experience to conform to the quasi-periodic characteristics of the respiratory impedance signal. For the time-frequency difference between the ECG signal (sampling rate 250Hz) and the respiratory impedance signal (sampling rate 50Hz), the algorithm sets the dynamic window width , extracts the characteristic points of the QRS complex as alignment anchor points through sliding window Fourier transform, and finally outputs a composite physiological signal matrix with unified time resolution.

[0035] Furthermore, as Figure 3 shown, the constructing of the transfer learning model according to the patient's historical medical records includes the following steps: Step 31: Extract multi-modal physiological parameter feature data from the patient's historical medical records, and use feature extraction technology to generate a standardized sequence of feature vectors; It can be understood that due to the limitations of heterogeneity and noise in the multi-modal physiological parameter data, feature extraction and standardization processing are required to obtain a sequence of feature vectors that can be used for subsequent analysis.

[0036] The historical medical records are stored in a relational database in the form of a time series, and each record contains the voltage value sequence of the ECG waveform sampling points , the voltage value sequence of the ECG waveform sampling points Specifically refers to the digital sampling data of the ECG signal continuously collected through flexible electrodes, where each element represents the millivolt-level potential difference generated during the depolarization and repolarization of myocardial cells measured at time ; Specifically, the historical medical records also contain the photoplethysmogram signal of blood oxygen saturation , where the continuous variable represents the time axis, and the measured value is the ratio of the absorbance of red light (660nm) to infrared light (940nm) after normalization, reflecting the instantaneous concentration change of oxyhemoglobin in peripheral capillaries; and the measured value of the respiratory impedance phase difference , which specifically refers to the change in thoracic impedance measured by the four-electrode method in the frequency range of 10 - 100kHz, and its positive real number attribute stems from the modulus calculation of the complex impedance of biological tissues.

[0037] The feature extraction process uses a multi-scale analysis method to detect the QRS complex of the ECG signal and extract the standard deviation of the R-R interval as the heart rate variability index. This standard deviation of the R-R interval As an indicator of the function of the autonomic nervous system, it quantifies the intensity of the heartbeat interval fluctuations in which the sinoatrial node pacemaker is regulated by both sympathetic and vagus nerves; Calculate the frequency of the zero-crossing points of the second derivative of the blood oxygen signal Characterize the peripheral circulation state, the frequency of the zero-crossing points of the second derivative of the blood oxygen signal By calculating The number of time points reflects the change in the pulse wave propagation velocity caused by the change in peripheral resistance in the peripheral blood vessels; Perform power spectral density analysis on the respiratory impedance signal and take the energy proportion in the frequency band of 0.15 - 0.4 Hz As an index of respiratory pattern stability, the energy proportion of the respiratory impedance signal in the frequency band of 0.15 - 0.4 Hz Corresponds to the low-frequency oscillation component formed by the rhythmic discharge of the respiratory center driving the diaphragmatic movement under the resting state of the human body.

[0038] Each characteristic parameter forms a feature vector after min-max standardization , where the asterisk represents the standardized value, and the standardization formula is:

[0039] In the formula is the mean value of this feature in the historical dataset, is the standard deviation, and the coefficient 3 ensures that most data falls within the interval [0, 1].

[0040] Step 32: Construct a feature distribution model for the patient population. Input the standardized feature vector sequence, use the clustering analysis method to identify the feature distribution rules of different patient populations, and output the patient population feature distribution model.

[0041] It can be understood that due to the limitations of individual differences and population diversity in the standardized feature vector sequence, clustering analysis is required to obtain a model that can reflect the feature distribution rules of different patient populations.

[0042] Specifically, input the standardized feature vector sequence into the Gaussian mixture model (GMM) for clustering analysis. Assume that the patient population is divided into subclasses.

[0043] Define the observed dataset , and solve the parameter set through the expectation-maximization algorithm, where is the mixing coefficient, which represents the occurrence probability of a certain type of patient population in the overall population; is the feature mean vector of the th subclass. This mean vector contains the typical numerical combinations of the electrocardiogram, blood oxygen, and respiratory characteristics of this type of patient; is the covariance matrix, and this covariance matrix characterizes the co-variation characteristics among various physiological parameters. The clustering validity is evaluated by the silhouette coefficient as follows:

[0044] where is the average distance from the sample to other samples in the same cluster, and is the minimum average distance from the sample to the nearest different cluster. When it is determined that the subclass division is reasonable.

[0045] The finally constructed feature distribution model can be expressed as: .

[0046] Step 33: Construct a transfer learning model, input the patient group feature distribution model and the individual feature vector sequence, and use the transfer learning algorithm to map the common feature distribution to the individual feature distribution, and output the individualized physical sign fluctuation threshold interval.

[0047] It can be understood that due to the limitation of the distribution difference between the patient group feature distribution model and the individual feature vector sequence, transfer learning is required to obtain an individualized physical sign fluctuation threshold interval that can adapt to the individual feature distribution.

[0048] Specifically, the transfer learning model adopts a domain adaptation network architecture, defines the source domain as the patient group feature distribution , and the target domain as the individual patient feature sequence . The network includes a shared feature extraction layer and a domain discriminator , and realizes distribution alignment through the maximum mean discrepancy (MMD) loss:

[0049] where is the mapping function of the reproducing kernel Hilbert space, and the Gaussian kernel is selected, and the Gaussian kernel parameter controls the similarity decay rate, and the reciprocal square root thereof corresponds to the effective comparison range in the feature space.

[0050] The individualized threshold interval is determined by quantile regression. Let be the value of the target domain feature at the quantile, then the threshold interval is:

[0051] where IQR is the interquartile range, represents the interval expansion operation, and in the quantile regression specifically refers to that for the target patient in his historical monitoring data, the physiological parameters have The probability is lower than this critical value, where and correspond to the lower and upper limits of normal fluctuations respectively.

[0052] Step 34: Update the normal range of the heart rate variability coefficient in real time. Input the individualized physical sign fluctuation threshold interval and the physiological parameter data monitored in real time, and use the dynamic adjustment algorithm to update the normal range of the heart rate variability coefficient.

[0053] It can be understood that, due to the time-varying limitations of the individualized physical sign fluctuation threshold interval, dynamic adjustment is required to obtain the normal range of the heart rate variability coefficient that can reflect the current physiological state.

[0054] Specifically, for the dynamic update of the heart rate variability coefficient , the exponentially weighted moving average method is adopted. Let be the real-time monitored value at time , and the update formula is ;

[0055] where the smoothing coefficient , and the exponentially weighted moving average coefficient reflects the time decay characteristic. When the sampling interval minutes, means that the weight of the data 24 hours ago decays to . The normal range is updated to ; When the real-time exceeds this interval three times continuously, the early warning review mechanism is triggered, and joint probability verification is carried out through the Bayesian network composite event detection model. The Bayesian network composite event detection model specifically refers to the joint probability inference mechanism that integrates multi-dimensional indicators such as abnormal heart rate variability coefficient, blood oxygen fluctuation pattern change, and respiratory rhythm disorder through the conditional probability table.

[0056] It can be understood that, since the abnormality of a single physiological parameter may not be sufficient to accurately judge the occurrence of a composite event, at least two physiological parameters need to be monitored simultaneously, and the Bayesian network composite event detection model is used for comprehensive analysis, so as to improve the accuracy and reliability of detection.

[0057] When at least two physiological parameters (such as the electrocardiogram waveform amplitude threshold and the blood oxygen saturation photoelectric intensity) in the aligned signal sequence simultaneously touch the individualized physical sign fluctuation threshold interval updated in real time, the Bayesian network composite event detection model is triggered. The individualized physical sign fluctuation threshold interval is the individualized physical sign fluctuation threshold interval output in step three, and the numerical range updated in real time through the dynamic adjustment algorithm, which is used to reflect the abnormal critical value of the patient's current physiological state.

[0058] Specifically, the method for triggering the Bayesian network composite event detection model is as follows: Based on the at least two physiological parameter data, the Bayesian network composite event detection model calculates the multi-parameter joint anomaly probability and outputs the multi-parameter joint anomaly probability.

[0059] It can be understood that since the anomaly of a single parameter may be caused by multiple factors, the calculation of the multi-parameter joint anomaly probability can more comprehensively reflect the occurrence possibility of the composite event, thereby providing a more accurate early warning basis.

[0060] Specifically, first, preprocess the input data, standardize the physiological parameter data such as electrocardiogram waveforms and blood oxygen saturation to a unified dimension. For example, normalize the electrocardiogram amplitude to the interval [0, 1], and convert the blood oxygen saturation to a percentage form to ensure the compatibility of the input data.

[0061] Then, for the model triggering mechanism, the real-time monitoring module determines whether multiple physiological parameters simultaneously cross the threshold interval. For example, if the electrocardiogram amplitude exceeds 0.8 (normalized value) and the blood oxygen saturation is lower than 90%, the model is triggered to run. Then, load the network topology, call the predefined Bayesian network structure file, which contains nodes (representing the abnormal states of physiological parameters) and directed edges (representing the causal relationships between parameters). For example, node A represents "electrocardiogram abnormality", node B represents "blood oxygen abnormality", and the edge A→B means that electrocardiogram abnormality may lead to blood oxygen abnormality. After that, the Bayesian network composite event detection model calculates the multi-parameter joint anomaly probability based on the formula where: (P(A)) represents the prior probability of the electrocardiogram parameter being abnormally alone, obtained by statistical analysis of historical data (such as the historical abnormality rate of patients is 0.15). (P(B|A)) represents the conditional probability of blood oxygen abnormality under the condition of electrocardiogram abnormality, provided by the clinical knowledge base (such as set to 0.6).

[0062] First, according to the abnormal states of the input parameters (such as A = 1, B = 1), retrieve the corresponding joint probability value from the conditional probability table of the model. For example, if the joint probability is defined as 0.75 when both A and B are abnormal in the table, directly output this value; if the deviation between the real-time data and the historical distribution exceeds the preset threshold (such as exceeding ±10%), then use the Bayesian update formula to correct the conditional probability, (P(A|B)) is the currently observed joint anomaly frequency, and is the original blood oxygen abnormality probability.

[0063] Furthermore, the method for generating the hierarchical warning signal is as follows: According to the multi-parameter joint anomaly probability and the preset hierarchical rules, generate a hierarchical warning signal, and the hierarchical warning signal includes warning information of different levels.

[0064] Specifically, according to the multi-parameter combined anomaly probability (P(A,B)), the hierarchical warning signals are generated according to the following rules: when (0.3 ≤ P < 0.6), a yellow warning is triggered, prompting medical staff to check the equipment connection or the patient's position; when (0.6 ≤ P < 0.8), an orange warning is triggered, and the recommended disposal measures (such as increasing the oxygen flow rate) are automatically pushed; when (P ≥ 0.8), a red warning is triggered, and the emergency call system is activated and the patient's position is displayed. In the specific implementation of generating the hierarchical warning signals, on the one hand, it is necessary to combine the individual physical sign fluctuation threshold range of the patient. For example, for high-risk patients, the red warning threshold is lowered from 0.8 to 0.7 to improve sensitivity; on the other hand, the warning signals are synchronously output through the flashing indicator light with a frequency increasing with the risk on the monitoring terminal, the buzzer with a frequency increasing with the risk, and the pop-up message containing the disposal guidance link to ensure that the information is conveyed in a timely manner.

[0065] It should be noted that in the embodiments of the present invention, the hierarchical warning signals include warning levels (1-3 levels), trigger parameter combinations, and timestamp fields. The keyword fields are extracted through a regular expression matching algorithm. The regular expression matching algorithm refers to a text processing method for extracting key information such as warning levels and parameter combinations from unstructured text. In this embodiment, the PCRE specification is used to implement field extraction. After extraction, it is converted into a unified dimensional vector Furthermore, the pushing of decision support information including the rescue priority label to the monitoring terminal includes the following steps: Step 51: Dynamically load the clinical knowledge graph through the microservice architecture, and retrieve the disposal plan associated with the warning signal from the clinical knowledge graph; It can be understood that since the clinical knowledge graph may need to be dynamically adjusted according to specific warning signals, it is necessary to dynamically load it through the microservice architecture to ensure that the retrieved disposal plan is relevant to the current warning signal.

[0066] Specifically, the clinical knowledge graph is stored in the Neo4j graph database. The Neo4j graph database is a non-relational database that stores data using a node-relationship structure. In this system, it is used to store disease entity nodes and their associated disposal plan nodes. The disease entity nodes contain the diagnostic criteria for several cardiovascular and cerebrovascular diseases, and the disposal plan nodes store several clinically verified first aid procedures. When receiving it, the plan retrieval is realized through the Cypher query statement MATCH (d:Disease)-[r:HAS_PROTOCOL]->(p:Protocol) WHERE d.symptom IN P RETURN p.weight,p.content.

[0067] Among them, the weight parameter , α = 0.6, β = 0.4 are coefficients optimized and determined by the receiver operating characteristic curve (ROC curve). The optimal coefficient combination is determined through the ROC curve optimization to balance the weights of the warning level and the parameter coverage rate. The indicator function takes the value of 1 when the parameter s exists in the set of pre-plan keywords, and 0 otherwise. Through the above operations, the TOP5 pre-plan set can be returned .

[0068] Step 52: Associate the warning signal with the retrieved disposal pre-plan to generate a rescue priority annotation; It can be understood that since the warning signal and the disposal pre-plan may need to be prioritized according to the specific clinical situation, an association operation is required to generate a rescue priority annotation that meets the clinical needs.

[0069] Specifically, when is associated with , a multi-objective optimization model is established, where represents the weight coefficient of the parameter matching similarity, represents the weight coefficient of the historical success rate, and this coefficient combination is determined by cardiologists based on clinical experience.

[0070] The similarity function adopts an improved algorithm of the Jaccard coefficient to calculate the geometric mean of the proportion of the intersection of the trigger parameter set and the pre-plan keyword set , that is .

[0071] After obtaining the optimal pre-plan , its priority annotation , where represents rounding up, generating an annotation set . The priority annotation in represents the normalized weight value of the optimal pre-plan, with a value range of [0, 1]. After multiplying by 10, it is converted into an integer priority scale of 1 - 10 through the rounding-up operator ⌈·⌉. Among them, ζ≥8 corresponds to a critical condition, 5≤ζ<7 corresponds to a warning condition, and ζ<5 corresponds to an observation condition.

[0072] Step 53: Push the decision support information containing the rescue priority annotation to the monitoring terminal; It can be understood that since the decision support information needs to be conveyed to the monitoring terminal in a timely manner to guide clinical operations, a push operation is required to ensure the timeliness and accuracy of the information.

[0073] Specifically, it is pushed to the monitoring terminal through the RabbitMQ message queue. The data packet is encapsulated into a message in HL7 format, which is an observation result message type in the medical information exchange standard and contains structured fields such as patient ID, device code, and warning data.

[0074] The terminal parsing module performs color coding according to the value: when ζ≥8, it flashes red, corresponding to a critical condition, requiring immediate intervention by medical staff; when 5≤ζ<8, it lights yellow constantly, corresponding to a warning condition, indicating enhanced monitoring; when ζ<5, it gives a blue prompt, corresponding to an observation condition, and the changes in vital signs need to be recorded. The push response time ≤200ms. Through the timestamp verification mechanism to ensure real-time performance, where ε = 300ms is the system tolerance threshold. The ε = 300ms in this timestamp verification mechanism is the upper limit of the system response time set according to the golden rescue time for cardiac arrest. When the absolute difference between the message sending time and the receiving time exceeds this threshold, the retransmission mechanism is triggered.

[0075] In addition, for the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art, no detailed description is given to avoid excessive elaboration.

[0076] As described above in the specific embodiments, the purpose, technical solutions, and beneficial effects of the present invention have been further elaborated in detail. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

[0077] The above preset parameters or preset thresholds are all set by those skilled in the art according to the actual situation or obtained through a large amount of data simulation.

[0078] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A physiological parameter monitoring and intelligent warning system for critically ill patients, characterized in that, It includes a multimodal data collection module, a signal alignment module, a physiological theory range update module, a hierarchical warning module, and a decision support module; among them, each module is electrically connected. The multimodal data collection module collects multimodal physiological parameter data of critically ill patients through the sensor array of the flexible electronic skin, and sends the multimodal physiological parameter data to the signal alignment module. The signal alignment module aligns the signal sequences with different sampling rates in the multimodal physiological parameter data, and sends the signal sequences to the hierarchical warning module. The physiological theory range update module constructs a transfer learning model based on the patient's historical medical records. The transfer learning model automatically generates an individualized sign fluctuation threshold interval, updates the physiological theory range of the heart rate variability coefficient in real time, and sends the physiological theory range to the hierarchical warning module. The hierarchical warning module triggers a Bayesian network composite event detection model when at least two physiological parameters in the signal sequence simultaneously touch outside the physiological theory range. The Bayesian network composite event detection model calculates the multi-parameter joint anomaly probability, generates a hierarchical warning signal, and sends the hierarchical warning signal to the decision support module. The decision support module dynamically loads a clinical knowledge graph through a microservice architecture. The clinical knowledge graph associates the hierarchical warning signal with a preset disposal plan, and pushes decision support information with rescue priority markings to the monitoring terminal.

2. The physiological parameter monitoring and intelligent warning system for critically ill patients according to claim 1, wherein The multimodal physiological parameter data includes electrocardiogram waveforms, blood oxygen saturation, and respiratory impedance signals.

3. The physiological parameter monitoring and intelligent warning system for critically ill patients according to claim 2, characterized in that, The steps of collecting the multimodal physiological parameter data of critically ill patients include: Step 11: Design and deploy the sensor array of the flexible electronic skin. The sensor array includes multiple signal acquisition channels for separately collecting electrocardiogram waveforms, blood oxygen saturation, and respiratory impedance signals. Step 12: Calibrate the sensitivity parameters of the flexible electronic skin sensor array. The sensitivity parameters include the amplitude threshold of the electrocardiogram waveform, the photoelectric signal intensity of the blood oxygen saturation, and the frequency response characteristics of the respiratory impedance. Step 13: Synchronously collect the multimodal physiological parameter data on the body surface of critically ill patients. The multimodal physiological parameter data includes electrocardiogram waveforms, blood oxygen saturation, and respiratory impedance signals. Step 14: Perform preliminary processing on the multimodal physiological parameter data. The preliminary processing includes signal denoising and normalization processing, and outputs a standardized multimodal physiological parameter data set.

4. The physiological parameter monitoring and intelligent warning system for critically ill patients according to claim 3, characterized in that, The method of using the dynamic time warping algorithm to align the signal sequences with different sampling rates in the multimodal physiological parameter data is as follows: Use the dynamic time warping algorithm to align the signal sequences with different sampling rates for the synchronized multimodal physiological parameter data.

5. The physiological parameter monitoring and intelligent warning system for critically ill patients according to claim 4, wherein, The steps of constructing the transfer learning model according to the patient's historical medical records include: Step 31: Extract multimodal physiological parameter feature data from the patient's historical medical records, and use feature extraction technology to generate a standardized feature vector sequence. Step 32: Construct a feature distribution model of the patient group, input the standardized feature vector sequence, and use the clustering analysis method to identify the feature distribution rules of different patient groups, and output the patient group feature distribution model. Step 33: Construct a transfer learning model, input the patient population feature distribution model and the individual feature vector sequence, and use the transfer learning algorithm to map the common feature distribution to the individual feature distribution, and output the individualized physical sign fluctuation threshold interval; Step 34: Update the normal range of the heart rate variability coefficient in real time. Input the individualized physical sign fluctuation threshold interval and the physiological parameter data monitored in real time, and use the dynamic adjustment algorithm to update the normal range of the heart rate variability coefficient.

6. The physiological parameter monitoring and intelligent warning system for critically ill patients according to claim 5, wherein The historical medical records are stored in a relational database in the form of a time series, and each record contains a sequence of voltage values of electrocardiogram waveform sampling points; The historical medical records also include photoplethysmogram signals of blood oxygen saturation.

7. The physiological parameter monitoring and intelligent warning system for critically ill patients according to claim 6, characterized in that, The feature extraction process adopts a multi-scale analysis method, detects QRS complexes in the electrocardiogram signal, calculates the frequency of zero-crossing of the second derivative of the blood oxygen signal to characterize the peripheral circulation state, and performs power spectral density analysis on the respiratory impedance signal.

8. The physiological parameter monitoring and intelligent warning system for critically ill patients according to claim 7, characterized in that, The method for triggering the Bayesian network composite event detection model is: The Bayesian network composite event detection model calculates the multi-parameter joint abnormality probability based on the at least two physiological parameter data and outputs the multi-parameter joint abnormality probability.

9. The physiological parameter monitoring and intelligent warning system for critically ill patients according to claim 8, wherein The method for generating a hierarchical warning signal is: Generate a hierarchical warning signal according to the multi-parameter joint abnormality probability and the preset hierarchical rules, and the hierarchical warning signal includes warning information of different levels.

10. The physiological parameter monitoring and intelligent warning system for critically ill patients according to claim 9, wherein, The pushing of the decision support information including the rescue priority annotation to the monitoring terminal includes the following steps: Step 51: Dynamically load the clinical knowledge graph through a microservice architecture, and retrieve the treatment plan associated with the warning signal from the clinical knowledge graph; Step 52: Associate the warning signal with the retrieved treatment plan to generate a rescue priority annotation; Step 53: Push the decision support information including the rescue priority annotation to the monitoring terminal.

Images