A multi-modal fusion-based infant anesthesia respiratory depression early warning system and method
By using a multimodal fusion system and model, notch filters and time series tensors are dynamically constructed, solving the problem of inaccurate risk assessment of respiratory depression during infant anesthesia and enabling early perception and dynamic warning of respiratory depression risk during infant anesthesia.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN MATERNITY & CHILD HEALTHCARE HOSPITAL
- Filing Date
- 2026-02-24
- Publication Date
- 2026-05-29
AI Technical Summary
During anesthesia for infant surgery, existing technologies struggle to effectively integrate multiple physiological signals, leading to inaccurate assessments of respiratory depression risk, a lack of targeted early warnings, and a lack of strategic support for ventilation control.
By collecting various physiological signals, a multimodal fusion system is constructed, including data acquisition, signal processing, and a multimodal fusion model. The system generates the risk probability of respiratory depression, dynamically constructs a notch filter parameter set, and combines time-division window filtering and weighted splicing to generate artifact-free respiratory signals. The respiratory rhythm characteristics and drug-stimulus balance index are calculated, and a time series tensor is constructed for risk assessment.
It enables early perception and dynamic warning of respiratory depression risk during infant anesthesia, improves the optimization of ventilation matching status and the ability to identify abnormal rhythm changes, and enhances the modeling accuracy and early warning capability of ventilation risks.
Smart Images
Figure CN122097769A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, specifically to an early warning system and method for respiratory depression in infants and young children under anesthesia based on multimodal fusion. Background Technology
[0002] In the perioperative management of infants and young children, ventilation support and sedation control are two highly interconnected and mutually influential core aspects. Infants and young children under anesthesia exhibit physiological characteristics such as incomplete airway development and sensitive neural regulation, making their response to respiratory interventions more drastic and their tolerance to ventilatory disturbances narrower. Artifacts caused by the aliasing of respiratory and electrocardiographic signals, ventilation mismatch between tidal volume and airway pressure, and dynamic fluctuations in intraoperative drug concentrations can all induce varying degrees of respiratory regulation imbalance. Conventional assessment methods often rely on single physiological indicators for status determination, failing to consider the fusion of multiple signals and temporal dynamics, thus hindering early risk identification and systematic perception.
[0003] Currently, during ventilation monitoring in infants and young children under surgical anesthesia, there is often a problem of impaired signal continuity of respiratory impedance signals due to electrocardiographic interference, which leads to decreased accuracy of subsequent cycle detection and unstable rhythm feature extraction. The respiratory rhythm status cannot accurately match the interaction between tidal volume and airway pressure, making it difficult to identify potential ventilation imbalance risks and resulting in a lack of targeted strategy support during ventilation regulation.
[0004] Secondly, when assessing the risk of respiratory depression tendencies in infants and young children during surgery, the current method cannot effectively integrate the differences between the real-time concentration of anesthetic drugs and the intensity of stimulation caused by the surgical stage, resulting in the inability to accurately perceive the individual's anesthesia-stimulation response state. Furthermore, the lack of an effective integration path between respiratory rhythm characteristics and key kinetic parameters in the end-tidal carbon dioxide waveform makes it difficult for the current method to model the inhibitory tendency in a multidimensional way, thus limiting the overall effectiveness of the intraoperative early warning mechanism. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention provides an early warning system and method for respiratory depression during infant anesthesia based on multimodal fusion.
[0006] A multimodal fusion-based early warning system for respiratory depression in infants and young children under anesthesia, the system comprising: Data acquisition module S11: used to acquire respiratory impedance signals, electrocardiogram signals, end-tidal carbon dioxide waveforms, tidal volume, airway pressure, real-time concentration of anesthetic drugs, and surgical stage markers; Signal acquisition module S12: used to calculate the instantaneous heart rate based on the continuous R wave interval in the electrocardiogram signal, construct a comb notch filter with the instantaneous heart rate as the fundamental frequency, and filter the respiratory impedance signal to obtain the artifact-free respiratory signal; Data analysis module S13: used to calculate respiratory rhythm characteristics based on artifact-free respiratory signals, tidal volume and airway pressure, and to generate the rising slope, plateau oscillation value and α angle value from the end-expiratory carbon dioxide waveform; Probability generation module S14: is used to calculate the drug-stimulus balance index based on the stimulation intensity coefficient mapped from the real-time concentration of anesthetic drugs to the surgical stage markers. It constructs a time series tensor based on the drug-stimulus balance index, respiratory rhythm characteristics and the slope of the rising branch, the plateau oscillation value and the α angle value, and inputs it into the multimodal fusion model to generate the probability of respiratory depression risk. Furthermore, the steps to obtain the artifact-free breathing signal are as follows: S121 calculates the instantaneous heart rate sequence at each moment by using the continuous R-wave intervals in the electrocardiogram signal to generate instantaneous heart rate data; S122, Construct a comb notch filter with instantaneous heart rate as the fundamental frequency based on instantaneous heart rate data, and obtain the notch filter parameter set; S123 uses the notch filter parameter set to apply to the respiratory impedance signal, performs notch filtering, and generates a respiratory signal after removing ECG artifact interference. S124 outputs the respiratory signal after removing ECG artifact interference as the artifact-free respiratory signal.
[0007] Furthermore, the steps of notch filtering are as follows: S123.1, Based on the center frequency and bandwidth data recorded in the notch filter parameter set, construct the frequency response function of the notch filter and generate the frequency domain model of the filter; S123.2, divides the respiratory impedance signal into multiple overlapping time window segments in the time dimension to generate a time window segmentation signal sequence; S123.3, the frequency domain model of the filter is applied to each time window to divide the signal sequence, and frequency domain filtering operation is performed to obtain a set of filtered result sequences; S123.4, based on the set of filtered result sequences, performs time window splicing and reconstruction processing to generate a continuous respiratory signal after removing ECG artifact interference.
[0008] Furthermore, the steps for generating respiratory rhythm characteristics, rising limb slope, plateau oscillation value, and α angle value are as follows: S131, Based on the continuous change sequence of artifact-free respiratory signals, calculate the start and end times of each complete respiratory cycle to obtain the respiratory cycle time series set; S132, Based on the tidal volume and airway pressure data of each cycle segment in the respiratory cycle time series, calculate the inspiratory time, expiratory time and average pressure change rate of each cycle to generate a respiratory rhythm feature sequence. S133, perform fitting operation based on the curve slope change of each exhalation segment in the end-expiratory carbon dioxide waveform to obtain the set of inflection point positions of the slope change of each segment; S134, calculate the slope of the ascending branch, the oscillation value of the plateau period, and the α angle value according to the set of inflection point positions of slope change, and generate the end-tidal carbon dioxide waveform characteristic sequence.
[0009] Furthermore, the steps for calculating the characteristic sequence of respiratory rhythm are as follows: S132.1, Perform integration calculation based on the tidal volume curve within each respiratory cycle segment to obtain a sequence of tidal volume integral values; S132.2 Calculate the duration of inhalation and exhalation by the start and end times of each respiratory cycle segment, and generate a set of time parameters; S132.3, calculate the ratio between the duration of each period and the integral value of tidal volume in the time parameter set to generate a periodic tidal flow velocity sequence; S132.4 performs covariance calculation based on the periodic tidal flow rate sequence and airway pressure curve to generate a respiratory rhythm characteristic sequence.
[0010] Furthermore, the steps for generating the end-tidal carbon dioxide waveform feature sequence are as follows: S134.1, perform linear fitting based on the concentration change curve of the rising segment of the end-tidal carbon dioxide waveform over time to obtain slope data and record it as the rising branch slope. S134.2 Extract the high-frequency oscillation interval from the carbon dioxide waveform plateau segment, perform mean and fluctuation amplitude analysis, and generate the plateau period oscillation value; S134.3, calculate the angle between the rising segment and the plateau segment based on the point of change in the slope of the carbon dioxide waveform, and obtain the angle value α; S134.4 assembles the rising branch slope, plateau oscillation value and α angle value into a set of triplets in chronological order to generate the end-tidal carbon dioxide waveform characteristic sequence.
[0011] Furthermore, the steps for generating the probability of respiratory depression risk are as follows: S141, perform intensity mapping processing based on the real-time concentration of anesthetic drugs and surgical stage markers to obtain the stimulus intensity coefficient sequence corresponding to each moment; S142, calculates the difference between the stimulus intensity coefficient sequence and the real-time concentration of anesthetic drugs to generate a drug-stimulus balance index sequence; S143, align the drug-stimulation balance index sequence, respiratory rhythm characteristic sequence, rising branch slope sequence, plateau oscillation value sequence and α angle value sequence according to the time axis to construct a time series tensor; S144: Input the time series tensor into the multimodal fusion model, perform model inference operations, and generate the probability of respiratory depression risk.
[0012] Furthermore, the steps for generating the drug-stimulus balance index sequence are as follows: S142.1, Concentration-intensity aligned sequence is constructed by aligning the stimulus intensity coefficient sequence with the real-time concentration sequence of anesthetic drugs on the time axis; S142.2, obtain the corresponding equilibrium offset value sequence by calculating the numerical difference of each pair of data points in the concentration-intensity alignment sequence; S142.3, arrange the difference results of each time point in the balance offset value sequence in chronological order to generate the drug-stimulus balance index sequence.
[0013] Furthermore, the steps for constructing the time series tensor are as follows: S143.1, timestamp-align the drug-stimulus balance index sequence with the respiratory rhythm feature sequence to generate a drug-respiratory feature parallel sequence; S143.2, based on the time axis of the parallel drug-respiratory feature sequence, add the aligned rising branch slope sequence, plateau oscillation value sequence and α angle value sequence to construct a multidimensional feature matrix sequence; S143.3 transforms a multidimensional feature matrix sequence into a multidimensional time series tensor with a uniform step size and time window structure.
[0014] A method for early warning of respiratory depression in infants and young children under anesthesia based on multimodal fusion, applied to any of the aforementioned early warning systems for early warning of respiratory depression in infants and young children under anesthesia based on multimodal fusion, the method comprising: S21: Acquire respiratory impedance signals, electrocardiogram signals, end-tidal carbon dioxide waveform, tidal volume, airway pressure, real-time concentration of anesthetic drugs, and surgical stage markers; S22: Calculate the instantaneous heart rate based on the continuous R wave intervals in the electrocardiogram signal, construct a comb notch filter with the instantaneous heart rate as the fundamental frequency, and filter the respiratory impedance signal to obtain an artifact-free respiratory signal; S23: Calculate respiratory rhythm characteristics based on artifact-free respiratory signals, tidal volume and airway pressure, and generate the rising slope, plateau oscillation value and α angle value from the end-expiratory carbon dioxide waveform; S24: Calculate the drug-stimulus balance index based on the stimulation intensity coefficient mapped from the real-time concentration of anesthetic drugs to the surgical stage markers. Construct a time series tensor based on the drug-stimulus balance index, respiratory rhythm characteristics, rising limb slope, plateau oscillation value, and α angle value. Input the tensor into the multimodal fusion model to generate the respiratory depression risk probability.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention dynamically constructs a set of notch filter parameters based on heart rate changes, and processes respiratory impedance signals using time-division window filtering and weighted splicing mechanisms. This effectively suppresses electrocardiographic interference while preserving the intrinsic characteristics of respiratory motion, thereby improving the continuity and interpretability of the artifact-free respiratory signal. Based on this artifact-free respiratory signal and combined with tidal volume and airway pressure data, a respiratory rhythm feature sequence is further constructed, making the rhythm status assessment stable and responsive. This optimizes the ventilation matching status of infants and young children during anesthesia ventilation and enables the identification of abnormal rhythm changes. Furthermore, this invention introduces a drug-stimulus balance index and combines it with multiple time-series features reflecting respiratory regulation mechanisms to construct a time-series tensor as a fusion input to achieve continuous output of the probability of respiratory depression risk, thereby enhancing the modeling accuracy and early warning capability of ventilation risk during anesthesia, thus improving the risk management strategy for infant anesthesia.
[0016] In summary, this invention achieves early perception and dynamic warning of respiratory depression risk during anesthesia in infants and young children by constructing a time-series tensor that integrates drug concentration, surgical stimulation, and multi-source respiratory dynamics indicators. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0018] Figure 1 This is a block diagram of an infant anesthesia respiratory depression early warning system based on multimodal fusion provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of an early warning method for respiratory depression in infants and young children under anesthesia based on multimodal fusion, provided in Embodiment 2 of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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. Example
[0020] Please see Figure 1As shown in the figure, this embodiment discloses an early warning system for respiratory depression in infants and young children under anesthesia based on multimodal fusion. The system includes: Data acquisition module S11: used to acquire respiratory impedance signals, electrocardiogram signals, end-tidal carbon dioxide waveforms, tidal volume, airway pressure, real-time concentration of anesthetic drugs, and surgical stage markers; First, the connection status of the chest electrode pads in infant patients was obtained. After connecting the electrodes, respiratory impedance signals and electrocardiogram (ECG) signals were acquired in the chest leads using a physiological parameter monitoring module. The respiratory impedance signal was used to reflect the trend of changes in thoracic volume, and the ECG signal was used to extract R-wave position and heart rate changes. The sampling frequency was set to 200Hz, and the sampling duration was consistent with the surgical procedure.
[0021] Next, the end-expiratory carbon dioxide waveform was acquired using a gas sensor located at the end of the respiratory circuit in the respiratory monitoring module. The carbon dioxide waveform was used for subsequent analysis of respiratory dynamic characteristics such as the slope of the ascending limb, the plateau oscillation value, and the α angle. The gas sensor sampling frequency was set to 50Hz, and the waveform period was recorded in segments using a tidal triggering method.
[0022] Subsequently, tidal volume and airway pressure curves were acquired in real time in the ventilator control module. Tidal volume was used to calculate changes in inspiratory and expiratory flow rates, while airway pressure was used to analyze the match between mechanical ventilation and the patient's spontaneous breathing. Tidal volume was measured in milliliters (mL), and airway pressure was measured in cmH2O; both were recorded synchronously for each respiratory cycle.
[0023] Furthermore, the surgical information integration module connects to the operating room anesthesia recording system to obtain real-time concentration information of anesthetic drugs output by the anesthesia pump, including but not limited to the target control concentrations of drugs such as propofol and remifentanil. The concentration unit is μg / mL, and the acquisition cycle is no more than once every 5 seconds.
[0024] Simultaneously, surgical stage markers recorded in the surgical monitoring system, including the trigger times of events such as skin incision, debridement, and suturing, are collected synchronously with drug concentration data via timestamps. These surgical stage markers are used to subsequently deduce the stimulation intensity level of the patient at different surgical stages, supporting drug-stimulus balance analysis.
[0025] This step involves the complete acquisition of seven types of physiological and surgical-related signal data required for this method, and the establishment of a unified time reference for each signal channel to ensure comparability and alignment in the subsequent multimodal feature extraction process.
[0026] Signal acquisition module S12: used to calculate the instantaneous heart rate based on the continuous R wave interval in the electrocardiogram signal, construct a comb notch filter with the instantaneous heart rate as the fundamental frequency, and filter the respiratory impedance signal to obtain the artifact-free respiratory signal; Specifically, the steps to obtain artifact-free breathing signals are as follows: S121 calculates the instantaneous heart rate sequence at each moment by using the continuous R-wave intervals in the electrocardiogram signal to generate instantaneous heart rate data; First, the peak position of the R wave is extracted from the acquired raw electrocardiogram signal. The peak can be located using the fixed threshold method or the derivative difference method. Then, the time interval between any two adjacent R waves is calculated to obtain the instantaneous heart rate value corresponding to each time point. The generated heart rate data forms a time-stamped sequence that can accurately reflect the heart rate level of infants and young children during anesthesia, and has high timeliness.
[0027] This instantaneous heart rate sequence serves as the basic data source for the subsequent design of filter frequency parameters, and its changing trend directly determines the dynamic adjustment capability of the notch filter's fundamental frequency.
[0028] S122, Construct a comb notch filter with instantaneous heart rate as the fundamental frequency based on instantaneous heart rate data, and obtain the notch filter parameter set; In one specific embodiment, based on each heart rate value in the aforementioned time series, a set of notch filter parameters is constructed at its corresponding time point, including parameters such as center frequency, bandwidth range, and number of filter cascades. Each set of parameters revolves around the fundamental frequency represented by the current instantaneous heart rate, forming a comb filter characteristic with the heart rate as the main frequency and its integer multiples as harmonics.
[0029] This step involves adjusting the filter parameter set in real time so that the filter can adapt to the fluctuation characteristics of heart rate, thereby dynamically tracking and suppressing interference signals related to heart rate frequency components, which is suitable for the rapid fluctuations in heart rate of infants and young children.
[0030] The generated notch filter parameter set is arranged in chronological order, providing continuous parameter support for subsequent time-division window filtering operations.
[0031] S123 uses the notch filter parameter set to apply to the respiratory impedance signal, performs notch filtering, and generates a respiratory signal after removing ECG artifact interference. In one specific embodiment, the original respiratory impedance signal is filtered using the generated set of filter parameters. This process employs a time-division window filtering structure to achieve a balance between real-time performance and robustness, and specifically includes the following steps: Specifically, the steps of notch filtering are as follows: S123.1, Based on the center frequency and bandwidth data recorded in the notch filter parameter set, construct the frequency response function of the notch filter and generate the frequency domain model of the filter; Within each time window, notch parameters for the corresponding time period are read, and a set of response functions for frequency domain filtering is constructed to define the degree of signal suppression and amplitude preservation range in different frequency regions. This frequency domain model will serve as the filter template for the current time window to suppress artifact interference caused by frequency components approaching heart rate and its harmonic components.
[0032] S123.2, divides the respiratory impedance signal into multiple overlapping time window segments in the time dimension to generate a time window segmentation signal sequence; To avoid edge effects during filtering and to improve responsiveness to rapidly fluctuating interference, the entire respiratory impedance signal is divided into segments with fixed lengths and overlap ratios. Each time window contains a data segment that partially overlaps with the previous time window to ensure a smooth transition and information extension.
[0033] S123.3, the frequency domain model of the filter is applied to each time window to divide the signal sequence, and frequency domain filtering operation is performed to obtain a set of filtered result sequences; For each time window segment of the signal, the frequency domain filter model constructed for the corresponding time period is called to perform the filtering calculation process and remove the frequency components affected by ECG interference.
[0034] After each calculation is completed, the filtered output of that time window is retained and used as part of the set of filtered result sequences.
[0035] S123.4, based on the set of filtered result sequences, performs time window splicing and reconstruction processing to generate a continuous respiratory signal after removing ECG artifact interference.
[0036] The filtered outputs of all time windows are spliced and reconstructed using a weighted splicing mechanism for overlapping parts, with the weights set based on historical experimental data; This reconstruction method ensures the continuity of the amplitude and trend of the output signal at the splicing point, avoiding the problems of jumps or noise amplification caused by artificial boundaries, thus forming a continuous, stable, and clearly structured artifact-free respiratory signal trajectory with higher physiological interpretability and modeling reliability.
[0037] S124 outputs the respiratory signal after removing ECG artifact interference as the artifact-free respiratory signal.
[0038] Data analysis module S13: used to calculate respiratory rhythm characteristics based on artifact-free respiratory signals, tidal volume and airway pressure, and to generate the rising slope, plateau oscillation value and α angle value from the end-expiratory carbon dioxide waveform; Specifically, the steps for generating respiratory rhythm characteristics, ascending limb slope, plateau oscillation value, and α angle value are as follows: S131, Based on the continuous change sequence of artifact-free respiratory signals, calculate the start and end times of each complete respiratory cycle to obtain the respiratory cycle time series set; In one specific embodiment, the continuous artifact-free respiratory signal is first divided into time segments using a fixed sampling frequency. By analyzing the peak-valley trends of the signal, the initial rise and fall points in the respiratory signal waveform are detected. The start and end times of each respiratory cycle are determined by identifying the boundaries between the inspiratory and expiratory phases. The cycle identification strategy is based on the zero-crossing points and local extrema of amplitude changes. In the presence of irregular fluctuations, a threshold is set based on the number of extrema and the amplitude of fluctuations within a sliding window to filter cycle stability, ultimately forming a set of well-structured respiratory cycle time sequences.
[0039] S132, Based on the tidal volume and airway pressure data of each cycle segment in the respiratory cycle time series, calculate the inspiratory time, expiratory time and average pressure change rate of each cycle to generate a respiratory rhythm feature sequence. In one specific embodiment, based on the start and end time intervals of each respiratory cycle, the tidal volume curve and airway pressure curve for the corresponding time period are extracted to obtain the basic data for rhythm feature calculation. Tidal volume measures the volume of gas exchange during inspiration and expiration, while airway pressure reflects changes in pulmonary ventilation resistance. Respiratory rhythm features are constructed by combining and analyzing these two parameters within the cycle.
[0040] Specifically, the steps for calculating the characteristic sequence of respiratory rhythm are as follows: S132.1, Perform integration calculation based on the tidal volume curve within each respiratory cycle segment to obtain a sequence of tidal volume integral values; In a specific embodiment, the tidal volume change curve over time in each cycle is numerically integrated, and the sampling point data is segmentally integrated using the trapezoidal method. The integration result reflects the actual ventilation volume change in that cycle, and the output is the tidal volume integral value, which constitutes the periodic tidal volume sequence.
[0041] S132.2 Calculate the duration of inhalation and exhalation by the start and end times of each respiratory cycle segment, and generate a set of time parameters; In a specific embodiment, based on the respiratory cycle boundary point in S131, the monotonic change segment of the tidal volume curve within the cycle is analyzed. The time period corresponding to the rise of tidal volume from the lowest point to the highest point is defined as the inspiratory phase, and the time period from the highest point to the lowest point is defined as the expiratory phase. The duration of both is calculated to generate a set of time parameters for the duration of cyclic inspiratory and expiratory phases.
[0042] S132.3, calculate the ratio between the duration of each period and the integral value of tidal volume in the time parameter set to generate a periodic tidal flow velocity sequence; In one specific embodiment, the tidal volume integral value of each cycle is divided by the corresponding inspiratory or expiratory duration to obtain the gas flow rate per unit time, which is expressed as the periodic tidal flow rate. The periodic tidal flow rate sequence can reflect the intensity and rate of change of respiratory rhythm, providing a basis for subsequent variability determination.
[0043] S132.4 performs covariance calculation based on the periodic tidal flow rate sequence and airway pressure curve to generate a respiratory rhythm characteristic sequence.
[0044] In one specific embodiment, a covariance analysis is performed on the tidal flow rate and its corresponding airway pressure curve for each cycle to determine the synchronous fluctuation relationship between the two. If the breathing process is accompanied by a synchronous rise or fall in pressure, the covariance value tends to be positively correlated; if the flow rate fluctuation is opposite to the pressure direction, it is negatively correlated.
[0045] This sequence is used to reflect the degree of matching between ventilation efficiency and respiratory resistance, serving as one of the core indicators of respiratory rhythm characteristics.
[0046] S133, perform fitting operation based on the curve slope change of each exhalation segment in the end-expiratory carbon dioxide waveform to obtain the set of inflection point positions of the slope change of each segment; In one specific embodiment, firstly, based on the expiratory time segment of the respiratory cycle, the corresponding end-expiratory waveform curve is extracted from the carbon dioxide waveform. For the carbon dioxide concentration change over time curve within each expiratory segment, the slope of that segment is calculated using a sliding window linear fitting method.
[0047] The trend of slope value change within a continuous window is used as the fitting curve. The locations where the first derivative changes significantly in the curve are extracted as inflection points of slope change. The set of inflection point locations reflects the key structural changes of the waveform trend at different stages of exhalation, providing a segmentation basis for subsequent feature extraction.
[0048] S134, calculate the slope of the rising branch, the oscillation value of the plateau period and the α angle value respectively according to the set of inflection point positions of slope change, and generate the end-tidal carbon dioxide waveform characteristic sequence; In one specific embodiment, based on the inflection point position extracted in step S133, the end-expiratory waveform curve is divided into typical structural segments such as the rising segment, plateau segment, and falling segment. Target feature parameters are extracted in each sub-segment.
[0049] Specifically, the steps for generating the end-tidal carbon dioxide waveform feature sequence are as follows: S134.1, perform linear fitting based on the concentration change curve of the rising segment of the end-tidal carbon dioxide waveform over time to obtain slope data and record it as the rising branch slope. In one specific embodiment, the least squares method is used to fit a linear model to the rising segment curve defined from the start of exhalation to inflection point 1, and the slope of the fitted line is used as the rising rate of that segment.
[0050] The slope of the rising branch reflects the rate of carbon dioxide release, indirectly corresponding to gas expulsion efficiency. A gentler rising slope typically indicates airway obstruction or delayed ventilation.
[0051] S134.2 Extract the high-frequency oscillation interval from the carbon dioxide waveform plateau segment, perform mean and fluctuation amplitude analysis, and generate the plateau period oscillation value; In one specific embodiment, a waveform plateau interval between the rising inflection point and the falling inflection point is selected, and a bandpass filter is used to extract the residual signal with a frequency higher than the basic respiratory rhythm in this segment.
[0052] The mean and variance of the residual signal are calculated as oscillation amplitude indicators to characterize the stability of gas flow in the plateau segment. Higher oscillation values during the plateau phase indicate stronger disturbances in the end-tidal gas expulsion stage, which may be related to airway instability or instrument interference.
[0053] S134.3, calculate the angle between the rising segment and the plateau segment based on the point of change in the slope of the carbon dioxide waveform, and obtain the angle value α; In a specific embodiment, the linear trend vectors of the rising segment and the plateau segment are first fitted separately, and the angle difference between the two segments is calculated using the vector angle formula. The result is the α angle value.
[0054] The larger the α angle, the more significant the difference between the rate of ascent and the plateau fluctuation trend, which is often used to identify abnormal expiratory resistance or gas exchange disorders.
[0055] S134.4 assembles the rising branch slope, plateau oscillation value and α angle value into a set of triplets in chronological order to generate the end-tidal carbon dioxide waveform characteristic sequence.
[0056] In one specific embodiment, the three waveform feature values generated in each expiratory cycle are assembled into a feature triplet in the order of occurrence. The triplet structure includes: the slope of the rising branch of the cycle, the oscillation value of the plateau period, and the α angle value.
[0057] The feature triples of all expiratory cycles are arranged in chronological order to form a sequence structure, which serves as the feature sequence of end-expiratory carbon dioxide waveform and is used as the input for subsequent construction of the multimodal risk assessment tensor.
[0058] Probability generation module S14: is used to calculate the drug-stimulus balance index based on the stimulation intensity coefficient mapped from the real-time concentration of anesthetic drugs to the surgical stage markers. It constructs a time series tensor based on the drug-stimulus balance index, respiratory rhythm characteristics and the slope of the rising branch, the plateau oscillation value and the α angle value, and inputs it into the multimodal fusion model to generate the probability of respiratory depression risk. In one specific embodiment, drug concentration curves and surgical operation log data from the anesthesia monitoring system are collected in real time and mapped as inputs to a stimulus intensity coefficient; combined with previously calculated respiratory-related features, a multimodal time-series data structure is formed; finally, a fusion inference model is used to generate the risk probability of respiratory depression in infants at the current moment.
[0059] Specifically, the steps for generating the probability of respiratory depression risk are as follows: S141, perform intensity mapping processing based on the real-time concentration of anesthetic drugs and surgical stage markers to obtain the stimulus intensity coefficient sequence corresponding to each moment; In one specific embodiment, based on the timestamp marked at the surgical stage (such as skin incision, traction, suturing, etc.) of the infant and the current concentration data of the anesthetic drug, a stimulation intensity coefficient is assigned to each moment according to a preset stage-intensity mapping rule.
[0060] This rule can be developed based on clinical experience, for example, setting a higher stimulus intensity coefficient during the traction phase and a lower level during the anesthesia maintenance phase. During the mapping process, the standard stimulus intensity value is read using the surgical phase label at the current time point as an index, and then arranged in a time sequence to generate a stimulus intensity coefficient sequence.
[0061] S142, calculates the difference between the stimulus intensity coefficient sequence and the real-time concentration of anesthetic drugs to generate a drug-stimulus balance index sequence; In one specific embodiment, to analyze the degree of matching between the intensity of drug sedation and the intensity of surgical stimulation, a balance index is established by the numerical difference between the two to quantify the relative balance of the anesthesia control state at the current moment.
[0062] Specifically, the steps for generating the drug-stimulus balance index sequence are as follows: S142.1, Concentration-intensity aligned sequence is constructed by aligning the stimulus intensity coefficient sequence with the real-time concentration sequence of anesthetic drugs on the time axis; In one specific embodiment, drug concentration and stimulation intensity are synchronized according to millisecond-level timestamps, and missing points are handled by forward padding to ensure that each moment corresponds to a complete set of concentration and intensity values.
[0063] S142.2, obtain the corresponding equilibrium offset value sequence by calculating the numerical difference of each pair of data points in the concentration-intensity alignment sequence; In one specific embodiment, a time-by-time difference calculation is performed: if the drug concentration is higher than the stimulation intensity, the corresponding difference is positive, indicating sufficient sedation; if the drug concentration is lower than the stimulation intensity, the difference is negative, indicating a possible risk of insufficient sedation.
[0064] S142.3, arrange the difference results of each time point in the balance offset value sequence in chronological order to generate the drug-stimulus balance index sequence.
[0065] In one specific embodiment, the balance offset values at each time point are organized according to time cues to obtain a continuous drug-stimulus balance index sequence, which will serve as an important component in the subsequent construction of temporal features.
[0066] S143, align the drug-stimulation balance index sequence, respiratory rhythm characteristic sequence, rising branch slope sequence, plateau oscillation value sequence and α angle value sequence according to the time axis to construct a time series tensor; In one specific embodiment, to fuse signal information from various sources, physiological indicators from multiple sources are merged into a structured time series tensor in a unified format through time axis alignment and feature integration operations, which is then used as multidimensional input for subsequent models.
[0067] Specifically, the steps for constructing a time series tensor are as follows: S143.1, timestamp-align the drug-stimulus balance index sequence with the respiratory rhythm feature sequence to generate a drug-respiratory feature parallel sequence; In one specific embodiment, the two sequences are first aligned using linear interpolation based on the timestamp sampling period to ensure that each time point has synchronized drug balance indices and respiratory rhythm features. This parallel sequence serves as the basis for subsequent splicing of multimodal features.
[0068] S143.2, based on the time axis of the parallel drug-respiratory feature sequence, add the aligned rising branch slope sequence, plateau oscillation value sequence and α angle value sequence to construct a multidimensional feature matrix sequence; In one specific embodiment, for each time step, five types of feature terms are concatenated: drug-stimulus balance index, inspiratory-to-expiratory ratio, upward slope, plateau oscillation amplitude, and α angle value. The concatenated multidimensional feature matrix sequence is indexed by time, with rows representing time steps and columns representing various feature channels.
[0069] S143.3 transforms a multidimensional feature matrix sequence into a multidimensional time series tensor with a uniform step size and time window structure.
[0070] In one specific embodiment, a sliding step of 5 seconds is used, and a time window of 20 seconds is set to extract continuous segments from the matrix sequence and construct them into a fixed-length tensor structure, which serves as the standard input data format for the multimodal fusion model.
[0071] It should be noted that the tensor dimension structure is maintained as "time steps × number of feature channels", which supports temporal relationship modeling.
[0072] S144: Input the time series tensor into the multimodal fusion model, perform model inference operations, and generate the probability of respiratory depression risk.
[0073] In one specific embodiment, the constructed time-series tensor is input into a pre-trained multimodal fusion neural network model that integrates recurrent neural units and attention mechanisms to capture the interactions and importance between different modal features.
[0074] The model outputs a set of probability values, representing the likelihood of respiratory depression events occurring in the present and short future. These probability values can serve as the triggering basis for clinical early warning systems, initiating further manual interventions or alarm mechanisms. It should be noted that the multimodal fusion model is a deep neural network model with temporal modeling and multi-channel feature interaction capabilities. It can utilize bidirectional gated recurrent networks based on attention mechanisms, temporal convolutional networks, Transformer structures with multi-head attention mechanisms, or lightweight fusion architectures integrating the outputs of multiple sub-models. This model can effectively align and assign weights between features across different modalities of the drug-stimulus balance index, respiratory rhythm characteristics, and end-tidal carbon dioxide in the temporal dimension. Through parameter fitting and optimization during the training phase, it outputs probability values reflecting the risk of respiratory depression in infants under anesthesia during actual inference, supporting real-time assessment of risk trends over continuous time periods. Example
[0075] Please see Figure 2 As shown, based on a unified inventive concept, this embodiment discloses a method for early warning of respiratory depression in infants and young children under anesthesia based on multimodal fusion, the method comprising: S21: Acquire respiratory impedance signals, electrocardiogram signals, end-tidal carbon dioxide waveform, tidal volume, airway pressure, real-time concentration of anesthetic drugs, and surgical stage markers; S22: Calculate the instantaneous heart rate based on the continuous R wave intervals in the electrocardiogram signal, construct a comb notch filter with the instantaneous heart rate as the fundamental frequency, and filter the respiratory impedance signal to obtain an artifact-free respiratory signal; S23: Calculate respiratory rhythm characteristics based on artifact-free respiratory signals, tidal volume and airway pressure, and generate the rising slope, plateau oscillation value and α angle value from the end-expiratory carbon dioxide waveform; S24: Calculate the drug-stimulus balance index based on the stimulation intensity coefficient mapped from the real-time concentration of anesthetic drugs to the surgical stage markers. Construct a time series tensor based on the drug-stimulus balance index, respiratory rhythm characteristics, rising limb slope, plateau oscillation value, and α angle value. Input the tensor into the multimodal fusion model to generate the respiratory depression risk probability.
[0076] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.
Claims
1. A multimodal fusion-based early warning system for respiratory depression in infants and young children under anesthesia, characterized in that, The system includes: Data acquisition module S11: used to acquire respiratory impedance signals, electrocardiogram signals, end-tidal carbon dioxide waveforms, tidal volume, airway pressure, real-time concentration of anesthetic drugs, and surgical stage markers; Signal acquisition module S12: used to calculate the instantaneous heart rate based on the continuous R wave interval in the electrocardiogram signal, construct a comb notch filter with the instantaneous heart rate as the fundamental frequency, and filter the respiratory impedance signal to obtain the artifact-free respiratory signal; Data analysis module S13: used to calculate respiratory rhythm characteristics based on artifact-free respiratory signals, tidal volume and airway pressure, and to generate the rising slope, plateau oscillation value and α angle value from the end-expiratory carbon dioxide waveform; Probability generation module S14: It is used to calculate the drug-stimulus balance index based on the stimulation intensity coefficient mapped from the real-time concentration of anesthetic drugs to the surgical stage markers. Based on the drug-stimulus balance index, respiratory rhythm characteristics and the slope of the rising branch, the plateau oscillation value and the α angle value, a time series tensor is constructed and input into the multimodal fusion model to generate the probability of respiratory depression risk.
2. The infant anesthesia respiratory depression early warning system based on multimodal fusion according to claim 1, characterized in that, The steps to obtain artifact-free breathing signals are as follows: S121 calculates the instantaneous heart rate sequence at each moment by using the continuous R-wave intervals in the electrocardiogram signal to generate instantaneous heart rate data; S122, Construct a comb notch filter with instantaneous heart rate as the fundamental frequency based on instantaneous heart rate data, and obtain the notch filter parameter set; S123 uses the notch filter parameter set to apply to the respiratory impedance signal, performs notch filtering, and generates a respiratory signal after removing ECG artifact interference. S124 outputs the respiratory signal after removing ECG artifact interference as the artifact-free respiratory signal.
3. The infant anesthesia respiratory depression early warning system based on multimodal fusion according to claim 2, characterized in that, The steps of notch filtering are as follows: S123.1, Based on the center frequency and bandwidth data recorded in the notch filter parameter set, construct the frequency response function of the notch filter and generate the frequency domain model of the filter; S123.2, divides the respiratory impedance signal into multiple overlapping time window segments in the time dimension to generate a time window segmentation signal sequence; S123.3, the frequency domain model of the filter is applied to each time window to divide the signal sequence, and frequency domain filtering operation is performed to obtain a set of filtered result sequences; S123.4, based on the set of filtered result sequences, performs time window splicing and reconstruction processing to generate a continuous respiratory signal after removing ECG artifact interference.
4. The infant anesthesia respiratory depression early warning system based on multimodal fusion according to claim 3, characterized in that, The steps for generating respiratory rhythm characteristics, rising limb slope, plateau oscillation value, and α angle value are as follows: S131, Based on the continuous change sequence of artifact-free respiratory signals, calculate the start and end times of each complete respiratory cycle to obtain the respiratory cycle time series set; S132, Based on the tidal volume and airway pressure data of each cycle segment in the respiratory cycle time series, calculate the inspiratory time, expiratory time and average pressure change rate of each cycle to generate a respiratory rhythm feature sequence. S133, perform fitting operation based on the curve slope change of each exhalation segment in the end-expiratory carbon dioxide waveform to obtain the set of inflection point positions of the slope change of each segment; S134, calculate the slope of the ascending branch, the oscillation value of the plateau period, and the α angle value according to the set of inflection point positions of slope change, and generate the end-tidal carbon dioxide waveform characteristic sequence.
5. The infant anesthesia respiratory depression early warning system based on multimodal fusion according to claim 4, characterized in that, The steps for calculating the characteristic sequence of respiratory rhythm are as follows: S132.1, Perform integration calculation based on the tidal volume curve within each respiratory cycle segment to obtain a sequence of tidal volume integral values; S132.2 Calculate the duration of inhalation and exhalation by the start and end times of each respiratory cycle segment, and generate a set of time parameters; S132.3, calculate the ratio between the duration of each period and the integral value of tidal volume in the time parameter set to generate a periodic tidal flow velocity sequence; S132.4 performs covariance calculation based on the periodic tidal flow rate sequence and airway pressure curve to generate a respiratory rhythm characteristic sequence.
6. The infant anesthesia respiratory depression early warning system based on multimodal fusion according to claim 5, characterized in that, The steps for generating the end-tidal carbon dioxide waveform feature sequence are as follows: S134.1, perform linear fitting based on the concentration change curve of the rising segment of the end-tidal carbon dioxide waveform over time to obtain slope data and record it as the rising branch slope. S134.2 Extract the high-frequency oscillation interval from the carbon dioxide waveform plateau segment, perform mean and fluctuation amplitude analysis, and generate the plateau period oscillation value; S134.3, calculate the angle between the rising segment and the plateau segment based on the point of change in the slope of the carbon dioxide waveform, and obtain the angle value α; S134.4 assembles the rising branch slope, plateau oscillation value and α angle value into a set of triplets in chronological order to generate the end-tidal carbon dioxide waveform characteristic sequence.
7. The infant anesthesia respiratory depression early warning system based on multimodal fusion according to claim 6, characterized in that, The steps to generate the probability of respiratory depression risk are as follows: S141, perform intensity mapping processing based on the real-time concentration of anesthetic drugs and surgical stage markers to obtain the stimulus intensity coefficient sequence corresponding to each moment; S142, calculates the difference between the stimulus intensity coefficient sequence and the real-time concentration of anesthetic drugs to generate a drug-stimulus balance index sequence; S143, align the drug-stimulation balance index sequence, respiratory rhythm characteristic sequence, rising branch slope sequence, plateau oscillation value sequence and α angle value sequence according to the time axis to construct a time series tensor; S144: Input the time series tensor into the multimodal fusion model, perform model inference operations, and generate the probability of respiratory depression risk.
8. The infant anesthesia respiratory depression early warning system based on multimodal fusion according to claim 7, characterized in that, The steps for generating the drug-stimulus balance index sequence are as follows: S142.1, Concentration-intensity aligned sequence is constructed by aligning the stimulus intensity coefficient sequence with the real-time concentration sequence of anesthetic drugs on the time axis; S142.2, obtain the corresponding equilibrium offset value sequence by calculating the numerical difference of each pair of data points in the concentration-intensity alignment sequence; S142.3, arrange the difference results of each time point in the balance offset value sequence in chronological order to generate the drug-stimulus balance index sequence.
9. The infant anesthesia respiratory depression early warning system based on multimodal fusion according to claim 8, characterized in that, The steps to construct a time series tensor are as follows: S143.1, timestamp-align the drug-stimulus balance index sequence with the respiratory rhythm feature sequence to generate a drug-respiratory feature parallel sequence; S143.2, based on the time axis of the parallel drug-respiratory feature sequence, add the aligned rising branch slope sequence, plateau oscillation value sequence and α angle value sequence to construct a multidimensional feature matrix sequence; S143.3 transforms a multidimensional feature matrix sequence into a multidimensional time series tensor with a uniform step size and time window structure.
10. A method for early warning of respiratory depression in infants and young children under anesthesia based on multimodal fusion, applied to the early warning system for early warning of respiratory depression in infants and young children under anesthesia based on multimodal fusion as described in any one of claims 1-9, characterized in that, The method includes: S21: Acquire respiratory impedance signals, electrocardiogram signals, end-tidal carbon dioxide waveform, tidal volume, airway pressure, real-time concentration of anesthetic drugs, and surgical stage markers; S22: Calculate the instantaneous heart rate based on the continuous R wave intervals in the electrocardiogram signal, construct a comb notch filter with the instantaneous heart rate as the fundamental frequency, and filter the respiratory impedance signal to obtain an artifact-free respiratory signal; S23: Calculate respiratory rhythm characteristics based on artifact-free respiratory signals, tidal volume and airway pressure, and generate the rising slope, plateau oscillation value and α angle value from the end-expiratory carbon dioxide waveform; S24: Calculate the drug-stimulus balance index based on the stimulation intensity coefficient mapped from the real-time concentration of anesthetic drugs to the surgical stage markers. Construct a time series tensor based on the drug-stimulus balance index, respiratory rhythm characteristics, rising limb slope, plateau oscillation value, and α angle value. Input the tensor into the multimodal fusion model to generate the respiratory depression risk probability.