Risk Trend Prediction System and Method Based on Multimodal Physiological Data

Abstract

This invention belongs to the field of risk prediction technology and discloses a risk trend prediction system and method based on multimodal physiological data. The method includes: acquiring real-time dual-frequency impedance data of the patient and calculating the ratio of the dual-frequency impedance data to obtain the dual-frequency ratio; outputting a primary warning result based on temporal changes; acquiring the patient's heart sound signals, and when the primary warning result is 1, preprocessing and feature analysis of the heart sound signals to obtain heart sound features, and performing pathological judgment on the heart sound features to obtain a pathological confirmation result; acquiring the patient's triaxial acceleration data, and when the pathological confirmation result is 1, calculating the patient's postural state, and combining real-time dual-frequency impedance and heart sound signals to analyze and obtain the patient's respiratory rate sequence, detecting and analyzing respiratory markers, and obtaining the respiratory compensation level based on continuous monitoring and analysis of respiratory markers; and finally comprehensively determining the final cardiopulmonary risk level. This invention can highly specifically identify slow physiological changes in physical condition and the early development trend of acute cardiogenic pulmonary edema.

Description

Technical Field

[0001] This invention relates to the field of risk prediction technology, and more specifically, to a risk trend prediction system and method based on multimodal physiological data. Background Technology

[0002] Existing cardiopulmonary risk warning systems primarily focus on immediate alerts for acute events, but their ability to accurately predict early, slow, and progressive deterioration trends is insufficient. In everyday home settings, physiological changes in body composition due to fat loss or muscle growth, as well as increased pleural effusion, both drive signal changes. The signal changes caused by these two conditions are highly similar. Furthermore, other non-heart failure conditions such as pneumonia and kidney failure can also cause changes in pleural effusion, thus driving signal changes. Existing cardiopulmonary risk warning systems cannot accurately isolate the pathological component purely stemming from cardiac decompensation from the signal changes caused by physiological changes in body composition due to fat loss or muscle growth and increased pleural effusion. This significantly reduces the reliability and clinical guidance value of the final cardiopulmonary risk trend prediction results.

[0003] In view of this, the present invention proposes a risk trend prediction system and method based on multimodal physiological data to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a risk trend prediction method based on multimodal physiological data, comprising:

[0005] The system collects real-time dual-frequency impedance data from patients, calculates the ratio of the dual-frequency impedances to obtain the dual-frequency ratio, monitors the temporal changes in the dual-frequency ratio, and outputs a primary warning result based on the temporal changes.

[0006] The patient's heart sound signals are collected. When the primary warning result is 1, the heart sound signals are preprocessed and feature analyzed to obtain heart sound features. The heart sound features are then pathologically determined to obtain pathological confirmation results.

[0007] Triaxial acceleration data of the patient was collected. When the pathological confirmation result was 1, the angle was calculated based on the triaxial acceleration data to obtain the patient's positional status. The patient's respiratory rate sequence was extracted and generated by combining real-time dual-frequency impedance, heart sound signal and triaxial acceleration data. The respiratory rate sequence was scanned and identified by a sliding time window. The respiratory markers were identified and obtained by combining the patient's positional status. The respiratory markers were continuously monitored, comprehensively evaluated and graded to obtain the respiratory compensation level.

[0008] The final cardiopulmonary risk level is determined based on the initial warning results, pathological confirmation results, and respiratory compensation level.

[0009] Furthermore, methods for obtaining preliminary early warning results include:

[0010] The first impedance and the second impedance are collected, and the ratio of the first impedance and the second impedance is calculated to obtain the dual-frequency ratio. When the dual-frequency ratio continues to rise during the monitoring period and exceeds the static threshold, the primary warning result is output as 1; otherwise, the primary warning result is output as 0.

[0011] Furthermore, methods for obtaining heart sound characteristics include:

[0012] The heart sound signal is bandpass filtered to obtain the filtered audio. The filtered audio is used as the input to the heart sound segmentation model to obtain the first heart sound position and the second heart sound position. The filtered audio, the first heart sound position, the second heart sound position, and the detection duration are used as the input to the anomaly analysis model to obtain the probability of the third heart sound and the indicator of the fourth heart sound. The probability of the third heart sound is used as the heart sound feature.

[0013] Furthermore, methods for obtaining pathological confirmation results include:

[0014] Within the warning period corresponding to a primary warning result of 1, the total number of times the probability of the third heart sound exceeds the probability threshold is counted. The ratio of the total number of times to the warning period is calculated to obtain the frequency of the probability of the third heart sound exceeding the probability threshold per unit time. If the frequency of the probability of the third heart sound exceeding the probability threshold per unit time exceeds the clinical threshold, the pathological confirmation result is output as 1; otherwise, the pathological confirmation result is output as 0.

[0015] Furthermore, methods for obtaining respiratory rate sequences include:

[0016] Real-time dual-frequency impedance, heart sound signal and triaxial acceleration data were detected and analyzed to obtain the first time series, the second time series and the third time series.

[0017] The time interval between adjacent time points in the three time series is calculated as the instantaneous respiratory cycle of the corresponding channel. Taking the time point t of any one signal as the reference signal, it is checked whether there are peak time points in the other two signals within the detection time window of the reference signal. If at least two of the three signals have peak time points within the detection time window, it is determined that the three signals have captured the same valid respiratory event. The average value of all time points within the detection time window is taken as a fused respiratory event time point. Any misaligned isolated time points are removed to obtain the fused respiratory event time point sequence. The time difference between two adjacent fused respiratory event time points is calculated to obtain the respiratory cycle. The ratio of one minute duration to the time difference between two adjacent fused respiratory event time points is calculated to obtain the respiratory rate sequence.

[0018] Furthermore, methods for obtaining respiratory markers include:

[0019] The impedance-filtered signal corresponding to the respiratory rate sequence is obtained and normalized to obtain the respiratory waveform amplitude sequence. A sliding time window is used to identify troughs and peaks, and the evaluation period is marked. If the evaluation period that meets the first quantization rule appears consecutively, it is marked as Cheyne-Stokes respiration. If the average value of the respiratory rate sequence and the average value of the respiratory waveform amplitude sequence meet the second quantization rule, it is marked as rapid shallow breathing. The impedance-filtered signal, filtered heart sound signal, and filtered kinetic acceleration signal are filtered, and apnea events are obtained by time overlap determination. The difference between the end time and the start time of all apnea events is calculated to obtain the event duration. Based on the event duration and the event waveform amplitude, it is marked as apnea or hypoventilation. The pitch and roll angles are calculated using triaxial acceleration data to determine the body position. When the body position changes directly from supine to upright, it is marked as orthopnea. The body position includes supine, left lateral, right lateral, upright, and active phases.

[0020] Furthermore, methods for obtaining apnea events include:

[0021] The signal segment whose absolute value of the respiratory waveform amplitude of any signal is consistently lower than the amplitude threshold of the corresponding signal is taken as a candidate low amplitude segment. The start and end times of the candidate low amplitude segments are recorded to obtain a list of time intervals.

[0022] Based on the time synchronization tolerance, determine whether the candidate low amplitude segments of the three signals overlap in time; only when at least two signals have candidate low amplitude segments overlapping in time during the screening period, the overlapping period is determined as a valid pause event.

[0023] If the interval between two valid pause events is less than the interval threshold, the two valid pause events are merged into one apnea event. The smaller start time of the two valid pause events is marked as the start time of the apnea event, and the larger end time of the two valid pause events is marked as the end time of the apnea event; otherwise, the valid pause event is directly treated as an apnea event.

[0024] Furthermore, methods for obtaining respiratory compensation levels include:

[0025] If Cheyne-Stokes respiration is detected and lasts for more than Q minutes; or if the first average respiratory rate within the supine window before the switching time is calculated; and the second average respiratory rate within the detection window after the switching time is calculated; and the difference between the first and second averages is calculated, and then the ratio of the difference to the first average is calculated to obtain the respiratory relief rate, and the respiratory relief rate is greater than the relief rate threshold, and within the detection window, an assessment sequence is constructed by collecting respiratory rates within the assessment window during the detection period, and the respiratory rates in the assessment sequence greater than the first average exceed the proportion threshold; then it is determined to be a Class A mode.

[0026] If rapid, shallow breathing is detected, with a respiratory rate exceeding X breaths per minute and a duration exceeding M minutes; or if a breathing apnea event is detected, it is classified as Level B mode.

[0027] Otherwise, it will be classified as Level C mode;

[0028] If any Grade A pattern is detected during the assessment period, it is directly determined as severe respiratory compensation;

[0029] If no Class A mode is detected, but one or more Class B modes are detected, and the total cumulative duration of Class B modes exceeds the percentage threshold, then moderate respiratory compensation is determined to exist.

[0030] Otherwise, it is determined as no respiratory compensation.

[0031] Furthermore, methods for obtaining cardiopulmonary risk levels include:

[0032] When the initial warning result is 1, the pathological confirmation result is 1, and the respiratory compensation level is severe or moderate, it is judged as high risk, and the corresponding respiratory compensation level is output.

[0033] When the initial warning result of the test is 1, the pathological confirmation result is 1, and the respiratory compensation level is no respiratory compensation, it is judged as medium risk;

[0034] When only the initial warning result is 1, it is judged as low risk;

[0035] Otherwise, it is considered normal.

[0036] The risk trend prediction system based on multimodal physiological data, implementing the risk trend prediction method based on multimodal physiological data, includes:

[0037] Anomaly detection module: Acquires real-time dual-frequency impedance of the patient, calculates the ratio of the dual-frequency impedance to obtain the dual-frequency ratio; monitors the temporal changes of the dual-frequency ratio, and outputs a primary warning result based on the temporal changes;

[0038] Pathological confirmation module: Collects the patient's heart sound signals. When the primary warning result is 1, the heart sound signals are preprocessed and feature analyzed to obtain heart sound features. The heart sound features are then used for pathological determination to obtain the pathological confirmation result.

[0039] Pattern detection module: Collects triaxial acceleration data of the patient. When the pathological confirmation result is 1, it calculates the angle based on the triaxial acceleration data to obtain the patient's position. Combined with real-time dual-frequency impedance, heart sound signal and triaxial acceleration data, it extracts and generates the patient's respiratory rate sequence. It performs pattern scanning and recognition on the respiratory rate sequence through a sliding time window. Combined with the patient's position, it identifies and marks respiratory markers. Based on the respiratory markers, it performs continuous monitoring, comprehensive evaluation and grade determination to obtain the respiratory compensation level.

[0040] Risk Analysis Module: Determines the final cardiopulmonary risk level based on the initial warning results, pathological confirmation results, and respiratory compensation level.

[0041] The technical effects and advantages of the risk trend prediction system and method based on multimodal physiological data of this invention are as follows:

[0042] This invention collects real-time dual-frequency impedance data from patients and calculates its ratio. Monitoring the temporal changes in this ratio outputs a primary warning result. The continuous upward trend of the dual-frequency impedance ratio captures the early physical trajectory of a slow increase in pleural fluid load. When the primary warning result is 1, this trend signal automatically triggers a second layer of verification, namely, pathological assessment of heart sound signals to confirm whether the fluid trend is accompanied by a specific pathological mechanism of heart failure. Heart sound signals are collected, preprocessed, feature-analyzed, and pathologically assessed to obtain a pathological confirmation result. When the pathological confirmation result is 1, triaxial acceleration data is collected to calculate the patient's position. Simultaneously, real-time dual-frequency impedance, heart sound signals, and triaxial acceleration data are integrated to extract and generate a respiratory rate sequence. A sliding time window scan identifies respiratory markers and determines the respiratory compensation level. Finally, combining the primary warning result, pathological confirmation result, and respiratory compensation level, cardiopulmonary risk is classified into four categories: high, medium, low, and normal. This highly specific approach can identify slow physiological changes in physical condition and the early development trend of acute cardiogenic pulmonary edema, improving the reliability of risk prediction and its clinical guidance value. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the risk trend prediction method based on multimodal physiological data of the present invention;

[0044] Figure 2 This is a schematic flowchart of the method for obtaining respiratory rate sequences according to the present invention;

[0045] Figure 3 This is a schematic diagram of the method for obtaining respiratory markers according to the present invention;

[0046] Figure 4 This is a schematic diagram of the risk trend prediction system based on multimodal physiological data according to the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1:

[0049] Please see Figure 1 As shown, this embodiment provides a risk trend prediction method based on multimodal physiological data, including:

[0050] The system collects real-time dual-frequency impedance data from patients, calculates the ratio of the dual-frequency impedances to obtain the dual-frequency ratio, monitors the temporal changes in the dual-frequency ratio, and outputs a primary warning result based on the temporal changes.

[0051] Methods for obtaining preliminary warning results include:

[0052] Low-frequency and high-frequency sinusoidal alternating currents are injected into the patient's pleural cavity via a chest patch. First and second impedances are collected, with the low-frequency sinusoidal alternating current preferably at 30 kHz and the high-frequency sinusoidal alternating current preferably at 90 kHz. The ratio of the first and second impedances is calculated to obtain the dual-frequency ratio. The values ​​of the first and second impedances can be determined through sensitivity simulation and experiments to identify candidate frequencies. For example, a simplified electrical model of the pleural cavity can be constructed, based on the thoracic anatomy, establishing a two-dimensional axisymmetric or three-dimensional simplified layered model. This model, from the outside in, should contain at least the following concentric layered structures:

[0053] A skin layer with fixed conductivity and dielectric constant;

[0054] The subcutaneous fat layer used to simulate changes in body constitution such as obesity and emaciation has a variable thickness.

[0055] The thickness of the chest wall muscle layer used to simulate changes in muscle mass is variable.

[0056] Lung tissue layer: The lung tissue layer is modeled as a composite material in which a variable proportion of liquid components are uniformly mixed in a background medium; the liquid proportion is an adjustable key parameter; the background medium has low electrical conductivity to simulate healthy air-containing lung tissue; the variable proportion of liquid components has high electrical conductivity to simulate pulmonary edema effusion.

[0057] For tissues such as fat, muscle, and skin, the frequency-varying parameters corresponding to the tissue type can be directly used from publicly available databases of biological tissue electrical properties.

[0058] For the lung tissue layer, a mixed media theory is used for modeling, including:

[0059] The electrical conductivity of healthy lung tissue is very low. The composite conductivity of the matrix phase corresponding to healthy lung tissue is fitted using the electrical conductivity of healthy lung tissue and its relative permittivity. For example, the composite conductivity of the matrix phase corresponding to healthy lung tissue... ;in, The electrical conductivity of healthy lung tissue; The imaginary unit; Angular frequency; It is the vacuum permittivity; The relative permittivity of healthy lung tissue;

[0060] The composite conductivity of the encapsulated phase corresponding to the body fluid is fitted by the body fluid conductivity and the relative permittivity of the body fluid; for example, the composite conductivity of the encapsulated phase corresponding to the body fluid. ;in, For body fluid conductivity; The relative permittivity of body fluids;

[0061] Set a liquid volume fraction parameter between 0 and 1. , Represents healthy lungs. The increase represents the gradual accumulation of liquid; the equivalent complex conductivity is calculated using the effective medium theory formula; such as the equivalent complex conductivity. Separate the real part from the equivalent complex conductivity as the equivalent conductivity of the model, and divide the imaginary part by... Thus, the equivalent relative permittivity is obtained;

[0062] Equivalent conductivity and equivalent relative permittivity are assigned to the lung tissue layers in the simplified electrical model of the thoracic cavity.

[0063] In a simplified electrical model of the thoracic cavity, an excitation port corresponding to the thoracic cavity patch electrode was set. Electromagnetic field simulation was used to calculate parameters at different frequencies f and different fluid volume fractions. And the complex impedance Z(f) of the excitation port was measured under different fat / muscle layer thicknesses.

[0064] The simplified electrical model of the thoracic cavity is configured with electrically insulating or radiative boundary conditions to simulate the body's state in air. Two small regions on the model surface corresponding to the locations of the thoracic patch electrodes are applied with alternating current sources of equal amplitude but opposite phase to simulate the injection of current into the thoracic cavity by the patch. The port impedance between the two excitation electrodes or in a four-electrode mode is then calculated; the port impedance is the ratio of the measured voltage to the injected current.

[0065] Set scan variables, including:

[0066] Frequency: Take multiple points logarithmically evenly within the target range, such as 10kHz-200kHz.

[0067] Physiological / pathological parameters:

[0068] Simulated pulmonary edema: Gradually increasing the fluid volume fraction parameter of the lung tissue layer For example, increasing from 0.05 to 0.30.

[0069] Simulate obesity / muscle growth: Proportionally increase the thickness of the fat layer and / or muscle layer.

[0070] Run the simulation: For each group (f, Calculate the port impedance Z(f) using the thickness parameter.

[0071] For each frequency, calculate when The rate of change of port impedance when it increases; such as when As the impedance increases, the rate of change of impedance ;in, For a fixed frequency f, when The change in the impedance of the thoracic port, calculated through electromagnetic field simulation, when the impedance increases; For when Increased port impedance; for each frequency, calculate the rate of change of port impedance as tissue thickness increases: e.g., the rate of change of impedance as tissue thickness increases. ;in, The change in the impedance of the thoracic cavity port, calculated by electromagnetic field simulation, is the amount of change in the tissue thickness at a fixed frequency f. This represents the port impedance as the tissue thickness increases.

[0072] Find two frequencies and This makes the liquid sensitivity ratio Maximize, while ensuring the ratio of tissue interference Approaching 1. Systematic scanning analysis revealed that at 30kHz, the current path flows more through the extracellular fluid, showing a clear response to changes in overall tissue structure and composition, such as fat layer thickening; at 90kHz, the current penetrates the cell membrane more easily, becoming extremely sensitive to changes in the total fluid content within the tissue, especially the increase in free ions. When pulmonary edema occurs, the increased ion concentration in the fluid leads to a sharp decrease in impedance at 90kHz, while the impedance change at 30kHz is relatively small, thus significantly increasing the impedance ratio. This leads to the conclusion that... and It is an effective and engineering-friendly combination that satisfies this optimization objective.

[0073] When the dual-frequency ratio continuously rises and exceeds the static threshold during the monitoring period, a primary warning result of 1 is output; otherwise, a primary warning result of 0 is output. The static threshold is set based on the average dual-frequency ratio of a normal healthy population, which is collected with the consent of the participants. It can be adjusted later based on individual patient baseline data. Utilizing the difference in frequency response between 30kHz and 90kHz currents to the conductivity of different tissue components such as fat and muscle, and body fluids, the ratio is calculated to physically separate the trends of increased impedance due to tissue thickening and decreased impedance due to increased fluid volume. By filtering out the interfering components of physiological changes such as fat reduction or muscle growth from the original, highly similar composite signals, a specific physical trend signal indicating an increase in pleural fluid volume is output, providing a clean data starting point for subsequent pathological analysis.

[0074] Heart sound signals from patients are collected. When the initial warning result is 1, the heart sound signals are preprocessed and feature analyzed to obtain heart sound characteristics. These characteristics are then pathologically assessed to obtain pathological confirmation results. Even with confirmed fluid overload, the patient's etiology may still be multifaceted, such as impedance changes due to non-cardiac causes like pneumonia, renal failure, or heart failure. This step involves detecting the S3 gallop rhythm, a pathological feature specific to heart failure and directly reflecting left ventricular diastolic dysfunction, and its frequency, to attach an etiological label to the aforementioned fluid signals. This strongly correlates the phenomenon of fluid overload with the specific pathological mechanism of cardiac decompensation, effectively excluding interfering signals caused by pleural effusion due to other non-heart failure diseases such as pneumonia and renal failure. At this point, the system has preliminarily isolated the target component of increased cardiac fluid load.

[0075] Methods for obtaining heart sound characteristics include:

[0076] The heart sound signal is bandpass filtered to obtain the filtered audio. The upper and lower limits of the bandpass filter are set based on empirical values ​​of heart sounds, with the preferred bandpass filter range being 20-150Hz, which can be adjusted according to actual conditions. The filtered audio is used as input to the heart sound segmentation model to obtain the positions of the first and second heart sounds. The filtered audio, the positions of the first and second heart sounds, and the detection duration are used as input to the anomaly analysis model to obtain the probability of the presence of the third heart sound and the indicator of the presence of the fourth heart sound. The probability of the presence of the third heart sound is used as a heart sound feature. The first heart sound is the systolic start sound, specifically the vibration produced by the sudden closure of the atrioventricular valves at the beginning of ventricular contraction, marking the start of systole. The intensity variation of the first heart sound can reflect myocardial contractility or valvular condition and serves as a benchmark anchor point for dividing the cardiac cycle. The second heart sound is the diastolic start sound, specifically the vibration produced by the closure of the semilunar valves at the beginning of ventricular diastole, marking the start of diastole. The second heart sound is present under normal conditions. The splitting of the second heart sound can reflect pulmonary artery pressure or cardiac conduction, and is a key point in distinguishing systole from diastole. The third heart sound is an early diastolic gallop rhythm, specifically during early ventricular diastole, when blood rapidly fills the ventricles from the atria, impacting the poorly relaxed and less compliant ventricular walls, causing vibrations. If the third heart sound occurs within 0.12-0.18 seconds after the second heart sound, its occurrence in adults is often pathological. The third heart sound is a specific acoustic marker of ventricular dysfunction leading to increased ventricular filling resistance and hemodynamic abnormalities. The fourth heart sound is a late diastolic gallop rhythm / atrial sound, specifically during end-diastole, when atrial contraction forces blood into the stiffer and less compliant ventricles, producing vibrations immediately before the first heart sound of the next cycle. The fourth heart sound usually indicates impaired ventricular diastolic function, myocardial hypertrophy, or ischemia. The presence of the fourth heart sound can help assess the state of cardiac diastolic function.

[0077] Training methods for heart sound segmentation models include:

[0078] Heart sound training data for group B1 was collected in advance. The heart sound training data included filtered audio and the corresponding first and second heart sound positions.

[0079] The filtered audio is used as input to the heart sound segmentation model, and the corresponding first and second heart sound positions are used as outputs. The goal is to minimize the error between the output first and second heart sound positions and their actual values. The network parameters of the heart sound segmentation model are iteratively optimized to obtain the network parameters that minimize the error between the output first and second heart sound positions and their actual values. The heart sound segmentation model constructed using these network parameters is then used as the trained heart sound segmentation model. A CNN network is preferred as the heart sound segmentation model.

[0080] Training methods for anomaly analysis models include:

[0081] Pre-collected abnormal training data from group B2, which included filtered audio, the location of the first heart sound, the location of the second heart sound and the detection duration, as well as the corresponding probability of the presence of the third heart sound and the presence flag of the fourth heart sound.

[0082] The filtered audio, the positions of the first and second heart sounds, and the detection duration are used as inputs to the anomaly analysis model. The corresponding probability of the third and fourth heart sounds is used as the output. The goal is to minimize the error between the output probability and the actual probability and value of the third and fourth heart sounds. The network parameters of the anomaly analysis model are iteratively optimized to obtain the network parameters that minimize the error between the output probability and the actual probability and value of the third and fourth heart sounds. The anomaly analysis model constructed using these network parameters is then used as the trained anomaly analysis model. An LSTM network is preferred as the anomaly analysis model.

[0083] By focusing the key frequency band of heart sounds through a 20-150Hz bandpass filter, locating the first and second heart sounds using a heart sound segmentation model, and then extracting the probability of the existence of the third heart sound through an anomaly analysis model, impedance abnormalities are associated with cardiac pathological features, which can preliminarily rule out non-cardiac impedance changes.

[0084] Methods for obtaining pathological confirmation results include:

[0085] Within the warning period corresponding to a primary warning result of 1, the total number of times the probability of the third heart sound exceeding the probability threshold is counted. The ratio of the total number of times to the warning period is calculated to obtain the frequency of the probability of the third heart sound exceeding the probability threshold per unit time. If the frequency of the probability of the third heart sound exceeding the probability threshold per unit time exceeds the clinical threshold, the pathological confirmation result is output as 1; otherwise, the pathological confirmation result is output as 0. The probability threshold is preferably 0.8. The clinical threshold is obtained based on clinical data statistics and can be adjusted later according to the patient's individual baseline data.

[0086] By statistically analyzing the frequency of the third heart sound exceeding 0.8 during the warning period, the qualitative pathological heart sound characteristics are transformed into quantitative indicators. Combined with clinical thresholds, the presence of cardiac pathological signs is determined, and simple changes in tissue composition are initially ruled out, i.e., impedance changes caused by the absence of a third heart sound abnormality.

[0087] The patient's triaxial acceleration data was collected. When the pathological confirmation result was 1, the angle was calculated based on the triaxial acceleration data to obtain the patient's position. The patient's respiratory rate sequence was extracted and generated by combining real-time dual-frequency impedance, heart sound signal and triaxial acceleration data. The respiratory rate sequence was scanned and identified by sliding time window. Respiratory markers were identified and marked based on the patient's position. The respiratory markers were continuously monitored, comprehensively evaluated and graded to obtain the respiratory compensation level.

[0088] The third heart sound arises directly from elevated end-diastolic pressure in the left ventricle and a significant decrease in ventricular wall compliance, which is the core pathophysiological change in acute left ventricular failure, i.e., cardiogenic pulmonary edema. When heart failure occurs, the impaired relaxation function of the ventricle reduces its ability to passively expand during rapid filling, causing low-frequency vibrations from the impact of blood flow on the ventricular wall. Therefore, the appearance of the third heart sound is not accidental noise, but rather a direct acoustic fingerprint of cardiac decompensation. In clinical medicine, the discovery of a third heart sound gallop rhythm during auscultation is an important physical examination criterion for diagnosing acute heart failure and assessing its severity. Its presence greatly increases the likelihood of cardiogenic dyspnea.

[0089] Reference Figure 2 Methods for obtaining respiratory rate sequences include:

[0090] The first impedance in the real-time dual-frequency impedance is bandpass filtered; the impedance-filtered signal is obtained, and the peak time point in the impedance-filtered signal is detected based on the peak detection algorithm to obtain the first time series; wherein, the bandpass filtering range is preferably 0.1-2.0Hz, corresponding to a respiratory rate of 1-120 breaths / minute.

[0091] The moment of the first heart sound is extracted, the time difference between adjacent first heart sounds is calculated as the heartbeat interval, and a heartbeat interval sequence is obtained. The heartbeat interval sequence is converted into a uniform sequence by interpolation, and a bandpass filter is applied to the uniform sequence to obtain a filtered heart sound signal. The peak time point sequence of the filtered heart sound signal is detected by a peak detection algorithm to obtain a second time sequence. The preferred bandpass filtering range is 0.1-2.0 Hz, corresponding to a respiratory rate of 1-120 breaths / minute.

[0092] The axial acceleration signal perpendicular to the chest wall is extracted from the triaxial acceleration data and labeled as the dynamic acceleration signal, usually the Z-axis acceleration. The dynamic acceleration signal is bandpass filtered to obtain the filtered dynamic acceleration signal. The peak time point sequence in the filtered dynamic acceleration signal is detected based on the peak detection algorithm to obtain the third time series.

[0093] The time interval between adjacent time points in the three time series is calculated and used as the instantaneous respiratory cycle of the corresponding time series.

[0094] Using any one signal's time point t as the reference signal, check whether the other two signals have peak time points within the reference signal's detection time window [t-ΔT, t+ΔT]; ΔT is the detection duration, which is set according to the actual situation, preferably 0.5 seconds, and can be adjusted according to the actual situation.

[0095] If at least two of the three signals have peak time points within the detection time window, it is determined that the three signals have jointly captured the same valid respiratory event. The average value of all time points within the detection time window is taken as a fused respiratory event time point. Any isolated time points that are not aligned are removed to obtain the fused respiratory event time point sequence.

[0096] Calculate the time difference between two adjacent fusion respiratory events to obtain the respiratory cycle. Calculate the ratio of one minute (60 seconds) to the time difference between two adjacent fusion respiratory events to obtain the respiratory rate sequence.

[0097] By integrating respiratory characteristics from three signals—dual-frequency impedance, heart sound signal, and triaxial acceleration—and eliminating isolated noise points through a 0.5-second time window, the reliability of respiratory rate detection can be improved, providing accurate data support for subsequent pathological respiratory pattern recognition.

[0098] Reference Figure 3 Methods for obtaining respiratory markers include:

[0099] Obtain the impedance-filtered signal corresponding to the respiratory frequency sequence, and normalize the corresponding impedance-filtered signal to obtain the respiratory waveform amplitude sequence. Using a preset sliding time window as the unit, identify all troughs and peaks in the respiratory waveform amplitude sequence.

[0100] The period from the start of any trough to the end of the next trough is defined as a period to be evaluated.

[0101] If, within a preset sliding time window, E consecutive evaluation cycles that satisfy the first quantification rule are detected, then the preset sliding time window is marked as Cheyne-Stokes respiration; E is preferably 2-3.

[0102] The first quantification rule is:

[0103] The duration of the evaluation period is within the respiratory duration range, wherein the lower limit of the respiratory duration range is preferably 30 seconds and the upper limit is preferably 120 seconds.

[0104] During the evaluation period, from the initial trough to the peak, the amplitude difference between any two adjacent points is greater than the negative tolerance; a small negative tolerance is allowed to tolerate measurement noise, but an overall amplitude decrease is not allowed; from the peak to the end trough, the amplitude difference between any two adjacent points is less than the positive tolerance; a small positive tolerance is allowed, but an overall amplitude increase is not allowed; the tolerance values ​​corresponding to the negative and positive tolerances are set based on empirical values, and can be preferably 0.05 times the patient's recent average resting amplitude.

[0105] Calculate the peak-to-trough amplitude difference between the peak and the initial trough within the period to be evaluated. The peak-to-trough amplitude difference should be greater than the baseline threshold, which is preferably half of the patient's recent average resting amplitude.

[0106] During the last quarter of the evaluation period, the duration of the respiratory waveform amplitude that is consistently below the first amplitude threshold is not less than the first duration threshold. The first amplitude threshold is K1 times the patient's recent average resting amplitude, and K1 is preferably 0.2. The first duration threshold is preferably 10-15 seconds. The patient's recent average resting amplitude is calculated from the respiratory waveform amplitude sequence collected when the patient's postural state signal is 0 and the patient is in the resting phase and core sleep time, such as 0-6 am.

[0107] Within the sliding time window, the average value of the respiratory rate sequence and the average value of the respiratory waveform amplitude sequence are calculated. If the average value of the respiratory rate sequence and the average value of the respiratory waveform amplitude sequence satisfy the second quantization rule within the sliding time window, the sliding time window is marked as rapid shallow breathing.

[0108] The second quantification rule includes:

[0109] The average value of the respiratory rate sequence is not lower than the frequency threshold; the frequency threshold is preferably 25 breaths / minute.

[0110] The average value of the respiratory waveform amplitude of the respiratory rate sequence is not higher than the second amplitude threshold; the second amplitude threshold is preferably K2 times the statistical median of the respiratory waveform amplitude of the patient in the past 24 hours at rest, and K2 is preferably 0.6.

[0111] The three signals—impedance filtered signal, filtered heart sound signal, and filtered dynamic acceleration signal—are filtered separately. Signal segments in which the absolute value of the respiratory waveform amplitude of any one signal is consistently lower than the amplitude threshold of the corresponding channel are selected as candidate low-amplitude segments. The start and end times of the candidate low-amplitude segments are recorded to obtain a list of time intervals. The amplitude threshold of the corresponding channel is set based on the standard deviation of the signal amplitude of the corresponding channel during the patient's most recent resting reference period, such as when the user is awake and sitting still.

[0112] Based on the time synchronization tolerance, determine whether the candidate low amplitude segments of the three signals overlap in time; that is, the difference between the start time and the end time of the candidate low amplitude segments of two signals should not be greater than the time synchronization tolerance; the time synchronization tolerance is set based on empirical values, preferably 0.5 seconds.

[0113] An overlapping period is considered a valid pause event only if at least two candidate low-amplitude segments of signals overlap in time during the selection period.

[0114] If the interval between two valid pause events is less than the interval threshold, the two valid pause events are merged into one apnea event. The smaller start time of the two valid pause events is marked as the start time of the apnea event, and the larger end time of the two valid pause events is marked as the end time of the apnea event. Otherwise, the valid pause event is directly treated as an apnea event. The interval threshold is preferably 0.5 seconds.

[0115] The difference between the end and start times of all apnea events is calculated to obtain the event duration. If the event duration is not less than a second duration threshold and the corresponding event waveform amplitude is less than the apnea amplitude threshold, it is marked as apnea. If the event duration is not less than the second duration threshold and the corresponding event waveform amplitude is within the ventilation amplitude range, it is marked as hypoventilation. The event waveform amplitude is the average respiratory waveform amplitude of the signal corresponding to the apnea event. The second duration threshold is set based on industry standards and is preferably 10 seconds. The apnea amplitude threshold is statistically analyzed based on the patient's historical data and is preferably 10%–20% of the patient's personal baseline amplitude. The ventilation amplitude range is statistically analyzed based on the patient's historical data and is preferably 30%–50% of the patient's personal baseline amplitude.

[0116] Based on individual baseline amplitude, quantitative rules are set to accurately identify different pathological breathing patterns. At the same time, orthopnea is marked. According to clinical data, orthopnea is generally a typical manifestation of cardiogenic pulmonary edema, providing respiratory-level evidence for distinguishing physiological / pathological impedance changes.

[0117] In the triaxial acceleration data, the positive z-axis is defined as the direction perpendicular to the chest wall and outward; the positive y-axis is defined as the direction from the right side of the body to the left side; and the positive x-axis is defined as the direction from the head to the feet when standing upright.

[0118] The variances of the triaxial acceleration data are calculated separately. If the variances are all below a variance threshold, the system is considered static; otherwise, it is considered dynamic. During the static period, the acceleration data for each axis is filtered to obtain the corresponding gravitational acceleration component. During the dynamic period, the triaxial gravitational acceleration components from the previous moment are obtained and used as the state vector for the Kalman filter algorithm. A 3×3 identity matrix is ​​set as the state transition matrix. The process noise follows a Gaussian distribution with a mean of 0 and a covariance matrix of C. The triaxial acceleration data is used as measurement noise. The vector sum of the triaxial gravitational acceleration components from the previous moment and the triaxial acceleration data is calculated to obtain the observation moment. The covariance matrix, based on the Kalman filter algorithm, uses the triaxial gravitational acceleration components of the previous moment as initial values ​​to predict the gravitational acceleration components corresponding to each axis. The variance threshold is set based on empirical values, preferably 0.05g²-0.1g², where g is the gravitational acceleration. C is estimated based on the system sampling period and the expected maximum angular velocity of the patient's body in the monitoring scenario. The expected maximum angular velocity is determined based on the typical postural change speed of the patient during rest, sleep, and daily light activities, and is usually taken as 0.1 to 0.2 radians / second. The value of the covariance matrix can reasonably reflect the small changes that may occur in the direction of gravity in a short period of time, while maintaining the stability of the state prediction.

[0119] Calculate the ratio of the x-axis gravitational acceleration component to the magnitude of the gravity vector, and then calculate the arcsine of the ratio to obtain the pitch angle.

[0120] Calculate the arctangent values ​​of the y-axis gravitational acceleration component Gy and the z-axis gravitational acceleration component Gz in four quadrants to obtain the roll angle. The four quadrant division is based on the signs of Gy and Gz: Gy > 0 and Gz > 0 in the first quadrant; Gy > 0 and Gz < 0 in the second quadrant; Gy < 0 and Gz < 0 in the third quadrant; and Gy < 0 and Gz > 0 in the fourth quadrant. When Gy and Gz are in the first quadrant, the arctangent value of the ratio of Gy to Gz is used as the roll angle. When Gy and Gz are in the second quadrant, calculate the arctangent of the ratio of Gy to Gz, and then calculate the sum of the arctangent and π as the roll angle. When Gy and Gz are in the third quadrant, calculate the arctangent of the ratio of Gy to Gz, and then calculate the difference between the arctangent and π as the roll angle. When Gy and Gz are in the fourth quadrant, calculate the arctangent of the ratio of Gy to Gz as the roll angle. Wherein, if Gz=0 and Gy>0, the roll angle is equal to π / 2, and if Gy<0, the roll angle is equal to -π / 2. If Gy=0, the roll angle is equal to 0.

[0121] By combining pitch and roll angles, the body position signal is determined according to the judgment rules.

[0122] The judgment rules include:

[0123] When the pitch angle is less than 30° and the roll angle is less than 45°, the patient is judged to be in a supine position, and the body position status signal is marked as 0.

[0124] When the pitch angle is less than 30° and the roll angle is greater than 45°, the patient is judged to be in left lateral decubitus position, and the body position status signal is marked as 1.

[0125] When the pitch angle is less than 30° and the roll angle is less than -45°, the patient is judged to be in right lateral decubitus position, and the body position status signal is marked as 2.

[0126] When the pitch angle is greater than 60°, it is determined to be sitting / standing, and the body position signal is marked as 3.

[0127] If any of the above judgment rules are not met, it is judged to be in the active period, and the body position status signal is marked as 4.

[0128] By using Kalman filtering to distinguish the gravitational acceleration components during the static and dynamic phases, and combining pitch and roll angles to accurately determine body position, this provides a crucial basis for identifying position-related physiological impedance changes.

[0129] When the body position signal changes directly from 0 to 3, it is marked as sitting breathing. The switching time is recorded, and the first average respiratory rate within the supine window before the switching time is calculated. The second average respiratory rate within the detection window after the switching time is calculated. The difference between the first and second average values ​​is calculated, and then the ratio of the difference to the first average value is calculated to obtain the respiratory relief rate.

[0130] Methods for obtaining a respiratory compensation grade include:

[0131] If Cheyne-Stokes respiration is detected and its duration exceeds Q minutes; or if the respiratory relief rate is greater than the relief rate threshold and within the detection window, an assessment sequence is constructed by collecting respiratory frequencies within the assessment window, and the respiratory frequencies in the assessment sequence greater than the first average value exceed the proportion threshold; then it is determined to be a Grade A mode; Q is obtained based on clinical data statistics, preferably 20 minutes, which can be adjusted according to actual conditions; the relief rate threshold is obtained based on clinical data statistics, preferably 15%; the detection window can be set according to clinical data, such as 300 seconds; the length of the assessment window is shorter than the detection window, and can be set to 30 seconds, which can be adjusted according to actual conditions; the proportion threshold is set based on empirical values, preferably 70%.

[0132] If rapid, shallow breathing is detected, with a respiratory rate exceeding X breaths / minute and a duration exceeding M minutes; or if an apnea event is detected, it is classified as Level B mode; X is set according to physiological limits, preferably 25 breaths / minute; M is obtained based on clinical data statistics, preferably 30 minutes.

[0133] Otherwise, it will be classified as Level C.

[0134] If any Grade A pattern is detected during the assessment period, it is directly determined as severe respiratory compensation;

[0135] If no Grade A mode is detected, but one or more Grade B modes are detected, and the total cumulative duration of Grade B modes exceeds the percentage threshold, then moderate respiratory compensation is determined to exist; the percentage threshold is obtained based on clinical data statistics, and is preferably 30%.

[0136] Otherwise, it is determined as no respiratory compensation.

[0137] Combining respiratory markers with clinical thresholds can quantify the degree of respiratory compensation and further exclude physiological impedance changes without respiratory compensation.

[0138] The final cardiopulmonary risk level is determined based on the initial warning results, pathological confirmation results, and respiratory compensation level.

[0139] Methods for obtaining a cardiopulmonary risk level include:

[0140] When the initial warning result is 1, the pathological confirmation result is 1, and the respiratory compensation level is severe or moderate, it is judged as high risk, and the corresponding respiratory compensation level is output.

[0141] When the initial warning result of the test is 1, the pathological confirmation result is 1, and the respiratory compensation level is no respiratory compensation, it is judged as medium risk;

[0142] When only the initial warning result is 1, it is judged as low risk;

[0143] Otherwise, it is considered normal.

[0144] By employing a three-tiered verification process involving primary warning, pathological confirmation, and respiratory compensation, impedance changes caused solely by physiological factors such as fat / muscle growth are eliminated, thereby achieving accurate early warning of cardiogenic pulmonary edema and eradicating false alarms from a mechanistic perspective.

[0145] Changes in pleural impedance caused by simple fat or muscle growth do not trigger the respiratory compensation pattern specific to heart failure. Therefore, even if a few borderline cases appear in the preceding steps, the above steps will output a low-risk level due to normal corresponding respiratory markers, ultimately ruling out interference from physiological changes at the clinical level and eliminating the possibility of being falsely reported as critical risk. Pleural effusion caused by non-cardiac causes such as pneumonia or renal failure has its pathological core in infection or metabolic problems, rather than an acute increase in left ventricular end-diastolic pressure. Therefore, these diseases usually do not present with typical periodic breathing caused by central instability in heart failure, such as Cheyne-Stokes respiration, or the orthopnea pattern specific to heart failure patients. However, when the heart sound analysis steps may fail to completely rule out these diseases for some reason, such as low signal quality, the above steps can provide final confirmation or skepticism by assessing whether the respiratory pattern presents typical cardiogenic decompensation characteristics. The above steps provide the final clinical anchor for the entire differential diagnosis chain by verifying whether the respiratory system exhibits a unique pattern of cardiogenic decompensation. It ensures that the system's final warnings are not limited to suspected heart failure effusion, but also include cardiogenic pulmonary edema that has led to clear physiological decompensation and poses an extremely high acute risk. This allows for the final and crucial purification and confirmation of signal changes into clinical emergency risk within the complex context of daily family life, eliminating false alarms caused by various non-target conditions such as physiological changes and other diseases. It ensures that the system only issues high-level warnings for cardiogenic pulmonary edema that has already produced clinical consequences and truly indicates a risk of acute deterioration. This achieves risk stratification between stable fluid retention and precursors to acute deterioration, completing the final transformation from pathological diagnosis to trend prediction, and giving the warnings clinically urgent guidance value.

[0146] Example 2:

[0147] Please see Figure 4 As shown, this embodiment provides a risk trend prediction system based on multimodal physiological data, including:

[0148] Anomaly detection module: Acquires real-time dual-frequency impedance of the patient, calculates the ratio of the dual-frequency impedance to obtain the dual-frequency ratio; monitors the temporal changes of the dual-frequency ratio, and outputs a primary warning result based on the temporal changes;

[0149] Pathological confirmation module: Collects the patient's heart sound signals. When the primary warning result is 1, the heart sound signals are preprocessed and feature analyzed to obtain heart sound features. The heart sound features are then used for pathological determination to obtain the pathological confirmation result.

[0150] Pattern detection module: Collects triaxial acceleration data of patients. When the pathological confirmation result is 1, the angle is calculated based on the triaxial acceleration data to obtain the patient's position. Combined with real-time dual-frequency impedance, heart sound signal and triaxial acceleration data, the respiratory rate sequence of patients is obtained by comparison and analysis. The respiratory rate sequence is detected and analyzed through a sliding time window to obtain respiratory markers. The respiratory markers are continuously monitored and analyzed to obtain the respiratory compensation level.

[0151] Risk Analysis Module: Determines the final cardiopulmonary risk level based on the initial warning results, pathological confirmation results, and respiratory compensation level.

[0152] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0153] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A risk trend prediction method based on multimodal physiological data, characterized in that, include: The patient's real-time dual-frequency impedance was collected, and the ratio of the dual-frequency impedances was calculated to obtain the dual-frequency ratio. Monitor the time-series changes in the dual-frequency ratio and output primary early warning results based on the time-series changes; The patient's heart sound signals are collected. When the primary warning result is 1, the heart sound signals are preprocessed and feature analyzed to obtain heart sound features. The heart sound features are then pathologically determined to obtain pathological confirmation results. Triaxial acceleration data of the patient was collected. When the pathological confirmation result was 1, the angle was calculated based on the triaxial acceleration data to obtain the patient's positional status. The patient's respiratory rate sequence was extracted and generated by combining real-time dual-frequency impedance, heart sound signal and triaxial acceleration data. The respiratory rate sequence was scanned and identified by a sliding time window. Respiratory markers were identified and marked based on the patient's positional status. Continuous monitoring, comprehensive evaluation and grade determination were carried out based on the respiratory markers to obtain the respiratory compensation level. The final cardiopulmonary risk level is determined based on the initial warning results, pathological confirmation results, and respiratory compensation level.

2. The risk trend prediction method based on multimodal physiological data according to claim 1, characterized in that, Methods for obtaining preliminary warning results include: The first impedance and the second impedance are collected, and the ratio of the first impedance and the second impedance is calculated to obtain the dual-frequency ratio. When the dual-frequency ratio continues to rise during the monitoring period and exceeds the static threshold, the primary warning result is output as 1; otherwise, the primary warning result is output as 0.

3. The risk trend prediction method based on multimodal physiological data according to claim 1, characterized in that, Methods for obtaining heart sound characteristics include: The heart sound signal is bandpass filtered to obtain the filtered audio. The filtered audio is used as the input to the heart sound segmentation model to obtain the first heart sound position and the second heart sound position. The filtered audio, the first heart sound position, the second heart sound position, and the detection duration are used as the input to the anomaly analysis model to obtain the probability of the third heart sound and the indicator of the fourth heart sound. The probability of the third heart sound is used as the heart sound feature.

4. The risk trend prediction method based on multimodal physiological data according to claim 3, characterized in that, Methods for obtaining pathological confirmation results include: Within the warning period corresponding to a primary warning result of 1, the total number of times the probability of the third heart sound exceeds the probability threshold is counted. The ratio of the total number of times to the warning period is calculated to obtain the frequency of the probability of the third heart sound exceeding the probability threshold per unit time. If the frequency of the probability of the third heart sound exceeding the probability threshold per unit time exceeds the clinical threshold, the pathological confirmation result is output as 1; otherwise, the pathological confirmation result is output as 0.

5. The risk trend prediction method based on multimodal physiological data according to claim 1, characterized in that, Methods for obtaining respiratory rate sequences include: Real-time dual-frequency impedance, heart sound signal and triaxial acceleration data were detected and analyzed to obtain the first time series, the second time series and the third time series. The time interval between adjacent time points in the three time series is calculated as the instantaneous respiratory cycle of the corresponding channel. Taking the time point t of any one signal as the reference signal, it is checked whether there are peak time points in the other two signals within the detection time window of the reference signal. If at least two of the three signals have peak time points within the detection time window, it is determined that the three signals have captured the same valid respiratory event. The average value of all time points within the detection time window is taken as a fused respiratory event time point. Any misaligned isolated time points are removed to obtain the fused respiratory event time point sequence. The time difference between two adjacent fused respiratory event time points is calculated to obtain the respiratory cycle. The ratio of one minute duration to the time difference between two adjacent fused respiratory event time points is calculated to obtain the respiratory rate sequence.

6. The risk trend prediction method based on multimodal physiological data according to claim 1, characterized in that, Methods for obtaining respiratory markers include: The impedance-filtered signal corresponding to the respiratory rate sequence is obtained and normalized to obtain the respiratory waveform amplitude sequence. A sliding time window is used to identify troughs and peaks, and the evaluation period is marked. If the evaluation period that meets the first quantization rule appears consecutively, it is marked as Cheyne-Stokes respiration. If the average value of the respiratory rate sequence and the average value of the respiratory waveform amplitude sequence meet the second quantization rule, it is marked as rapid shallow breathing. The impedance-filtered signal, filtered heart sound signal, and filtered kinetic acceleration signal are filtered, and apnea events are obtained by time overlap determination. The difference between the end time and the start time of all apnea events is calculated to obtain the event duration. Based on the event duration and the event waveform amplitude, it is marked as apnea or hypoventilation. The pitch and roll angles are calculated using triaxial acceleration data to determine the body position. When the body position changes directly from supine to upright, it is marked as orthopnea. The body position includes supine, left lateral, right lateral, upright, and active phases.

7. The risk trend prediction method based on multimodal physiological data according to claim 6, characterized in that, Methods for obtaining apnea events include: The signal segment whose absolute value of the respiratory waveform amplitude of any signal is consistently lower than the amplitude threshold of the corresponding signal is taken as a candidate low amplitude segment. The start and end times of the candidate low amplitude segments are recorded to obtain a list of time intervals. Based on the time synchronization tolerance, determine whether the candidate low amplitude segments of the three signals overlap in time; only when at least two signals have candidate low amplitude segments overlapping in time during the screening period, the overlapping period is determined as a valid pause event. If the interval between two valid pause events is less than the interval threshold, the two valid pause events are merged into one apnea event. The smaller start time of the two valid pause events is marked as the start time of the apnea event, and the larger end time of the two valid pause events is marked as the end time of the apnea event; otherwise, the valid pause event is directly treated as an apnea event.

8. The risk trend prediction method based on multimodal physiological data according to claim 6, characterized in that, Methods for obtaining a respiratory compensation grade include: If Cheyne-Stokes respiration is detected and lasts for more than Q minutes; or if the first average respiratory rate within the supine window before the switching time is calculated; and the second average respiratory rate within the detection window after the switching time is calculated; and the difference between the first and second averages is calculated, and then the ratio of the difference to the first average is calculated to obtain the respiratory relief rate, and the respiratory relief rate is greater than the relief rate threshold, and within the detection window, an assessment sequence is constructed by collecting respiratory rates within the assessment window during the detection period, and the respiratory rates in the assessment sequence greater than the first average exceed the proportion threshold; then it is determined to be a Class A mode. If rapid, shallow breathing is detected, with a respiratory rate exceeding X breaths per minute and a duration exceeding M minutes; or if a breathing apnea event is detected, it is classified as Level B mode. Otherwise, it will be classified as Level C mode; If any Grade A pattern is detected during the assessment period, it is directly determined as severe respiratory compensation; If no Class A mode is detected, but one or more Class B modes are detected, and the total cumulative duration of Class B modes exceeds the percentage threshold, then moderate respiratory compensation is determined to exist. Otherwise, it is determined as no respiratory compensation.

9. The risk trend prediction method based on multimodal physiological data according to claim 1, characterized in that, Methods for obtaining a cardiopulmonary risk level include: When the initial warning result is 1, the pathological confirmation result is 1, and the respiratory compensation level is severe or moderate, it is judged as high risk, and the corresponding respiratory compensation level is output. When the initial warning result of the test is 1, the pathological confirmation result is 1, and the respiratory compensation level is no respiratory compensation, it is judged as medium risk; When only the initial warning result is 1, it is judged as low risk; Otherwise, it is considered normal.

10. A risk trend prediction system based on multimodal physiological data, implementing the risk trend prediction method based on multimodal physiological data as described in any one of claims 1-9, characterized in that, include: Anomaly detection module: Collects real-time dual-frequency impedance data from the patient and calculates the ratio of the dual-frequency impedances to obtain the dual-frequency ratio. Monitor the time-series changes in the dual-frequency ratio and output primary early warning results based on the time-series changes; Pathological confirmation module: Collects the patient's heart sound signals. When the primary warning result is 1, the heart sound signals are preprocessed and feature analyzed to obtain heart sound features. The heart sound features are then used for pathological determination to obtain the pathological confirmation result. Pattern detection module: Collects triaxial acceleration data of the patient. When the pathological confirmation result is 1, it calculates the angle based on the triaxial acceleration data to obtain the patient's position. Combined with real-time dual-frequency impedance, heart sound signals and triaxial acceleration data, it extracts and generates the patient's respiratory rate sequence. It performs pattern scanning and recognition on the respiratory rate sequence through a sliding time window. Combined with the patient's position, it identifies and marks respiratory markers. Based on the respiratory markers, it performs continuous monitoring, comprehensive evaluation and grade determination to obtain the respiratory compensation level. Risk Analysis Module: Determines the final cardiopulmonary risk level based on the initial warning results, pathological confirmation results, and respiratory compensation level.

Images