Blood pressure dynamic monitoring system integrating overall risk and local anomaly detection

Abstract

The invention provides a blood pressure dynamic monitoring system integrating overall risk and local anomaly detection, and relates to the field of electric digital data processing. Comprising a multi-source blood pressure data fusion and acquisition module, a dual-path feature learning and abnormity pre-detection module, a local and overall interactive abnormity accurate identification module and a dynamic risk assessment and intelligent early warning decision module, and the multi-source blood pressure data fusion and acquisition module is used for acquiring blood pressure and context data; the dual-path feature learning and anomaly pre-detection module is used for extracting time sequence features and performing preliminary anomaly screening, and the local overall interactive anomaly accurate identification module is used for detecting and analyzing local anomaly. The dynamic risk assessment and intelligent early warning decision module is used for comprehensively assessing the risk, forming a feedback optimization mechanism and outputting a personalized early warning decision; the system can accurately identify abnormal blood pressure, realizes accurate assessment and prediction of risks, and provides effective support for clinical decision and personalized health management.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and specifically to a dynamic blood pressure monitoring system that integrates overall risk and local anomaly detection. Background Technology

[0002] Blood pressure, as an important physiological parameter reflecting the health status of the cardiovascular system, is crucial for the diagnosis, treatment, and management of hypertension. Traditional blood pressure measurement methods mainly include invasive intra-arterial measurement and non-invasive cuff measurement, but both methods have significant limitations. While invasive measurement is accurate, it requires inserting a catheter into an artery, posing a risk of infection and making it unsuitable for routine monitoring. Traditional cuff measurement, although safe, only provides intermittent measurement data and cannot capture dynamic changes in blood pressure, easily missing abnormal blood pressure events.

[0003] In recent years, with the rapid development of sensor technology, signal processing technology, and artificial intelligence algorithms, continuous non-invasive blood pressure monitoring technology has received widespread attention. These technologies are mainly based on principles such as photoplethysmography (PPG), pulse wave transit time, and cardiac impact signals, attempting to achieve real-time continuous blood pressure monitoring. Simultaneously, anomaly detection algorithms have been introduced into blood pressure monitoring systems to identify abnormal fluctuations in blood pressure and potential risks.

[0004] The commonly used blood pressure anomaly detection algorithms in existing technologies mainly include the following three types. The first type is threshold-based anomaly detection algorithms. This method judges anomalies by setting fixed upper and lower blood pressure thresholds, triggering an alarm when the measured value exceeds the preset range. This method is simple, direct, computationally inexpensive, and easy to implement, but it has significant shortcomings. It cannot adapt to individual differences and physiological fluctuations in blood pressure, easily generates a large number of false alarms, and has insufficient detection capability for slow trend changes. The second type is statistical anomaly detection algorithms. This method uses historical blood pressure data to build a statistical model, calculates the mean and standard deviation, and judges data points that deviate from the statistical distribution as anomalies. This method can adapt to individual characteristics to some extent, but its ability to model non-stationary time series is limited, it is difficult to handle periodic changes and long-term trends in blood pressure, and the model stability is poor when the amount of data is insufficient. The third type is machine learning-based anomaly detection algorithms, including support vector machines, random forests, neural networks, etc., which identify abnormal patterns by training classification models. These methods have strong pattern recognition capabilities, but require a large amount of labeled data for training. The models have poor interpretability and are difficult to provide a reliable basis for clinical decision-making. Furthermore, they usually treat anomaly detection as an independent task and fail to integrate it with risk assessment.

[0005] Regarding patented technology, foreign patent EP3102097 discloses a non-invasive continuous blood pressure monitoring system. This system monitors multiple physiological indicators, including blood pressure, by estimating pulse wave velocity, and uses a combination of photoplethysmography (PPG) sensors and electrocardiogram (ECG) sensors to acquire signals. The shortcomings of this patent are that the system primarily focuses on the accuracy of blood pressure measurement, with limited ability to detect anomalies in the blood pressure data. It lacks intelligent identification and correction mechanisms for measurement errors and motion artifacts, making it prone to inaccurate measurements in practical applications due to connection errors. More importantly, the system fails to establish a correlation analysis mechanism between local anomalies and overall risk, cannot distinguish between physiological and pathological abnormalities in blood pressure changes, lacks context-awareness based on contextual information, and struggles to achieve personalized anomaly detection and risk assessment.

[0006] Domestic patent CN108186000A discloses a real-time blood pressure monitoring system based on cardiac impact signals and photoelectric signals. This system calculates blood pressure values ​​by simultaneously acquiring cardiac impact signals and photoplethysmography (PPG) pulse wave signals, utilizing the phase difference between the two signals. The main shortcomings of this patent are: the system requires an accelerometer to be placed on the bed beam to collect cardiac impact signals, severely limiting its application scenarios and making it unsuitable for daily dynamic monitoring, resulting in poor portability. Regarding anomaly detection, the system only mentions signal preprocessing and filtering, lacking a systematic anomaly identification mechanism and adaptive correction capabilities for data quality issues such as motion artifacts and sensor loosening. Furthermore, the system does not consider individualized baseline modeling and dynamic update mechanisms, failing to adapt to long-term individual blood pressure evolution trends, and it does not integrate anomaly detection with risk assessment, thus failing to provide comprehensive health management support. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings by proposing a dynamic blood pressure monitoring system that integrates overall risk and local anomaly detection.

[0008] The present invention adopts the following technical solution:

[0009] A dynamic blood pressure monitoring system that integrates overall risk and local anomaly detection includes a multi-source blood pressure data fusion acquisition module, a dual-path feature learning and anomaly pre-detection module, a local and overall interactive anomaly accurate identification module, and a dynamic risk assessment and intelligent early warning decision module.

[0010] The multi-source blood pressure data fusion acquisition module is used to acquire blood pressure and context data; the dual-path feature learning and anomaly pre-detection module is used to extract temporal features and perform preliminary anomaly screening; the local and overall interactive anomaly accurate identification module is used to detect and analyze local anomalies; and the dynamic risk assessment and intelligent early warning decision module is used to comprehensively assess risks, form a feedback optimization mechanism, and output personalized early warning decisions.

[0011] The multi-source blood pressure data fusion acquisition module includes a multi-channel blood pressure signal synchronous acquisition unit, a physiological behavior context association unit, and a data quality adaptive correction unit. The multi-channel blood pressure signal synchronous acquisition unit is used to synchronously acquire blood pressure data. The physiological behavior context association unit is used to synchronously acquire physiological behavior information and align it with the blood pressure data in time. The data quality adaptive correction unit is used to identify data quality problems and perform intelligent correction and compensation based on context information.

[0012] The dual-path feature learning and anomaly pre-detection module includes a multi-scale temporal feature automatic extraction unit, a personalized dynamic baseline continuous learning unit, and an anomaly candidate event initial screening unit. The multi-scale temporal feature automatic extraction unit is used to discover and extract key temporal features of blood pressure data. The personalized dynamic baseline continuous learning unit is used to build and continuously update the personalized baseline model. The anomaly candidate event initial screening unit is used to perform coarse-grained anomaly detection during the feature extraction stage to quickly screen out suspicious anomaly candidate sets.

[0013] The local-to-global interactive anomaly accurate identification module includes a context-aware anomaly detection unit, an anomaly risk causal correlation analysis unit, and a multi-level anomaly classification and tracing unit. The context-aware anomaly detection unit is used to dynamically adjust the threshold and strategy of anomaly detection. The anomaly risk causal correlation analysis unit is used to analyze the causal relationship between detected local anomalies and overall risks. The multi-level anomaly classification and tracing unit is used to classify the detected anomalies by type and trace their triggering factors.

[0014] The dynamic risk assessment and intelligent early warning decision-making module includes a multi-dimensional risk dynamic modeling and prediction unit, a risk anomaly bidirectional feedback optimization unit, and a personalized graded early warning and intervention suggestion unit. The multi-dimensional risk dynamic modeling and prediction unit is used to construct a dynamic risk model and predict future risk evolution trends. The risk anomaly bidirectional feedback optimization unit is used to optimize anomaly detection parameters in reverse using risk assessment results, and update the risk assessment model with anomaly detection results. The personalized graded early warning and intervention suggestion unit is used to generate differentiated risk levels, early warning strategies, and actionable intervention suggestions.

[0015] Furthermore, the individualized dynamic baseline continuous learning unit includes a baseline model initialization processor, an adaptive weight decay update processor, and a seasonal periodic modeling processor. The baseline model initialization processor is used to construct an individualized blood pressure baseline model based on the user's initial historical blood pressure data. The adaptive weight decay update processor is used to update the baseline model online as new data is continuously collected. The seasonal periodic modeling processor is used to identify periodic components in blood pressure data, extract periodic parameters and amplitude parameters of regular fluctuation patterns such as diurnal rhythm, weekly pattern, and seasonal effect, and incorporate these regular changes into the baseline model to avoid misjudging normal periodic fluctuations as abnormal.

[0016] The adaptive weight decay update processor calculates the individualized baseline value according to the following formula:

[0017] ;

[0018] Where B(t) represents the individualized baseline value at time t, BP i This represents the i-th historical blood pressure measurement, where n is the total number of historical data points, and w i (t) represents the adaptive time weight of the i-th historical data at time t, M i (t) represents the memory retention factor of the i-th historical data;

[0019] The adaptive time weights are calculated according to the following formula:

[0020] ;

[0021] Among them, c i t represents the confidence score of the i-th data point. i This indicates the time of data collection for the i-th data point. S is the instantaneous decay function. i Let i be the local stability index for the i-th data. For stability gain parameters, This represents the central blood pressure value of the cluster to which the i-th data belongs.

[0022] Furthermore, the context-aware anomaly detection unit includes a multi-dimensional feature weighted deviation calculation processor, a context feature fusion processor, and a detection strategy switching processor. The multi-dimensional feature weighted deviation calculation processor is used to calculate the degree of deviation of the current blood pressure measurement value from the individualized baseline in multiple feature dimensions. The context feature fusion processor is used to jointly analyze the current physiological behavior context information with the blood pressure data to determine whether the blood pressure change is consistent with the current context. The detection strategy switching processor is used to select an appropriate anomaly detection algorithm from the strategy library according to the identified current context type.

[0023] The multi-dimensional feature weighted deviation calculation processor calculates the comprehensive deviation D according to the following formula. total (t):

[0024] ;

[0025] Where m is the total number of feature dimensions. This represents the i-th eigenvalue at time t. This represents the individualized baseline value of the i-th feature at time t. This represents the time-varying standard deviation of the i-th feature. This represents the main effect weight of the i-th feature. The nonlinear exponent represents the i-th feature. This represents the interaction effect weight between the i-th feature and the j-th feature. Let represent the interaction deviation term between the i-th feature and the j-th feature at time t, and q be the norm parameter. This is a context-sensitive dynamic correction factor.

[0026] Furthermore, the multidimensional risk dynamic modeling and prediction unit includes a multi-factor risk integration processor, a local-to-overall risk time-series fusion modeling processor, and a risk prediction and deduction processor. The multi-factor risk integration processor is used to construct a comprehensive risk assessment vector at the current moment by integrating multidimensional information. The local-to-overall risk time-series fusion modeling processor is used to construct a dynamic risk model that integrates local abnormal event sequences and the long-term trend of overall blood pressure and calculate the comprehensive risk evolution trajectory. The risk prediction and deduction processor is used to predict the risk development trajectory within a preset time period in the future by extrapolation or time series prediction model based on the current risk status and historical risk evolution trajectory.

[0027] Furthermore, the local-to-global risk time-series fusion modeling processor calculates the comprehensive risk score according to the following formula. :

[0028] ;

[0029] in, Accumulate risks for local anomaly time series. This represents a long-term risk to the overall trend. For time-varying local-to-global trade-off coefficients, To integrate nonlinear parameters, This is the synchronization amplification factor. This is an index of synchronous deterioration between the local and the overall system. This is a risk gating factor.

[0030] The beneficial effects achieved by this invention are:

[0031] This system employs a dual mechanism of adaptive time weighting and memory retention factors, simultaneously considering data timeliness, confidence level, local stability, and pattern reproducibility. This allows the baseline model to quickly respond to recent changes in individual blood pressure while preserving long-term stable trends, effectively avoiding baseline drift or oversensitivity issues caused by traditional fixed-weight methods. In anomaly detection, the system not only considers the independent deviation of each feature dimension but also introduces interaction effects between features and context-based dynamic correction factors. This enables accurate identification of complex abnormal patterns caused by multi-feature synergistic changes and dynamically adjusts the anomaly judgment criteria based on the current physiological context, significantly reducing the false alarm rate caused by physiological blood pressure fluctuations and improving the detection sensitivity of pathological abnormalities. In terms of risk assessment, by establishing a fusion model of the cumulative risk of local anomalies and the long-term risk of overall trends, the system organically integrates short-term abnormal events with long-term trend changes. The fast-slow dual-memory mechanism simultaneously captures the immediate impact and chronic cumulative effects of acute abnormal events; the sudden enhancement factor identifies and amplifies the risk amplification effect of abnormal events occurring in a short period; the synchronous deterioration index quantifies the degree of synergistic change between local anomalies and overall trends, enabling timely detection of warning signals of accelerated risk deterioration; and the time-varying trade-off coefficient adaptively adjusts the weights of local and overall information entropy based on the dynamic balance, ensuring that risk assessment neither overemphasizes short-term anomalies at the expense of long-term trends nor overlooks urgent abnormal events due to an overemphasis on long-term trends. This multi-level, multi-scale mathematical modeling method enables the system to accurately quantify the cumulative effects and risk evolution trends of blood pressure abnormalities, providing clinicians with reliable decision-making basis and patients with precise personalized early warnings and intervention suggestions, significantly improving the clinical application value and health management effectiveness of dynamic blood pressure monitoring.

[0032] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall structural framework of the present invention;

[0034] Figure 2 This is a schematic diagram of the multi-source blood pressure data fusion acquisition module of the present invention;

[0035] Figure 3 This is a schematic diagram of the dual-path feature learning and anomaly pre-detection module of the present invention;

[0036] Figure 4 This is a schematic diagram of the local and overall interactive anomaly accurate identification module of the present invention;

[0037] Figure 5 This is a schematic diagram of the dynamic risk assessment and intelligent early warning decision-making module of the present invention;

[0038] Figure 6 This is a schematic diagram comparing the false alarm rate and false negative rate of the present invention with those of traditional methods;

[0039] Figure 7 This is a schematic diagram comparing the individualized baseline modeling effects of the present invention with those of traditional methods;

[0040] Figure 8 This is a schematic diagram of the interactive interface (UI) of the present invention. Detailed Implementation

[0041] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0042] Example 1: A dynamic blood pressure monitoring system integrating overall risk and local anomaly detection, combined with... Figure 1 It includes a multi-source blood pressure data fusion acquisition module, a dual-path feature learning and abnormal pre-detection module, a local and overall interactive abnormal accurate identification module, and a dynamic risk assessment and intelligent early warning decision-making module;

[0043] The multi-source blood pressure data fusion acquisition module is used to acquire high-quality, multi-dimensional blood pressure and context data; the dual-path feature learning and anomaly pre-detection module is used to extract temporal features and perform preliminary anomaly screening; the local-to-global interactive anomaly accurate identification module is used to accurately detect and analyze local anomalies in the context of overall risk; and the dynamic risk assessment and intelligent early warning decision module is used to comprehensively assess risks, form a feedback optimization mechanism, and output personalized early warning decisions.

[0044] Combination Figure 2The multi-source blood pressure data fusion acquisition module includes a multi-channel blood pressure signal synchronous acquisition unit, a physiological behavior context association unit, and a data quality adaptive correction unit. The multi-channel blood pressure signal synchronous acquisition unit is used to synchronously acquire blood pressure data through various measurement methods such as cuff, photoelectric, and pulse wave, to achieve cross-validation and fusion of multi-source data and improve measurement reliability. The physiological behavior context association unit is used to synchronously acquire physiological behavior information such as heart rate, body temperature, activity status, and sleep stage, and align it with blood pressure data in time to provide multi-dimensional data support for subsequent context-aware analysis. The data quality adaptive correction unit is used to identify data quality problems caused by movement, sensor loosening, or signal drift, and to perform intelligent correction and compensation based on context information, rather than simply removing abnormal data.

[0045] Combination Figure 3 The dual-path feature learning and anomaly pre-detection module includes a multi-scale time-series feature automatic extraction unit, a personalized dynamic baseline continuous learning unit, and an anomaly candidate event initial screening unit. The multi-scale time-series feature automatic extraction unit is used to automatically discover and extract key time-series features of blood pressure data at different time scales through adaptive algorithms, rather than relying on preset statistical indicators, thereby enhancing the flexibility and representativeness of feature expression. The personalized dynamic baseline continuous learning unit is used to build and continuously update a personalized baseline model based on historical blood pressure data, enabling the model to adapt to long-term dynamic characteristics such as seasonal changes and age-related evolution of individual blood pressure. The anomaly candidate event initial screening unit is used to perform coarse-grained anomaly detection during the feature extraction stage, quickly screening out suspicious anomaly candidate sets, reducing the computational burden and providing a target set for subsequent refined analysis.

[0046] Combination Figure 4 The local-to-global interactive anomaly accurate identification module includes a context-aware anomaly detection unit, an anomaly risk causal association analysis unit, and a multi-level anomaly classification and tracing unit. The context-aware anomaly detection unit is used to dynamically adjust the threshold and strategy of anomaly detection based on the current overall risk level, physiological context state, and individual characteristics, so as to realize real-time interactive fusion of local anomaly detection and overall risk state. The anomaly risk causal association analysis unit is used to analyze the causal relationship between the detected local anomalies and the overall risk, determine whether the anomaly is the cause of the risk or the manifestation of the risk state, and establish a two-way association mechanism. The multi-level anomaly classification and tracing unit is used to classify the detected anomalies by type and trace their triggering factors, identify whether they are caused by drug reaction, dietary influence, emotional fluctuations, or potential pathological changes, and provide richer information dimensions for clinical decision-making.

[0047] Combination Figure 5The dynamic risk assessment and intelligent early warning decision-making module includes a multi-dimensional risk dynamic modeling and prediction unit, a risk anomaly bidirectional feedback optimization unit, and a personalized graded early warning and intervention suggestion unit. The multi-dimensional risk dynamic modeling and prediction unit is used to integrate short-term abnormal history, long-term blood pressure trends, and individual health characteristics to construct a dynamic risk model and predict future risk evolution trends, rather than just assessing the current risk status. The risk anomaly bidirectional feedback optimization unit is used to use the risk assessment results to optimize the anomaly detection parameters in reverse, and at the same time use the anomaly detection results to update the risk assessment model, forming a closed-loop feedback mechanism to continuously improve the overall performance of the system. The personalized graded early warning and intervention suggestion unit is used to generate differentiated risk levels, early warning strategies, and actionable intervention suggestions based on the user's age, underlying diseases, compliance, and other individual characteristics, and outputs accurate early warning information and personalized health management plans when a preset threshold is reached.

[0048] The multi-channel blood pressure signal synchronous acquisition unit includes a signal source interface processor, a timestamp synchronization processor, and a multi-source data fusion processor. The signal source interface processor is used to establish communication connections with different types of blood pressure measurement devices, receive raw blood pressure signals from various devices such as cuff blood pressure monitors, photoplethysmography (PPG) sensors, and pulse wave velocity meters, and perform standardized signal format conversion. The timestamp synchronization processor is used to add a unified high-precision timestamp to blood pressure data from different measurement devices to ensure accurate alignment of multi-source data in the time dimension and eliminate timing deviations caused by differences in device clocks. The multi-source data fusion processor is used to perform cross-validation and weighted fusion of blood pressure data from different measurement channels at the same time. By comparing the consistency of the measurement values ​​of each channel and allocating fusion weights according to the historical accuracy of each device, it outputs a fused blood pressure measurement value that has undergone multi-source verification.

[0049] The physiological behavior context association unit includes a multimodal sensor interface processor, a spatiotemporal association mapping processor, and a context feature encoding processor. The multimodal sensor interface processor is used to collect data from various physiological and behavioral sensors such as heart rate, body temperature, acceleration, gyroscope, and GPS positioning, and establish a collaborative acquisition mechanism with the blood pressure monitoring device. The spatiotemporal association mapping processor is used to accurately match the collected physiological behavior data with blood pressure data in the time and space dimensions, and construct a complete physiological behavior state snapshot at the time of blood pressure measurement through timestamp alignment and sliding window association. The context feature encoding processor is used to convert the raw physiological behavior data into a structured context feature vector, encode acceleration data into activity intensity levels through threshold determination, classify sleep stages according to heart rate variability and body movement signals, and provide standardized input for subsequent context-aware analysis.

[0050] The adaptive data quality correction unit includes a quality anomaly identification processor, a context-driven correction processor, and a data confidence assessment processor. The quality anomaly identification processor is used to detect measurement anomalies in blood pressure data, identify sensor loosening by calculating the signal amplitude mutation rate, detect motion artifacts through spectrum analysis, monitor baseline drift through sliding window variance, and mark abnormal data segments and their types. The context-driven correction processor is used to intelligently compensate for the identified quality anomalies based on context information. For example, when vigorous exercise is detected, adaptive filtering is applied to the photoelectric signal to eliminate motion artifacts, and interpolation repair is performed based on the linear trend of the preceding and following normal data segments when the sensor is temporarily loose. The data confidence assessment processor is used to calculate a quality confidence score for each blood pressure measurement value, and provides data confidence weights in the range of 0 to 1 for subsequent analysis by weighted accumulation of measurement stability score, multi-source consistency score, and context reasonableness score.

[0051] The multi-scale time-series feature automatic extraction unit includes a time window adaptive segmentation processor, a feature importance evaluation processor, and a multi-scale feature fusion processor. The time window adaptive segmentation processor is used to automatically determine the analysis windows of multiple time scales such as short-term, medium-term, and long-term through sliding window variance analysis based on the fluctuation characteristics of blood pressure data and individual behavioral patterns, avoiding insufficient feature expression caused by using a fixed time window. The feature importance evaluation processor is used to extract candidate features such as the mean, standard deviation, coefficient of variation, and peak frequency of blood pressure at each time scale, and evaluate the contribution of each feature to anomaly detection and risk assessment through correlation analysis and information gain calculation. The multi-scale feature fusion processor is used to concatenate the features extracted from different time scales according to the time dimension, and assign fusion weights to the features of each dimension according to the feature importance score, thereby constructing a hierarchical time-series feature vector.

[0052] The individualized dynamic baseline continuous learning unit includes a baseline model initialization processor, an adaptive weight decay update processor, and a seasonal periodic modeling processor. The baseline model initialization processor is used to construct an individualized blood pressure baseline model based on the user's initial historical blood pressure data by calculating the statistical distribution of blood pressure at different time periods. This model includes multi-dimensional baseline parameters such as the mean baseline at rest, the mean baseline during daily activities, and the amplitude of diurnal rhythm fluctuations. The adaptive weight decay update processor is used to update the baseline model online using a time decay weighting mechanism as new data is continuously collected. This allows the model to adapt to recent changes in individual blood pressure while preserving long-term stable trends. The baseline is smoothly evolved by dynamically adjusting the weights of historical data. The seasonal periodic modeling processor is used to identify periodic components in blood pressure data through Fourier analysis or wavelet decomposition. It extracts periodic parameters and amplitude parameters of regular fluctuation patterns such as diurnal rhythm, weekly patterns, and seasonal effects, and incorporates these regular changes into the baseline model to avoid misjudging normal periodic fluctuations as abnormalities.

[0053] The adaptive weight decay update processor calculates the individualized baseline value according to the following formula:

[0054] ;

[0055] Where B(t) represents the individualized baseline value at time t, BP i This represents the i-th historical blood pressure measurement, where n is the total number of historical data points, and w i (t) represents the adaptive time weight of the i-th historical data at time t, M i (t) represents the memory retention factor of the i-th historical data;

[0056] The adaptive time weights are calculated according to the following formula:

[0057] ;

[0058] Among them, c i t represents the confidence score of the i-th data point. i This indicates the time of data collection for the i-th data point. S is the instantaneous decay function. i Let i be the local stability index for the i-th data. For stability gain parameters, This represents the central blood pressure value of the cluster to which the i-th data belongs;

[0059] The memory retention factor is calculated according to the following formula:

[0060] ;

[0061] in, For memory enhancement coefficient, This represents the distance metric between the blood pressure pattern at the i-th data point and the pattern at the current time t. For distance scaling parameters, This represents the number of times a pattern similar to the i-th data point has been repeated in history. The baseline number of recurrences;

[0062] The pattern distance is obtained by processing the feature components using the common Euclidean distance calculation method;

[0063] The initial screening unit for abnormal candidate events includes a fast deviation detection processor, an abnormal candidate clustering processor, and a candidate priority ranking processor. The fast deviation detection processor is used to quickly identify significant deviation points exceeding a preset threshold by calculating the standardized deviation distance between the blood pressure measurement value and the individualized baseline. These deviations include preliminary abnormal patterns such as single-point mutations, short-term trend anomalies, and abnormal fluctuation amplitudes, forming a set of abnormal candidate events. The abnormal candidate clustering processor is used to cluster the detected abnormal candidate events according to their temporal proximity. By setting a time window threshold, adjacent candidate points are merged into persistent abnormal events, identifying isolated abnormal points and persistent abnormal events to reduce redundancy. The candidate priority ranking processor is used to calculate the priority score of each candidate event. By weighted combination of features such as abnormal amplitude, duration, and historical recurrence frequency, the candidate events are sorted in descending order to determine the key candidate set that needs to be refined for analysis.

[0064] The context-aware anomaly detection unit includes a multi-dimensional feature weighted deviation calculation processor, a context feature fusion processor, and a detection strategy switching processor. The multi-dimensional feature weighted deviation calculation processor is used to calculate the degree of deviation of the current blood pressure measurement value from the individualized baseline in multiple feature dimensions. By weighting and fusing different feature deviations, a comprehensive deviation score is generated as the core quantitative indicator for anomaly determination and risk assessment. The context feature fusion processor is used to jointly analyze the current physiological behavioral context information with blood pressure data. By comparing the matching degree between the current activity intensity, emotional state, and blood pressure changes, it determines whether the blood pressure change is consistent with the current context. For example, it identifies that the increase in blood pressure after exercise is a physiological response rather than a pathological abnormality. The detection strategy switching processor is used to select an appropriate anomaly detection algorithm from the strategy library according to the identified current context type. For example, it switches to a detection strategy that focuses more on slow trend changes during sleep and switches to a detection strategy that focuses more on rapid fluctuations during strenuous activity.

[0065] The multi-dimensional feature weighted deviation calculation processor calculates the comprehensive deviation D according to the following formula. total (t):

[0066] ;

[0067] Where m is the total number of feature dimensions. This represents the i-th eigenvalue at time t. This represents the individualized baseline value of the i-th feature at time t. This represents the time-varying standard deviation of the i-th feature. This represents the main effect weight of the i-th feature. The nonlinear exponent represents the i-th feature. This represents the interaction effect weight between the i-th feature and the j-th feature. Let represent the interaction deviation term between the i-th feature and the j-th feature at time t, and q be the norm parameter. Context-sensitive dynamic correction factor;

[0068] When the overall deviation is greater than the threshold, it is judged as a local abnormal event. Based on the judgment result, a corresponding abnormal verification flag is generated, with 1 indicating abnormal and 0 indicating normal.

[0069] The feature interaction deviation term is calculated according to the following formula:

[0070] ;

[0071] in, This represents the interaction strength coefficient between the i-th feature and the j-th feature;

[0072] The context dynamic correction factor is calculated according to the following formula:

[0073] ;

[0074] Where C represents the total number of context types. This represents the maximum attenuation coefficient for the i-th context. Let be the intensity value of the i-th context at time t. This represents the resting baseline value for the i-th context. Let be the sensitivity parameter for the i-th context;

[0075] The abnormal risk causal association analysis unit includes a time-series causal inference processor, an abnormal risk contribution quantification processor, and a reverse verification processor. The time-series causal inference processor is used to analyze the temporal sequence relationship between local abnormal events and changes in the overall risk state. By comparing the time difference between the time of the abnormality occurrence and the time of risk change, a sliding window correlation analysis is used to determine whether the abnormality is the cause of the risk increase or the result of the risk state deterioration. The abnormal risk contribution quantification processor is used to evaluate the contribution of each detected local abnormality to the current and future overall risk. By calculating the change in risk score before and after the abnormality occurs and considering the impact of the duration of the abnormality, high-impact abnormalities and low-impact abnormalities are distinguished. The reverse verification processor is used to trace back the possible abnormal triggering factors from the current risk state. By searching for cases similar to the current risk pattern in historical abnormality records, the rationality of the abnormality-risk association hypothesis is tested and the association confidence is calculated.

[0076] The multi-level anomaly classification and tracing unit includes an anomaly type discrimination processor, a triggering factor tracing processor, and an anomaly knowledge base matching processor. The anomaly type discrimination processor is used to classify detected abnormal events according to predefined pathological type classification rules. By comparing the amplitude features, duration features, and fluctuation pattern features of the anomaly with typical feature templates of each type, it distinguishes different categories such as hypertensive emergencies, hypotensive events, excessive blood pressure fluctuations, and measurement errors. The triggering factor tracing processor is used to extract contextual information, medication records, diet logs, emotion markers, and other data within a preset time window before the anomaly occurs. It infers possible triggering factors of the anomaly through association rule mining and frequent itemset analysis. The anomaly knowledge base matching processor is used to calculate the similarity between the feature vector of the current abnormal event and the feature vector of historical abnormal cases stored in the knowledge base. It identifies the most similar historical cases through Euclidean distance or cosine similarity measurement and extracts verified anomaly-cause association rules to assist in the current tracing analysis.

[0077] The multidimensional risk dynamic modeling and prediction unit includes a multi-factor risk integration processor, a local-to-overall risk time-series fusion modeling processor, and a risk prediction and extrapolation processor. The multi-factor risk integration processor integrates multi-dimensional information such as short-term abnormal event history, long-term blood pressure trend indicators, individual basic risk factors, and medication adherence. By normalizing and weighting the information in each dimension, it constructs a comprehensive risk assessment vector for the current moment. The local-to-overall risk time-series fusion modeling processor constructs a dynamic risk model that integrates local abnormal event sequences and overall long-term blood pressure trends. It calculates the comprehensive risk evolution trajectory through a time-series weighted fusion mechanism, reflecting both the cumulative effect of short-term abnormal events and the continuous impact of long-term trends. The risk prediction and extrapolation processor predicts the risk development trajectory within a preset time period based on the current risk status and historical risk evolution trajectory, using extrapolation or time series prediction models, and estimates the probability of serious consequences such as hypertensive emergencies and cardiovascular events.

[0078] The local-to-global risk time-series fusion modeling processor calculates the comprehensive risk score according to the following formula. :

[0079] ;

[0080] in, Accumulate risks for local anomaly time series. This represents a long-term risk to the overall trend. For time-varying local-to-global trade-off coefficients, To integrate nonlinear parameters, This is the synchronization amplification factor. This is an index of synchronous deterioration between the local and the overall system. As a risk gating factor;

[0081] The risk of local anomaly time series accumulation is calculated according to the following formula:

[0082] ;

[0083] Where, N w v represents the length of the backtracking time window. i Let X be the anomaly verification flag at time ti, and let X be the weight allocation for fast and slow memory. To reduce the rate of forgetting, For slow forgetting rate, H burst (t, i) represents the burst enhancement factor;

[0084] The burst enhancement factor is calculated according to the following formula:

[0085] ;

[0086] in, For sudden intensity coefficient, Indicates the time before and after time ti The number of abnormal events within the time window, N exp Indicates the expected normal abnormal frequency. Represents the event constant for the decay of the sudden effect;

[0087] In this article, t refers to the time point, and the duration between two consecutive time points is fixed.

[0088] The synchronous deterioration index is calculated according to the following formula:

[0089] ;

[0090] Where Cov represents the covariance function, Var represents the variance function, and T sync To synchronize the analysis time window;

[0091] The time-varying tradeoff coefficient is calculated according to the following formula:

[0092] ;

[0093] in, To adjust the steepness parameter, H local (t) represents the local anomaly information entropy, H global (t) represents the overall trend information entropy, H scale For entropy scaling parameters, These are bias parameters;

[0094] The risk anomaly bidirectional feedback optimization unit includes an anomaly detection parameter optimization processor, a risk model calibration processor, and a closed-loop performance evaluation processor. The anomaly detection parameter optimization processor is used to adjust the threshold parameters of the anomaly detection module in reverse according to the risk assessment results. By calculating the difference between the current risk level and the preset risk level, it lowers or raises the anomaly detection threshold according to a predefined adjustment strategy to optimize the detection sensitivity. The risk model calibration processor is used to calibrate the risk assessment model using the anomaly detection results. When the detected anomaly frequency deviates significantly from the risk prediction value, it adjusts the feature weight parameters in the risk model to make the model output more consistent with the actual observation. The closed-loop performance evaluation processor is used to continuously monitor the effect of the bidirectional feedback mechanism by statistically analyzing key performance indicators such as anomaly detection accuracy, risk prediction accuracy, and false alarm rate, and adjusts the feedback intensity coefficient and model update frequency according to the changing trends of the performance indicators.

[0095] The personalized tiered early warning and intervention suggestion unit includes a user profile building processor, a tiered early warning strategy generation processor, and an intervention plan recommendation processor. The user profile building processor integrates multi-dimensional features such as the user's age, gender, underlying diseases, medication history, health literacy, and adherence records. It constructs a personalized user profile model through feature encoding and vectorization processing to provide a decision-making basis for differentiated early warning and intervention. The tiered early warning strategy generation processor matches suitable early warning plans from the early warning strategy library based on the current risk level, user profile features, and contextual information. This includes determining the early warning trigger threshold, selecting the early warning urgency level, and setting the early warning information presentation method. The intervention plan recommendation processor retrieves matching intervention measures from the intervention knowledge base based on the anomaly type, risk level, and user characteristics. It sorts the intervention plans by calculating their feasibility score and expected effect score and generates a recommendation list of tiered intervention measures, including immediate medical advice, medication adjustment reminders, and lifestyle improvement plans.

[0096] The i, j, and k mentioned above are ordinal numbers used to represent sequence numbers and have no actual meaning.

[0097] To verify the advantages of the proposed blood pressure dynamic monitoring system, which integrates overall risk and local anomaly detection, compared with traditional methods in terms of anomaly detection accuracy, risk assessment precision, and early warning effect, the following experiment was conducted: 500 hypertensive patients were selected for continuous monitoring, with an average of 8,000 blood pressure data points collected from each patient, including various physiological and behavioral contextual information. The patients' ages ranged from 45 to 75 years, covering different underlying diseases and medication situations.

[0098] Traditional methods: employ fixed threshold anomaly detection and simple statistical analysis, lacking context awareness and individualized baselines.

[0099] Data acquisition phase: Blood pressure data is simultaneously collected using multiple devices such as cuff-type, photoelectric, and pulse wave sensors, while also recording contextual information such as heart rate, body temperature, activity level, and sleep stage. The method of this invention employs adaptive data quality correction to identify and correct quality issues such as motion artifacts and sensor loosening.

[0100] Baseline modeling phase: Traditional methods set fixed thresholds based on population statistics; the method of this invention constructs an individualized dynamic baseline model for each patient, including parameters such as resting baseline, circadian rhythm pattern, and seasonal variations.

[0101] Feature learning stage: The method of this invention uses multi-scale time series feature automatic extraction to discover key patterns at different time scales such as short-term (hourly), medium-term (daily), and long-term (weekly); traditional methods only extract basic statistical features such as mean and standard deviation.

[0102] Anomaly Detection Comparison: Two methods were run simultaneously, recording detected abnormal events. The method of this invention dynamically adjusts the detection strategy based on the overall risk level and the current context, distinguishing between physiological fluctuations and pathological abnormalities; traditional methods use fixed thresholds for judgment.

[0103] Risk assessment validation: The accuracy of risk assessment of the two methods was compared. The method of this invention integrates short-term abnormal history, long-term blood pressure trends, and individual characteristics to predict future risk evolution; traditional methods assess risk based solely on current blood pressure values.

[0104] Early warning effectiveness evaluation: The accuracy, lead time, and user compliance of the two methods were statistically analyzed. The method of this invention generates differentiated early warning strategies and actionable intervention suggestions based on individual user characteristics; traditional methods use a uniform early warning threshold.

[0105] The data was organized and obtained Figure 6 and Figure 7 .

[0106] Example 2: A specific implementation of a dynamic blood pressure monitoring system that integrates overall risk and local anomaly detection. This system is deployed in a wearable health monitoring device and works in conjunction with a mobile terminal to provide 24 / 7 continuous monitoring and intelligent early warning for patients with chronic hypertension. This example uses a 65-year-old male hypertensive patient as a case study to illustrate the implementation details and technical effects of the system.

[0107] The multi-source blood pressure data fusion acquisition module adopts a three-channel parallel acquisition architecture. The first channel uses a digital blood pressure monitor with oscillation measurement method to perform four timed measurements at fixed times each day as a reference, with a measurement accuracy of ±2 mmHg and a repeatability error of less than 3 mmHg. The second channel uses a photoplethysmography (PPG) sensor to continuously acquire waveforms at a sampling rate of 250 Hz, and infers the blood pressure value by analyzing the pulse wave transit time. The sampling window is 8 seconds of continuous sampling, automatically triggered every 10 minutes or immediately triggered when a change in activity is detected. The third channel uses a piezoelectric pulse wave sensor to monitor the radial artery pulse wave at a sampling rate of 100 Hz, estimating blood pressure through waveform characteristic parameters. This sensor consumes only 12 mW, making it suitable for long-term wear, and its immunity to motion interference is about 40% better than that of photoelectric sensors. The three channels of data are transmitted via a blue... The 5.2 low-power protocol synchronously transmits data to the main control chip, with transmission delay controlled within 15ms to ensure timing synchronization. The main control chip adopts a dual-core ARM architecture, with the main core running at a frequency of 528MHz for data processing and algorithm calculation, and the auxiliary core running at a frequency of 200MHz for sensor management and communication protocol. As a technical variation replacement, the photoplethysmography (PPG) sensor can be replaced with an impedance blood pressure sensor, which calculates blood pressure by measuring changes in bioimpedance in the upper arm or wrist. The advantage of this solution is that it is less affected by skin color and ambient light, but it requires more frequent calibration and increases power consumption by about 25%. The piezoelectric sensor can also be replaced with a capacitive or strain gauge sensor. The capacitive sensor has higher sensitivity but increases the cost by about 60%, while the strain gauge sensor has a lower cost but slightly lower long-term stability and requires recalibration every 90 days.

[0108] The signal source interface processor in the multi-channel blood pressure signal synchronous acquisition unit integrates hardware interfaces for three communication protocols, including a serial UART interface for connecting to a digital blood pressure monitor with a baud rate set to 115200bps; analog differential input interfaces ADC1 and ADC2 for connecting to the photoelectric sensor and piezoelectric sensor respectively, both using 16-bit resolution and differential input mode to improve the signal-to-noise ratio to 82dB; all analog signals are first filtered by a second-order Butterworth low-pass filter for anti-aliasing before acquisition, with a cutoff frequency set to 50Hz to effectively suppress power frequency interference and high-frequency noise; the data format transmitted by the digital blood pressure monitor is ASCII, requiring string parsing and numerical conversion; and the raw voltage signals from the photoelectric sensor and piezoelectric sensor need to be amplified, zero-point calibrated, and linearized before being converted into physical blood pressure values. The signal source interface processor integrates... A hardware clock synchronization module was introduced, which uses a real-time clock chip with a crystal oscillator frequency of 32.768kHz to provide a unified time base. The timestamp accuracy of all sampling moments reaches 1ms. To address the potential clock drift issue of the three different measurement devices, the system automatically performs clock synchronization calibration every 6 hours, aligning with the network clock of the mobile terminal. The actual test results show that after calibration, the cumulative clock deviation over 24 hours is less than 50ms, meeting the requirements for multi-source data timing alignment. As a technical variation, timestamp synchronization can also use GPS or BeiDou timing schemes. GPS timing accuracy can reach 50ns, but power consumption increases and outdoor environment is required. A more economical solution is to use a software synchronization algorithm, which performs cross-correlation calculations by analyzing the characteristic points of the three signals to achieve post-event timing alignment. This solution does not require additional hardware, but the computational complexity increases and the requirements for signal quality are higher.

[0109] The multi-source data fusion processor employs an adaptive weighted fusion algorithm. First, it performs a consistency check on three blood pressure measurements taken at the same time, calculating the standard deviation of each measurement. If the standard deviation is less than 5 mmHg, the three data points are considered to have good consistency. In this case, a weighted average method is used to calculate the fusion value. The weight allocation is dynamically adjusted based on the historical accuracy of each device. The initial weights are set as follows: digital blood pressure monitor 0.5, photoelectric sensor 0.3, and piezoelectric sensor 0.2. After seven consecutive days of adaptive learning, the weights for this patient are optimized to 0.48, 0.35, and 0.17, indicating that the photoelectric sensor's measurement accuracy for this patient is better than initially expected. When the standard deviation is greater than 5 mmHg but less than 10 mmHg, the system initiates an outlier removal mechanism, using the Grubbs test to identify outliers. Significantly deviating measurements are marked as low confidence and their fusion weight is reduced to 30% of the original weight. When the standard deviation exceeds 10 mmHg, the data is considered... In cases of severe measurement error or when the patient is engaged in strenuous exercise, fusion is paused, and only the timed measurement value from the digital blood pressure monitor is retained as a reference, triggering a data quality alarm. Actual data shows that within a 30-day monitoring period, 76.3% of the three data streams showed good consistency, 18.2% required outlier removal, and 5.5% exhibited severe deviations. Comparison of the fused blood pressure measurements with those from a hospital ambulatory blood pressure monitor reduced the mean absolute error from 6.8 mmHg for a single sensor to 3.2 mmHg, a reduction of 52.9%. Technical variations include using Kalman filtering for fusion, which can fuse measured and predicted values ​​but requires establishing a state-space model of blood pressure and has high computational complexity; another option is Bayesian fusion, which calculates the posterior probability distribution using prior probability distribution and likelihood function. This method is theoretically superior but requires a large amount of historical data to train the prior distribution and has poor real-time performance.

[0110] The physiological behavior context association unit employs a multimodal sensor array, including a photoelectric heart rate sensor continuously monitoring heart rate changes at a 25Hz sampling rate, a temperature sensor sampling once per minute to monitor body surface temperature, a three-axis accelerometer and a three-axis gyroscope sampling at a 50Hz sampling rate to monitor limb movement and posture changes, a barometric pressure sensor sampling 10 times per second to monitor altitude changes for identifying vertical movements such as climbing stairs, and a GPS module updating location information every 5 seconds in outdoor environments to identify activity patterns such as commuting and walking. All sensor data and blood pressure data are aligned with a unified timestamp. The heart rate sensor uses a green LED and photodiode to detect changes in blood volume. After bandpass filtering, the heart rate value is extracted using a peak detection algorithm, and heart rate variability indices, including SDNN and RMSSD, are calculated to assess the state of autonomic nervous system function. Accelerometer and gyroscope data are used to calculate the device's three-dimensional posture angles using a posture calculation algorithm to determine whether the user is sitting, standing, walking, running, or sleeping. Activity intensity is divided into 5 levels. The system categorizes blood pressure into four levels: resting state, mild activity, moderate activity, vigorous activity, and extreme exercise. Each level corresponds to a different normal blood pressure fluctuation range. Actual data showed that the patient's baseline systolic blood pressure was 132 mmHg at rest, rising to 141 mmHg during mild activity, 156 mmHg during moderate activity, and reaching 178 mmHg during vigorous activity. These context-related baseline values ​​were incorporated into the individualized model, enabling the system to distinguish between physiological blood pressure elevation and pathological hypertension after exercise. During the 30-day monitoring period, contextual awareness reduced false alarms by 87 cases, accounting for 62.1% of the total false alarms. Technical modifications include adding a skin conductance sensor to monitor stress and emotional states. Elevated skin conductance values ​​are usually accompanied by sympathetic nerve excitation and increased blood pressure. This sensor increases the cost by approximately 15 yuan but provides contextual information related to emotional states. Another option is to integrate a blood oxygen saturation sensor, as low blood oxygen may lead to compensatory increases in blood pressure. This sensor shares a photoelectric detection unit with the heart rate sensor, requiring only the addition of a red LED, increasing the cost by approximately 8 yuan.

[0111] The data quality adaptive correction unit employs a multi-level detection strategy for its quality anomaly identification processor. The first level is amplitude anomaly detection, where a reasonable range for blood pressure measurements is set as 60-220 mmHg systolic and 40-130 mmHg diastolic. Measurements outside this range are directly marked as invalid. The second level is abrupt change detection, calculating the difference between two adjacent measurements. If the difference exceeds 30 mmHg and the time interval is less than 5 minutes, a measurement abrupt change is considered to exist. Analysis of the simultaneous acceleration signal determines whether it is caused by motion. If the acceleration abrupt change is synchronized with the blood pressure abrupt change, it is marked as a motion artifact; if the acceleration is stable, it is marked as a suspicious pathological event requiring further analysis. The third level is signal quality assessment, which evaluates the photoelectric sensor... Frequency domain analysis was performed on the original waveform of the piezoelectric sensor to calculate the energy percentage of the dominant frequency component in the signal power spectrum. When the dominant frequency energy percentage was below 65%, the signal quality was considered poor, possibly due to motion interference or poor sensor contact. The fourth level was a consistency check, comparing the three measurement values. If the deviation of a measurement value from the other two exceeded 15 mmHg, the data quality of that value was marked as questionable. During the 30-day monitoring period, a total of 17,856 blood pressure data points were collected, of which 312 were amplitude abnormalities, 156 were sudden events, 892 were poor signal quality, and 467 were consistency anomalies. The overall data quality problem rate was 10.22%. The context-driven correction processor addressed different types of quality problems. A targeted correction strategy was employed to address motion artifacts using an adaptive notch filter. The filter's center frequency dynamically adjusted to follow the motion frequency, which was extracted using the Fourier transform of the acceleration signal. The notch depth was set to -40 dB, and the bandwidth was set to ±10% of the motion frequency. This filter effectively suppressed periodic motion interference while preserving the true fluctuations in the blood pressure signal. Actual test results showed that the signal quality score improved from an average of 52 points to 78 points after filtering. For short-term data loss due to sensor loosening, a piecewise linear interpolation method was used. Stable data segments were extracted 10 minutes before and after the missing data segment, and the linear trend of blood pressure change was calculated. The missing segment was then interpolated and filled according to this trend. When the missing duration exceeded 30 minutes, [the method was not implemented]. Instead of interpolation, data was marked as blank. During the 30-day monitoring period, 73 sensor loosening events were successfully repaired, with an average missing duration of 8.3 minutes. For baseline drift, a high-pass filtering method was used, with the filter cutoff frequency set to 0.0017Hz, which can remove slow baseline shifts while preserving the true long-term trend of blood pressure. The data confidence assessment processor calculates a confidence score for each blood pressure measurement, ranging from 0 to 1, and includes three sub-dimensions. The measurement stability score is obtained by calculating the consistency of multiple measurements within the same sampling window. When the standard deviation is <2 mmHg, the score is 1.0; when the standard deviation is 2-5 mmHg, the score linearly decays to 0.7; and when the standard deviation is >5 mmHg, the score is 0.4. The multi-source consistency score is obtained by comparing the deviations of three measurement values. A score of 1.0 is given when the maximum deviation is <3 mmHg, a score that linearly decays to 0.6 when the deviation is between 3-8 mmHg, and a score of 0.3 when the deviation is >8 mmHg. The contextual rationality score is obtained by judging the degree of matching between the blood pressure value and the current activity state. For example, blood pressure should be 10-20% lower than the waking state during sleep. If the blood pressure value meets expectations, the score is 1.0; if it deviates from expectations by 10-20%, the score is 0.8; and if the deviation exceeds 20%, the score is 0.9. The confidence level was divided into three weighted averages of 0.5, with weights of 0.4, 0.4, and 0.2 respectively. Actual results showed that 82.3% of the data had high confidence, 12.7% had medium confidence, and 5.0% had low confidence. A possible technical variation involves training a data quality classifier using machine learning methods, learning quality discrimination rules from a large amount of labeled data through support vector machines or random forest algorithms. This approach can improve accuracy to 92%, but requires thousands of labeled samples and has a long training time.

[0112] The multi-scale temporal feature automatic extraction unit of the dual-path feature learning and anomaly pre-detection module adopts a sliding window analysis method, defining five time scales: the ultra-short-term scale for capturing immediate blood pressure changes such as postprandial and post-exercise changes; the short-term scale for capturing blood pressure fluctuation patterns within half a day; the medium-term scale for analyzing circadian rhythms; the long-term scale for identifying interweekly variations and weekend effects; and the ultra-long-term scale for tracking seasonal and medication-adjusted chronic variations. At each time scale, the extracted feature set includes mean, standard deviation, maximum value, minimum value, coefficient of variation, skewness, kurtosis, upward trend slope, downward trend slope, peak frequency, and trough value. A total of 23 features, including frequency, zero-crossing rate, autocorrelation coefficient, and power spectral density distribution, were extracted across five scales, resulting in a 115-dimensional feature vector. The feature importance assessment processor employed an information gain-based feature selection algorithm to calculate the contribution of each feature to anomaly detection and risk assessment. Information gain was defined as the reduction in system entropy before and after using the feature. Statistical analysis of 30 days of data revealed the top 10 features with the highest information gain to be: 24-hour standard deviation, 6-hour maximum value, 24-hour coefficient of variation, 1-hour upward trend slope, 7-day mean drift, 6-hour peak frequency, 24-hour autocorrelation coefficient, and 30-day standard deviation. The cumulative information gain of these 10 features—information gain rate, the difference between the maximum and minimum values ​​in one hour, and the average value over seven weekend days—reached 73.6%, indicating that they contained most of the useful information. To reduce computational complexity, the system ultimately selected 42 features with information gain > 0.15 to form a simplified feature set. The multi-scale feature fusion processor adopted a hierarchical concatenation method. First, the feature vectors of the five time scales were aligned along the time dimension. Then, fusion weights were assigned to each dimension of features based on their importance scores. After weight normalization, the sum of the weights was equal to 1. Features with high information gain received greater weights and thus dominated the fused feature vector. The final output fused feature vector has a dimension of 42, which is 63.5% less than the original 115-dimensional features. The computational complexity is reduced by about 68%, while the classification performance only decreases by 2.3%. Technical variations include using principal component analysis for feature dimensionality reduction, projecting the 115-dimensional features onto a 30-dimensional principal component space. This method can eliminate the linear correlation between features but has poor interpretability and requires offline training of the transformation matrix. Another approach is to use a deep learning autoencoder for feature learning, which automatically learns the low-dimensional representation of the data through a neural network. This method has strong feature extraction capabilities but requires GPU acceleration and has high power consumption, making it unsuitable for wearable devices.

[0113] The baseline model initialization processor in the individualized dynamic baseline continuous learning unit requires collecting the patient's historical blood pressure data from the previous 14 days as initialization samples. A total of 2016 blood pressure data points were collected during the initialization phase. First, the data were divided into five groups according to activity level: sleep group, resting group, mild activity group, moderate activity group, and vigorous activity group. Statistical distribution parameters were calculated for each group. The mean systolic blood pressure for the sleep group was 118.6 mmHg with a standard deviation of 7.2 mmHg; the mean systolic blood pressure for the resting group was 132.4 mmHg with a standard deviation of 9.8 mmHg; and the mean systolic blood pressure for the mild activity group was 141.7 mmHg with a standard deviation of 12.3 mmHg. The mean blood pressure in the moderate activity group was 156.2 mmHg with a standard deviation of 15.6 mmHg, while the mean blood pressure in the vigorous activity group was 178.4 mmHg with a standard deviation of 18.9 mmHg. These statistical parameters constituted the initial parameter set for the individualized baseline model. In addition, circadian rhythm parameters were calculated. The patient's blood pressure exhibited a typical dipper pattern, with nighttime blood pressure decreasing by 18.3% compared to daytime blood pressure. The rate of blood pressure rise was 3.2 mmHg / hour from 6:00 AM to 9:00 AM, and the rate of blood pressure decrease was 2.1 mmHg / hour from 6:00 PM to 9:00 PM. The adaptive weighted decay update processor used an exponential decay weighted method to update the baseline model online. The decay coefficient is set to 0.95, meaning that the weight of historical data decays to 95% of its original value every day. After 20 days, the weight decays to 35.8%, and after 50 days, it decays to 7.7%. This decay mechanism allows the baseline model to focus more on recent data while preserving long-term trends. In addition to time decay, data quality weighting is also introduced, with data with higher confidence levels receiving greater update weights. Specifically, data with a confidence level > 0.8 has a weight of 1.0, data with a confidence level between 0.6 and 0.8 has a weight of 0.6, and data with a confidence level < 0.6 has a weight of 0.3. Experimental results show that after 30 days of continuous learning, the resting baseline weights decrease from the initial 1. The blood pressure was adjusted from 32.4 mmHg to 130.8 mmHg, and the diurnal rhythm decrease was adjusted from 18.3% to 17.6%. These adjustments reflect the long-term trend of gradually improving blood pressure control after medication and lifestyle improvements. The seasonal periodicity modeling processor used Fast Fourier Transform to identify periodic components and performed frequency domain analysis on the 30-day blood pressure time series, detecting three significant periodic components: a 24-hour cycle, a 7-day cycle, and a 3.5-day cycle. Further analysis revealed that the patient's average blood pressure on weekdays (Monday to Friday) was 137.2 mmHg, while the average blood pressure on weekends (Saturday and Sunday) was 132.6 mmHg, a difference of 4 mmHg.A 6 mmHg reading, this interweekly pattern, was incorporated into the baseline model to avoid misinterpreting a normal Monday blood pressure elevation as abnormal. Technical variations included using wavelet transform for multi-resolution analysis. Wavelet transform offers better time-frequency localization than Fourier transform and can detect non-stationary periodic signals, but its computational complexity increases by approximately 40%. Another approach was to use an autoregressive moving average model for time series modeling. This method can predict future blood pressure trends but requires assumptions of data stationarity and periodic re-estimation of model parameters.

[0114] The rapid deviation detection processor in the abnormal candidate event screening unit employs a standardized deviation calculation method to calculate the degree of deviation of each blood pressure measurement relative to an individualized baseline. The deviation is defined as the difference between the measured value and the baseline value divided by the baseline standard deviation. A deviation absolute value >1.5 is marked as mild deviation, >2.5 as moderate deviation, >3.5 as severe deviation, and >5.0 as extremely severe deviation. During the 30-day monitoring period, a total of 1872 deviation events were detected in this patient, including 1456 mild deviations, 312 moderate deviations, and 89 severe deviations. There are 15 extremely severe deviations. The anomaly candidate clustering processor performs time-series clustering on the detected deviation events, setting a time window threshold of 30 minutes. If the time interval between two deviation events is less than 30 minutes, they are merged into a single persistent anomaly event. After clustering, 532 independent anomaly candidate events are formed, including 378 isolated deviation points, 98 short-term anomalies, 43 medium-term anomalies, and 13 long-term anomalies. The candidate priority ranking processor calculates a priority score for each candidate event, with the scoring formula comprehensively considering the anomaly magnitude, duration, and historical recurrence. The analysis considered three factors: frequency, magnitude of abnormality (represented by the maximum deviation), duration (in hours), and historical recurrence frequency (defined as the number of similar abnormalities occurring within the past 7 days). The weights of these three factors were 0.5, 0.3, and 0.2, respectively. After sorting the candidates in descending order of priority scores, the top 50 candidate events were selected for the key analysis set. These 50 events included 15 with extremely severe deviations, 13 with long-term abnormalities, and 22 with frequent recurrences. Through this priority ranking mechanism, the system focused computational resources on the abnormal events most likely to have clinical significance, avoiding detailed analysis of all candidate events. The analysis revealed computational waste. Actual test results showed that 42 out of 50 events in the key analysis set were ultimately confirmed as real anomalies, with an accuracy rate of 84%. However, only 38 out of the remaining 482 low-priority candidates were confirmed as real anomalies, with an accuracy rate of only 7.9%. This indicates that the priority ranking strategy effectively improved the initial screening efficiency. Technical variations include using clustering algorithms to spatially cluster anomaly candidates, considering not only temporal proximity but also similarity in feature space, which can identify anomaly clusters with similar feature patterns. However, this increases computational complexity and requires manual setting of distance and density thresholds.

[0115] The context-aware anomaly detection unit of the local-to-global interactive anomaly precise identification module employs a dynamic threshold adjustment mechanism. Based on the current overall risk level and context type, it queries a threshold strategy table for the appropriate deviation threshold. The strategy table contains 5 risk levels and 8 context types, totaling 40 strategy combinations. For example, in a low-risk, resting, and awake context, the deviation threshold is set to 3.0, while in the same context in a high-risk state, the threshold decreases to 1.8. This means that high-risk patients are subject to stricter anomaly criteria to improve detection sensitivity. Similarly, in a low-risk, strenuous exercise context, the threshold is set to 5.5, allowing for a significant increase in blood pressure after exercise, while in a high-risk, strenuous exercise context, the threshold decreases to 4.0 and will trigger exercise. Risk warning: The patient's overall risk level dynamically changed during the 30-day monitoring period. Days 1-8 were classified as medium-high risk, days 9-18 decreased to medium risk, and days 19-30 decreased to medium-low risk. Corresponding abnormality detection thresholds were adjusted accordingly. The contextual feature fusion processor jointly analyzed blood pressure deviation with contextual features to establish a situation-blood pressure response expectation model. This model defines the normal response range for blood pressure under different situations. For example, the expected blood pressure increase during moderate activity is 20-40 mmHg, the expected blood pressure increase one hour after a meal is 5-15 mmHg, and the expected blood pressure decrease during deep sleep is 15-25%. When the actual measured blood pressure change falls within the expected range, even if it deviates from the individual's resting blood pressure... A large baseline is not considered abnormal but rather a normal physiological response. Conversely, if blood pressure changes significantly exceed the expected range, an abnormal alarm will be triggered even if the deviation does not exceed a fixed threshold. Actual results show that during a 30-day monitoring period, the context-aware mechanism avoided 62 false alarms after exercise, 19 after meals, and 11 false alarms due to emotional stress. The total number of false alarms decreased from 148 using the traditional fixed threshold method to 56, a 62.2% reduction in the false alarm rate. It also detected 8 abnormalities due to context mismatch, which would be missed by the fixed threshold method. The detection strategy switching processor selects the appropriate detection algorithm based on the context type. The system has five built-in detection algorithms: threshold discrimination, trend analysis, fluctuation detection, and pattern recognition. Matching and statistical tests were used. In the patient's sleep state, trend analysis detected two instances of persistently rising nocturnal blood pressure. Such slow changes are easily missed in threshold discrimination. In a moderate activity state, fluctuation detection detected three instances of excessive blood pressure fluctuations during exercise. Technical variations include using machine learning classifiers for anomaly detection, learning anomaly patterns from large amounts of labeled data by training support vector machines or neural networks. This method can achieve an accuracy of over 90%, but requires thousands of labeled samples and has poor model interpretability. Another approach is to use the isolated forest algorithm. This algorithm does not require labeled samples and is unsupervised learning. It identifies isolated points by randomly segmenting the data space, making it suitable for detecting unknown types of anomalies, but it is sensitive to parameter settings.

[0116] The temporal causal inference processor in the abnormal risk causal association analysis unit employs the Granger causality test to analyze the causal relationship between local abnormal events and changes in overall risk status. A time window of 6 hours before and after the event is set. For each detected local abnormal event, the overall risk score sequence for the 6 hours before and after its occurrence is extracted. An autoregressive model is used to test whether the occurrence of the abnormal event has a significant predictive effect on the risk score for the next 6 hours. If the p-value is <0.05, it is determined that the abnormality leads to increased risk; otherwise, the risk score sequence for the 6 hours before the abnormal event is extracted, and the effect of the risk score on the risk score is tested. Whether the occurrence of abnormal events has predictive value is determined by p-value < 0.05, indicating that the abnormality is caused by risk. Among the 80 confirmed abnormal events identified during the 30-day monitoring period, causality tests showed that 32 abnormalities were risk-triggered (abnormalities leading to increased risk), 26 were risk-manifested (risk-increased risk leading to abnormality), and 22 were bidirectionally correlated (abnormality and risk mutually influence each other). This finding has guiding significance for clinical intervention. For risk-triggered abnormalities, targeted treatment of the triggering factors is needed; for risk-manifested abnormalities, adjustments to the overall treatment plan are required. The abnormal risk contribution quantification processor calculates... The contribution of each abnormality to the risk score is defined as the magnitude of the change in the risk score before and after the abnormality multiplied by the duration of the abnormality. In this case, a severe hypertension event was detected on day 12, lasting 3.2 hours. Before the event, the overall risk score was 0.58, and 6 hours after the event, the risk score rose to 0.79, a change of 0.21. The contribution was calculated as 0.21 × 3.2 = 0.672. This event was marked as a high-impact abnormality and triggered an immediate medical attention alert. After the patient promptly went to the hospital and their medication was adjusted, their blood pressure was controlled. Similarly, a mild hypertension event was detected on day 24, lasting... Within 0.8 hours, the risk score changed by only 0.03, with a contribution of 0.024. This event was marked as a low-impact abnormality, recorded but not warned. Through the contribution quantification mechanism, the system can distinguish between high-risk abnormalities that require immediate clinical intervention and low-risk abnormalities that can be continuously observed. The reverse verification processor retrieves similar cases from historical abnormality records and uses the k-nearest neighbor algorithm to search for the top 5 historical cases most similar to the current abnormality in the feature space. The Euclidean distance between the feature vectors of the current abnormality and the historical cases is calculated. A distance <0.3 is considered highly similar, 0.3-0.6 is considered moderately similar, and >0.6. For cases with low similarity, the system extracts validated triggering factors and intervention effects as references for the current anomaly. For example, if an abnormal blood pressure is detected on day 18, the system retrieves two highly similar historical anomalies from days 7 and 11. The triggering factor for both anomalies was a high-salt diet, and the intervention was a reduction in salt intake followed by a return to normal blood pressure. Therefore, the system infers that the current anomaly may also be caused by a high-salt diet and suggests the patient check their recent diet. The patient confirms they ate pickled food the previous night, and their blood pressure returned to normal after limiting salt intake the next day. The association confidence score is calculated by dividing the number of similar cases by the total number of searches. This case has a confidence score of 2 / 5 = 0.4, indicating moderate confidence. Technical variations include using decision trees or random forests to establish anomaly-risk association rules, clearly representing the impact paths of different types of anomalies on risk through a tree structure. This method is highly interpretable but requires a large amount of training data. Another approach is to use Bayesian networks to model causal relationships, representing the conditional dependencies between variables through directed acyclic graphs. This method is theoretically rigorous but the network structure learning is complex.

[0117] The multi-level anomaly classification and tracing unit's anomaly type discrimination processor categorizes detected anomalies into six types: hypertensive emergency, hypertensive urgency, stage 1 hypertension, stage 2 hypertension, hypotensive events, and excessive blood pressure fluctuations. The classification criteria include multi-dimensional features such as absolute blood pressure value, deviation amplitude, duration, fluctuation rate, and accompanying symptoms. The patient's anomaly distribution during the 30-day monitoring period was: 1 instance of hypertensive urgency, 4 instances of stage 2 hypertension, 23 instances of stage 1 hypertension, 3 hypotensive events, and 12 instances of excessive blood pressure fluctuations. The triggering factor tracing processor extracts potential triggering factors from data collected 8 hours prior to the anomaly, including medication records, dietary logs, activity records, sleep quality, emotional state, and environmental factors. Association rule mining algorithms were used to analyze the co-occurrence patterns of anomalies and triggering factors. With a minimum support of 15% and a minimum confidence of 60%, 12 strong association rules were identified. For example, the rule "missed antihypertensive medication → elevated blood pressure within 6 hours" had a support of 28.6% and a confidence of 83.3%; the rule "high-salt diet → elevated blood pressure within 12 hours" had a support of 18.2% and a confidence of 71.4%; and the rule "insufficient rest after strenuous exercise → excessive blood pressure fluctuations within 2 hours" had a support of 15.8% and a confidence of 66.7%. These rules provide clear action guidelines for clinical intervention. The triggering factor analysis results for the patient's 80 abnormal events showed that medication-related events occurred 22 times, diet-related events 18 times, and exercise-related events 15 times. The anomaly knowledge base matching processor maintains a knowledge base containing more than 3,000 anomaly cases. Each case records anomaly feature vector, triggering factors, intervention measures, and prognostic results. The cosine similarity between the current anomaly and all cases in the knowledge base is calculated. The similarity calculation is based on a 42-dimensional feature vector, and the top 10 cases with the highest similarity are selected as references. The patient's hypertensive urgency event on day 12 has a similarity of 0.91 with case A-1547 in the knowledge base. The triggering factors for this historical case were missed antihypertensive medication and emotional stress. The intervention was immediate oral administration of short-acting antihypertensive medication and bed rest. Blood pressure dropped to [a certain level] within 2 hours. Within a safe range, referring to this case, the system recommended the same intervention to the patient. After the patient's blood pressure dropped to a safe level within 1.5 hours. The knowledge base adopts an online learning mechanism, adding new abnormalities to the knowledge base whenever they are clinically validated, so that the system's traceability capability continuously improves with the extension of usage time. Technical variations include using deep learning models for abnormality classification, automatically learning the temporal pattern features of abnormalities through convolutional neural networks or long short-term memory networks. This method can achieve a classification accuracy of over 95%, but requires GPU computing resources and a large amount of training data. Another option is to use fuzzy logic for classification, allowing an abnormality to belong to multiple categories with different membership degrees. This method is more in line with clinical practice, but the reasoning process is complex.

[0118] The multidimensional risk dynamic modeling and prediction unit of the dynamic risk assessment and intelligent early warning decision-making module integrates four categories of risk factors: short-term abnormal factors, including the frequency of abnormal events in the past 7 days, total duration of abnormalities, maximum deviation, and average deviation; the patient's short-term abnormality score on day 30 was 0.42; long-term trend factors, including the deviation of the 30-day mean blood pressure from the baseline, the 30-day standard deviation of blood pressure, the rate of blood pressure control, and blood pressure variability; the patient's long-term trend score on day 30 was 0.38; individual baseline factors, including age, gender, BMI, smoking history, alcohol consumption history, family history, comorbidities, and target organ damage; the individual baseline score was 0.61; and medication adherence factors, including on-time medication rate, frequency of medication adjustments, and incidence of adverse reactions. The medication adherence score was 0.88. A weighted summation of the four factors yielded a comprehensive risk score with weights of 0.3, 0.3, 0.25, and 0.15, respectively. The patient's comprehensive risk score on day 30 was 0.42×0.3 + 0.38×0.3 + 0.61×0.25 + 0.88×0.15 = 0.525, classifying it as medium risk. The local-to-global risk temporal fusion modeling processor constructed a dynamic risk evolution model, which tracked the temporal changes in the risk score. The patient's 30-day risk score trajectory was: Day 1 0.78, Day 5 0.71, Day 10 0.63, Day 15 0.58, Day 20 0.52, Day 25 0.49, and Day 30 0.525, showing an overall downward trend. The treatment was effective, but a slight rebound occurred between days 25 and 30, requiring attention. The model also analyzed the cumulative effect of local abnormal events on overall risk, employing a decaying cumulative mechanism. The impact of recent abnormalities has a higher weight, while the impact of long-term abnormalities has a lower weight. Specifically, the impact of an abnormality on the risk on day t+k has a weight of 0.9^k, meaning the weight decreases by 10% each day. This mechanism allows the impact of short-term abnormalities to gradually diminish without affecting the risk score in the long term. The risk prediction and extrapolation processor uses an autoregressive model to predict the risk evolution over the next 24 and 72 hours. The model is based on the risk score sequence of the past 7 days and estimates the autoregressive parameters using the least squares method. The patient's predicted risk score for day 31 on day 30 is 0.53, and the predicted risk score for day 32 is... The risk score for the day was 0.52, and the prediction results showed that the risk continued to decrease slightly and the fluctuation range was within an acceptable range. The system also predicted that the probability of hypertensive emergencies in the next 7 days was 2.3% and the probability of cardiovascular events was 0.8%. These predictions provided quantitative basis for doctors to formulate treatment plans. Technical variations include using Long Short-Term Memory (LSTM) networks for risk prediction. LSTM can learn the long-term dependencies of time series and the prediction accuracy is about 15% higher than that of autoregressive models, but it requires GPU acceleration for training and has high model complexity. Another approach is to use Markov models, which discretize the risk state into a finite number of levels and predict the future risk state through the state transition probability matrix. This method is computationally simple, but discretization will lose information accuracy.

[0119] The anomaly detection parameter optimization processor in the risk anomaly bidirectional feedback optimization unit dynamically adjusts the anomaly detection threshold based on the risk assessment results. The adjustment strategy is as follows: when the risk score increases, the anomaly detection threshold is lowered to improve detection sensitivity; specifically, the deviation threshold is lowered by 20% and the priority score threshold by 15%. When the risk score decreases, the anomaly detection threshold is increased to reduce the false alarm rate; the deviation threshold is increased by 15% and the priority score threshold by 10%. For example, in this patient's case, the risk score decreased from 0.63 to 0.58 on days 9-10, and the system automatically increased the deviation threshold from 2.5 to 2.875. On days 25-26, the risk score... The error rebounded from 0.49 to 0.53. The system then lowered the deviation threshold from 3.0 to 2.4. The effect of this parameter adjustment was validated between days 26 and 30, with all six detected anomalies being true anomalies, achieving 100% accuracy. Without adjusting the parameters and maintaining the original threshold of 3.0, two anomalies would have been missed, resulting in a false negative rate of 33.3%. The risk model calibration processor monitored the consistency between the anomaly detection results and risk predictions. When the frequency of detected anomalies was significantly higher than the risk model's predicted value, it indicated that the model underestimated the risk, requiring an increase in the risk score or adjustment of the model parameters. Conversely, when the frequency of anomalies was significantly lower than the predicted value, it indicated that the model overestimated the risk. The risk score needs to be lowered. Specifically, a consistency threshold of 20% is set. Calibration is triggered when the actual abnormal frequency deviates from the predicted value by more than 20%. For this patient, the risk model predicted an abnormal frequency of 1.2 times / day for the next 7 days on day 15, but the actual detected abnormal frequency was 0.6 times / day, a deviation of 50% exceeding the threshold. The system judged the risk to be overestimated. By reducing the weight of short-term abnormal factors, the model was calibrated, and the calibrated risk score decreased from 0.58 to 0.52, which is more in line with the actual situation. The closed-loop performance evaluation processor calculates system performance indicators every 7 days, including abnormal detection accuracy, sensitivity, specificity, and risk. Performance indicators for risk prediction accuracy and timely warning rate were as follows: Week 1: Detection accuracy 78.3%, Sensitivity 82.1%, Specificity 74.6%, Risk prediction accuracy 71.2%; Week 2: Detection accuracy 83.7%, Sensitivity 86.4%, Specificity 81.2%, Risk prediction accuracy 76.8%; Week 3: Detection accuracy 88.9%, Sensitivity 91.2%, Specificity 86.7%, Risk prediction accuracy 82.5%; Week 4: Detection accuracy 92.4%, Sensitivity 94.6%, Specificity 90.3%, Risk prediction accuracy 87%.The 3% improvement in performance metrics validates the effectiveness of the bidirectional feedback optimization mechanism. The system gradually adapts to individual patient characteristics and disease progression patterns through continuous learning and adjustment. Technical variations include using reinforcement learning for parameter optimization, modeling anomaly detection and risk assessment as Markov decision processes, and training the optimal policy through a reward function. While theoretically capable of achieving global optimum, this method is complex and has slow convergence. Another approach is to use Bayesian optimization for hyperparameter tuning, modeling the relationship between parameters and performance through Gaussian processes and efficiently searching for the optimal configuration in the parameter space. This method is highly efficient but requires extensive performance evaluation experiments.

[0120] The user profile building processor in the personalized graded early warning and intervention suggestion unit integrates the patient's multi-dimensional features to construct a user profile vector. Demographic characteristics include age 65 years, male, high school education, retired worker occupation, and urban residence. Health characteristics include BMI 28.3, waist circumference 96cm, smoking history (quit smoking 5 years ago), occasional light alcohol consumption, exercise habits (brisk walking 3 times a week for 30 minutes each time), and a diet that is high in salt and oil. Disease characteristics include a 2-year history of hypertension, a 5-year history of diabetes, left ventricular hypertrophy, no renal impairment, no history of cerebrovascular disease, and no peripheral vascular disease. Medication characteristics include antihypertensive medication being a calcium channel blocker combined with an ACE inhibitor, hypoglycemic medication being metformin, good medication adherence, and no adverse drug reactions. Psychological characteristics include moderate health literacy, good disease awareness, mild anxiety score, and high adherence score. All characteristics were encoded and converted into a 121-dimensional feature vector, and then reduced to 18 dimensions through principal component analysis, retaining 92% of the information content. The tiered early warning strategy generator matches an early warning plan based on the user profile characteristics. This patient belongs to the middle-aged and elderly male type with good disease awareness and high adherence. The system has configured a detailed data-driven early warning plan for him / her. The features of this plan are... The warning information includes specific numerical values, the timing and frequency of the warnings are appropriate, and the warning method is mobile application push + SMS backup. The warning content combines text and images, and the urgency level of the warnings is divided into 5 levels: alert, attention, warning, severe, and emergency. This patient triggered 12 attention-level warnings, 4 warning-level warnings, and 1 severe-level warning within 30 days. The intervention program recommendation processor recommends intervention measures based on the abnormality type and risk level. For the hypertensive subacute event on day 12, the system recommended a three-level intervention program: immediate measures, short-term measures, and long-term measures. After the patient implemented the immediate measures, their blood pressure... The blood pressure dropped to a safe level within 1.5 hours. The next day, after a medical visit, the doctor changed the antihypertensive medication from monotherapy to combination therapy. On the 24th day, a mild hypertensive event occurred. The system recommended a primary intervention plan, which included lifestyle modifications, medication reminders, and continuous observation. The feasibility score of the intervention plan was based on the patient's adherence record and intervention cost. For example, the feasibility score for "seeking immediate medical attention" was 0.95, and the feasibility score for "strictly limiting salt intake" was 0.72. The expected effect score was based on historical intervention effect statistics. For example, the expected effect score for "taking short-acting antihypertensive drugs" was 0.88, and the expected effect score for "increasing exercise" was 0.65. Intervention plans were ranked based on a comprehensive feasibility and expected outcome score, prioritizing those with high feasibility and effectiveness. During the 30-day monitoring period, the system recommended 47 interventions, of which 39 were actually implemented by the patient. Of these, 32 resulted in improved blood pressure. Six of the eight not implemented interventions were long-term measures such as strict salt restriction and weight loss. Technical variations included personalized recommendations using a collaborative filtering algorithm, which predicts the optimal plan for the current patient by analyzing the intervention effects of similar patients. This method has high accuracy but requires a large patient database. Another approach is online learning using a multi-armed slot machine algorithm, treating each intervention as an arm and balancing new and known effective interventions through an explore-and-utilize strategy. This method is highly adaptable but has a slower convergence speed and requires a longer learning period.

[0121] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. A dynamic blood pressure monitoring system integrating overall risk and local anomaly detection, characterized in that, It includes a multi-source blood pressure data fusion acquisition module, a dual-path feature learning and anomaly pre-detection module, a local and global interactive anomaly accurate identification module, and a dynamic risk assessment and intelligent early warning decision-making module; The multi-source blood pressure data fusion acquisition module is used to acquire blood pressure and context data; the dual-path feature learning and anomaly pre-detection module is used to extract temporal features and perform preliminary anomaly screening; the local and overall interactive anomaly accurate identification module is used to detect and analyze local anomalies; and the dynamic risk assessment and intelligent early warning decision module is used to comprehensively assess risks, form a feedback optimization mechanism, and output personalized early warning decisions. The multi-source blood pressure data fusion acquisition module includes a multi-channel blood pressure signal synchronous acquisition unit, a physiological behavior context association unit, and a data quality adaptive correction unit. The multi-channel blood pressure signal synchronous acquisition unit is used to synchronously acquire blood pressure data. The physiological behavior context association unit is used to synchronously acquire physiological behavior information and align it with the blood pressure data in time. The data quality adaptive correction unit is used to identify data quality problems and perform intelligent correction and compensation based on context information. The dual-path feature learning and anomaly pre-detection module includes a multi-scale temporal feature automatic extraction unit, a personalized dynamic baseline continuous learning unit, and an anomaly candidate event initial screening unit. The multi-scale temporal feature automatic extraction unit is used to discover and extract key temporal features of blood pressure data. The personalized dynamic baseline continuous learning unit is used to build and continuously update the personalized baseline model. The anomaly candidate event initial screening unit is used to perform coarse-grained anomaly detection during the feature extraction stage to quickly screen out suspicious anomaly candidate sets. The local-to-global interactive anomaly accurate identification module includes a context-aware anomaly detection unit, an anomaly risk causal correlation analysis unit, and a multi-level anomaly classification and tracing unit. The context-aware anomaly detection unit is used to dynamically adjust the threshold and strategy of anomaly detection. The anomaly risk causal correlation analysis unit is used to analyze the causal relationship between detected local anomalies and overall risks. The multi-level anomaly classification and tracing unit is used to classify the detected anomalies by type and trace their triggering factors. The dynamic risk assessment and intelligent early warning decision-making module includes a multi-dimensional risk dynamic modeling and prediction unit, a risk anomaly bidirectional feedback optimization unit, and a personalized graded early warning and intervention suggestion unit. The multi-dimensional risk dynamic modeling and prediction unit is used to construct a dynamic risk model and predict future risk evolution trends. The risk anomaly bidirectional feedback optimization unit is used to optimize anomaly detection parameters in reverse using risk assessment results, and update the risk assessment model with anomaly detection results. The personalized graded early warning and intervention suggestion unit is used to generate differentiated risk levels, early warning strategies, and actionable intervention suggestions.

2. The blood pressure dynamic monitoring system integrating overall risk and local anomaly detection as described in claim 1, characterized in that, The individualized dynamic baseline continuous learning unit includes a baseline model initialization processor, an adaptive weight decay update processor, and a seasonal periodic modeling processor. The baseline model initialization processor is used to construct an individualized blood pressure baseline model based on the user's initial historical blood pressure data. The adaptive weight decay update processor is used to update the baseline model online as new data is continuously collected. The seasonal periodic modeling processor is used to identify periodic components in blood pressure data, extract periodic parameters and amplitude parameters of regular fluctuation patterns such as diurnal rhythm, weekly pattern, and seasonal effect, and incorporate these regular changes into the baseline model to avoid misjudging normal periodic fluctuations as abnormal. The adaptive weight decay update processor calculates the individualized baseline value according to the following formula: ； Where B(t) represents the individualized baseline value at time t, BP i This represents the i-th historical blood pressure measurement, where n is the total number of historical data points, and w i (t) represents the adaptive time weight of the i-th historical data at time t, M i (t) represents the memory retention factor of the i-th historical data; The adaptive time weights are calculated according to the following formula: ； Among them, c i t represents the confidence score of the i-th data point. i This indicates the time of data collection for the i-th data point. S is the instantaneous decay function. i Let i be the local stability index for the i-th data. For stability gain parameters, This represents the central blood pressure value of the cluster to which the i-th data belongs.

3. The blood pressure dynamic monitoring system integrating overall risk and local anomaly detection as described in claim 2, characterized in that, The context-aware anomaly detection unit includes a multi-dimensional feature weighted deviation calculation processor, a context feature fusion processor, and a detection strategy switching processor. The multi-dimensional feature weighted deviation calculation processor is used to calculate the degree of deviation of the current blood pressure measurement value from the individualized baseline in multiple feature dimensions. The context feature fusion processor is used to jointly analyze the current physiological behavioral context information with the blood pressure data to determine whether the blood pressure change is consistent with the current context. The detection strategy switching processor is used to select an appropriate anomaly detection algorithm from the strategy library according to the identified current context type. The multi-dimensional feature weighted deviation calculation processor calculates the comprehensive deviation D according to the following formula. total (t): ； Where m is the total number of feature dimensions. This represents the i-th eigenvalue at time t. This represents the individualized baseline value of the i-th feature at time t. This represents the time-varying standard deviation of the i-th feature. This represents the main effect weight of the i-th feature. The nonlinear exponent represents the i-th feature. This represents the interaction effect weight between the i-th feature and the j-th feature. Let represent the interaction deviation term between the i-th feature and the j-th feature at time t, and q be the norm parameter. This is a context-sensitive dynamic correction factor.

4. The blood pressure dynamic monitoring system integrating overall risk and local anomaly detection as described in claim 3, characterized in that, The multidimensional risk dynamic modeling and prediction unit includes a multi-factor risk integration processor, a local-to-overall risk time-series fusion modeling processor, and a risk prediction and extrapolation processor. The multi-factor risk integration processor is used to construct a comprehensive risk assessment vector for the current moment by integrating multi-dimensional information. The local-to-overall risk time-series fusion modeling processor is used to construct a dynamic risk model that integrates local abnormal event sequences and the long-term trend of overall blood pressure and calculate the comprehensive risk evolution trajectory. The risk prediction and extrapolation processor is used to predict the risk development trajectory within a preset time period in the future by extrapolation or time series prediction model based on the current risk status and historical risk evolution trajectory.

5. A dynamic blood pressure monitoring system integrating overall risk and local anomaly detection as described in claim 4, characterized in that, The local-to-global risk time-series fusion modeling processor calculates the comprehensive risk score according to the following formula. : ； in, Accumulate risks for local anomaly time series. This represents a long-term risk to the overall trend. For time-varying local-to-global trade-off coefficients, To integrate nonlinear parameters, This is the synchronization amplification factor. This is an index of synchronous deterioration between the local and the overall system. This is a risk gating factor.

Images