Patient vital sign abnormity detection method based on artificial intelligence technology

Abstract

The invention provides a patient vital sign anomaly detection method based on an artificial intelligence technology, and relates to the technical field of data processing, and the method comprises the steps: collecting original sign data of a patient; calculating multi-dimensional characteristic parameters; establishing an individual baseline model, and determining a comprehensive reference interval in the model; the vital sign features monitored in real time are constructed into multi-dimensional feature vectors, the multi-dimensional feature vectors are input into the individual baseline model, and the deviation degree of the real-time multi-dimensional feature vectors in the comprehensive reference interval is calculated; performing a clustering analysis to identify an anomalous aggregation region; performing trend analysis, calculating change direction consistency and continuous change amplitude of the multi-dimensional feature vector, and generating a trend analysis result; calculating an accumulated change index in a continuous time window according to a trend analysis result to obtain a dynamic confidence score; when the dynamic confidence score continuously exceeds an adaptive threshold value, determining that a vital sign abnormal event exists; the autonomy and accuracy of the method for detecting the vital sign abnormity of the patient are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for detecting abnormal vital signs in patients based on artificial intelligence technology. Background Technology

[0002] In existing technologies, the detection of abnormal patient vital signs typically relies on wearable sensor devices and threshold-based algorithms. Abnormalities are determined based on predefined thresholds (such as a heart rate exceeding 120 beats per minute or blood oxygen levels below 90%). Once an abnormality is detected, an alarm is triggered, and medical staff or family members are notified. While this method has some practicality in basic monitoring, its identification capabilities largely depend on preset rules or simple statistical models, making it difficult to handle individual differences among patients, noise interference, and complex correlations between multidimensional signals.

[0003] In practical applications, such threshold-based or simple rule-based detection methods may lead to false alarms or missed alarms. For example, for elderly patients in the postoperative recovery period, their vital signs often fluctuate significantly, and temporary changes in heart rate or respiratory rate may be misinterpreted as abnormal, thus frequently triggering alarms and distracting medical staff. Conversely, for patients with chronic arrhythmia or intermittent hypoxemia, if the abnormal pattern does not exceed the set threshold range, potential risks may not be identified, delaying intervention. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting abnormal vital signs in patients based on artificial intelligence technology, in order to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A method for detecting abnormal vital signs in patients based on artificial intelligence technology, the method comprising: The patient's raw vital signs data are collected and then time-synchronized, artifact detected, and anomaly interpolation compensation is performed to generate vital signs data. Based on vital sign data, multidimensional feature parameters are calculated, including short-term fluctuation amplitude, cycle stability, acceleration of change, and cross-index correlation, to obtain a multidimensional feature dataset. Based on the patient's historical monitoring records, multidimensional distribution learning is performed on the multidimensional feature dataset to establish an individual baseline model, and a comprehensive reference interval reflecting normal physiological state is determined in the model. The vital signs obtained from real-time monitoring are constructed into multi-dimensional feature vectors and input into the individual baseline model. The deviation of the real-time multi-dimensional feature vectors in the comprehensive reference interval is calculated to generate deviation feature data. Based on the deviation feature data, perform cluster analysis to identify anomalous clustering regions that occur within a continuous time window and generate anomalous labeling results; Based on the anomaly labeling results, trend analysis is performed to calculate the consistency of the direction of change and the magnitude of continuous change of the multidimensional feature vectors, generating trend analysis results; based on the trend analysis results, the cumulative change index within the continuous time window is calculated to obtain the dynamic confidence score; When the dynamic confidence score continuously exceeds the adaptive threshold, an abnormal vital sign event is identified, and the corresponding feature change trajectory is recorded; based on the characteristics of the abnormal event, the feature weights and comprehensive reference interval of the individual baseline model are updated.

[0006] The above-described solution of the present invention has at least the following beneficial effects: First, this invention overcomes the limitations of traditional single-threshold-based judgments by using multi-dimensional feature extraction and time series analysis, achieving a transformation from "static numerical judgment" to "dynamic pattern recognition." By collecting heart rate, respiratory rate, blood oxygen saturation, body temperature, and blood pressure signals, and performing time synchronization, artifact detection, and interpolation compensation, the signals maintain temporal consistency, reducing noise interference at the source and ensuring the accuracy of subsequent feature modeling and anomaly identification.

[0007] Secondly, this invention achieves a quantitative description of the dynamic changes in vital signs by calculating multidimensional features such as short-term fluctuation amplitude, periodic stability, acceleration of change, and cross-indicator correlation. This mechanism can simultaneously reflect the volatility and synergy of multiple physiological indicators, and can identify abnormal coupling relationships between different vital signs, rather than relying solely on the judgment of a single indicator exceeding its limit, thereby significantly improving the sensitivity and robustness of anomaly detection.

[0008] Furthermore, this invention establishes an individual baseline model through multidimensional distribution learning of historical data, generating a comprehensive reference range that reflects normal physiological state. This model can adapt to individual differences among different patients, dynamically adjust the reference range, reduce false alarms and false negatives, and is particularly suitable for chronic disease or postoperative rehabilitation monitoring scenarios.

[0009] Furthermore, this invention achieves dynamic quantitative assessment of patient status by calculating the deviation of real-time multidimensional feature vectors from a comprehensive reference interval. It can identify trend abnormalities in advance and trigger early warnings before deviations reach traditional thresholds, improving the timeliness of risk identification.

[0010] Finally, this invention achieves hierarchical identification and confidence assessment of abnormal events through progressive calculation of anomaly marking, trend analysis, and dynamic confidence scoring. Trend analysis extracts the direction and duration of multidimensional feature changes, while dynamic confidence scoring combines time weighting and deviation fusion to generate a risk value. When the score continuously exceeds an adaptive threshold, abnormal vital signs events are automatically identified and the model is updated.

[0011] In summary, this invention achieves sensitive and stable detection of abnormal vital signs through multidimensional modeling, individualized learning, and dynamic confidence assessment. It maintains high accuracy in noisy environments and can identify potential health risks in advance, demonstrating significant clinical application value and practicality. Attached Figure Description

[0012] Figure 1 This is a flowchart of a method for detecting abnormal vital signs in patients based on artificial intelligence technology, provided in an embodiment of the present invention. Detailed Implementation

[0013] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0014] like Figure 1 As shown, embodiments of the present invention propose a method for detecting abnormal patient vital signs based on artificial intelligence technology, the method comprising: The patient's heart rate, respiratory rate, blood oxygen saturation, body temperature and blood pressure signals are collected to form raw vital signs data; the raw vital signs data are then synchronized with time, artifacts are detected and outlier interpolation is performed to generate vital signs data. Based on vital sign data, multidimensional feature parameters are calculated, including short-term fluctuation amplitude, cycle stability, acceleration of change, and cross-index correlation, to obtain a multidimensional feature dataset. Based on the patient's historical monitoring records, multidimensional distribution learning is performed on the multidimensional feature dataset to establish an individual baseline model, and a comprehensive reference interval reflecting normal physiological state is determined in the model. The vital signs obtained from real-time monitoring are constructed into multi-dimensional feature vectors and input into the individual baseline model. The deviation of the real-time multi-dimensional feature vectors in the comprehensive reference interval is calculated to generate deviation feature data. Based on the deviation feature data, perform cluster analysis to identify anomalous clustering regions that occur within a continuous time window and generate anomalous labeling results; Based on the anomaly labeling results, trend analysis is performed to calculate the consistency of the direction of change and the magnitude of continuous change of the multidimensional feature vectors, generating trend analysis results; based on the trend analysis results, the cumulative change index within the continuous time window is calculated to obtain the dynamic confidence score; When the dynamic confidence score continuously exceeds the adaptive threshold, an abnormal vital sign event is identified, and the corresponding feature change trajectory is recorded; based on the characteristics of the abnormal event, the feature weights and comprehensive reference interval of the individual baseline model are updated.

[0015] In this embodiment of the invention, continuous monitoring and intelligent identification of changes in the patient's physiological state are achieved by collecting patient vital signs signals, extracting features, creating individualized models, and detecting dynamic anomalies.

[0016] First, the original vital sign signals are time-synchronized and artifacts are detected to remove outliers and artifacts. Then, interpolation compensation is used to maintain the continuity of the time series, thereby obtaining stable vital sign data that can be used for analysis, providing reliable input for feature calculation.

[0017] In the multidimensional feature extraction stage, short-term fluctuation amplitude, periodic stability, acceleration of change, and cross-indicator correlation are calculated to obtain a multidimensional feature dataset describing the dynamic characteristics of vital signs. This dataset integrates local changes and overall trends, reflecting the synergistic relationship between multiple indicators such as heart rate, respiration, and blood pressure, providing high-dimensional input for individualized modeling.

[0018] In the personalized modeling phase, based on patients' historical monitoring records, distribution learning is performed on multidimensional feature data to establish an individual baseline model and determine a comprehensive reference interval reflecting normal physiological state. This model can adaptively adjust the judgment criteria according to individual characteristics, improving the adaptability of the detection to individual differences.

[0019] In real-time monitoring, newly acquired vital signs are converted into multi-dimensional feature vectors and input into a baseline model. By calculating their deviation from the comprehensive reference range, real-time deviation feature data is obtained, which is used to identify the difference between the current state and the normal range. Subsequently, cluster analysis is performed on the deviation feature data to identify abnormal clusters that are continuous in time and show significant deviations, avoiding misjudgments based on single-point fluctuations.

[0020] When a persistent deviation is detected, trend analysis is performed to calculate the directional consistency and magnitude of continuous changes in multidimensional features, and a dynamic confidence score is generated to quantify the risk of anomalies. When the score consistently exceeds an adaptive threshold, an abnormal vital sign event is identified, and the feature change trajectory is recorded for model updates.

[0021] For example, in clinical monitoring scenarios, a patient's heart rate, respiratory rate, blood pressure, and blood oxygen saturation signals are continuously collected and input into this method. First, an individualized multidimensional baseline model is established to form characteristic distribution intervals under normal physiological conditions. When a patient's heart rate and respiratory rate continuously increase due to underlying disease, and the correlation between indicators strengthens, the method can identify deviations from the clustering region and calculate trend analysis results with high consistency in the direction of change. As the trend continues to develop, if the dynamic confidence score exceeds the adaptive threshold, the method automatically identifies the abnormal event and outputs a prompt, thus providing healthcare professionals with real-time health status assessment data.

[0022] In a preferred embodiment of the present invention, updating the feature weights and comprehensive reference interval of the individual baseline model based on the characteristics of the abnormal event specifically includes: When the dynamic confidence score consistently exceeds the adaptive threshold, an abnormal vital sign event is identified within the current time period, and the corresponding multidimensional feature change trajectory is recorded. Subsequently, based on the time span, influencing feature components, and deviation direction of the abnormal event, core feature data of the abnormal event is extracted. This core feature data includes the average deviation magnitude of the feature components, the duration of the change, and the deviation ratio from the historical reference interval, reflecting the feature contribution of the abnormal event in the multidimensional space.

[0023] First, the deviation of each feature component in the abnormal event is statistically analyzed, and the weight change trend of each feature during the abnormality process is calculated. For example, when the heart rate feature shows a high deviation amplitude and a long duration in multiple abnormalities, it is determined that this feature is more sensitive to changes in physiological state, thus increasing its feature weight. This process can be achieved through a normalized proportional allocation method, so that the adjustment amplitude of each feature weight increases proportionally to its performance in the abnormal event.

[0024] Next, the comprehensive reference interval is dynamically updated based on the new feature weight information. During the update process, newly collected normal monitoring data is fused with historical normal data to calculate the feature distribution center and distribution boundary under the new weight conditions. When the weight change of a feature component exceeds a set threshold, the reference interval boundary of the corresponding distribution dimension is adaptively adjusted to make the comprehensive reference interval more closely match the latest individual physiological characteristics. Through the periodic update mechanism, the model can continuously learn the patient's physiological change trends, improving its adaptability to new states while maintaining historical stability. Finally, the updated feature weights and comprehensive reference interval are synchronously written into the individual baseline model to form a new physiological state benchmark for subsequent deviation detection and trend analysis.

[0025] In a preferred embodiment of the present invention, multidimensional feature parameters are calculated based on vital sign data, including short-term fluctuation amplitude, periodic stability, acceleration of change, and cross-index correlation, to obtain multidimensional feature data, including: Based on vital sign data, continuous time windows are divided, and the signal sequences within each time window are averaged and differentiated to generate fluctuation data representing the degree of local change. Based on the fluctuation data, calculate the change ratio between adjacent time windows to generate short-term fluctuation amplitude data; Based on the fluctuation data, correlation analysis is performed on the periodic trends between adjacent windows to generate periodic stability data; Based on the fluctuation data, calculate the rate of change of the signal at each time point and generate rate of change data; The rate of change data is smoothed, and continuous difference operations are performed on the smoothed time series. The variance of the difference results is calculated to generate the acceleration data. A feature set data is constructed based on short-term fluctuation amplitude data, cycle stability data, and change acceleration data. Calculate the time delay correlation between different vital sign signals based on feature set data, and generate cross-index correlation data; By performing time alignment and feature fusion on short-term fluctuation data, cycle stability data, change acceleration data, and cross-indicator correlation data, a multidimensional feature dataset is generated.

[0026] In this embodiment of the invention, a multi-dimensional feature dataset reflecting the changing patterns of physiological signals can be obtained through multi-level calculation and step-by-step feature extraction of vital sign data. This process first divides the original vital sign signals into continuous time windows, and then performs averaging and differencing on the signal sequences within each window to generate fluctuation data that represents the degree of local change. By introducing time window division, the temporal continuity of the signal is preserved while effectively suppressing the interference of instantaneous fluctuations on the overall analysis, providing a structured data foundation for subsequent feature extraction.

[0027] Subsequently, the rate of change between adjacent time windows is calculated based on the fluctuation data to generate short-term fluctuation amplitude data. This data reflects the severity of changes in vital signs within a short timescale and can be used to determine whether an individual is in a state of acute physiological fluctuation. Furthermore, by calculating the periodic trend correlation between time windows, periodic stability data is obtained, enabling the identification of whether the signal maintains regular fluctuations, such as the periodic patterns observed in respiratory rate or heart rate. By smoothing and differentiating the rate of change of the signal, and then calculating the variance, acceleration data can be obtained, which is used to describe the stability and abruptness of changes in vital signs.

[0028] Furthermore, the method calculates the time-delay correlation between multiple physiological signals to obtain cross-index correlation data, thereby revealing the dynamic coupling relationship between different vital signs. This index can reflect the degree of coordination of multiple physiological activities, such as the synchronous changes in heart rate and respiratory rate. Finally, short-term fluctuation amplitude, periodic stability, acceleration of change, and cross-index correlation are time-aligned and fused to form a multi-dimensional feature dataset. This process realizes the transformation from a single time-series signal to multi-parameter dynamic features, enabling a comprehensive description of the changing trends of vital signs in a multi-dimensional space, laying a quantitative foundation for the establishment of individual baseline models and subsequent anomaly detection.

[0029] In a preferred embodiment of the present invention, correlation analysis is performed on the periodic trends between adjacent windows based on fluctuation data to generate periodic stability data, specifically including: First, the periodic characteristics of the signal within each time window are extracted, such as the peak-to-trough interval, trough-to-peak interval, and average amplitude. Then, these periodic characteristics are compared between adjacent time windows to calculate their similarity, thereby reflecting the periodic repetition of the physiological signal.

[0030] This invention analyzes the signal by combining the proportion of change in adjacent periodic features with the phase difference. When the proportion of change in periodic features within two time windows is small and the phase difference is relatively stable, it indicates that the periodicity of the physiological signal is well maintained, and this result is represented as high periodic stability. Conversely, if the proportion of change in periodic features is large or the phase difference changes frequently, it indicates that the periodic features of the signal are disturbed, which will correspondingly reduce the periodic stability score.

[0031] A sliding window mechanism can be used to iteratively compare multiple adjacent time periods, calculate the average value of periodic stability over a period, and generate periodic stability data. This data not only reflects the temporal stability of a single physiological signal but also provides a temporal continuity reference for subsequent cross-index correlation analysis. Finally, the periodic stability data is integrated into a multidimensional feature dataset to characterize the rhythmic features of an individual's physiology.

[0032] In a preferred embodiment of the present invention, calculating the time delay correlation between different vital sign signals based on feature set data to generate cross-index correlation data specifically includes: First, time series of different vital signs, such as heart rate, respiratory rate, and blood oxygen saturation, are extracted from the feature set data. Then, these signal sequences are subjected to sliding alignment to calculate their similarity under different time offset conditions, thereby identifying the time delay relationships between the signals.

[0033] By gradually shifting the time window of a signal sequence, the trend consistency of two signals under different offset conditions can be observed. When the fluctuation direction and amplitude of the two signals reach optimal matching at a certain time offset, this time offset value is recorded as the delay time between the two. Subsequently, the value corresponding to the maximum matching degree is used as the cross-index correlation to measure the dynamic coupling strength between different vital sign signals.

[0034] In a preferred embodiment of the present invention, short-term fluctuation amplitude data, periodic stability data, acceleration of change data, and cross-indicator correlation data are time-aligned and feature-fused to generate a multi-dimensional feature dataset, specifically including: First, the time axes of different features are aligned to ensure that all feature data have a consistent time reference within the same time window. After time alignment, the data from different feature dimensions are normalized to eliminate differences in numerical units among different physiological signals. Next, a weighted fusion strategy is used to combine multiple feature data according to their importance weights to form a multidimensional feature vector. During the fusion process, short-term fluctuation amplitude reflects sensitivity to local changes, periodic stability reflects temporal rhythm, acceleration of change describes the dynamic smoothness of the signal, and cross-index correlation represents overall synergy. By weighted integration of these four types of features, a set of multidimensional feature vectors that can comprehensively characterize an individual's physiological state is generated.

[0035] Finally, the multidimensional feature vectors generated within multiple time windows are arranged in chronological order to construct a complete multidimensional feature dataset. This dataset serves as the core input for subsequent individual baseline model training and anomaly detection, establishing a statistical relationship between temporal continuity and feature synergy, thus providing a foundation for high-precision modeling of individual physiological states.

[0036] In a preferred embodiment of the present invention, based on the patient's historical monitoring records, multidimensional distribution learning is performed on the multidimensional feature dataset to establish an individual baseline model, and a comprehensive reference interval reflecting the normal physiological fluctuation range is determined in the model, including: Based on the patient's historical monitoring cycle, extract multidimensional feature data for the corresponding time period to generate a historical feature sample set; Multidimensional distribution clustering is performed on the historical feature sample set to calculate the joint distribution law of vital signs in the multidimensional feature space and generate statistical results describing the multi-parameter coupling relationship. Based on the statistical results, the correlation structure and time distribution trend between multidimensional feature vectors are extracted to generate modeling input data; Based on the modeling input data, the influence of each feature component on the overall physiological state change is calculated, and feature weight information is generated. Based on the modeling input data and feature weight information, an individual baseline model is established to form a multidimensional distribution region that represents the patient's normal physiological state. The multidimensional distribution region constitutes a comprehensive reference interval that reflects the normal physiological state. The integrated reference interval, baseline data, and feature weight information are stored together in the individual baseline model.

[0037] In this embodiment of the invention, a personalized baseline model is constructed by performing multidimensional distribution learning on historical patient monitoring data, enabling the establishment of personalized reference standards based on the physiological characteristics of different patients. First, the method extracts multidimensional feature data from the patient's historical monitoring cycles, generating a historical feature sample set. This sample set covers information on changes in the patient's physiological state over different time periods, providing a statistical basis for subsequent modeling. Next, multidimensional distribution clustering is performed on this sample set to calculate the joint distribution pattern of each vital sign feature in the multidimensional space, thereby revealing the coupling patterns and temporal correlations between different physiological indicators. This processing method enables the model to distinguish between sporadic changes and persistent fluctuations, improving the model's adaptability to individual differences.

[0038] By analyzing clustering results, the correlation structure and temporal distribution trends among multidimensional feature vectors are extracted to generate modeling input data. Subsequently, the influence of each feature component on overall physiological state changes is calculated, yielding feature weight information and enabling a quantitative assessment of the importance of different physiological indicators. This information reflects the dominant role of each indicator in individual physiological regulation, providing a weighted basis for subsequent anomaly detection. Based on the modeling input data and feature weight information, the method establishes an individual baseline model, forming a multidimensional distribution region reflecting normal physiological state in the multidimensional feature space. The boundary range of this region is defined as the comprehensive reference interval, representing the statistical fluctuation limits of the combination of indicators under normal conditions.

[0039] This comprehensive reference interval not only preserves the cooperative distribution characteristics of individual physiological signals but also reflects their dynamic stability over time. By storing this interval along with baseline data and feature weight information in the model, deviation calculations can be performed on new data during real-time monitoring to determine the degree of difference between the current state and the individual's normal range. This method effectively improves detection accuracy through an individualized multidimensional learning mechanism, avoiding misjudgments caused by group average standards in traditional models.

[0040] In a preferred embodiment of the present invention, multidimensional distribution clustering is performed on the historical feature sample set to calculate the joint distribution law of vital sign features in the multidimensional feature space, generating statistical results describing the multi-parameter coupling relationship, specifically including: First, based on the patient's historical monitoring records, cleaned and synchronized vital sign data are extracted from multiple time periods, and corresponding multidimensional feature vectors are generated for each time window. Each vector contains parameters reflecting physiological characteristics, such as short-term fluctuation amplitude, periodic stability, acceleration of change, and cross-indicator correlation. The resulting historical feature sample set reflects the changes in the patient's physiological state during long-term monitoring.

[0041] In cluster analysis, unsupervised learning is employed to spatially aggregate multidimensional feature vectors from a historical feature sample set. By calculating the similarity between samples (e.g., using feature distance or dissimilarity measures), samples with high similarity and consistent feature trends are grouped into the same cluster unit. This process is not based on a single metric but comprehensively considers the overall distribution relationship of features across different dimensions, thereby forming multiple feature distribution clusters.

[0042] Next, statistical analysis is performed on the data within each cluster unit to extract the central features of that type of sample (such as the average eigenvector, feature variance, and collaborative variation patterns among features), and the boundary distribution between different cluster units is calculated. By summarizing the feature centers and distribution boundaries of all cluster units, the joint distribution law of vital signs in the multidimensional feature space is obtained.

[0043] This joint distribution pattern reflects the dynamic coupling characteristics between various physiological parameters, such as the synchronization pattern of heart rate fluctuations and blood oxygenation changes, and the interdependence between respiratory rate and body temperature changes. Based on this result, statistical data describing the coupling relationship of multiple parameters are generated for subsequent modeling stages, thereby providing a quantifiable multidimensional distribution structure for individual baseline models.

[0044] In a preferred embodiment of the present invention, based on statistical results, the correlation structure and time distribution trend among multidimensional feature vectors are extracted to generate modeling input data, specifically including: First, the feature center vectors and covariance distributions of each cluster unit are analyzed. By comparing the direction and magnitude of fluctuations in different feature dimensions across various samples, the positive, negative, or independent relationships between features are determined. For example, if the trend of decreasing blood oxygen saturation with increasing heart rate is repeatedly observed in most cluster units, then a negative correlation between the two can be determined.

[0045] Subsequently, the trend of feature changes at the time series level is modeled. To this end, based on the time stamp order of each cluster unit, the migration path of the cluster centers in the time dimension is calculated, yielding the distribution trend of features over time. In this way, the temporal dependency structure between features can be identified, such as whether the change of one physiological signal precedes or lags the change of another signal. The analysis results form a temporal correlation matrix to describe the temporal coordination relationship between the feature components.

[0046] After extracting correlations and temporal trends, the two types of information are integrated to generate modeling input data. This input data includes: correlation characteristics between multidimensional features, feature variance, feature coupling strength, and time delay patterns. All data, after normalization and dimensionality compression, are uniformly organized into a standardized modeling input vector set. This vector set can be directly input into the individual baseline model to construct a statistical model that reflects the individual's physiological synergistic characteristics. By introducing temporal distribution trends and structural correlations between features, the model not only possesses the ability to describe instantaneous states but also captures the dynamic evolution patterns of individual physiological laws, providing a stable data foundation for constructing a comprehensive reference interval.

[0047] In a preferred embodiment of the present invention, based on the modeling input data, the degree of influence of each feature component on the overall physiological state change is calculated, and feature weight information is generated, specifically including: First, sensitivity analysis is performed on different feature components to determine the weight of each feature in the overall physiological state. This process is achieved by calculating the fluctuation range of each feature component under different clustering states and its contribution to the migration of the overall distribution center. Specifically, when a change in a feature causes a significant shift in the overall multidimensional distribution center, that feature is identified as a high-influence component; conversely, if the change in a feature has a small impact on the overall state, its weight is correspondingly lower.

[0048] Next, weight adjustments are made based on the synergistic effects between features. By analyzing the correlation matrix between feature components, feature groups that synergistically amplify or counteract changes in the overall state are identified. For example, when respiratory rate and heart rate increase simultaneously for most time periods and jointly cause an overall distribution shift, their joint weight value is increased to reflect this synergistic effect. If two features are negatively correlated, their weight allocation is adjusted to prevent them from having a repetitive impact on the model.

[0049] After obtaining the corrected initial weights, weighted statistics are performed on samples from different time periods to calculate the stability factor of each feature component over time. This stability factor reflects the reliability and consistency of the feature in long-term monitoring. When a feature stably reflects changes in physiological state across different monitoring periods, its weight is further increased, making the model more robust in long-term applications.

[0050] Finally, the combined influence value of each feature component, the synergy correction coefficient, and the time stability factor are integrated into feature weight information. This feature weight information not only reflects the physiological importance of each feature but also provides an accurate weight reference for subsequent deviation calculations and dynamic scoring of the model.

[0051] In a preferred embodiment of the present invention, cluster analysis is performed based on deviation feature data to identify anomalous clustering regions occurring within a continuous time window, and anomaly labeling results are generated, including: Based on the deviation characteristic data, calculate the deviation magnitude and time distribution characteristics within each time window, and generate deviation distribution data; Based on the deviation distribution data, similarity calculation is performed on adjacent time windows, and windows with deviation magnitude within a preset deviation range and continuous time intervals are aggregated into candidate clustering units to generate candidate clustering data; Based on the candidate clustering data, the cluster density and average distance value are calculated. When the cluster density and average distance value both exceed the preset threshold conditions, abnormal clustering area data are generated. Based on the data of abnormal clustering areas, the time center point and the feature component with the highest deviation contribution in that area are extracted to generate anomaly labeling results.

[0052] In this embodiment of the invention, by performing cluster analysis on deviation feature data, abnormal clustering areas of vital signs within continuous time periods can be identified, thereby achieving early identification and dynamic localization of potential abnormal events. The process first calculates the deviation magnitude and its temporal distribution characteristics for each time window based on the deviation feature data, forming deviation distribution data and establishing a statistical description of the degree of deviation changing over time. Subsequently, by calculating the similarity between adjacent time windows, windows with deviation magnitudes within a preset deviation range and continuous time intervals are aggregated into candidate clustering units, effectively extracting abnormal fluctuation segments with temporal continuity. This step avoids false alarms caused by single-point deviations and emphasizes the persistence and temporal correlation of anomalies.

[0053] Next, cluster density and average distance are calculated based on the candidate clustering data to assess the spatial concentration and distributional tightness of anomalous events. By setting preset threshold conditions, density and distance values ​​are jointly judged. When the cluster density is higher than the upper bound of normal and the average distance is lower than the lower bound, the time period is identified as an anomalous clustering area. This dual-threshold judgment mechanism makes anomaly identification more robust, filtering out noise interference and random fluctuations simultaneously. Furthermore, based on the clustering results, the time center point and the feature component with the highest deviation contribution are extracted to generate anomaly labeling results, thereby locating the dominant physiological indicators that cause the anomaly.

[0054] This method transforms single-point deviation behavior into temporally continuous abnormal pattern recognition through joint analysis of cluster density and average distance, achieving hierarchical judgment from "local deviation" to "clustering trend." Its technical advantage lies in its ability to rapidly identify the formation stages of potential physiological abnormalities, especially in the case of coordinated changes in multidimensional vital signs, enabling the detection of early abnormal signs and providing stable data support for subsequent trend analysis and dynamic scoring.

[0055] In a preferred embodiment of the present invention, deviation amplitude and time distribution characteristics within each time window are calculated based on deviation feature data to generate deviation distribution data, specifically including: First, the average deviation of each feature component within the current time window is calculated to characterize the overall physiological deviation level. Simultaneously, the difference between the maximum and minimum deviation of each feature component is calculated to measure the intensity and instability of fluctuations within the window.

[0056] To capture the characteristics of temporal distribution, trend analysis was performed on the deviation values ​​within each time window. By observing the direction of increase or decrease of the deviation values ​​over time, the persistence and consistency of the direction of deviation changes within that time window were determined. For example, when the deviation value continues to increase, it indicates that the patient's physiological state is gradually moving away from the normal range; conversely, when the deviation value gradually returns to the normal range, it indicates that the physiological state is tending to recover.

[0057] After obtaining the magnitude and temporal trend of the deviation, these two are combined to form deviation distribution descriptive data. This data reflects the overall deviation intensity, direction of change, and stability of fluctuations in physiological characteristics within a specific time window. Finally, a corresponding deviation distribution vector is generated for each time window, and the results of all time windows are constructed into a continuous deviation distribution data sequence. This deviation distribution data not only reflects the abnormal change patterns of individuals over time but also provides a basic input for subsequent cluster analysis, enabling the identification of potential anomalous clustering areas.

[0058] In a preferred embodiment of the present invention, based on the deviation distribution data, similarity calculation is performed on time-adjacent windows, and windows with deviation magnitudes within a preset deviation range and continuous time intervals are aggregated into candidate clustering units to generate candidate clustering data, specifically including: First, a similarity index is calculated between adjacent time windows. This index is based on a comprehensive evaluation of the difference in the magnitude of deviation between windows and the consistency of the direction of the time distribution trend. When two time windows have small differences in the magnitude of deviation and the trend of deviation changes is consistent, they are considered to have a high degree of similarity.

[0059] Subsequently, a preset deviation range is set as a similarity filtering condition. This preset deviation range is derived from the normal fluctuation range of the individual baseline model and is used to exclude slight random fluctuations. For example, if the average deviation of two time windows is between the upper and lower bounds of the baseline model deviation, and the time interval is less than a preset continuous threshold (such as 5 minutes or 10 sampling points), these two time windows are aggregated into a candidate clustering unit.

[0060] Next, a sliding aggregation approach is used to group all consecutive time windows in the time series that meet the conditions. Each group of consecutive time windows is considered a candidate clustering region, reflecting the potential accumulation of anomalous trends. The aggregation duration, average deviation intensity, and internal deviation consistency of each candidate clustering unit are calculated to describe the overall properties of the clustering characteristics.

[0061] Finally, all candidate clustering units are recorded in chronological order and output as candidate clustering data. This data reflects the degree and persistence of anomalous concentration of multidimensional features over time, providing a basis for further judgment on whether anomalous clustering regions have formed.

[0062] In a preferred embodiment of the present invention, based on the abnormal clustering region data, the time center point and the feature component with the highest deviation contribution within that region are extracted to generate anomaly labeling results, specifically including: First, the time location and corresponding deviation magnitude of all time windows within the region are statistically analyzed. The time-weighted average of the deviation intensity is then calculated and used as the time center point of the anomaly region. This time center point represents the focal point of the anomaly in time and can serve as a reference node for subsequent trend analysis and alarm triggering.

[0063] Next, the deviation contribution of all multidimensional feature components within the anomaly cluster region is calculated. Specifically, the average deviation and variance of each feature component within the region are statistically analyzed, and its proportion relative to the overall deviation intensity is calculated. When the deviation of a certain feature in the anomaly region remains consistently high, and its fluctuation amplitude is significantly greater than that of other feature components, that feature is considered to have the largest contribution to the anomaly.

[0064] After identifying the feature component with the highest deviation contribution, this feature component is marked as the dominant anomalous feature, and the time center point, deviation intensity information, and dominant feature type are combined to form the anomaly labeling result. This anomaly labeling result is used to describe the spatial and temporal attributes of the anomalous event, clearly indicating the time period of the anomaly, the main anomalous features, and their intensity of change.

[0065] Finally, the anomaly labeling results are used to calculate the direction and persistence of the anomaly. In clinical applications, these results can directly alert medical staff to the specific source of the abnormal signal, such as "abnormally elevated heart rate accompanied by a slight decrease in blood oxygen," and provide a reference for the time period in which the anomaly occurred, thereby achieving precise anomaly localization and early warning based on time-series characteristics.

[0066] In a preferred embodiment of the present invention, based on the anomaly marking results, trend analysis is performed to calculate the consistency of the direction of change and the magnitude of continuous change of the multidimensional feature vectors, generating trend analysis results, including: Based on the anomaly labeling results, multidimensional feature data within the corresponding time window are extracted to form anomaly time series data; Perform time sorting and smoothing on abnormal time series data, calculate the direction information of changes in each feature component between adjacent time windows, and generate change direction data; Based on the change direction data, the consistency ratio of the change directions of different feature components is statistically analyzed to generate directional consistency data that reflects the synergy of multi-dimensional features. Based on the multidimensional feature differences between adjacent time windows in abnormal time series data, the cumulative value of the change amplitude is calculated to generate continuous change amplitude data; Data on directional consistency and data on the magnitude of continuous change are used as the results of trend analysis.

[0067] In this embodiment of the invention, trend analysis of anomaly marking results enables the identification of the direction and persistence of vital signs changes from a time-series perspective, forming trend analysis results that reflect the patterns of multidimensional physiological state changes. The method first extracts multidimensional feature data within the corresponding time window based on the anomaly marking results, constructing anomaly time-series data to track the dynamic changes of each feature on the time axis. Subsequently, the anomaly time-series data undergoes time sorting and smoothing to remove short-term noise and sudden interference, extracting more representative trend information, thereby enhancing the continuity and stability of the analysis.

[0068] During trend calculation, the direction of change of each characteristic component is compared in adjacent time windows to generate directional change data, which describes the upward or downward trend of vital signs. Furthermore, by statistically analyzing the consistency ratio of the direction of change of different characteristic components, directional consistency data is obtained, thereby quantifying the synergy of multiple physiological indicators in terms of direction of change. For example, when heart rate, respiratory rate, and blood pressure all show an upward trend simultaneously, the directional consistency is high, reflecting the patient's overall physiological state and stress response.

[0069] Furthermore, based on the multidimensional feature differences between adjacent windows in the abnormal time series, the cumulative value of the change amplitude is calculated to generate persistent change amplitude data, which is used to assess the intensity and persistence of abnormal changes. The joint analysis of directional consistency and persistent change amplitude allows for the simultaneous assessment of both the "directional consistency" and "intensity persistence" of abnormal changes, thereby more accurately identifying the formation trend of physiological abnormalities. Finally, the directional consistency data and persistent change amplitude data are combined and output to form trend analysis results, providing time-series input for subsequent dynamic scoring calculations. This process realizes a transformation from static deviation identification to dynamic trend judgment, enabling the detection results to not only focus on instantaneous abnormal points but also reflect the overall direction and evolution of physiological changes, providing quantitative evidence for early clinical warning and trend intervention.

[0070] In a preferred embodiment of the present invention, the abnormal time series data is sorted and smoothed over time, and the direction information of the changes of each feature component between adjacent time windows is calculated to generate change direction data, specifically including: First, the data is strictly sorted according to the sampling timestamp to ensure the continuity of each feature component over time. This sorting process ensures that the trend of feature changes is consistent with the actual physiological process, avoiding analytical errors caused by temporal discrepancies.

[0071] Subsequently, noise in the time series data is smoothed. Since individual sensors may generate abnormal data points due to external interference or instantaneous fluctuations in vital sign monitoring, this invention uses moving average or weighted smoothing to correct the signal and prevent it from biasing the characteristic trend analysis. By calculating the weighted average of adjacent sample points within a certain time window, a continuous and smooth characteristic sequence is obtained, ensuring greater stability and continuity in subsequent analysis.

[0072] After smoothing, the feature values ​​of adjacent time windows are differentially analyzed to extract the direction of change of each feature component over time. The sign of the difference result indicates the trend: a positive result indicates that the feature value increases over time; a negative result indicates that the feature value decreases; and a result close to zero indicates that the feature remains relatively stable. In this way, complex time series signals can be converted into discrete trend direction sequences.

[0073] Finally, the direction of change of each feature component within a continuous time window is encoded, generating change direction data. This change direction data not only reflects the short-term trends of different physiological characteristics, but also provides a time-series logical basis for subsequent calculations of feature synergy and the magnitude of sustained changes.

[0074] In a preferred embodiment of the present invention, based on the change direction data, the consistency ratio of the change directions of different feature components is statistically analyzed to generate direction consistency data reflecting the synergy of multidimensional features, specifically including: First, calculate whether the directions of change of different feature components are consistent within each time window. For example, if heart rate, respiratory rate, and blood pressure all show an upward trend within a certain time period, the window is considered to have high directional consistency; if the directions of change of the features are inconsistent, such as heart rate increasing while blood oxygen decreases, it is considered a low consistency window.

[0075] To improve the stability of the analysis, sliding statistics were further performed across multiple adjacent time windows to calculate the proportion of time windows with consistent direction. That is, if most windows show the same direction of change within a certain time interval, this proportion is used as a measure of directional consistency. This proportion reflects the degree of coordinated change among individual vital signs within a certain period; a higher proportion indicates a more holistic physiological change.

[0076] In addition, the weighting relationships between features are considered. When calculating the proportion of directional consistency, features with stronger clinical significance or greater fluctuations are given higher weights. For example, among cardiopulmonary features, if the changes in heart rate and respiratory rate are in the same direction and increase significantly at the same time, the directional consistency score for that window is increased.

[0077] Finally, the directional consistency measurement results for each time window are combined chronologically to form a continuous directional consistency data sequence. This data not only characterizes the synergistic change patterns among features but also reveals the dynamic interactions between different physiological processes.

[0078] In a preferred embodiment of the present invention, the cumulative value of the change amplitude is calculated based on the multidimensional feature difference between adjacent time windows in the abnormal time series data to generate continuous change amplitude data, specifically including: The difference operation is performed on the multidimensional feature vectors within adjacent time windows to calculate the absolute change of each feature component in adjacent time periods. This absolute change is used to measure the intensity of the feature change, and is not affected by the direction of change.

[0079] Subsequently, the magnitude of changes within consecutive time windows is accumulated over time to assess the persistence of feature changes. If a feature maintains a high magnitude of change across multiple adjacent windows, it indicates that the physiological changes of that feature are highly persistent. By weighted summing of the changes across all time windows, the cumulative magnitude of change for each feature component within that time period is obtained.

[0080] To prevent sudden increases in amplitude caused by individual outliers, this invention introduces a smoothing mechanism during the accumulation process. This involves applying a moving average to the amplitude changes over a continuous window, ensuring that the calculation results reflect both trend strength and suppress noise interference. Furthermore, the amplitude changes of different physiological signals are weighted according to feature weight information. For example, the impact of changes in features with high sensitivity within the comprehensive reference interval (such as blood oxygen saturation) on the overall physiological state is amplified.

[0081] Finally, the cumulative changes of each feature component are integrated along the time dimension to generate continuous change amplitude data. This data reflects the intensity of continuous changes in an individual's vital signs within a certain time range and is an important input for subsequent dynamic confidence score calculation.

[0082] In a preferred embodiment of the present invention, the cumulative change index within a continuous time window is calculated based on the trend analysis results to obtain a dynamic confidence score, including: Based on the trend analysis results, extract data on directional consistency and data on the magnitude of continuous change to form trend feature input data; The trend feature input data is processed with time weighting to calculate the cumulative change value within a continuous time window and generate cumulative change index data. Based on the cumulative change index data and combined with the comprehensive reference interval in the individual baseline model, the deviation of the multidimensional feature vector from the comprehensive reference interval is calculated, and standardized deviation data is generated. Based on the standardized deviation data and the feature weight information in the individual baseline model, the deviation degree of different feature components is weighted and fused to generate comprehensive index data; Based on the deviation between the comprehensive index data and the comprehensive reference interval, the confidence probability of the current abnormal state is calculated, and a normalized dynamic confidence score is generated based on the confidence probability.

[0083] In this embodiment of the invention, by performing time-weighted and deviation calculations on the trend analysis results, the cumulative effect of vital sign changes can be dynamically assessed, generating a dynamic confidence score for anomaly identification. The method first extracts directional consistency data and continuous change magnitude data from the trend analysis results to form trend feature input data, reflecting multidimensional change trends under time continuity. Subsequently, the trend feature input data undergoes time-weighted accumulation processing, giving higher weight to recent changes and relatively lower weight to earlier changes, thereby enabling the model to respond more sensitively to current physiological fluctuations and achieving dynamic adaptive detection.

[0084] During the calculation, cumulative change index data is generated based on the cumulative processing results. This data describes the total amount and directional consistency of physiological signal changes within a continuous time window. Next, the cumulative change index is compared with the comprehensive reference interval in the individual baseline model. The deviation of the multidimensional feature vector relative to the comprehensive reference interval is calculated, generating standardized deviation data. This standardization process eliminates dimensional differences between different vital signs, allowing parameters to be compared on a uniform scale and improving the scientific rigor of the overall fusion calculation.

[0085] Subsequently, based on standardized deviation data and feature weight information from individual baseline models, a weighted fusion of deviations from different feature components is performed to generate comprehensive index data. This gives key physiological indicators a higher weight in dynamic assessment, while relatively weakening the influence of non-key indicators. By calculating the deviation relationship between the comprehensive index data and the comprehensive reference interval, the confidence probability of the current abnormal state can be obtained. After normalization, the confidence probability forms a dynamic confidence score, which is used to measure the degree of abnormal risk of the patient's current physiological state. This method, through multi-layer data fusion and time-weighted mechanisms, achieves dynamic, continuous, and individualized abnormal risk assessment, providing quantifiable decision support for the early detection of potential crises.

[0086] In a preferred embodiment of the present invention, based on the cumulative change index data and combined with the comprehensive reference interval in the individual baseline model, the deviation of the multidimensional feature vector from the comprehensive reference interval is calculated to generate standardized deviation data, specifically including: First, extract comprehensive reference interval parameters representing normal physiological state from the individual baseline model, including the distribution center (i.e. normal mean) and upper and lower limits (i.e. acceptable fluctuation range) of each feature component.

[0087] Subsequently, the multidimensional feature vectors obtained from real-time monitoring are compared with the comprehensive reference interval. For each feature component, the difference between its real-time feature value and the corresponding reference interval center value is calculated to reflect the current deviation direction and magnitude of the feature. If the difference is positive, it indicates that the feature value is higher than the individual's normal level; if it is negative, it indicates that the feature value is lower than the normal range.

[0088] To enable a unified comparison of the deviations of different features, the difference results are standardized. The standardization process scales the feature deviation value to the width of a reference interval, mapping the deviation of each feature to a uniform scale (e.g., between 0 and 1). This allows for comparative analysis under a unified standard, even if features such as heart rate and body temperature have different physical dimensions.

[0089] After standardizing each feature, the standardized deviations of each feature component are combined into a multidimensional standardized deviation vector. This deviation vector describes the overall deviation of an individual from its baseline model at the current moment. The temporal trend of this deviation vector is further recorded for use in subsequent dynamic confidence calculations.

[0090] In a preferred embodiment of the present invention, the confidence probability of the current abnormal state is calculated based on the deviation relationship between the comprehensive index data and the comprehensive reference interval, and a normalized dynamic confidence score is generated based on the confidence probability, specifically including: First, identify the position of the composite index value within the reference interval. When the value is within the interval, the patient is considered to be in a normal state; as the value gradually approaches or exceeds the boundary, its abnormal probability is calculated based on the magnitude of the deviation. Specifically, the closer to the interval boundary, the higher the confidence probability; when the deviation exceeds the boundary threshold, the state is marked as a high-confidence abnormality.

[0091] Subsequently, the calculated confidence probabilities are dynamically normalized. The normalization process compares the confidence probabilities within the current time window with the probability distributions during historical monitoring periods to ensure the comparability of the scoring results across different monitoring weeks. All confidence probabilities are mapped to a continuous interval between 0 and 1, where values ​​closer to 1 represent high anomaly risk and values ​​closer to 0 represent stable physiological states.

[0092] Furthermore, a time-weighted factor is introduced to smooth the confidence probability within consecutive windows. When an anomalous trend persists across multiple consecutive windows, the confidence score for the corresponding window is increased to reflect the temporal consistency of the anomalous state. The final dynamic confidence score not only reflects the instantaneous intensity of the anomalous situation but also comprehensively embodies its persistence and trend.

[0093] In a preferred embodiment of the present invention, the method for determining the adaptive threshold includes: Based on the comprehensive reference interval data in the individual baseline model, the boundary distribution features of the multidimensional reference interval are extracted, the probability density boundary values ​​corresponding to each feature combination are calculated, and the initial static threshold data are generated. Based on confirmed abnormal events in historical monitoring records, the distribution density changes of multidimensional deviation trajectories before and after the abnormality are statistically analyzed, and feature sensitivity data are calculated. Based on the feature sensitivity data and real-time monitoring conditions, dynamic correction coefficient data is generated; The initial static threshold data is corrected, and the multidimensional threshold boundary is adjusted according to the dynamic correction coefficient to generate adaptive threshold data.

[0094] In this embodiment of the invention, an adaptive threshold determination mechanism is designed to dynamically adjust the anomaly detection threshold based on individual baseline characteristics and real-time monitoring conditions, thereby improving the sensitivity and stability of anomaly detection. The method first extracts the boundary distribution features of the multidimensional reference interval based on the comprehensive reference interval data stored in the individual baseline model, calculates the probability density boundary values ​​corresponding to each feature combination, and forms initial static threshold data. This step ensures that the initial threshold setting reflects the statistical regularity of an individual under normal conditions, rather than using a fixed empirical threshold, thus guaranteeing individual suitability for detection.

[0095] Based on this, the method statistically analyzes the changes in the distribution density of multidimensional deviation trajectories before and after the occurrence of anomalies, using confirmed abnormal events from historical monitoring records, to generate feature sensitivity data. By quantifying the sensitivity of each feature to anomaly formation, it is possible to determine which features have higher reference value in anomaly identification. Next, dynamic correction coefficients are generated based on the feature sensitivity data and real-time monitoring conditions (such as sampling frequency, data noise level, and fluctuations in the patient's current state) to reflect the impact of environmental factors or short-term physiological changes on the threshold.

[0096] Finally, the initial static threshold is corrected based on a dynamic correction coefficient to obtain real-time updated adaptive threshold data. This adaptive threshold automatically adjusts the detection sensitivity as the monitoring data changes: when the patient's condition is stable, the threshold range is appropriately tightened to improve detection accuracy; when physiological fluctuations are large, the threshold range is appropriately widened to reduce the false alarm rate. Through this bidirectional self-adjusting mechanism, balanced detection performance can be maintained under different conditions, thereby achieving flexible identification and stable judgment of abnormal events. This method is particularly advantageous in long-term continuous monitoring scenarios, as it can continuously optimize the threshold strategy, making the abnormal detection results both accurate and robust.

[0097] In a preferred embodiment of the present invention, based on the comprehensive reference interval data in the individual baseline model, the boundary distribution features of the multidimensional reference interval are extracted, the probability density boundary values ​​corresponding to each feature combination are calculated, and initial static threshold data are generated, specifically including: First, the multidimensional distribution parameters of the comprehensive reference interval are read, including the mean, variance, and correlation characteristics between each feature component. These statistical parameters allow us to construct the joint probability distribution of the multidimensional features in space.

[0098] Subsequently, boundary extraction processing is performed on the joint distribution. By analyzing the joint probability density of each feature combination, locations where the probability density decreases significantly are identified, and these locations are considered the dividing points between normal and abnormal states. In other words, when the probability of a feature combination occurring is lower than the typical distribution range of the baseline model, the feature combination is considered to be in the boundary region.

[0099] To transform this distribution boundary into a quantifiable threshold, a probability density boundary value is calculated for each feature combination. This boundary value represents the lower limit of the feature distribution probability under normal conditions. For example, when the probability of a certain feature combination occurring is lower than this boundary value, it indicates that the patient's condition has exceeded the normal fluctuation range and there may be a potential abnormality.

[0100] Finally, the boundary probability density values ​​of all feature combinations are integrated to form initial static threshold data. This threshold data is used as a reference in subsequent real-time monitoring to help quickly identify anomalous features that deviate from the baseline model.

[0101] In a preferred embodiment of the present invention, based on confirmed abnormal events in historical monitoring records, the distribution density changes of multidimensional deviation trajectories before and after the abnormality are statistically analyzed, and feature sensitivity data is calculated, specifically including: First, extract multidimensional feature vector sequences from historical data within a certain time window before and after the abnormal event (e.g., 30 minutes before and after the abnormality occurs). These data reflect the transition process from the normal state to the abnormal state.

[0102] Subsequently, the changes in the distribution density of each feature component within these time windows are calculated. Specifically, for each feature dimension, the difference in its probability distribution before the abnormality (i.e., the physiologically stable period) and after the abnormality (i.e., the physiologically disordered period) is compared. For example, if the heart rate distribution is concentrated in a lower range before the abnormality, but shifts significantly upward after the abnormality, the distribution shift direction and magnitude of this feature during the abnormal period are recorded.

[0103] Further statistical analysis was conducted on the distribution patterns of different feature components across multiple abnormal events, and the significance of the change in distribution density of each feature before and after the abnormality was calculated. Features with higher significance are more sensitive to the occurrence of abnormal events. For example, in the monitoring of patients with arrhythmias, the distribution changes in heart rate fluctuation amplitude and blood oxygen saturation are often significantly greater than those in body temperature or blood pressure, thus being assigned higher feature sensitivity in the model.

[0104] Finally, the distribution density changes of all features are combined to generate feature sensitivity data. This data reflects the responsiveness of different physiological features to abnormal states, providing a quantitative reference for subsequent adaptive threshold correction.

[0105] In a preferred embodiment of the present invention, dynamic correction coefficient data is generated based on feature sensitivity data and real-time monitoring conditions, specifically including: First, real-time monitoring of the patient's external and internal conditions is crucial, including activity level (resting or moving), measurement environment (postural changes, sleep status, etc.), and data acquisition stability (e.g., signal interference level, sampling integrity). These conditions directly impact the reasonable range of vital sign fluctuations.

[0106] Subsequently, based on the response strength of each feature to abnormalities recorded in the feature sensitivity data, the correction ratio under the current conditions is calculated. For example, when the patient is exercising, the sensitivity to heart rate fluctuations should be reduced, while the sensitivity to respiratory rate and body temperature should be increased accordingly; conversely, in a resting state, heart rate deviations should receive higher weighting.

[0107] To achieve dynamic adjustment, a correction coefficient matrix is ​​generated based on the matching relationship between the sensitivity values ​​of different feature components and real-time state parameters. Each element of this matrix represents the correction ratio of a specific feature relative to the baseline model under the current monitoring conditions. Through this state-adaptive correction mechanism, the threshold boundary has the ability to be flexibly adjusted under different monitoring scenarios.

[0108] Finally, the generated correction coefficient data is stored together with historical sensitivity information to form a dynamic correction coefficient dataset. This data is used as a correction factor in subsequent threshold adjustments, thereby ensuring that the judgment results remain reasonable and accurate under different physiological states.

[0109] In a preferred embodiment of the present invention, the initial static threshold data is corrected by adjusting the multidimensional threshold boundary according to the dynamic correction coefficient to generate adaptive threshold data, specifically including: First, the boundary distribution values ​​corresponding to each feature combination are read from the initial static threshold data. These boundaries reflect the normal fluctuation range of the patient's multidimensional features in a stable state.

[0110] Subsequently, the dynamic correction coefficients are matched feature-by-feature with the static thresholds. For each feature dimension, the corresponding threshold boundary is adjusted based on the value of the correction coefficient. When the correction coefficient is greater than 1, it indicates that the tolerance range of the feature should be expanded, allowing for greater fluctuations; when the correction coefficient is less than 1, it indicates that the threshold range of the feature should be tightened to more sensitively capture potential abnormalities. For example, when a patient is in the early postoperative recovery period, the dynamic threshold range of heart rate may be wider than usual, while after recovery stabilizes, the heart rate threshold will automatically tighten to improve detection accuracy.

[0111] Furthermore, a multi-dimensional boundary coordination mechanism is employed to ensure that the adjustment results of all features maintain overall statistical consistency. Specifically, the joint distribution density of each feature boundary after adjustment is calculated. If distribution distortion or excessive overlap occurs, the correction ratio is normalized twice to maintain the morphological stability of the comprehensive reference interval.

[0112] Finally, the adjusted boundary distribution values ​​are recorded as new adaptive threshold data and stored in the individual baseline model for subsequent real-time anomaly detection.

[0113] In a preferred embodiment of the present invention, an individual baseline model is established based on modeling input data and feature weight information to form a multidimensional distribution region characterizing the patient's normal physiological state. This multidimensional distribution region constitutes a comprehensive reference interval reflecting the normal physiological state, including: Based on the modeling input data, extract the mean, variance and correlation features of each feature component in the time series, and generate feature statistical parameter data; Based on the feature statistical parameter data, calculate the covariance matrix of each feature component in the multidimensional space to generate feature covariance data; Based on the feature covariance data and feature weight information, calculate the comprehensive distribution center and distribution radius of the multidimensional feature vector in the multidimensional feature space, and generate multidimensional baseline distribution data. Based on multidimensional baseline distribution data, the average physiological state point corresponding to the distribution center is determined, and a comprehensive reference interval characterizing the normal fluctuation range is constructed with the distribution radius as the boundary. The comprehensive reference interval and its corresponding distribution parameters are stored in the individual baseline model to form model reference data that can be used for subsequent deviation calculation.

[0114] In this embodiment of the invention, by collaboratively calculating the modeling input data and feature weight information, an individual baseline model is established, forming a multidimensional distribution region reflecting normal physiological state, thus realizing spatial modeling and statistical representation of individualized physiological characteristics. This method first extracts the mean, variance, and correlation features of each feature component in the time series based on the modeling input data, generating feature statistical parameter data, thereby describing the fluctuation stability and changing trend of each feature in the time dimension. This process effectively reduces the impact of short-term outliers on the overall distribution, making the model better reflect the stable physiological patterns of the individual.

[0115] Subsequently, the covariance matrix of each feature component in the multidimensional space is calculated based on the feature statistical parameter data to obtain feature covariance data, thereby describing the interdependencies between various physiological indicators. The introduction of the covariance matrix enables the model to identify coupling patterns between different vital sign signals, such as the synchronicity changes in heart rate and respiration, thus forming a continuous and interconnected feature structure in the multidimensional feature space. Based on the feature covariance data and feature weight information, the comprehensive distribution center and distribution radius of the multidimensional feature vectors in space are further calculated to form multidimensional baseline distribution data. The distribution center represents the patient's average physiological state point, while the distribution radius reflects the fluctuation boundary under normal conditions.

[0116] Based on this, a comprehensive reference interval characterizing the normal fluctuation range is constructed using multidimensional baseline distribution data. This reference interval serves as the basis for subsequent deviation analysis and can dynamically adapt to the physiological differences among individuals. By storing the comprehensive reference interval and its distribution parameters in the individual baseline model, the model can quickly match newly acquired feature vectors in real-time detection, enabling immediate calculation of the degree of deviation. The technical advantage of this process lies in establishing a multidimensional distribution structure that reflects the long-term physiological patterns of individuals through multidimensional statistical modeling and feature weighted fusion. This provides a highly stable and interpretable reference for subsequent anomaly detection, enabling individual self-adaptation and long-term updability.

[0117] In a preferred embodiment of the present invention, based on the modeling input data, the mean, variance, and correlation features of each feature component in the time series are extracted to generate feature statistical parameter data, specifically including: For each vital sign component (such as heart rate, respiratory rate, body temperature, blood pressure, etc.), the historical monitoring period is divided into time windows. The central trend of the characteristic is reflected by calculating the average value of the data in each time window, and the variance is calculated to reflect the degree of characteristic fluctuation.

[0118] Subsequently, correlation analysis was performed on different feature components to extract their synergistic relationships over time. For example, the temporal correlation characteristics of heart rate and respiratory rate were calculated to determine whether there was a synchronous trend in their changes across different time windows. A high correlation indicates that these features may be regulated by the same factor in terms of physiological mechanisms.

[0119] To avoid interference from short-term outliers in the statistical results, outlier removal and smoothing preprocessing are performed before extracting statistical parameters. Methods such as moving averages and median filtering are used to reduce noise. The preprocessed data is more representative of an individual's true physiological state in terms of stability.

[0120] Finally, the mean, variance, and correlation features of each feature are combined to form a feature statistical parameter dataset. This dataset serves as the basis for subsequent feature distribution modeling and covariance matrix calculation, providing input data for constructing individualized multidimensional baseline models.

[0121] In a preferred embodiment of the present invention, based on the feature statistical parameter data, the covariance matrix of each feature component in the multidimensional space is calculated to generate feature covariance data, specifically including: First, based on the mean, variance, and correlation characteristics contained in the feature statistical parameter data, the joint distribution of each feature component in the multidimensional space is constructed.

[0122] Subsequently, the covariance between the feature components is calculated to measure the joint direction and magnitude of change of the two features within the same time period. For example, if the covariance of heart rate and respiratory rate is positive, it means that both usually rise or fall simultaneously; if it is negative, it means that they change in opposite directions.

[0123] To improve the reliability of the modeling, a time-weighted mechanism is introduced when calculating the covariance, assigning higher weight to data from the most recent time window to reflect the dynamic trends in the patient's physiological state. In this way, the model not only considers the overall stable relationships between features but also adapts to recent changes in the patient's condition.

[0124] Finally, the covariance results between each feature pair are stored in matrix form to form feature covariance data. Each element in this matrix reflects the strength of the co-variance of a feature pair, thus forming the core input for multidimensional distribution learning.

[0125] In a preferred embodiment of the present invention, based on feature covariance data and feature weight information, the comprehensive distribution center and distribution radius of the multidimensional feature vector in the multidimensional feature space are calculated to generate multidimensional baseline distribution data, specifically including: First, the distribution center of each feature component in the multidimensional feature space is calculated based on the feature covariance matrix. This distribution center represents the average steady state of the feature during long-term monitoring and can be regarded as a benchmark point for individual physiology.

[0126] Next, based on the overall structure of the covariance matrix, the outer boundary range of the multidimensional distribution is calculated to determine the acceptable range of feature fluctuations. The size of the distribution radius reflects the variable amplitude of each feature component in space, and its calculation incorporates feature weight information. Features with higher weights have a greater impact on the overall distribution, and therefore their fluctuation range has a correspondingly stronger influence on the distribution radius.

[0127] Furthermore, the scale differences of features across different dimensions are normalized to ensure that each feature component can be compared under a unified metric in the multidimensional space. This process yields multidimensional baseline distribution data containing the coordinates of the distribution center and the distribution radius parameters.

[0128] In a preferred embodiment of the present invention, based on multidimensional baseline distribution data, the average physiological state point corresponding to the distribution center is determined, and a comprehensive reference interval characterizing the normal fluctuation range is constructed using the distribution radius as the boundary, specifically including: First, the coordinates of the distribution center point are extracted. These coordinates represent the multidimensional feature combination that most closely approximates the patient's normal physiological state during long-term monitoring. Then, the distribution radius is used as the boundary of the fluctuation range to define the patient's normal physiological state interval. The distribution radius refers to the weighted distance threshold of each feature dimension in the normalized multidimensional space, used to represent the normal fluctuation range.

[0129] In the specific implementation, the distribution center point is used as the core, and upper and lower boundary values ​​are set for each feature dimension. For example, when the average value of a certain feature (such as body temperature) in a normal individual state is 36.7℃ and the distribution radius is 0.3℃, 36.4℃ to 37.0℃ is taken as the normal range for this feature. For features such as heart rate and blood oxygen, the same logic is used to determine their upper and lower limits.

[0130] Meanwhile, considering the correlation between different features, multidimensional cross-correction is used to ensure the overall consistency of the feature boundaries. When multiple features are strongly correlated (such as heart rate and respiratory rate), the range will be appropriately adjusted when the boundary is determined to prevent abnormal misjudgments caused by excessive expansion of a single dimension.

[0131] Finally, the set of boundary values ​​for all feature dimensions is defined as a comprehensive reference interval to describe the multidimensional fluctuation space of a patient under normal conditions. This interval not only reflects the individual's physiological stability range but can also dynamically adapt to subsequent model updates, enabling personalized health monitoring and anomaly identification.

[0132] In a preferred embodiment of the present invention, based on candidate clustering data, cluster density and average distance are calculated. When both cluster density and average distance exceed a preset threshold, abnormal clustering region data are generated, including: Based on the candidate clustering data, calculate the number of samples and spatial distribution range within each time window to generate window density data; Based on the window density data, calculate the sample overlap and sample center distance between adjacent windows to generate window distance data; Based on the window density data and the distance data between windows, calculate the mean cluster density and the mean average distance, and generate cluster statistics. Based on cluster statistics and preset threshold conditions, when the cluster density is higher than the upper limit threshold and the average distance is lower than the lower limit threshold, the current time window is determined as an abnormal clustering window, and abnormal clustering area data is generated; the preset threshold conditions are calculated based on the historical normal distribution characteristics stored in the individual baseline model to determine the upper limit threshold of the normal interval of cluster density and the lower limit threshold of the normal interval of average distance. Based on the data of abnormal clustering areas, the duration of the abnormality and the stability of the clustering are calculated, and the abnormality labeling results are generated.

[0133] In this embodiment of the invention, by jointly calculating the density and distance of candidate clustering data, accurate identification of anomalous clustering regions can be achieved, thereby distinguishing between genuine anomalies and random fluctuations. The method first calculates the number of samples and their spatial distribution range within each time window based on the candidate clustering data, generating window density data to characterize the degree of aggregation of local data. The calculation of window density reflects the concentration characteristics of data on the time axis, providing a basis for judging the persistence of anomalies. Subsequently, by calculating the sample overlap and sample center distance between adjacent time windows, inter-window distance data is generated to analyze the continuity and stability of anomalous changes within different time periods.

[0134] Based on cluster density and distance, the method calculates the mean cluster density and mean average distance to generate cluster statistics, thus forming a comprehensive description of the continuous clustering pattern over time. Next, the method judges the cluster statistics according to preset threshold conditions. When the cluster density exceeds the upper threshold and the average distance is lower than the lower threshold, the time window is determined to be an anomalous clustering window. This dual judgment mechanism effectively avoids false alarms caused by isolated fluctuations or occasional deviations, making anomaly identification more robust and reliable.

[0135] It is important to note that the preset threshold conditions are not fixed empirical parameters, but are dynamically determined based on the historical normal distribution characteristics stored in the individual baseline model. This allows for automatic adjustment of the recognition sensitivity according to the characteristic distribution of different patients. Once an abnormal clustering area is detected, the duration of the abnormality and the stability of the clustering are further calculated to generate anomaly labeling results. Through the collaborative judgment of cluster density and distance, this method can accurately capture abnormal clustering patterns of vital sign changes and locate the time center and main influencing features of the abnormality, achieving cross-time period detection from point-like abnormalities to continuous abnormalities. This technology has high practicality in long-term health monitoring and can provide early warning of sub-health trends or early physiological abnormalities.

[0136] In a preferred embodiment of the present invention, based on candidate clustering data, the number of samples and spatial distribution range within each time window are calculated to generate window density data, specifically including: First, spatial statistical analysis is performed on the deviation feature samples within each time window to quantify the degree of clustering in the deviation dimension for each window. Specifically, in the multidimensional deviation feature space, each time window is treated as an independent observation unit, and the number of sample points it contains is counted, i.e., the number of abnormal deviation records occurring within that time period. The larger the sample size, the more concentrated the fluctuations in the patient's vital signs are within that time period.

[0137] Subsequently, the spatial distribution range of samples in each window is calculated. To do this, the maximum and minimum deviation values ​​of the samples within each time window are determined, and the width of the distribution interval is calculated based on the distance between them. This distribution width reflects the dispersion of the deviation data within that window. If the distribution width is small and the number of samples is large, it indicates that the abnormal deviations are strongly clustered within that window; conversely, it indicates that the fluctuations are dispersed and the anomalies are not significant.

[0138] After calculating the sample size and distribution range, these two factors are combined to generate window density data. Window density can be understood as the sample concentration per unit deviation in space; the larger the value, the more concentrated the anomalies are within that time window. The results of this step not only provide input data for subsequent cluster density analysis but can also be used to preliminarily identify active regions of anomalous signals. This method can characterize the distribution features of anomalies from both temporal and spatial dimensions, laying the foundation for anomaly cluster identification.

[0139] In a preferred embodiment of the present invention, the sample overlap and sample center distance between adjacent windows are calculated based on the window density data to generate window distance data, specifically including: After obtaining the window density data, the spatial similarity between adjacent time windows is further analyzed to determine the continuity of the anomalous distribution across different time periods. First, sample sets from adjacent windows are extracted, and the overlap ratio between them is calculated. This ratio reflects whether there are identical or similar anomalous feature points in the two windows. A high overlap ratio indicates that the anomalous phenomenon has temporal continuity.

[0140] Next, the center distance between the sample distributions of adjacent windows is calculated. Specifically, the spatial center point of each time window sample is determined (i.e., the weighted average position of the sample deviation values), and the distance between two adjacent center points in the multidimensional deviation space is measured. A shorter distance indicates strong spatial continuity of the anomaly, while a longer distance indicates that the anomaly has gradually recovered or shifted.

[0141] To reduce the interference of sporadic fluctuations, the overlap and center distance results are smoothed, and a time-weighted approach is used to emphasize the changing characteristics of continuous trends. Finally, the results of sample overlap and center distance are combined to generate inter-window distance data. This data reflects both the spatial continuity between temporally adjacent windows and describes the temporal stability of anomalous features, serving as an important basis for identifying continuous anomalous regions in cluster analysis.

[0142] In a preferred embodiment of the present invention, cluster density mean and average distance mean are calculated based on window density data and inter-window distance data to generate cluster statistics, specifically including: First, the density values ​​across all time windows are averaged to obtain the mean cluster density. The mean cluster density reflects the concentration of anomalous samples throughout the observation period and is an important indicator for judging the overall anomalous activity.

[0143] Simultaneously, the average distance between the center points of adjacent windows is calculated based on the window distance data. This average distance value describes the continuity and spatial stability of anomalous samples over time. A smaller average distance indicates that the anomalous features occur continuously over time; a larger average distance indicates that the anomalous events are more dispersed.

[0144] Based on this, the mean cluster density and the mean average distance are combined to construct cluster statistics. To improve analytical accuracy, this statistical data is normalized to make the clustering results comparable across different individuals. This data provides a quantitative basis for subsequent determination of whether a preset threshold has been reached. Once regions with high cluster density and low average distance are identified, it can be inferred that a significant anomalous clustering trend exists within that time period. Through this statistical calculation, a quantitative description of spatiotemporal clustering phenomena is achieved.

[0145] In a preferred embodiment of the present invention, the anomaly duration and cluster stability are calculated based on the anomaly clustering region data to generate anomaly labeling results, specifically including: First, the duration of the anomaly is calculated based on the number and time span of time windows within the anomaly cluster region. A longer duration indicates a longer-lasting abnormal state, potentially with higher clinical significance. Then, cluster stability is calculated by analyzing the density changes within each time window of the cluster region. Cluster stability reflects the fluctuation of the anomaly intensity over the entire time range. Small changes in cluster density across the region indicate a stable anomaly; drastic density fluctuations may indicate a short-term spurious anomaly or data perturbation. The duration and cluster stability are then weighted and fused to form an anomaly stability index. This index comprehensively reflects the intensity and persistence of the anomaly. Based on this index, the final anomaly labeling results are determined, including the time center of the anomaly occurrence, the main deviation feature type, and the anomaly level. Finally, the generated anomaly labeling results are compared with historical records to enable model correction and adaptive updates in future detections.

[0146] In a preferred embodiment of the present invention, based on standardized deviation data and feature weight information in the individual baseline model, the deviation degrees of different feature components are weighted and fused to generate comprehensive index data, including: Based on the standardized deviation data, the deviation values ​​of each feature component and their corresponding time series positions are extracted to generate time-paired deviation data; Based on the feature weight information in the time-paired deviation data and the individual baseline model, a weighted average is performed on each feature component to generate initial comprehensive deviation data. Based on the initial comprehensive deviation data, the average deviation trend and fluctuation amplitude within a continuous time window are calculated to generate trend-weighted deviation data. Based on the trend-weighted deviation data, the high deviation range is normalized to generate comprehensive index data.

[0147] In this embodiment of the invention, comprehensive index data is generated through multi-layer weighted fusion processing of standardized deviation data and feature weight information, thereby achieving unified quantification and dynamic evaluation of the degree of deviation of multi-dimensional vital signs. The method first extracts the deviation value of each feature component and its corresponding time series position based on the standardized deviation data, forming time-paired deviation data. This ensures that each deviation value corresponds to a specific time window, thus maintaining data consistency at the time-series level.

[0148] Subsequently, based on the feature weight information in the time-paired deviation data and the individual baseline model, a weighted average is performed on each feature component to generate initial comprehensive deviation data. The core of this step lies in introducing a weight allocation mechanism, enabling the model to distribute the proportion of influence based on the contribution of each physiological indicator to the overall state. For example, when heart rate or blood oxygen fluctuations are more sensitive to physiological abnormalities, their weighting values ​​are automatically increased, thus making the comprehensive assessment results more medically relevant. Next, based on the initial comprehensive deviation data, the average deviation trend and fluctuation amplitude within a continuous time window are calculated to generate trend-weighted deviation data, reflecting the overall deviation intensity and directional changes of multidimensional indicators over time.

[0149] In the final stage, segments with high deviations in the trend-weighted deviation data are normalized to generate a comprehensive index, ensuring that different feature components are compared on a uniform scale. This comprehensive index not only reflects the overall degree of deviation of an individual in multidimensional physiological characteristics but also dynamically responds to the cumulative changes in abnormal trends. Through this weighted fusion and normalization calculation mechanism, multidimensional feature deviation data can be transformed into a single interpretable index, providing a stable input for dynamic confidence scoring. This technical solution, by introducing a hierarchical weighting and temporal fusion structure, achieves nonlinear integration among multiple features, improving the overall consistency and clinical interpretability of anomaly identification, and has significant practical value in complex vital sign monitoring scenarios.

[0150] In a preferred embodiment of the present invention, based on trend-weighted deviation data, high deviation segments are normalized to generate comprehensive index data, specifically including: First, time segments with high deviations are extracted from the trend-weighted deviation data; these are continuous time windows where the deviation values ​​of multidimensional features exceed a set proportional threshold. Then, the extracted high-deviation segments undergo amplitude standardization. To avoid the unbalanced impact of extreme deviations in individual feature components on the overall results, the deviation values ​​of each feature component are proportionally scaled to unify their range to a comparable scale. For example, the deviation value of each feature is mapped proportionally to its maximum permissible fluctuation range within the comprehensive reference interval, thus ensuring that different features such as heart rate, blood oxygen, and respiration have the same evaluation weight in the comprehensive calculation.

[0151] Further normalization is performed on the time dimension. This process aims to balance the impact of short-term sharp fluctuations with long-term slight fluctuations. Specifically, the duration of each high deviation segment is calculated, and a time-weighted deviation value is generated based on the product of the duration and the deviation intensity. Subsequently, the weighted deviation values ​​of all segments are normalized as a whole, so that the final comprehensive index result is limited to a fixed range (e.g., between 0 and 1), thereby achieving comparability of results across different monitoring periods.

[0152] During the normalization process, an abnormal noise suppression mechanism is introduced to smooth and correct short-term abnormal spike signals, preventing abnormal expansion caused by instantaneous sensor interference. The generated comprehensive index data is stored as the final output, used as input for dynamic confidence scoring and anomaly detection. This comprehensive index not only reflects the overall deviation of each feature over different time periods, but also takes into account the deviation intensity, duration, and consistency of fluctuations, enabling the quantification of overall changes in vital signs on a unified scale.

[0153] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1.A method for detecting patient vital sign abnormalities based on artificial intelligence technology, characterized in that, The method comprises: Collecting the original vital sign data of the patient, and performing time synchronization, artifact detection and abnormal point interpolation compensation to generate the vital sign data; According to the vital sign data, calculate the multi-dimensional feature parameters, including short-term fluctuation amplitude, cycle stability, change acceleration and cross-index correlation, to obtain the multi-dimensional feature data set; Based on the patient's historical monitoring records, multi-dimensional distribution learning is performed on the multi-dimensional feature data set to establish an individual baseline model, and a comprehensive reference interval reflecting the normal physiological state is determined in the model; The vital sign features obtained by real-time monitoring are constructed into a multi-dimensional feature vector, which is input into the individual baseline model to calculate the deviation degree of the real-time multi-dimensional feature vector in the comprehensive reference interval, and the deviation feature data is generated; According to the deviation feature data, perform clustering analysis to identify abnormal aggregation areas appearing in a continuous time window, and generate an abnormal marking result; According to the abnormal marking result, perform trend analysis to calculate the change direction consistency and continuous change amplitude of the multi-dimensional feature vector, and generate a trend analysis result; according to the trend analysis result, calculate the cumulative change index in the continuous time window to obtain a dynamic confidence score; When the dynamic confidence score continuously exceeds the adaptive threshold, it is determined that there is a vital sign abnormal event, and the corresponding feature change trajectory is recorded; according to the characteristics of the abnormal event, the feature weight and the comprehensive reference interval of the individual baseline model are updated. 2.The patient vital sign anomaly detection method based on artificial intelligence technology according to claim 1, characterized in that, According to the vital sign data, calculate the multi-dimensional feature parameters, including short-term fluctuation amplitude, cycle stability, change acceleration and cross-index correlation, to obtain the multi-dimensional feature data, including: According to the vital sign data, divide the continuous time window, and perform average and difference processing on the signal sequence in each time window to generate fluctuation data representing the local change degree; According to the fluctuation data, calculate the change rate between adjacent time windows to generate short-term fluctuation amplitude data; According to the fluctuation data, perform correlation analysis on the periodic trend between adjacent windows to generate cycle stability data; According to the fluctuation data, calculate the change rate of the signal at each time point to generate change rate data; Smooth the change rate data, and perform continuous difference operation on the smoothed time series, and generate change acceleration data by calculating the variance of the difference result; According to the short-term fluctuation amplitude data, cycle stability data and change acceleration data, construct the feature set data; According to the feature set data, calculate the time delay correlation between different vital sign signals to generate cross-index correlation data; Time align and feature fuse the short-term fluctuation amplitude data, cycle stability data, change acceleration data and cross-index correlation data to generate the multi-dimensional feature data set. 3.The patient vital sign anomaly detection method based on artificial intelligence technology according to claim 1, characterized in that, Based on the patient's historical monitoring records, multi-dimensional distribution learning is performed on the multi-dimensional feature data set to establish an individual baseline model, and a comprehensive reference interval reflecting the normal physiological fluctuation range is determined in the model, including: According to the patient's historical monitoring period, extract the multi-dimensional feature data of the corresponding time period to generate a historical feature sample set; The historical feature sample set is subjected to multi-dimensional distribution clustering processing, the joint distribution law of vital sign features in a multi-dimensional feature space is calculated, and statistical results describing the multi-parameter coupling relationship are generated; According to the statistical results, the correlation structure and time distribution trend between the multi-dimensional feature vectors are extracted, and modeling input data are generated; According to the modeling input data, the influence degree of each feature component on the overall physiological state change is calculated, and feature weight information is generated; According to the modeling input data and the feature weight information, an individual baseline model is established, a multi-dimensional distribution area representing the normal physiological state of the patient is formed, and the multi-dimensional distribution area constitutes a comprehensive reference interval reflecting the normal physiological state; The comprehensive reference interval, the baseline data and the feature weight information are stored in the individual baseline model. 4.The patient vital sign anomaly detection method based on artificial intelligence technology according to claim 1, characterized in that, According to the deviation feature data, clustering analysis is performed to identify abnormal aggregation areas appearing in continuous time windows, and abnormal marking results are generated, including: According to the deviation feature data, the deviation amplitude and time distribution characteristics in each time window are calculated, and deviation distribution data are generated; According to the deviation distribution data, similarity calculation is performed on adjacent time windows, windows with deviation amplitudes within a preset deviation range and continuous time intervals are aggregated as candidate clustering units, and candidate clustering data are generated; According to the candidate clustering data, clustering density and average distance values are calculated, and when the clustering density and the average distance values exceed the preset threshold conditions at the same time, abnormal aggregation area data are generated; According to the abnormal aggregation area data, the time center point and the feature component with the highest deviation contribution in the area are extracted, and the abnormal marking results are generated. 5.The patient vital sign anomaly detection method based on artificial intelligence technology according to claim 1, characterized in that, According to the abnormal marking results, trend analysis is performed, the consistency of the change direction of the multi-dimensional feature vectors and the continuous change amplitude are calculated, and trend analysis results are generated, including: According to the abnormal marking results, the multi-dimensional feature data in the corresponding time window is extracted, and abnormal time series data are formed; The abnormal time series data are subjected to time sorting and smoothing processing, the direction information of the change of each feature component between adjacent time windows is calculated, and change direction data are generated; According to the change direction data, the consistent proportion of the change direction of different feature components is calculated, and direction consistency data reflecting the coordination of the multi-dimensional features are generated; According to the multi-dimensional feature difference values of adjacent time windows in the abnormal time series data, the cumulative value of the change amplitude is calculated, and continuous change amplitude data are generated; The direction consistency data and the continuous change amplitude data are used as the trend analysis results. 6.The patient vital sign anomaly detection method based on artificial intelligence technology according to claim 1, characterized in that, According to the trend analysis results, the cumulative change index in the continuous time window is calculated, and a dynamic confidence score is obtained, including: According to the trend analysis results, the direction consistency data and the continuous change amplitude data are extracted, and trend feature input data are formed; The trend feature input data are subjected to time weighted accumulation processing, the change accumulation value in the continuous time window is calculated, and cumulative change index data are generated; According to the cumulative change index data, in combination with the comprehensive reference interval in the individual baseline model, the deviation degree of the multi-dimensional feature vector relative to the comprehensive reference interval is calculated, and standardized deviation data are generated; According to the standardized deviation data and the feature weight information in the individual baseline model, the deviation degrees of different feature components are weighted and fused, and comprehensive index data are generated; According to the deviation relationship between the comprehensive index data and the comprehensive reference interval, the confidence probability of the current abnormal state is calculated, and a normalized dynamic confidence score is generated according to the confidence probability. 7.The patient vital sign anomaly detection method based on artificial intelligence technology according to claim 1, characterized in that, The method for determining the adaptive threshold value comprises: According to the comprehensive reference interval data in the individual baseline model, the boundary distribution characteristics of the multi-dimensional reference interval are extracted, the probability density boundary values corresponding to each characteristic combination are calculated, and initial static threshold value data is generated; According to the confirmed abnormal events in the historical monitoring records, the distribution density changes of the multi-dimensional deviation trajectory before and after the abnormality are counted, and the feature sensitivity data is calculated; According to the feature sensitivity data and the real-time monitoring conditions, dynamic correction coefficient data is generated; The initial static threshold value data is corrected, the multi-dimensional threshold boundary is adjusted according to the dynamic correction coefficient, and adaptive threshold value data is generated. 8.The patient vital sign anomaly detection method based on artificial intelligence technology according to claim 3, characterized in that, According to the modeling input data and the feature weight information, an individual baseline model is established to form a multi-dimensional distribution area representing the normal physiological state of the patient, and the multi-dimensional distribution area constitutes a comprehensive reference interval reflecting the normal physiological state, comprising: According to the modeling input data, the average value, variance and correlation characteristics of each feature component in the time sequence are extracted to generate feature statistical parameter data; According to the feature statistical parameter data, the covariance matrix of each feature component in the multi-dimensional space is calculated to generate feature covariance data; According to the feature covariance data and the feature weight information, the comprehensive distribution center and the distribution radius of the multi-dimensional feature vector in the multi-dimensional feature space are calculated to generate multi-dimensional baseline distribution data; According to the multi-dimensional baseline distribution data, the average physiological state point corresponding to the distribution center is determined, and the distribution radius is taken as the boundary to construct a comprehensive reference interval representing the normal fluctuation range; The comprehensive reference interval and the corresponding distribution parameters are stored in the individual baseline model to form model reference data that can be used for subsequent deviation degree calculation. 9.The patient vital sign anomaly detection method based on artificial intelligence technology according to claim 4, characterized in that, According to the candidate clustering data, the clustering density and the average distance value are calculated, and when the clustering density and the average distance value both exceed the preset threshold condition, the abnormal aggregation area data is generated, comprising: According to the candidate clustering data, the number of samples and the spatial distribution range in each time window are calculated to generate window density data; According to the window density data, the sample overlap degree and the sample center distance between adjacent windows are calculated to generate window distance data; According to the window density data and the window distance data, the clustering density mean value and the average distance mean value are calculated to generate clustering statistical data; According to the clustering statistical data and the preset threshold condition, when the clustering density is higher than the upper threshold of the density and the average distance is lower than the lower threshold of the distance, the current time window is determined as an abnormal aggregation window, and the abnormal aggregation area data is generated; the preset threshold condition is calculated according to the historical normal distribution characteristics stored in the individual baseline model, and the upper threshold of the normal interval of the clustering density and the lower threshold of the normal interval of the average distance; According to the abnormal aggregation area data, the abnormal duration and the aggregation stability are calculated to generate abnormal marking results. 10.The patient vital sign anomaly detection method based on artificial intelligence technology according to claim 6, characterized in that, According to the standardized deviation data and the feature weight information in the individual baseline model, the deviation degrees of different feature components are weighted and fused to generate comprehensive index data, comprising: According to the standardized deviation data, the deviation value of each feature component and the corresponding time sequence position are extracted, and time-paired deviation data are generated; According to the time-paired deviation data and the feature weight information in the individual baseline model, weighted average processing is performed on each feature component, and initial comprehensive deviation data are generated; According to the initial comprehensive deviation data, the average deviation trend and the fluctuation amplitude in the continuous time window are calculated, and trend-weighted deviation data are generated; According to the trend-weighted deviation data, the high deviation section is normalized, and comprehensive index data are generated.

Images