Method for identifying heart risk by using night vital sign data mutation

Abstract

The invention relates to the technical field of medical treatment and health data processing, and discloses a method for identifying heart risks by utilizing night vital sign data mutation, which comprises the following steps: carrying out scene classification on respiratory signals, and when the classification result is an ambiguous fluctuation artifact state, further acquiring a signal quality index of an ECG signal; the method comprises the following steps: acquiring a signal quality index, performing cross validation on a fluctuation artifact state by using the signal quality index, correcting the fluctuation artifact state into a physiological wave state or a signal artifact state, and adaptively selecting an ECG analysis model according to a final classification result. According to the method, the inherent ambiguity problem of the fluctuation artifact state is solved, and the method can effectively distinguish the real physiological fluctuation from the technical artifact, so that a targeted analysis model is scheduled for the two scenes with different properties, and the accuracy of risk identification is improved.

Description

Technical Field

[0001] This invention relates to a method for identifying cardiac risk using nighttime vital sign data mutations, belonging to the field of medical and health data processing technology. Background Technology

[0002] Currently, in the field of medical and health data processing, especially in nighttime monitoring, multi-channel vital sign data, such as electrocardiogram (ECG) and respiratory signals, are typically collected. Static thresholds are set for each independent signal channel to analyze abnormal changes in the data that may indicate risk. Existing data processing methods that use isolated channels and static thresholds analyze ECG and respiratory signals as independent channels at the data processing level. This approach has limitations when dealing with specific physiological phenomena. For example, in cases of compensatory tachycardia induced by obstructive sleep apnea and primary cardiac conduction abnormalities, although their clinical causes are different, the phenotypes of ECG signals may be highly similar. The aforementioned isolated channel processing method, because it fails to consider the state of respiratory signals simultaneously when analyzing ECG signals, cannot distinguish between these two events at the data level, triggering the same alarm for both. This results in a large number of invalid alarms or information confusion, interfering with subsequent medical judgment.

[0003] To address the challenge of distinguishing events with similar phenotypes but different causes, one approach in this field is to construct a complex multivariate fusion model, such as a Bayesian network or deep learning model, to calculate the posterior probability of events in real time. Some solutions, however, seek more sophisticated algorithmic models to solve this problem. For example, Chinese invention patent CN119851932A discloses a method and system for predicting cardiac arrest risk levels based on fuzzy logic. This approach attempts to use machine learning for feature selection and combines fuzzy C-means clustering and case-based reasoning to differentiate between multiple events. The approach of fusing data to calculate a comprehensive risk level still belongs to a complex multivariate fusion model. Essentially, it fuses multiple vital sign data after blurring them, but fails to identify the clearly associated physiological context, such as respiratory status, at the source of data processing, i.e., when analyzing electrocardiogram signals. It also has difficulty distinguishing events with different causes. However, in engineering practice, this approach usually requires a large amount of high-quality labeled data, and its model building and training process is relatively complex and costly. For monitoring systems that need to be widely deployed and require high reliability, it is insufficient in terms of universality and simplicity.

[0004] Therefore, the technical problem to be solved by this invention is how to use respiratory pattern analysis as a preliminary step in ECG signal analysis without relying on the aforementioned complex fusion model, through an adjustment of the data processing flow. That is, to use the classification results of respiratory scenarios to dynamically select or switch the set of analysis rules to be used for subsequent ECG signals, so as to associate an objective physiological scenario with the ECG signal analysis at the initial stage of data processing, thereby distinguishing events with different causes. Summary of the Invention

[0005] This invention provides a method for identifying cardiac risk using nocturnal vital sign data mutations. Its main purpose is to solve the problem in existing data processing methods that fail to effectively distinguish events with similar data phenotypes but different causes due to the failure to associate them with physiological context.

[0006] To achieve the above objectives, the present invention provides a method for identifying cardiac risk using nocturnal vital sign data mutations, the method comprising:

[0007] Step a: Acquire respiratory signals and ECG signals in real time;

[0008] Step b: Based on the respiratory signal, classify the current respiratory pattern into one of multiple preset respiratory scenarios. The multiple preset respiratory scenarios include at least: steady state scenario, respiratory event state scenario, and fluctuation artifact state scenario.

[0009] Step c: When the classification result of step b is a fluctuating artifact scenario, the cross-validation rule is executed. This rule includes: obtaining the signal quality index (SQI) of the ECG signal within the same time window as the respiratory signal; if the respiratory pattern is a fluctuating artifact scenario and the SQI is determined to be of high quality, the final classification of the scenario is corrected to a physiological wave dynamic scenario; if the respiratory pattern is a fluctuating artifact scenario and the SQI is determined to be of low quality, the final classification of the scenario is corrected to a signal artifact scenario.

[0010] Step d: Based on the results of the steady-state scenario or respiratory event scenario classified in step b, or based on the final classification results of the physiological wave dynamic scenario or signal artifact scenario corrected in step c, adaptively select an ECG analysis model corresponding to the scenario or final classification from multiple preset ECG analysis models.

[0011] Step e involves processing the ECG signal using only the ECG analysis model selected in step d to identify the presence of cardiac risk mutations.

[0012] Preferably, in step c, determining whether the SQI is of high or low quality includes: calculating the SQI value based on at least one of the baseline drift, signal saturation, or QRS wave morphology characteristics of the ECG signal; and comparing the SQI value with a preset SQI threshold to determine whether the SQI is of high or low quality.

[0013] Preferably, before step b, the method further includes: calculating the respiratory signal quality index Resp-SQI of the respiratory signal, and only executing steps b and c if the Resp-SQI meets a preset quality threshold. If the Resp-SQI does not meet the preset quality threshold, the execution of steps b and c is stopped, and a predefined high-sensitivity baseline model is used to process the ECG signal.

[0014] Preferably, the calculation of the respiratory signal quality index Resp-SQI includes: calculating the value of Resp-SQI based on at least one of the signal amplitude, saturation or baseline drift characteristics of the respiratory signal, with a preset quality threshold being a threshold used to characterize the reliability of the respiratory signal.

[0015] Preferably, the multiple preset ECG analysis models include: a high-sensitivity baseline model corresponding to the steady-state scenario, a respiratory compensation analysis model corresponding to the respiratory event scenario, a fluctuation-specific model that maintains high sensitivity to malignant morphological changes and corresponds to the physiological wave dynamic scenario, and a comprehensive suppression model corresponding to the signal artifact scenario.

[0016] Preferably, in step b, classifying the current breathing pattern into one of multiple preset breathing scenarios based on the breathing signal includes: extracting a set of breathing state feature vectors (RSV) from the breathing signal; inputting the breathing state feature vectors (RSV) into a rule-based breathing state classifier to output one of the multiple preset breathing scenarios.

[0017] Preferably, the method further includes: within a preset global time window, recording in real time the time series of the respiratory scenario output by step b or the final classification corrected by step c to form a scenario state log; at the end of the global time window, performing time-series statistical analysis on the scenario state log to generate a quantitative report characterizing the stability of the cardiopulmonary system; the time-series statistical analysis includes: counting the total number of scenario state transitions within the global time window. and according to Calculate the Scenario Fragmentation Index (CFI), where This represents the total duration of the global time window.

[0018] Preferably, the respiratory state feature vector RSV includes at least: respiratory rate, respiratory amplitude variation coefficient, and apnea or hypoventilation event flag AHE_Flag.

[0019] Preferably, before step b is executed, the method further includes: during the calibration period, combining the heart rate variability analysis of body motion signals and ECG signals, automatically identifying and capturing a respiratory signal segment characterizing the user's stable sleep, performing statistical analysis on the respiratory signal segment, calculating the individualized respiratory baseline parameters for the specific user, and using the individualized respiratory baseline parameters to dynamically adjust the judgment threshold used to determine the steady-state scenario within the rule-based respiratory state classifier.

[0020] Preferably, the method further includes: setting a primary risk counter and a respiratory-induced risk counter within a preset global time window; whenever a cardiac risk mutation is identified in step e, if the cardiac risk mutation occurs in a steady-state scenario, the primary risk counter is counted; if the cardiac risk mutation occurs in a respiratory event scenario, the respiratory-induced risk counter is counted; at the end of the global time window, a cardiac risk attribution report is generated based on the count values ​​of the primary risk counter and the respiratory-induced risk counter.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. By real-time characterization and classification of respiratory signals to determine the current physiological context, and using the classification results of this context, an ECG signal analysis model with different analysis rules is adaptively selected. Finally, only the selected model is used to process the ECG signal. This cascaded data processing flow, which uses respiratory analysis as a prerequisite for ECG analysis, provides a unique physiological context for ECG signal mutations from the very beginning of data processing. As a result, when faced with events such as secondary tachycardia induced by apnea and primary tachycardia caused by the heart itself, which are highly similar in ECG phenotypes, this method, because it operates under two different respiratory contexts, will schedule two ECG analysis models with different functions for processing. Thus, in terms of data processing mechanism, it has the ability to distinguish between these two different types of events, avoiding the problem of confusion between the two due to context blindness in existing technologies.

[0023] 2. When initially classifying the breathing pattern into a fluctuating artifact state, the system further acquires the electrocardiogram signal quality index (SQI) for the same time window. Based on the breathing pattern and the SQI, this fluctuating artifact state is further distinguished into physiological fluctuations or signal artifacts. This mechanism, which uses the quality information of one signal (ECG) to cross-validate and arbitrate the scenario classification of another signal (breathing), enables the system to distinguish between real physiological fluctuations and technical artifacts. The effect is that the system can schedule more targeted ECG analysis models for these two distinctly different scenarios, such as matching a dedicated model for physiological fluctuations and an artifact suppression model for artifacts. This avoids the safety hazard of underreporting malignant cardiac events during real physiological fluctuations, such as REM sleep, which may be caused by the generalized treatment in the main scheme.

[0024] 3. By recording the respiratory scenario classification results continuously generated in step b within a preset global time window, a scenario state sequence is formed. This method then performs time-series statistical analysis on this scenario state sequence to count the total number of scenario switching or the cumulative duration of a specific scenario. This method of redefining the instantaneous output of first-order processing as a second-order data stream that can be analyzed gives this method an additional ability to objectively and quantitatively assess the long-term cardiopulmonary stability of users, in addition to its ability to judge instantaneous events. For example, the scenario fragmentation index provides an objective data tool for assessing the stability of the patient's overall nighttime condition. Attached Figure Description

[0025] Figure 1 This is a diagram of the cascaded data processing architecture based on respiratory scenarios of the present invention;

[0026] Figure 2 This is a dynamic relationship diagram between the respiratory signal ECG-SQI and the context classification of the present invention;

[0027] Figure 3 This is a core functional logic block diagram for cardiac risk identification in this invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be noted that the specific embodiments described herein are only used to explain this invention and are not intended to limit the scope of protection of this invention.

[0029] The present invention discloses a method for identifying cardiac risk using nocturnal vital sign data mutations, executed by a data processing system. This method acquires respiratory signals and ECG signals in real time. Before analyzing the respiratory signals, a preferred implementation involves the system performing a quality gating step, i.e., calculating the respiratory signal quality index (Resp-SQI) of the respiratory signals, and only proceeding with subsequent classification if the Resp-SQI meets a preset quality threshold. If the Resp-SQI does not meet the threshold, classification is stopped and a predefined high-sensitivity baseline model is used to process the ECG signals. Another preferred implementation involves performing a pre-calibration step before the system reaches stable operation. The system automatically captures a segment of respiratory signal representing a user's stable sleep within the calibration period and calculates individualized respiratory baseline parameters based on this signal. These parameters are used to dynamically adjust the judgment threshold in the subsequent respiratory state classifier. After entering the routine processing flow, the system classifies the current respiratory mode into one of several preset respiratory scenarios based on the confirmed reliable respiratory signal. These scenarios include at least a steady-state scenario, a respiratory event scenario, and a fluctuation artifact scenario. If and only if the classification result is a fluctuation artifact scenario, the system triggers a cross-validation rule. This rule then acquires the signal quality index (SQI) of the ECG signal in the same time window and, based on the quality judgment result of the SQI, classifies the fluctuation artifact... The system corrects the steady-state scenario to a physiological wave dynamic scenario or a signal artifact scenario. Based on the results of the steady-state scenario or respiratory event scenario obtained in step b, or based on the final classification results corrected in step c, the system adaptively selects a corresponding ECG analysis model from multiple preset ECG analysis models. These models may include a high-sensitivity baseline model, a respiratory compensation analysis model, a fluctuation-specific model, and a comprehensive suppression model. Finally, the system uses only the selected ECG analysis model to process the ECG signal to identify the presence of cardiac risk mutations. In addition, this method may also include recording and performing time-series statistical analysis of the scenario classification results, and analyzing different... The identified cardiac risk mutations under the scenario are classified and counted to generate a stability report or risk attribution report. To ensure the reliability of subsequent scenario classification, the system preferably performs a quality gating on the respiratory signal before executing step b. This step aims to address the objective challenge in monitoring practice that respiratory sensors (such as chest impedance electrodes or pressure catheters) are prone to signal distortion due to body movement, sweat, or displacement. A specific implementation is that the system calculates the respiratory signal quality index Resp-SQI of the respiratory signal in real time in a preset time window of 5 or 10 seconds. The calculation of Resp-SQI can be based on the quantitative analysis of signal amplitude, saturation, or baseline drift characteristics.Taking amplitude characteristics as an example, the system calculates the peak-to-peak value of the signal within the window. If this peak-to-peak value is lower than a minimum amplitude threshold used to distinguish between weak breathing and sensor disconnection (this threshold can be obtained through offline calibration), the Resp-SQI of that window is determined to not meet the quality threshold. Based on this, the system executes a path decision: only if the Resp-SQI meets a preset quality threshold characterizing signal reliability (this threshold can be 0.7) will subsequent steps b and c be authorized for execution. When the Resp-SQI does not meet this threshold, the system immediately suspends the execution of steps b and c and forcibly bypasses to a fail-safe path, directly using a predefined high-sensitivity baseline model to process the ECG signal, thereby maintaining the system's sensitivity to primary cardiac risk monitoring in unknown scenarios where respiratory signals are unreliable.

[0030] To address the poor adaptability of static classifiers due to significant differences in physiological parameters, particularly respiratory baseline, among individuals, this method adds a pre-calibration step before step b. This step operates within an initial calibration cycle, typically the first 30 minutes of the monitoring task, and aims to automatically identify a respiratory signal segment representing stable sleep for a specific user. This stable sleep identification is achieved through a multimodal confirmation logic, one aspect of which involves analyzing the heart rate variability (HRV) of body motion and ECG signals. When the body motion signal is below an activity threshold and HRV analysis indicates a stable state based on indicators such as the LF / HF ratio, the system captures the respiratory signal segment corresponding to that time window. The system performs statistical analysis on the respiratory signal segment to calculate the user's individualized respiratory baseline parameters, which may include the user's median respiratory rate, mean respiratory amplitude, etc. Finally, the system uses these individualized respiratory baseline parameters to dynamically adjust or override the internal judgment thresholds used in the rule-based respiratory state classifier to determine steady-state scenarios. Taking respiratory rate as an example, its stable interval upper and lower limits will be reset from a fixed general value to a specific percentage of the user's individual baseline, in one instance being ±20% of the median, thus making subsequent scenario classification more consistent with the individual's physiological characteristics. In step b, i.e., classifying the current respiratory pattern based on the respiratory signal, a specific implementation method is that the system... From the respiratory signal, a set of respiratory state feature vectors (RSVs) is extracted. The RSV may include at least the respiratory rate, the coefficient of variation of respiratory amplitude, and an apnea or hypopnea event flag (AHE_Flag) obtained through amplitude or morphological analysis. The system inputs this RSV into a rule-based respiratory state classifier, which can be a hard-coded decision tree. Based on the RSV value, the current respiratory state is forcibly classified into one of the following scenarios: steady-state, respiratory event-state, or fluctuation artifact-state. When the classification result of step b is a fluctuation artifact-state scenario, the system recognizes that this scenario is ambiguous, meaning it may be caused by real physiological fluctuations, which could be triggered by REM sleep or body movement, or by sensor loosening. Technical artifacts can cause this; to resolve this ambiguity, the system immediately executes step c, the cross-validation rule. This rule acquires the ECG signal within the same time window as the respiratory signal and calculates its Signal Quality Index (SQI). The SQI calculation can be based on the analysis of baseline drift, signal saturation, or QRS wave morphology characteristics of the ECG signal. One implementation method is to compare the morphological similarity between the current QRS waveform and a dynamically updated standard QRS template to quantify the degree of signal distortion and output an SQI value between 0 and 1.0. The system compares this SQI value with a preset SQI threshold, which can be 0.5. If the SQI is determined to be of low quality, i.e., the SQI value is below 0.If the SQI is 5, the system determines that both respiratory and ECG signals are distorted simultaneously, indicating a systematic technical artifact. Therefore, the final classification of this scenario is revised to a signal artifact scenario. Conversely, if the SQI is determined to be high quality (i.e., the SQI value is not lower than 0.5), it indicates that while the respiratory signal fluctuates, the ECG signal remains clear and readable, indicating a genuine physiological fluctuation. Therefore, the final classification of this scenario is revised to a physiological wave dynamic scenario.

[0031] After obtaining a unique, ambiguity-neutralized final classification result, the system executes step d for adaptive selection. The system internally pre-defines multiple ECG analysis models, each with a functionally specialized set of analysis rules for a specific scenario. The system executes a scheduling logic that matches the final classification result with the unique analysis model: if the final classification is a steady-state scenario, a high-sensitivity baseline model is selected. This model has an extremely narrow heart rate and rhythm variability (HRV) safety threshold, used to detect the slightest primary abnormalities in a stable background. If the final classification is a respiratory event scenario, a respiratory compensation analysis model is selected. This model is designed to identify specific secondary arrhythmia patterns associated with respiratory events (an example being apnea), one of which is bradycardia-tachycardia syndrome. If the final classification is a physiological variability scenario, a fluctuation-specific model is selected. This model has high tolerance for benign HRV fluctuations (common during REM sleep) but is less tolerant of malignant ECG morphology. Changes (including ventricular tachycardia waveforms) maintain high sensitivity; if the final classification is a signal artifact state scenario, a comprehensive suppression model is selected, which will pause analysis and suppress alarms to prevent technical artifacts from causing invalid alarms; in step e, the system only uses the ECG analysis model selected in step d to process the ECG signal, and only events captured by the currently activated specific model are judged as cardiac risk mutations requiring attention; to provide long-term trend assessment beyond momentary alarms, this method may also include time-series statistical analysis of data generated during system operation; one implementation is that the system records the time series of the respiratory scenario output by step b or the final classification corrected by step c in real time within a preset global time window (which can be a complete nighttime monitoring cycle), forming a scenario status log; at the end of the global time window, the system performs time-series statistical analysis on the log, one analysis method is to count the total number of scenario state switching within the window, i.e. And according to a specific calculation formula To calculate the Context Fragmentation Index (CFI), where The CFI metric, representing the total duration of the global time window, objectively quantifies the overall stability of a user's cardiopulmonary system at night. Alternatively, the system can also generate cardiac risk attribution reports. To achieve this, the system sets up a primary risk counter and a respiratory-induced risk counter within the global time window. Whenever step e identifies a cardiac risk mutation, the system checks the context in which the mutation occurred: if the mutation occurred in a steady-state context, the primary risk counter is counted; if the mutation occurred in a respiratory event context, the respiratory-induced risk counter is counted. At the end of the window, the system generates an objective attribution report based on the final count values ​​of these two counters. This report can assist professionals in identifying the main drivers of cardiac risk.

[0032] Example 1: In a nighttime vital signs data processing system, the system receives two types of data streams within the same time window, both exhibiting a rapid increase in heart rate on ECG: User A's data stream comes from an individual with a history of obstructive sleep apnea (OSA), while User B's data stream comes from an individual with a history of primary arrhythmia. This situation, with similar ECG phenotypes but vastly different physiological causes, presents a challenge for the data processing system to differentiate between them. When processing User A's data stream, the system performs step a as described above, acquiring its respiratory and ECG signals; when performing step b, the system analyzes the feature vector of the respiratory signal... Upon detecting the presence of the apnea or hypoventilation event flag AHE_Flag, the system classifies user A's current breathing pattern as a respiratory event scenario. Based on this classification result, during step d, the system adaptively selects the respiratory compensation analysis model corresponding to this scenario. Finally, in step e, the system uses only the respiratory compensation analysis model to process user A's ECG signal. The model's built-in rule set is used to identify secondary cardiac responses related to respiratory events. Therefore, the system classifies this heart rate increase as respiratory-induced compensatory tachycardia and marks it as a respiratory-related event, rather than a primary cardiac risk mutation.

[0033] When processing user B's data stream within the same time window, the system also executes step a to acquire data. During step b, the system analyzes user B's respiratory signal, determining that both respiratory rate and amplitude are stable, and AHE_Flag is negative. Therefore, user B's current breathing pattern is classified as a steady-state scenario. Correspondingly, during step d, the system adaptively selects a high-sensitivity baseline model corresponding to the steady-state scenario. In step e, the system uses only this high-sensitivity baseline model for processing. Because this model has an analysis threshold for a stable background, it immediately identifies this sudden change in heart rate under a stable respiratory background as a high-risk cardiac event. The mutation triggers a high-priority alarm. The execution result of this data processing flow is that the system first executes step b to determine a respiratory scenario before analyzing the ECG signal, and then dynamically schedules a functionally specialized ECG analysis model in step d using the classification results of this scenario. This cascaded data processing method enables the system to distinguish between two events: user A (secondary response) and user B (primary mutation). This method couples respiratory analysis with ECG analysis, providing an objective physiological context for ECG signal mutations, thereby solving the problem of event confusion and invalid alarms caused by unclear scenarios at the data processing level.

[0034] Example 2: This example uses data from a publicly available, professionally annotated medical and health database to objectively verify the effectiveness of the cross-validation rule in step c of the method of the present invention. Two representative challenging data segments were selected: Data segment A, characterized by drastic fluctuations in respiratory signals, but with a calculated ECG signal quality index (SQI) of 0.85 (high quality), and containing an annotated cardiac risk mutation; and Data segment B, characterized by drastic fluctuations in both respiratory and ECG signals, with a calculated ECG signal quality index (SQI) of 0.23 (low quality). For comparison, three experimental groups were set up: the present invention sample group, which uses the complete method of the aforementioned specific implementation, including the cross-validation rule in step c; Control group 1, which simulates a technique for indiscriminately scheduling a fully suppressed model in a fluctuating artifact state scenario; and Control group 2, which simulates a technique for incorrectly continuing to use a high-sensitivity baseline model in a fluctuating artifact state scenario.

[0035] When processing data segment A, all three experimental groups initially classified it as a [fluctuation artifact state scenario] in step b. Control group 1, which used a fully suppressed model, failed to identify the cardiac risk mutation, resulting in a false negative event. Control group 2, which used a high-sensitivity baseline model, identified the mutation. The present invention sample group executed step c, and based on the high-quality judgment of an SQI of 0.85, corrected the scenario to a physiological wave dynamic scenario, and then used a fluctuation-specific model, which also identified the cardiac risk mutation. When processing data segment B, all three experimental groups again initially classified it as a fluctuation artifact state scenario in step b. Control group 1, which used a fully suppressed model, did not generate an alarm, resulting in a true negative event. Control group 2, which used a high-sensitivity baseline model, incorrectly identified the signal artifact as a cardiac risk mutation, resulting in a false positive event. The present invention sample group executed step c, which, based on the low-quality judgment of an SQI of 0.23, corrected the scenario to a signal artifact state scenario, and then used a fully suppressed model, which did not generate an alarm, also resulting in a true negative event. Table 1 shows the comparison results of the experimental groups in this embodiment. See Table 1.

[0036] Table 1: Comparison Results of Experimental Groups

[0037]

[0038] Experimental data show that although control group 1 avoided false positives, it produced false negatives in data segment A; although control group 2 identified risks in data segment A, it produced false positive alarms in data segment B; the sample group of this invention, by introducing the cross-validation rule in step c, that is, using the SQI information of the ECG signal to arbitrate the nature of the fluctuation artifact state, is the only solution that can achieve the expected treatment in both data segment A (risk identification) and data segment B (artifact suppression).

[0039] Example 3: This example combines Figures 1 to 3 The method for identifying cardiac risk using nocturnal vital sign data mutations is explained, such as... Figure 1As shown, the vital signs sensor is responsible for collecting ECG and respiratory signals. The respiratory signal enters the respiratory signal quality gating unit to calculate its Resp-SQI (Respiratory Signal Quality Index). Confident respiratory signals are sent to the respiratory scenario classification unit, which outputs a preliminary scenario. The steady-state and respiratory event states in the preliminary scenario are used for subsequent ECG model adaptive selection. The fluctuation artifact states in the preliminary scenario are sent to the cross-validation and correction module. This module simultaneously receives ECG signals for arbitration and outputs a corrected scenario, i.e., the physiological wave dynamics or signal artifact state. This corrected scenario is also sent to the ECG model adaptive selection unit. The CG model adaptive selection unit selects a chosen ECG analysis model for the current scenario based on models in the preset ECG analysis model library (D1) and schedules it to the ECG signal processing unit. The ECG signal processing unit uses the selected model to process the raw ECG signal to identify cardiac risk mutations and output classification results. The cardiac risk mutation information is delivered to the monitoring personnel, while the classification results are sent to the report and log generation module. This module reads and writes the scenario status log (D2) and the risk counter (D3) to finally generate a cardiac risk attribution report and a cardiopulmonary system stability report, which are also submitted to the monitoring personnel.

[0040] like Figure 2 As shown, the left vertical axis represents the respiratory signal amplitude, the right vertical axis represents the ECG signal quality index (SQI) and the level of respiratory scenario classification, and the horizontal axis represents the time points from T0 to T8. The respiratory signal amplitude curve, ECG signal quality index (SQI) curve, and respiratory scenario classification curve in the figure together reveal the classification logic. For example, at T2 and T6, the respiratory signal amplitude fluctuates, but the SQI remains high quality. At this time, the respiratory scenario is initially classified as a fluctuating artifact state. At T3, the respiratory signal amplitude fluctuates and the SQI drops to low quality. The scenario is also initially classified as a fluctuating artifact state. These two fluctuating artifact states will be cross-validated and corrected based on the level of SQI. At T7, the respiratory signal amplitude reaches its peak, the SQI is low, and the scenario is classified as a respiratory event state.

[0041] like Figure 3As shown, respiratory, ECG, and body movement signals collected by vital sign sensors are acquired in real time. This data stream is used to perform individualized self-calibration and respiratory scenario classification. The scenario classification results trigger the execution of cross-validation rules to arbitrate fluctuation artifact states and are used to generate a scenario status log. This log is then used to generate a cardiopulmonary stability report and a CFI index. At the same time, the results of individualized self-calibration, respiratory scenario classification, and cross-validation are jointly fed into an adaptive ECG analysis model module. The output of this module is used to identify cardiac risk mutations. The identified mutation information is finally used to generate a cardiac risk attribution report. Ultimately, both the cardiopulmonary stability report and the cardiac risk attribution report are submitted to the monitoring personnel.

[0042] Example 4: This example illustrates a reproducible, standardized engineering calibration and implementation method of the present invention, clarifying how the rule-based respiratory state classifier and multiple ECG analysis models, which are the core of the method, are objectively and individually constructed and deployed. This method requires acquiring calibration data for at least 30 minutes. This data can come from a publicly available database annotated by professionals or from the initial stage of the user's monitoring task, and must include synchronized respiratory signals, ECG signals, and body motion signals acquired by an accelerometer. After obtaining this calibration data, the system performs a pre-calibration step. To determine individualized respiratory baseline parameters, the system iterates through the calibration data in 1-minute windows, calculating the variance of the body motion signal and the heart rate variability (SDNN) index of the ECG signal within each window. When the variance of the body motion signal in a window is lower than a preset activity threshold used to distinguish between rest and activity... Furthermore, its SDNN is above a baseline threshold characterizing autonomic neural stability. At this time, the window is marked as a stable sleep segment; the system compiles all respiratory signals marked as stable sleep segments and calculates their median respiratory rate, which is used as one of the user's individualized respiratory baseline parameters, denoted as . Simultaneously, the median of their respiratory amplitude is calculated and denoted as . After completing the individualized parameter calibration, the system constructs the specific judgment logic of the rule-based respiratory state classifier. In the subsequent real-time operation of the system, this classifier will operate as follows: The system extracts the respiratory state feature vector RSV in real time using a 5-second sliding window. The coefficient of variation of respiratory amplitude is obtained by calculating the standard deviation of each respiratory amplitude within the window and dividing it by its average value. The apnea or hypopnea event flag AHE_Flag is generated through an amplitude discrimination logic; that is, when a respiratory amplitude is detected to be lower than 10 seconds for more than 10 seconds... When the respiratory rate reaches 30%, AHE_Flag is set to 1; this RSV is input into the classifier, which executes the following rules: if AHE_Flag is 1, the current breathing pattern is classified as a respiratory event state scenario; if AHE_Flag is 0, but the respiratory rate is greater than 30%, the current breathing pattern is classified as a respiratory event state scenario. If the respiratory amplitude variation coefficient is greater than 0.45 or 1.3 times the average respiratory amplitude, the current respiratory pattern is initially classified as a fluctuating artifact state. This state will be further arbitrated by the SQI cross-validation rule in step c. If the above conditions are not met, it is classified as a steady state.

[0043] This embodiment further clarifies the internal differentiation rules of multiple ECG analysis models. These rules also utilize the baseline parameters obtained in the aforementioned calibration steps. One example is... This refers to the mean heart rate during stable sleep segments; when the system classification result is [steady-state scenario], the scheduled high-sensitivity baseline model executes a set of strict decision rules, which may include: if the current heart rate is greater than... If the value is 1.2 times higher than 105 bpm, it is considered a risk mutation; or if the SDNN index in the 1-minute window is lower than 105 bpm, it is considered a risk mutation. A mutation is considered a risk mutation if the QRS waveform is 0.75 times the SDNN baseline of the stable sleep segment; or if the morphological similarity between the QRS waveform and the stable period template is less than 0.92. When the system classification result is a physiological wave dynamic scenario, i.e., a scenario corrected by high-quality SQI validation of the wave artifact scenario, the scheduled wave period-specific model executes a set of functionally specialized rules, which may include: if the current heart rate is greater than... If the heart rate is 1.5 times higher than 135 bpm, it is considered a risk mutation. This heart rate threshold is higher than the corresponding threshold of the high-sensitivity baseline model to tolerate benign heart rate accelerations caused by REM sleep, etc. The judgment rules related to heart rate variability (SDNN) are disabled or the threshold is significantly relaxed. However, the QRS waveform morphological similarity rule remains strict. That is, if the morphological similarity is less than 0.92, it is still considered a risk mutation to capture malignant morphological changes.

[0044] Example 5: The method of the present invention can also be used to generate cardiac risk attribution reports. Its data processing system executes within a preset global time window, which can be set to an 8-hour nighttime monitoring cycle. At the beginning of this window, the system initializes a primary risk counter and a respiratory evoked risk counter, both of which are initially set to 0. Within the global time window, the data processing system continues to run. Whenever the system identifies a cardiac risk mutation in step e, the system additionally executes an attribution determination step. This step checks the respiratory scenario provided by step b or step c when the mutation is triggered. If it is determined that the mutation was identified by the high-sensitivity baseline model in a steady-state scenario, the system increments the count value of the primary risk counter by 1. If it is determined that the mutation was identified by the respiratory compensation analysis model in a respiratory event-state scenario, the system increments the count value of the respiratory evoked risk counter by 1. If the mutation occurs in other scenarios, the counters in this procedure do not count.

[0045] At the end of the global time window, the system reads the final count values ​​of the two counters; in one data instance, at the end of the 8-hour cycle, the final count value of the primary risk counter is 3, and the final count value of the respiratory induced risk counter is 29; based on these two count values, the system generates a cardiac risk attribution report, which presents the following data: a total risk event count of 32, and a primary risk event percentage of [missing data]. That is, 9.375%; the proportion of respiratory-induced risk events was That is, 90.625%, the report provides caregivers with a quantitative indicator of the main drivers of a user's nocturnal cardiac risk.

[0046] Example 6: This example discloses a standardized engineering calibration procedure for key algorithms and parameters in the method of this invention. This procedure is executed in an offline calibration phase to ensure the reproducibility and accuracy of the method deployment. The procedure requires an offline calibration dataset containing multiple sources and labeled by professionals. This dataset contains synchronously acquired respiratory signals and ECG signals with a cumulative duration of not less than 100 hours, and provides clear labels for the data segments. The labels are at least divided into: Class A (clear respiratory signals and stable physiological state), Class B (clear respiratory signals accompanied by REM sleep or body movement), Class C (respiratory signals distorted due to sensor displacement or saturation), Class D (clear ECG signals and containing known primary cardiac risk mutations), and Class E (ECG signals of low quality due to baseline drift or artifacts). The system traverses all Class A data segments, extracts all high-quality QRS waveforms, and generates a standard QRS template by aligning and calculating the median waveform, which is used for subsequent SQI morphological similarity calculation.

[0047] When calibrating the algorithm threshold, the system iterates through the dataset, using class C data as negative samples and classes A and B data as positive samples, to evaluate the respiratory signal quality index. The receiver operating characteristic (ROC) curve was analyzed using the calculation method, and a value was determined that made both specificity and sensitivity exceed 0.95. A numerical value, such as 0.72, is set as the preset quality threshold. The system uses the same method, with E-class data as negative samples and D-class data as positive samples, to evaluate the signal quality index of the ECG signal. The calculation method is used to perform ROC analysis and determine an optimal equilibrium point. The value, such as 0.55, is set as the preset value. Thresholds; When calibrating the internal parameters of the rule-based respiratory state classifier and multiple ECG analysis models, the system employs a multi-parameter optimization process. This process uses a comprehensive F1 score as the objective function, which comprehensively considers the classification accuracy and recall of four key events: A, B, C, and D. The system iteratively searches within a preset parameter space, which includes multiples of respiratory rate (e.g., from 1.1 to 1.5, step size 0.05) and thresholds for the coefficient of variation of amplitude (e.g., from 0.30 to 0.60, step size 0.05). The system uses a combination of parameters, including a respiratory rate multiple of 1.3, an amplitude variation coefficient threshold of 0.45, a heart rate multiple of 1.2 for the high-sensitivity baseline model, a heart rate multiple of 1.5 for the fluctuation-specific model, and a morphological similarity threshold of 0.92. This procedure transforms the setting of all key parameters from empirical assignment into a reproducible, data-driven optimization process.

[0048] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions 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 solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying cardiac risk using night-time vital sign data mutation, characterized in that, The method comprises: Step a, acquiring a respiratory signal and an electrocardiogram (ECG) signal in real time; Step b, classifying a current respiratory mode into one of a plurality of preset respiratory scenarios based on the respiratory signal, the plurality of preset respiratory scenarios comprising at least a stable state scenario, a respiratory event state scenario, and a fluctuation artifact state scenario; Step c, when the classification result of step b is the fluctuation artifact state scenario, performing a cross-validation rule, the rule comprising: acquiring a signal quality index (SQI) of the ECG signal in a same time window as the respiratory signal, and according to the respiratory mode being the fluctuation artifact state scenario and the SQI being determined as high quality, modifying a final classification of the scenario to a physiological fluctuation state scenario, and according to the respiratory mode being the fluctuation artifact state scenario and the SQI being determined as low quality, modifying the final classification of the scenario to a signal artifact state scenario; Step d, according to a result of the stable state scenario or the respiratory event state scenario classified in step b, or according to a final classification result of the physiological fluctuation state scenario or the signal artifact state scenario modified in step c, adaptively selecting an ECG analysis model corresponding to the scenario or the final classification from a plurality of preset ECG analysis models; Step e, using only the ECG analysis model selected in step d to process the ECG signal to identify whether a cardiac risk mutation exists.

2. The method for identifying cardiac risk using night-time vital sign data surges as claimed in claim 1, wherein, In step c, determining the SQI as high quality or low quality comprises: calculating a value of the SQI based on at least one of baseline drift, signal saturation, or QRS waveform features of the ECG signal; and comparing the value of the SQI with a preset SQI threshold to determine whether the SQI is high quality or low quality.

3. The method of identifying cardiac risk using nocturnal vital sign data excursions of claim 1, wherein, Before step b, the method further comprises: calculating a respiratory signal quality index (Resp-SQI) of the respiratory signal, and only when the Resp-SQI satisfies a preset quality threshold, performing steps b and c, and when the Resp-SQI does not satisfy the preset quality threshold, stopping performing steps b and c, and using a predefined high-sensitivity baseline model to process the ECG signal.

4. The method for identifying cardiac risk using night-time vital sign data surges as claimed in claim 3, wherein, Calculating the respiratory signal quality index (Resp-SQI) of the respiratory signal comprises: calculating a value of the Resp-SQI based on at least one of signal amplitude, saturation, or baseline drift features of the respiratory signal, and the preset quality threshold is a threshold for characterizing a reliable respiratory signal.

5. The method for identifying cardiac risk using night-time vital sign data surges as recited in claim 1, wherein, The plurality of preset ECG analysis models comprise: a high-sensitivity baseline model corresponding to the stable state scenario, a respiratory compensation analysis model corresponding to the respiratory event state scenario, a fluctuation period special model corresponding to the physiological fluctuation state scenario and maintaining high sensitivity to malignant morphological changes, and a comprehensive suppression model corresponding to the signal artifact state scenario.

6. The method for identifying cardiac risk using night-time vital sign data surges as recited in claim 1, wherein, In step b, classifying the current respiratory mode into one of the plurality of preset respiratory scenarios based on the respiratory signal comprises: extracting a set of respiratory state feature vectors (RSVs) from the respiratory signal; and inputting the respiratory state feature vectors (RSVs) into a rule-based respiratory state classifier to output one of the plurality of preset respiratory scenarios.

7. The method for identifying cardiac risk using night-time vital sign data surges as recited in claim 1, wherein, The method further comprises: recording the time series of the respiratory scenarios output by step b or the final classification corrected by step c in real time within a preset global time window to form a scenario state log; performing time series statistical analysis on the scenario state log at the end of the global time window to generate a quantitative report representing the stability of the cardiopulmonary system; the time series statistical analysis comprises: counting the total number of times of scenario state switching within the global time window , and according to calculating a scenario fragmentation index CFI, wherein is the total duration of the global time window.

8. The method for identifying cardiac risk using night-time vital sign data surges as recited in claim 6, wherein, The respiratory status vector RSV comprises at least: a respiratory rate, a respiratory amplitude coefficient of variation, and an apnea or hypopnea event flag AHE_Flag.

9. The method for identifying cardiac risk using night-time vital sign data surges as recited in claim 6, wherein, Before step b is performed, the method further comprises: within a calibration period, automatically identifying and capturing a piece of respiratory signal segment representing stable sleep of the user by combining body motion signal and heart rate variability analysis of electrocardiogram ECG signal, performing statistical analysis on the respiratory signal segment, calculating individualized respiratory baseline parameters of the specific user, and dynamically adjusting the decision threshold for determining stable state scenario inside the rule-based respiratory status classifier using the individualized respiratory baseline parameters.

10. The method for identifying cardiac risk using night-time vital sign data surges as recited in claim 1, wherein, The method further comprises: within a preset global time window, setting a primary risk counter and a respiratory-induced risk counter, whenever a cardiac risk mutation is identified in step e, if the cardiac risk mutation occurs in a stable state scenario, counting the primary risk counter; if the cardiac risk mutation occurs in a respiratory event state scenario, counting the respiratory-induced risk counter; at the end of the global time window, generating a cardiac risk attribution report based on the count values of the primary risk counter and the respiratory-induced risk counter.

Images