A risk prediction method for the attack pattern of sleep apnea syndrome

Abstract

The present invention discloses a method for predicting the risk of apnea syndrome attack patterns, which relates to the field of health management. It calls the user status data of all age groups, mines the risk arrays based on status risk characteristics under different attack patterns, conducts lightweight monitoring sampling and risk prediction training based on single features and combined features, constructs a risk prediction model, collects the sleep monitoring data of the target user and performs risk prediction and attack pattern determination, determines the risk prediction result and issues a user status warning, so as to solve the technical problem in the prior art of how to effectively extract and utilize the user's status data for efficient and accurate health supervision, and thus provide early risk warnings in daily monitoring. Through multi-feature mining and lightweight monitoring models, it provides a personalized and low-power risk prediction and management solution, which can effectively improve the efficiency and accuracy of risk prediction.

Description

Technical Field

[0001] The present invention relates to the field of health management, and particularly to a method for predicting the risk of apnea syndrome attack patterns. Background Art

[0002] Sleep apnea syndrome is a common sleep breathing disorder, mainly characterized by repeated pauses or hypoventilation during sleep, which may lead to decreased oxygen saturation, cardiovascular diseases and other serious health problems. In the prior art, analysis is mainly carried out through means such as polysomnography. Although accurate, the equipment is expensive, the operation is complex, and the patient needs to complete the examination in a medical institution, which is difficult to meet the needs of daily monitoring. In addition, portable sleep monitoring devices are gradually popularized, but there are still deficiencies in data collection and risk prediction, especially in the lack of effective methods for fine-grained risk assessment of attack patterns.

[0003] Currently, the prior art often relies on single features or simple thresholds for risk determination, and it is difficult to comprehensively capture the personalized risk characteristics of different users. In addition, the resource consumption of the monitoring sampling and analysis process is relatively large, which limits the real-time performance and low-power requirements of the device. Therefore, how to effectively extract and utilize the status data of users for efficient and accurate health supervision, so as to provide early risk warnings in daily monitoring, is a technical problem that urgently needs to be solved at present. Summary of the Invention

[0004] This application provides a method for predicting the risk of apnea syndrome attack patterns, which is used to solve the technical problem in the prior art of how to effectively extract and utilize the status data of users for efficient and accurate health supervision, so as to provide early risk warnings in daily monitoring.

[0005] In view of the above problems, this application provides a method for predicting the risk of apnea syndrome attack patterns.

[0006] In a first aspect, this application provides a method for predicting the risk of apnea syndrome attack patterns. The method includes: calling the status data of users of all age stages, mining the risk matrix based on the status risk characteristics under different attack patterns, where the status risk characteristics include pattern common characteristics and pattern specific characteristics, and defining the boundary of the feature risk degree with absolute contraction; based on the risk matrix, performing lightweight monitoring sampling and risk prediction training based on single features and combined features, and constructing a risk prediction model, where the risk prediction model is built into the sleep monitoring device; connecting the sleep monitoring device, combining the risk prediction model, determining the sleep monitoring data of the target user through staggered sampling, performing data transmission back and executing risk prediction and attack pattern determination to determine the risk prediction result; based on the risk prediction result, performing status alarm on the target user, and generating a time series monitoring table for the monitoring time zone.

[0007] In a second aspect, the present application provides a risk prediction system for the attack mode of sleep apnea syndrome. The system includes: a data mining module, which is used to call the user status data of all age stages and mine the risk matrix based on the status risk characteristics under different attack modes. Among them, the status risk characteristics include mode common characteristics and mode specific characteristics, and the boundary of the feature risk degree is defined by absolute contraction; a training module, which is used to perform lightweight monitoring sampling and risk prediction training based on the single feature and combined features based on the risk matrix, and construct a risk prediction model. Among them, the risk prediction model is built into the sleep monitoring device; a risk prediction module, which is used to connect to the sleep monitoring device, combine the risk prediction model, determine the sleep monitoring data of the target user through staggered sampling, perform data transmission back and execute risk prediction and attack mode determination to determine the risk prediction result; a risk management module, which is used to perform status warning on the target user based on the risk prediction result and generate a time series monitoring table for the monitoring time zone.

[0008] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0009] A risk prediction method for the attack mode of sleep apnea syndrome provided by an embodiment of the present application calls the user status data of all age stages, mines the risk matrix based on the status risk characteristics under different attack modes. The status risk characteristics include mode common characteristics and mode specific characteristics, and the boundary of the feature risk degree is defined by absolute contraction; based on the risk matrix, perform lightweight monitoring sampling and risk prediction training based on single feature and combined features, construct a risk prediction model and build it into the sleep monitoring device; connect to the sleep monitoring device, combine the risk prediction model, determine the sleep monitoring data of the target user through staggered sampling, perform data transmission back and execute risk prediction and attack mode determination to determine the risk prediction result; based on the risk prediction result, perform status warning on the target user and generate a time series monitoring table for the monitoring time zone, which is used to solve the technical problem in the prior art of how to effectively extract and utilize the user's status data for efficient and accurate health supervision, so as to provide early risk warning in daily monitoring. Through multi-feature mining and lightweight monitoring models, a personalized and low-power risk prediction and management solution is provided, which can effectively improve the efficiency and accuracy of risk prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 FIG. is a schematic flow chart of a risk prediction method for the attack mode of sleep apnea syndrome provided by the present application;

[0011] Figure 2This application provides a schematic diagram of the determination process for the monitoring sampling method in a risk prediction method for the attack mode of sleep apnea syndrome;

[0012] Figure 3 This application provides a schematic diagram of the structure of a risk prediction system for the attack mode of sleep apnea syndrome.

[0013] Explanation of reference numerals: data mining module 11, training module 12, risk prediction module 13, risk management module 14. Detailed implementation manners

[0014] This application provides a risk prediction method for the attack mode of sleep apnea syndrome, which calls the user status data of all age stages, mines the risk matrix based on the status risk characteristics under different attack modes, performs lightweight monitoring sampling and risk prediction training based on single features and combined features, constructs a risk prediction model, collects the sleep monitoring data of the target user, executes risk prediction and attack mode determination, determines the risk prediction result, and issues a user status alarm, so as to solve the technical problem in the prior art of how to effectively extract and utilize the user's status data for efficient and accurate health supervision, so as to provide early risk warning in daily monitoring.

[0015] Example 1: As Figure 1 shown, this application provides a risk prediction method for the attack mode of sleep apnea syndrome, and the method includes:

[0016] S1: Call the user status data of all age stages, and mine the risk matrix based on the status risk characteristics under different attack modes, where the status risk characteristics include mode common characteristics and mode specific characteristics, and the boundary of the feature risk degree is defined by absolute contraction.

[0017] In the embodiment of this application, when calling the user status data of all age stages, it is first necessary to obtain a sleep-related data set covering multi-dimensional characteristics such as different age groups, genders, and physiques. This data should include basic status information such as the user's ventilation volume, heart rate, respiratory rate, oxygen saturation, and movement status to ensure its universality and representativeness. By mining this data, it can be identified that the attack modes of sleep apnea syndrome include obstructive sleep apnea hypopnea syndrome, central sleep apnea syndrome, and mixed apnea syndrome, and different modes are different in terms of etiology status and specific manifestations. Further, the status of the attack mode in different individuals and situations is also different, specifically including various manifestations such as single attack, periodic attack, and persistent risk. For further refined analysis, the status risk characteristics related to each attack mode need to be extracted.

[0018] The state risk characteristics can be divided into pattern common characteristics and pattern specific characteristics. Pattern common characteristics refer to risk indicators that commonly appear in different seizure patterns, such as the degree and frequency of oxygen saturation decrease. Pattern specific characteristics are unique indicators under specific seizure patterns, such as the change trend of respiratory rhythm within a specific time period during periodic seizures. Through the refined analysis of these characteristics, the internal laws of seizure patterns can be described more accurately.

[0019] Furthermore, after the feature extraction is completed, by constructing a risk matrix, the above state risk characteristics are classified and quantified according to patterns. The risk matrix is a multi-dimensional matrix structure, and each dimension represents the feature risk degree under different seizure patterns, which is used to describe the risk level of specific features. To ensure the scientificity and consistency of the definition of risk degree, the method of absolute contraction is introduced to determine the critical values where risks may exist. These boundary values not only represent the critical thresholds of risk characteristics but also provide a quantitative standard for subsequent pattern determination.

[0020] Through the above process, the invocation and analysis of user state data in all age stages are completed, which not only lay a foundation for the feature recognition of different seizure patterns but also provide multi-dimensional support for risk prediction through the establishment of a risk matrix, providing a basis for the precise risk management of sleep apnea syndrome.

[0021] Furthermore, to mine the risk matrix of state risk characteristics under different seizure patterns, step S1 of this application includes:

[0022] Obtain the first hypopnea feature, where the first hypopnea feature ∈ the state risk characteristics; traverse the seizure patterns, and based on the user state data, conduct feature mining and clustering to determine N eigenvalue groups based on the first hypopnea feature, where the seizure pattern corresponds to the eigenvalue group one by one; based on the eigenvalue groups, conduct absolute maximum contraction within the group to determine N risk degree boundary values of the first hypopnea feature, where the first hypopnea feature is an extremely small feature negatively correlated with the risk degree; add the N risk degree boundary values into the risk matrix.

[0023] Among them, for extremely large features, absolute minimum contraction is used as the absolute contraction method; for extremely small features, absolute maximum contraction is used as the contraction method.

[0024] In the embodiments of the present application, the first hypopnea feature is used to measure the degree of decrease in the ventilation volume of the user within a specific time period, and is one of the feature in the state risk features. Hypopnea feature data related to sleep apnea syndrome is screened from the user state data, and feature mining and clustering analysis are performed according to different attack modes, such as single attack, periodic attack, and complex mixed attack, etc. Specifically, through the statistical and clustering processing of the user state data, the user state data related to the first hypopnea feature is grouped, and N eigenvalue groups are integrally determined. Each eigenvalue group corresponds to an attack mode, and the eigenvalues within the group reflect the distribution law of the hypopnea feature in this mode.

[0025] Since the first hypopnea feature belongs to an extremely small feature, that is, the smaller its value, the higher the risk degree. By using the method of absolute maximum contraction within the group, the N eigenvalue groups are processed, which can ensure that the boundary values accurately reflect the full risk interval to determine the N risk degree boundary values of the first hypopnea feature. Exemplarily, through absolute maximum contraction, that is, contracting towards the upper limit of the data interval, the boundary of the eigenvalue distribution is obtained. For example, when the value distribution of an eigenvalue group of the first hypopnea feature is [50, 80], after absolute maximum contraction, the risk degree boundary value is 70, that is, 70 is the critical value where risk may exist. As long as it reaches 70, there may be a risk. The smaller the data, the greater the probability of risk. 70 is used as the standard for risk determination.

[0026] For extremely large features (that is, the larger the eigenvalue, the higher the risk degree), according to the determined eigenvalue groups, the method of absolute minimum contraction is adopted, that is, contracting towards the lower limit of the data interval, so as to obtain the risk boundary of the eigenvalue distribution, that is, the critical value where risk may exist, as the risk degree boundary value.

[0027] For each state risk feature in the state risk features, according to the above steps, feature mining and clustering based on the user state data are respectively performed, the feature type is determined, that is, extremely large or extremely small, and absolute contraction processing is performed to determine the corresponding risk degree boundary value.

[0028] Finally, the above risk degree boundary values are integrated to generate the risk matrix, which is used as a quantitative risk indicator for the corresponding attack mode. The update and improvement of the risk matrix provide accurate input basis for subsequent mode determination and risk prediction, and at the same time further improve the adaptability and operability of the risk assessment model.

[0029] Further, adding the N risk degree boundary values into the risk matrix, step S1 of the present application includes:

[0030] Based on the risk evolution trend of the first hypopnea feature, determine a preset tolerance scale, where the preset tolerance scale is positively correlated with the risk evolution speed; based on the preset tolerance scale, perform an out-of-range expansion process on the N risk degree boundary values to determine the N risk precursor feature values of the first hypopnea feature; establish a mapping between the attack mode and the N risk precursor feature values and add it to the risk matrix.

[0031] In the embodiment of the present application, the risk evolution trend of the first hypopnea feature, that is, the variation law over time under different attack modes, namely how the risk level gradually evolves from low to high. This trend is usually achieved through time series analysis, reflecting the dynamic change rate and direction of the first hypopnea feature in a specific attack mode. The specific manifestations of the risk evolution trend may include the feature value gradually decreasing to the critical point or dropping sharply within a certain time period, etc. The preset tolerance scale refers to the tolerance range set to allow normal fluctuations or avoid misjudgment, and its size is positively correlated with the risk evolution speed. For example, for a mode with a faster risk evolution speed, the tolerance scale needs to be enlarged accordingly to capture potential risks in advance.

[0032] Based on the above preset tolerance scale, perform an out-of-range expansion process on the N risk degree boundary values to determine the N risk precursor feature values of the first hypopnea feature. The out-of-range expansion process means adding an extension on the basis of the original boundary value to describe the possible early risks. For example, when the risk degree boundary value of the first hypopnea feature is 70, if the preset tolerance scale is 5, the risk degree boundary value after the out-of-range expansion process is 75, which can effectively enhance the sensitivity to risk precursors and avoid missing the judgment of early risks especially in practical applications.

[0033] Similarly, if it is an extremely small feature, an expansion in the decreasing direction is required, such as adjusting 70 to 65 as the risk precursor feature value.

[0034] After determining the above N risk precursor feature values, further establish a mapping relationship between the attack mode and its corresponding risk precursor feature values. This mapping relationship is a one-to-one logical binding, clearly indicating which precursor feature values may indicate the attack risk under a specific mode. For example, for a periodic attack mode, the range of its risk precursor feature values may show a periodic distribution, which is significantly different from the random distribution characteristics of other modes.

[0035] Finally, add the above mapping relationship together with the risk precursor feature values to the risk matrix, so that the risk matrix not only contains the risk degree boundary values, but also has the ability to predict precursor risks. It can more comprehensively support the early identification and mode determination of sleep apnea syndrome.

[0036] S2: Based on the risk matrix, perform lightweight monitoring sampling and risk prediction training under single features and combined features to construct a risk prediction model, where the risk prediction model is built into the sleep monitoring device.

[0037] First, by analyzing the risk characteristic values and precursor characteristic values of different seizure patterns in the risk matrix, through feature grouping and staggered sampling configuration, after determining the sampling method, perform lightweight monitoring sampling and training. The "lightweight" means that in a hardware environment with limited resources, by optimizing the sampling frequency, data volume, and computational complexity, low-power and high-efficiency sampling and risk prediction processing training are achieved. Specifically, adopt grouped staggered sampling and a sample training-based method to perform precise control training for feature sampling triggering and construct the first sampling unit.

[0038] Furthermore, taking single feature determination - seizure mode determination - combined feature determination as the benchmark, taking the risk degree boundary matching determination based on the risk matrix as the single feature determination method, taking the matching of pattern-specific features as the seizure mode determination method, and by calling training samples, that is, risk analysis records determined through big data retrieval, train until convergence in a sample supervised training manner. If the preset output accuracy is met, obtain the second decision unit. The sensitivity and reliability of risk decision-making are significantly improved through a multi-level judgment mechanism.

[0039] Then, connect the second decision unit to the back of the first sampling unit to generate the risk prediction model. The risk prediction model is built into the sleep monitoring device, enabling the device to have the ability of real-time data collection and risk prediction. Through the low-latency operation of the built-in model, users can obtain personalized risk warnings and pattern analysis results without relying on external devices or complex data transmission, providing convenient support for daily health management. At the same time, the construction and deployment of this model not only achieve efficient risk prediction but also achieve a balance among cost, power consumption, and usability.

[0040] Furthermore, to construct the risk prediction model, step S2 of this application includes:

[0041] Determine the monitoring sampling method, perform sensing sampling training, and construct the first sampling unit; perform sample-driven training based on the risk degree boundary value determination of a single detection feature and the collaborative determination of a detection feature group to construct the second decision unit, where the seizure mode is determined by the specific features; combine the first sampling unit and the second decision unit to generate the risk prediction model.

[0042] In the embodiment of the present application, a specific strategy for sensor sampling is formulated by analyzing the characteristic distribution and sampling requirements in the user status data. The core of the monitoring sampling method is to optimize the sampling frequency and sensor configuration so as to achieve comprehensive capture of risk characteristics with minimum data volume and power consumption. According to the monitoring sampling method, sensor sampling training is performed to form a functional module that focuses on the processing and storage of sampled data, namely the first sampling unit.

[0043] After completing the construction of the first sampling unit, further risk judgment training based on single detection features and detection feature groups is carried out to construct the second decision unit. The risk boundary value judgment of a single detection feature refers to comparing the single feature value obtained by sensor sampling with the boundary value in the risk array one by one to determine whether there is a potential risk. For example, when the hypoventilation feature value is less than its boundary value (such as the critical point 75), the risk flag is triggered immediately. Since there is a certain correlation between the features, the collaborative judgment of the detection feature group analyzes the relationship between multiple features and mines the risk information caused by the collaboration of multiple features in this mode. For example, if the respiratory rate decreases and the blood oxygen saturation decreases rapidly, the collaborative feature points to a higher risk level. Through the sample-driven training method, a large amount of historical data is learned and optimized to ensure that the second decision unit can accurately judge the risk situation of single features and combined features.

[0044] Among them, the second decision unit uses pattern-specific characteristics as key indicators, that is, the difference between different patterns is used as the judgment standard to further realize the accurate judgment of the attack pattern. For example, by identifying the repetitiveness of periodic characteristics or the rapid change trend of sudden characteristics, it is judged which attack mode (such as single attack or mixed attack) the current attack belongs to. This pattern recognition mechanism based on specific characteristics determines the corresponding pattern and conducts analysis under the attack mode, which can help improve the pertinence and practicality of risk prediction.

[0045] Finally, the first sampling unit and the second decision unit are combined to generate a complete risk prediction model. The first sampling unit is responsible for efficient data collection to ensure the integrity and timeliness of the input; the second decision unit achieves high-precision risk assessment and attack pattern determination through intelligent analysis and pattern matching. The organic combination of the two forms a closed-loop monitoring and prediction system that can process the user's sleep data in real time and output accurate risk prediction results, providing reliable technical support for the prevention and control of sleep apnea syndrome.

[0046] Further, such as Figure 2 As shown, the step S2 of determining the monitoring sampling method includes:

[0047] Based on feature correlation, combine the state risk features to determine M detection feature groups; serialize the M detection feature groups to determine a risk detection sequence, where each detection feature group corresponds to a sequence node; set a preset sampling interval, and based on the risk detection sequence, perform out-of-time sampling configuration of the state risk features to determine the monitoring sampling method.

[0048] In the embodiments of the present application, based on feature correlation, the state risk features are combined. According to the potential association and complementarity between different features, the accuracy of risk determination can be effectively improved, and the errors existing in single-feature analysis can be avoided. At the same time, through grouping, the sampling monitoring of one feature group is performed once, and out-of-time monitoring sampling is performed between different feature groups. On the basis of ensuring the analysis accuracy, a large amount of data analysis in full-feature monitoring sampling can be avoided, and the accuracy and efficiency of risk decision-making can be effectively balanced.

[0049] Specifically, the state risk features refer to the key indicators affecting the onset of sleep apnea syndrome, such as hypopnea features, blood oxygen saturation, respiratory rate, etc. Through combination, these features can reveal more complex risk patterns. For example, the changes in hypopnea features and respiratory rate may jointly indicate the occurrence of a risk state under a certain onset mode. Through feature correlation, it is identified which features have strong correlations, and these features with strong correlations are combined into one group to form M detection feature groups. The composition of the detection feature groups will be based on the complementarity and correlation between features, which can further improve the accuracy of risk determination. For example, hypopnea and blood oxygen saturation may form a group, while respiratory rate and pulse, heart rate may form another group, and each detection feature group corresponds to at least two state risk features.

[0050] Next, serialize the M detection feature groups. Exemplarily, a time series is constructed, and each sequence node corresponds to one of the M detection feature groups. Among them, the matching of the detection feature groups based on the time series is randomly assigned, so that each detection feature group can be monitored and analyzed in a certain order. The serialization process arranges the M feature groups in chronological order or according to their risk assessment priorities to ensure that key risk changes can be captured in a timely manner during the monitoring process. Preferably, high-frequency monitoring of strongly correlated features can be performed according to the sleep stage, while appropriately reducing the monitoring frequency of non-strongly correlated features. For example, the monitoring of hypopnea and blood oxygen saturation may need to be carried out in the initial and middle stages of sleep, while the monitoring of respiratory rate may focus on the deep stage of sleep. Through this serialization process, a risk detection sequence is formed, where each detection feature group corresponds to a specific sequence node, which means that in the actual monitoring process, these feature groups are sampled and analyzed in the preset order.

[0051] To ensure the rationality of data collection and monitoring efficiency, a preset sampling interval is set. Here, the preset sampling interval is the sampling interval time range set by those skilled in the art. For example, three minutes.

[0052] Furthermore, based on the risk detection sequence and the preset sampling interval, a misaligned sampling configuration for state risk characteristics is performed. Specifically, after the monitoring sampling of the detection feature group of the previous sequence node is completed according to the risk detection sequence, based on the preset sampling interval, the monitoring sampling of the monitoring feature group of the next sequence node is performed. One group of monitoring feature sampling is performed each time, and the state risk characteristics are sampled in batches, which can reduce the amount of data processed each time on the basis of ensuring accuracy, and reasonably balance the accuracy and efficiency of risk decision-making.

[0053] Preferably, the sampling frequency can be appropriately adjusted according to the sleep state characteristics. For example, in the deep sleep stage, the change of hypoventilation characteristics may be more obvious, so the sampling frequency of hypoventilation characteristics can be increased at this time; while in the light sleep stage, the sampling frequency is appropriately reduced to reduce power consumption and interference.

[0054] By considering feature correlation for feature grouping and performing misaligned sampling configuration for features, a sampling strategy is formed that can not only efficiently capture risk changes but also minimize energy consumption and data redundancy to the greatest extent.

[0055] S3: Connect the sleep monitoring device, combine the risk prediction model, determine the sleep monitoring data of the target user through misaligned sampling, perform data transmission back and execute risk prediction and seizure mode determination to determine the risk prediction result.

[0056] Among them, the sleep monitoring device usually includes a variety of sensors, such as a respiratory sensor, a blood oxygen sensor, a heart rate monitoring sensor, etc., for real-time collection of the user's physiological data. By connecting the sleep monitoring device, the state sensing monitoring of the user during sleep can be carried out in real time for subsequent analysis and processing.

[0057] Based on the first sampling unit, according to the monitoring sampling method, misaligned sampling based on the risk detection sequence is performed, that is, taking the detection feature group corresponding to the first sequence node as the benchmark for sensing acquisition. When the preset sampling interval is met, taking the detection feature group corresponding to the second sequence node as the benchmark for sensing acquisition, and taking the data of a single sampling as the sleep monitoring data of the target user.

[0058] Further transmit the sleep monitoring data to the second decision-making unit of the risk prediction model to provide real-time input for risk prediction. For example, the device records data such as the user's respiratory rate, blood oxygen saturation, and heart rate in real time through sensors and transmits it back to the risk prediction model. At this time, the built-in algorithm of the model processes the received data, combines the existing characteristic risk boundaries and pattern recognition rules, and performs risk prediction and seizure pattern determination.

[0059] Specifically, perform risk prediction and seizure pattern determination for single-feature determination and feature combination determination respectively. Specifically, when a certain feature (such as hypopnea) crosses the preset risk boundary value, the model will recognize the potential apnea risk and determine whether it belongs to a certain specific seizure pattern. Further, perform feature combination determination under this seizure pattern. For example, when the respiratory rate drops sharply and is accompanied by a decrease in blood oxygen saturation, the model may determine it as a severe apnea attack. In this case, the risk prediction result will be a high-risk warning and relevant countermeasures will be prompted. By performing complementary feature determination, the accuracy of risk analysis can be effectively guaranteed.

[0060] S4: Based on the risk prediction result, perform status warning for the target user and generate a time-series monitoring table for the monitoring time zone.

[0061] Specifically, based on the risk prediction result, perform status warning for the target user to ensure that the user can get a timely response when discovering potential health risks. Exemplarily, if the risk value exceeds the preset safety limit or a possible severe seizure pattern (such as aggravated apnea, hypoxemia, etc.) is predicted, the status warning mechanism will be triggered. The status warning can be notified in various ways, such as the sleep monitoring device alarming, the terminal displaying a warning, etc. The content of the warning includes the user's current risk status and possible health impacts. It can help the user take corresponding intervention measures in the first time and reduce the health risks caused by delayed treatment.

[0062] The time-series monitoring table of the monitoring time zone shows the detailed risk changes and corresponding health status in each time period during the user's sleep. The main purpose of this process is to show the health dynamics of the user in each sleep period through the time series of data. For example, in the table, the system will combine the monitoring data and risk prediction results in different time periods according to the time axis to generate a clear time-series monitoring table. Each monitoring time zone (such as the start of sleep, light sleep period, deep sleep period, etc.) will record the corresponding physiological data and risk prediction results, such as the changes in indicators such as hypopnea characteristics, blood oxygen saturation, and respiratory rate. The time-series monitoring table can show in detail whether a feature has crossed the set risk boundary or whether there is a precursor of a seizure pattern in each time period.

[0063] In addition, the time series monitoring table understands the health change trends at different stages by tracing the data trajectory. The combination of status alerts and the time series monitoring table not only improves the accuracy of individual health management but also effectively promotes disease prevention and early intervention, providing users with scientific and accurate health management services.

[0064] Furthermore, for the execution of risk prediction and seizure pattern determination, step S3 of this application includes:

[0065] Transmit the sleep monitoring data back to the second decision-making unit of the risk prediction model, traverse the risk matrix for matching, risk degree boundary value determination, and trend prediction to determine the first determination data; traverse the first determination data to identify boundary crossing features. If the boundary crossing features are empty, terminate the risk decision-making process. If the boundary crossing features are not empty, perform seizure pattern matching and risk determination.

[0066] In the embodiment of this application, the sleep monitoring data is transmitted back to the second decision-making unit of the risk prediction model, and the second decision-making unit is used to further analyze and make decisions on the input monitoring data. Matching is performed through the risk feature boundary values to help determine the risk status of the user.

[0067] Specifically, the risk matrix contains multi-dimensional risk feature data, and each risk feature has a corresponding risk degree boundary value. The sleep monitoring data collected in real time is compared based on the risk boundary values. Through the sleep monitoring data, that is, the sampling values of each feature in this detection feature group, matching is performed based on the risk matrix, and the corresponding risk degree boundary value is determined based on the matching result to perform single-feature crossing limit determination and trend prediction as the first determination data.

[0068] For example, if a certain monitoring data (such as respiratory rate) breaks through the preset risk degree boundary (for example, the respiratory rate is lower than the threshold), it will be marked as a potential risk. In addition to the determination of the risk degree boundary, at the same time, the change trend of the data is predicted. If the data shows a continuous change trend (such as a gradually decreasing oxygen saturation), further speculate on its possible development direction (such as a possible apnea attack) and generate a trend prediction result.

[0069] Further traverse the first determination data to identify boundary crossing features, that is, whether any feature has crossed its corresponding risk boundary. The boundary crossing feature refers to when a certain detection feature, that is, a physiological parameter (such as hypoventilation) exceeds the set risk degree boundary, this feature is marked as a boundary crossing feature. If the boundary crossing features are empty, it means that all monitoring data have not exceeded the safe range, and at this time, the system will stop the subsequent risk decision-making process, indicating that the user has no immediate health risk at the current moment.

[0070] If the boundary-crossing feature is not empty, it indicates that at least one physiological data point has exceeded the preset safety boundary, suggesting potential health risks. At this time, further risk determination is performed under seizure pattern matching and feature combination.

[0071] Through the above steps, with the minimum amount of data processing, the accuracy of risk detection is maximally guaranteed, potential health risks can be accurately identified, helping users take timely actions, and thus effectively preventing the occurrence of serious health problems.

[0072] Furthermore, for seizure pattern matching and risk determination, step S3 of this application includes:

[0073] If the boundary-crossing feature is not empty, determine the target seizure pattern by matching with the pattern-specific feature; for the target seizure pattern, perform feature combination determination and trend prediction to determine the second determination data; fuse the first determination data and the second determination data as the risk prediction result.

[0074] Specifically, if the boundary-crossing feature is not empty, it means that during sleep monitoring, some key physiological parameters (such as blood oxygen saturation, respiratory rate, etc.) have exceeded the preset safety threshold, indicating potential health risks. In this case, the system will determine the target seizure pattern by matching with the pattern-specific feature. The pattern-specific feature refers to the distinguishing features under different seizure patterns.

[0075] By matching the boundary-crossing feature with the pattern-specific feature, determine the current seizure pattern based on the matching result as the target seizure pattern.

[0076] Once the target seizure pattern is determined, perform feature combination determination and trend prediction for this pattern. Feature combination determination means that under the target seizure pattern, cooperate with the current detection feature group and comprehensively consider multiple relevant physiological features (such as heart rate fluctuations during apnea, blood oxygen decline, etc.) to further analyze health risks. For example, if in the seizure pattern of apnea, the system detects that the heart rate gradually rises and the blood oxygen saturation gradually decreases, this may indicate the aggravation of the condition. Trend prediction is to predict the change trend of these physiological features in the future period of time (such as continuous blood oxygen decline or aggravated apnea) based on comprehensive multi-dimensional data to evaluate the dynamic change of risks.

[0077] Through feature combination determination and trend prediction, determine the second determination data. The second determination data not only reflects the risk situation in the current state but also considers the possible health change trend. For example, if it is predicted that the user's blood oxygen will further decline in the next few minutes, a risk warning will be issued in advance according to this trend to ensure timely intervention measures are taken.

[0078] Finally, fuse the first determination data and the second determination data as the risk prediction result. The first determination data is derived from a preliminary risk assessment, including the determination of individual feature boundary values and a preliminary judgment on whether the limit is exceeded; while the second determination data provides a more accurate and dynamic risk prediction through deeper feature combination analysis and trend prediction. After combining the two, the system can generate a comprehensive risk prediction result, accurately reflecting the current health status of the target user and predicting possible future risks. This result will be used as the basis for subsequent alarm and intervention decisions to ensure that users can receive corresponding health management in a timely manner.

[0079] A method for predicting the risk of an apnea syndrome attack mode provided by this application has the following technical effects:

[0080] 1. By combining different attack modes, mining a risk matrix based on state risk features, dividing common features and mode-specific features, and using an absolute shrinkage method to define the boundary of feature risk degree, it can comprehensively cover the health risks of users of different age groups, fully consider the diversity of attack modes, and improve the accuracy and universality of risk prediction.

[0081] 2. Taking the hypoventilation feature as an example, determine the mining process of state risk features. Accurately defining the risk boundary of features can effectively identify risks related to health problems such as apnea, and improve the sensitivity of risk determination. By setting a preset tolerance scale, performing value expansion processing, and determining the risk precursor feature value, it can give a timely warning before the attack, enhance the early warning ability, and gain time for intervention measures.

[0082] 3. By dividing feature groups for misaligned sampling configuration, it can effectively balance the processing of data volume and analysis accuracy, achieve lightweight configuration, and combine feature risk degree boundary determination and feature group collaborative determination to enhance the accuracy and reliability of prediction.

[0083] In summary, it can provide a comprehensive, real-time and high-precision risk prediction of the apnea syndrome attack mode, significantly improve the early warning ability and intervention efficiency, and prevent the occurrence of health risks.

[0084] Embodiment 2: Based on the same inventive concept as the method for predicting the risk of an apnea syndrome attack mode in the foregoing embodiment, as Figure 3 shown, this application provides a system for predicting the risk of an apnea syndrome attack mode, and the system includes:

[0085] A data mining module 11, which is used to call user state data of all age stages, mine a risk matrix based on state risk features under different attack modes, where the state risk features include mode common features and mode-specific features, and define the boundary of feature risk degree by absolute shrinkage;

[0086] A training module 12, which is used to perform lightweight monitoring sampling and risk prediction training based on single features and combined features based on the risk matrix, and construct a risk prediction model, wherein the risk prediction model is built into the sleep monitoring device;

[0087] A risk prediction module 13, which is used to connect to the sleep monitoring device, combine the risk prediction model, determine the sleep monitoring data of the target user through staggered sampling, perform data backhaul and execute risk prediction and seizure mode determination to determine the risk prediction result;

[0088] A risk management module 14, which is used to perform status warning on the target user based on the risk prediction result and generate a time-series monitoring table for the monitoring time zone.

[0089] Wherein, the data mining module 11 is further used to perform the following steps: obtain the first hypopnea feature, wherein the first hypopnea feature ∈ the state risk feature; traverse the seizure mode, perform feature mining and clustering based on the user state data to determine N eigenvalue groups based on the first hypopnea feature, wherein the seizure mode corresponds to the eigenvalue group one by one; based on the eigenvalue group, perform the absolute maximum contraction within the group to determine N risk degree boundary values of the first hypopnea feature, wherein the first hypopnea feature is an extremely small feature negatively correlated with the risk degree; add the N risk degree boundary values to the risk matrix.

[0090] Wherein, for extremely large features, the absolute minimum contraction is used as the absolute contraction method; for extremely small features, the absolute maximum contraction is used as the contraction method.

[0091] Wherein, the data mining module 11 is further used to perform the following steps: determine a preset tolerance scale based on the risk evolution trend of the first hypopnea feature, wherein the preset tolerance scale is positively correlated with the risk evolution speed; based on the preset tolerance scale, perform value expansion processing on the N risk degree boundary values to determine N risk precursor eigenvalue of the first hypopnea feature; establish a mapping between the seizure mode and the N risk precursor eigenvalue and add it to the risk matrix.

[0092] Wherein, the training module 12 is further used to perform the following steps: determine the monitoring sampling method, perform sensing sampling training, and construct a first sampling unit; perform sample-driven training based on the risk degree boundary value determination of the single detection feature and the collaborative determination of the detection feature group to construct a second decision unit, wherein the seizure mode is determined by the specific feature; combine the first sampling unit and the second decision unit to generate the risk prediction model.

[0093] Among them, the training module 12 is further configured to perform the following steps: based on feature correlation, combine the state risk features to determine M detection feature groups; serialize the M detection feature groups to determine a risk detection sequence, where each detection feature group corresponds to a sequence node; set a preset sampling interval, and based on the risk detection sequence, perform out-of-time sampling configuration of the state risk features to determine a monitoring sampling method.

[0094] Among them, the risk prediction module 13 is further configured to perform the following steps: transmit the sleep monitoring data back to the second decision unit of the risk prediction model, traverse the risk matrix for matching, risk degree boundary value determination, and trend prediction to determine first determination data; traverse the first determination data to identify boundary crossing features. If the boundary crossing features are empty, terminate the risk decision process. If the boundary crossing features are not empty, perform attack mode matching and risk determination.

[0095] Among them, the risk prediction module 13 is further configured to perform the following steps: if the boundary crossing features are not empty, determine the target attack mode by matching with the pattern-specific features; for the target attack mode, perform feature combination determination and trend prediction to determine second determination data; fuse the first determination data and the second determination data as the risk prediction result.

[0096] Through the foregoing detailed description of a method for predicting the risk of an apnea syndrome attack mode in this specification, those skilled in the art can clearly know the method for predicting the risk of an apnea syndrome attack mode in this embodiment. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For related parts, refer to the description in the method part.

[0097] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the risk of an apnea syndrome attack mode, characterized in that, The method includes: Invoking user status data across all age groups, mining a risk matrix based on state risk characteristics under different seizure patterns, where the state risk characteristics include pattern common characteristics and pattern specific characteristics, defining the boundary of feature risk degree with absolute contraction. Pattern common characteristics refer to risk indicators that commonly appear in different seizure patterns, and pattern specific characteristics are unique indicators under specific seizure patterns. The risk matrix is a multi-dimensional matrix structure, and each dimension represents the feature risk degree under different seizure patterns, used to describe the risk level of specific features. The risk degree boundary is the critical threshold of the risk feature; Based on the risk matrix, performing lightweight monitoring sampling and risk prediction training based on single detection features and combined features, and constructing a risk multi-dimensional prediction model. The risk multi-dimensional prediction model is built into the sleep monitoring device, and the single detection feature is a single eigenvalue obtained by sensing sampling; Connecting the sleep monitoring device, combining the risk multi-dimensional prediction model, determining the sleep monitoring data of the target user by staggered sampling of different sequence nodes, performing data backhaul and executing risk prediction and seizure pattern determination to obtain the risk prediction result; Based on the risk prediction result, issuing a status alert for the target user and generating a time series monitoring table for the monitoring time zone; Among them, mining the risk matrix of state risk characteristics under different seizure patterns includes: Obtaining the first hypopnea feature, where the first hypopnea feature belongs to the state risk characteristics and is used to measure the degree of decrease in the ventilation volume of the user within a specific time period; Traversing the seizure patterns, performing feature mining and clustering based on the user status data, and determining N groups of eigenvalue corresponding to the first hypopnea feature, where each seizure pattern corresponds to a group of eigenvalues; Based on the group of eigenvalues, performing absolute maximum contraction within the group to determine N risk degree boundary values of the first hypopnea feature. The first hypopnea feature is an extremely small feature negatively correlated with the risk degree, and the absolute maximum contraction within the group is a contraction towards the upper limit of the data interval; Adding the N risk degree boundary values to the risk matrix; Among them, adding the N risk degree boundary values to the risk matrix includes: Determining a preset tolerance scale based on the risk evolution trend of the first hypopnea feature, where the preset tolerance scale is positively correlated with the risk evolution speed. The specific manifestation of the risk evolution trend includes the eigenvalue gradually decreasing to the critical point or dropping sharply within a certain time period; Based on the preset tolerance scale, performing value expansion processing on the N risk degree boundary values to determine N risk precursor eigenvalue of the first hypopnea feature. The preset tolerance scale is the tolerance range set to allow normal fluctuations or avoid misjudgment; Establishing the mapping between the seizure pattern and the N risk precursor eigenvalue and adding it to the risk matrix.

2. The risk prediction method for the attack mode of sleep apnea syndrome according to claim 1, wherein For extremely large features, use absolute minimum contraction as the absolute contraction method; for extremely small features, use absolute maximum contraction as the contraction method.

3. The risk prediction method for the attack mode of sleep apnea syndrome according to claim 1, wherein Constructing a risk multi-dimensional prediction model includes: Determine the monitoring sampling method, conduct sensing sampling training, and construct the first sampling unit; Based on the risk degree boundary value determination of a single detection feature and the collaborative determination with the detection feature group, conduct sample-driven training to construct the second decision-making unit, where the attack mode is determined by the specific feature; Combine the first sampling unit and the second decision-making unit to generate the risk multi-dimensional prediction model.

4. The risk prediction method for the attack mode of sleep apnea syndrome according to claim 3, wherein, The determination of the monitoring sampling method includes: Based on feature correlation, combine the state risk features to determine M detection feature groups; Serialize the M detection feature groups to determine the risk detection sequence, where each detection feature group corresponds to a sequence node; Set a preset sampling interval, and based on the risk detection sequence, perform out-of-time sampling configuration of the state risk features to determine the monitoring sampling method.

5. The risk prediction method for the attack mode of sleep apnea syndrome according to claim 1, characterized in that The execution of risk prediction and attack mode determination includes: Transmit the sleep monitoring data back to the second decision-making unit of the risk multi-dimensional prediction model, traverse the risk number array for matching, risk degree boundary value determination, and trend prediction to determine the first determination data; Traverse the first determination data to identify boundary crossing features. If the boundary crossing features are empty, terminate the risk decision-making process. If the boundary crossing features are not empty, perform attack mode matching and risk determination.

6. The risk prediction method for the attack mode of sleep apnea syndrome according to claim 5, characterized in that, Performing attack mode matching and risk determination includes: If the boundary crossing features are not empty, determine the target attack mode by matching with the mode specific features; For the target attack mode, perform feature combination determination and trend prediction to determine the second determination data; Fuse the first determination data and the second determination data as the risk prediction result.

7. A risk prediction system for the attack mode of sleep apnea syndrome, characterized in that, The system is used to execute the method for predicting the risk of the attack mode of sleep apnea syndrome according to any one of claims 1-6. The system includes: A data mining module, which is used to call the user state data of all age stages, mine the risk number array based on the state risk features under different attack modes, where the state risk features include mode common features and mode specific features, define the feature risk degree boundary with absolute contraction, the mode common features refer to the risk indicators that commonly appear in different attack modes, the mode specific features are the unique indicators under a specific attack mode, the risk number array is a multi-dimensional matrix structure, each dimension of which represents the feature risk degree under different attack modes and is used to describe the risk level of a specific feature, and the risk degree boundary is the critical threshold of the risk feature; A training module, which is used to conduct lightweight monitoring sampling and risk prediction training based on the risk number array under single detection features and combined features to construct a risk multi-dimensional prediction model, where the risk multi-dimensional prediction model is built into the sleep monitoring device, and the single detection feature is a single feature value obtained by sensing sampling; A risk prediction module, which is used to connect to the sleep monitoring device, combine the risk multi-dimensional prediction model, determine the sleep monitoring data of the target user through out-of-time sampling of different sequence nodes, transmit the data back and execute risk prediction and attack mode determination to determine the risk prediction result; A risk management module, which is used to perform status alerts for the target user based on the risk prediction results and generate a time-series monitoring table for the monitoring time zone.

Images