Method for identifying data outliers in a chronic disease automatic monitoring network

Abstract

The present application relates to the technical field of chronic disease monitoring, and discloses a data abnormal value identification method of a chronic disease automatic monitoring network. The method first acquires a physiological index time series data set of a target patient at multiple monitoring nodes of the chronic disease monitoring network; then divides time intervals according to data collection time sequence and data point quantity to obtain multiple sub-time series data segments; subsequently associates a first physiological index feature of a first patient node having the same disease feature as the target patient; and finally outputs abnormal identification data on a preset monitoring interface, which directly represents the abnormal probability of the physiological index of the target patient. This method can mine time series data time features and patient group features with the same disease features, optimize abnormal value identification effects, and present abnormal probability directly, which is convenient for medical personnel to quickly master abnormal conditions of patient physiological indexes and suitable for various chronic disease automatic monitoring scenarios.

Description

Technical Field

[0001] This invention relates to the field of chronic disease monitoring technology, specifically a method for identifying outlier data in an automated chronic disease monitoring network. Background Technology

[0002] In the current field of chronic disease management, with the integration of IoT technology and medical monitoring equipment, automated chronic disease monitoring networks are gradually becoming an important tool for real-time tracking of patients' physiological indicators. These monitoring networks typically consist of multiple monitoring nodes distributed across different regions, continuously collecting data on patients' blood pressure, blood glucose, heart rate, and other physiological indicators, generating massive amounts of time-series physiological data. However, in practical applications, accurately identifying outliers from this continuously generated time-series data remains a challenging problem for medical personnel and technology developers. Traditional methods for identifying abnormal physiological indicators mostly rely on isolated analysis of data collected from a single monitoring node, failing to fully consider the temporal correlation and continuity of time-series data. For example, some methods simply set fixed thresholds to determine whether data is abnormal; when a physiological indicator value exceeds the preset threshold range at a certain moment, it is directly identified as abnormal data. However, the physiological indicators of patients with chronic diseases often exhibit fluctuations, and the normal range of indicators may differ among different patients at different times. Relying solely on fixed thresholds for judgment can easily misclassify indicator data within the normal fluctuation range as abnormal, and may also miss some slightly exceeding thresholds but potentially risky abnormal data, resulting in low accuracy in anomaly identification. Current technologies for outlier identification typically analyze only the data of individual patients, neglecting the reference value of physiological indicators from other patients with similar disease characteristics. The onset and progression of chronic diseases exhibit certain group characteristics; patients with the same type and duration of chronic diseases often show common trends in their physiological indicators. For example, patients with type 2 diabetes for 5-10 years may show similar upward trends in blood glucose levels when dietary control is inadequate. If the physiological indicators of these patients with similar disease characteristics are not considered during outlier identification, and judgment is made solely based on the historical data of a single patient, it is difficult to comprehensively understand the abnormalities in the patient's physiological indicators. This is especially true for newly diagnosed patients or those with limited historical monitoring data, where the lack of sufficient personal data for comparison significantly reduces the reliability of outlier identification. Existing technologies typically present anomaly indicators in simple numerical or textual form. Medical personnel must then analyze these indicators in conjunction with the patient's monitoring time, historical data, and other information to determine the specific circumstances and potential risks associated with the anomalies. This presentation method not only increases the workload of medical staff but also makes the anomaly indicators less intuitive, hindering their ability to quickly and accurately grasp the patient's health status and consequently impacting the timeliness and effectiveness of subsequent diagnostic and treatment decisions. Summary of the Invention

[0003] The purpose of this invention is to provide a method for identifying outlier data in an automated chronic disease monitoring network, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides a method for identifying outliers in data from an automated chronic disease monitoring network, the method comprising: At multiple monitoring nodes in the chronic disease surveillance network, time-series data sets of physiological indicators corresponding to target patients are acquired; Based on the collection time sequence and number of data points of the physiological indicator time series data set, the time interval of the physiological indicator time series data set is divided to obtain multiple sub-time series data segments; Associate the first physiological indicator feature with the first patient node that has the same disease characteristics as the target patient; The abnormal identification data generated based on the multiple sub-time series data segments and the first physiological indicator features is output on the preset monitoring interface; The abnormality identification data represents the probability of abnormality in the target patient's physiological indicators.

[0005] Preferably, the step of dividing the time interval of the physiological indicator time series data set into multiple sub-time series data segments based on the collection time sequence and the number of data points includes: Based on the acquisition time sequence and the number of data points, the cumulative time of each sub-time series data segment is recorded based on the time interval to determine the initial time interval when there is no time gap in each sub-time series data segment. When the total number of data points in each of the sub-time series data segments is less than the expected number of data points in the time interval, the physiological indicator characteristics of adjacent sub-time series data segments are compared, and abnormal sub-data segment groups with time gaps are marked. The predicted time gap value of the abnormal sub-data segment group is determined based on the characteristic difference between the physiological indicator features; The predicted time gap value is inserted into the initial time interval corresponding to the abnormal sub-data segment group, and the cumulative time of each sub-time series data segment is recalculated to obtain the time interval of each sub-time series data segment.

[0006] Preferably, the step of comparing the physiological indicator characteristics of adjacent sub-time series data segments and marking abnormal sub-data segment groups with time gaps includes: Extract the physiological indicator change trend characteristics of adjacent sub-time series data segments; Obtain the peak points and trough points in the trend characteristics of each physiological indicator, and calculate the difference in the number of peak points and the difference in the number of trough points in the trend characteristics of each physiological indicator. When the difference in the number of peak points is greater than the first quantity threshold or the difference in the number of trough points is greater than the second quantity threshold, it is determined that there is a time gap in the abnormal sub-data segment group and it is marked. When the difference in the number of peak points is less than the first quantity threshold and the difference in the number of trough points is less than the second quantity threshold, obtain the set of matching points of the nearest peak points and trough points in the trend characteristics of each physiological indicator; The total distance value of each set of matching points is calculated. When the total distance value is greater than the trend fit threshold, it is determined that there is a time gap in the abnormal sub-data segment group and it is marked.

[0007] Preferably, the method further includes: Identify the blank time intervals in the chronic disease surveillance network where data is missing; Select the second sub-time series data segment corresponding to the second target patient node adjacent to the blank time interval as the reference data segment; Calculate the mean value of the data corresponding to the same physiological index type in each of the aforementioned reference data segments; The mean value of the data is determined to be the fill value of the corresponding physiological indicator type within the blank time interval, and the filled time series data of the blank time interval is generated based on the fill value.

[0008] Preferably, the first physiological indicator feature associated with the first patient node that has the same disease characteristics as the target patient includes: Construct a first feature map corresponding to the target patient, wherein the first feature map includes the identification features of the target patient and the physiological indicator features corresponding to the multiple sub-time series data segments; Obtain a second feature map corresponding to the first patient node, wherein the second feature map includes the identification features of the first patient node and the first physiological indicator features; Merge the first feature map and the second feature map to generate a comprehensive feature path map; Based on a preset temporal correlation strategy, physiological indicator features are selected in the comprehensive feature path map to obtain a set of key physiological indicator features.

[0009] Preferably, the step of selecting physiological indicator features from the comprehensive feature path map according to a preset temporal association strategy to obtain a key physiological indicator feature set includes: Based on the feature node transition probability defined by the preset temporal association strategy, the connected physiological indicator feature nodes are traversed in the comprehensive feature path map. Based on the temporal correlation between the physiological indicator feature nodes, target physiological indicator feature nodes that meet the preset correlation strength threshold are sampled. The physiological indicator features corresponding to the target physiological indicator feature nodes are aggregated to form a key physiological indicator feature set.

[0010] Preferably, the step of outputting abnormal identification data generated based on the plurality of sub-time series data segments and the first physiological indicator features on the preset monitoring interface includes: The three anomaly probability dimension data corresponding to the multiple sub-time series data segments are respectively converted into operation dimension data; The three anomaly probability dimensions include time offset dimension data, indicator deviation dimension data, and trend deviation dimension data. The operational dimension data includes time-dimensional data, intensity-dimensional data, and duration-dimensional data. Target time dimension representation data is generated based on the time offset dimension data transformation; Based on the deviation dimension data of the aforementioned indicators, target intensity dimension representation data is generated; Based on the aforementioned trend deviation dimension data transformation, target persistence dimension representation data is generated; The anomaly identification data is generated based on the target time dimension representation data, target intensity dimension representation data, and target duration dimension representation data.

[0011] Preferably, the larger the time offset value in the time offset dimension data, the lower the trend frequency in the corresponding target time dimension data. The larger the deviation value of the indicator in the indicator deviation dimension data, the higher the intensity value in the corresponding transformed target intensity dimension data. The faster the rate of decline in the trend deviation dimension data, the shorter the duration in the corresponding target persistence dimension data.

[0012] Preferably, the method further includes: The historical physiological index fluctuation range of the target patient was tested to obtain the patient data fluctuation range test results. Based on the test results of the patient data fluctuation range, the time interval length of the multiple sub-time series data segments is adjusted to obtain the time interval granularity adjustment result; The time intervals of the physiological indicator time series data set are redefined based on the time interval granularity adjustment results.

[0013] Preferably, the chronic disease monitoring network includes multiple monitoring nodes distributed across different parts of the body; The step of generating the anomaly identification data based on the target time dimension representation data, the target intensity dimension representation data, and the target duration dimension representation data includes: Generate independent anomaly identification data streams for each of the multiple monitoring nodes; The independent anomaly identification data streams corresponding to the multiple monitoring nodes are output synchronously to form comprehensive anomaly identification data.

[0014] Compared with the prior art, the beneficial effects of the present invention are: In the outlier identification process, the time-series data of physiological indicators of target patients acquired from multiple monitoring nodes in the chronic disease surveillance network are first divided into time intervals based on the collection time sequence and the number of data points, resulting in multiple sub-time-series data segments. This division method based on time dimension and data volume can fully consider the continuity and stage characteristics of physiological indicator time-series data, avoiding the problem of traditional methods that analyze time-series data as a whole and ignore the changing patterns of indicators within local time periods due to the large data span. By decomposing the overall time-series data into multiple sub-time-series data segments, the fluctuations of physiological indicators within each time period can be captured more meticulously, allowing subsequent anomaly analysis to focus on specific time intervals, thereby more accurately identifying potential anomalies and reducing the possibility of anomaly omissions or misjudgments caused by overall data analysis. After associating the first physiological indicator characteristics with the first patient node that shares the same disease characteristics as the target patient, more dimensions of reference are provided for outlier identification of the target patient. Since physiological indicator changes in chronic disease patients exhibit group-wide commonalities, the physiological indicator characteristics of patients with the same disease characteristics can reflect the typical indicator change patterns of this type of disease at a specific stage. Combining these characteristics with the target patient's sub-time series data segments for analysis can effectively compensate for the limitations of relying solely on the target patient's own data for anomaly judgment. For patients with limited historical monitoring data, the physiological indicator characteristics of patients with the same disease characteristics can serve as an important reference standard, helping to refine the basis for anomaly judgment; for patients with some historical data, the indicator characteristics of patients with the same disease characteristics can further verify the rationality of their own data anomalies, reducing misjudgments caused by individual data fluctuations, and making the anomaly identification results more reliable and comprehensive. The system outputs abnormal identification data based on multiple sub-time series data segments and primary physiological indicator characteristics on a preset monitoring interface. This abnormal identification data directly represents the probability of abnormality in the target patient's physiological indicators, offering a highly intuitive presentation. Medical personnel no longer need to manually integrate scattered monitoring data and reference information; they can directly obtain the abnormal probability corresponding to each sub-time series data segment through the preset monitoring interface, enabling them to quickly understand the degree of abnormal risk of the target patient's physiological indicators at different time periods. Furthermore, the quantitative presentation of abnormal probabilities allows medical personnel to clearly compare abnormalities across different time periods, accurately grasp the severity and trends of abnormal data, and thus more efficiently assess the patient's health status, reducing the time cost of information integration and analysis, and creating favorable conditions for timely implementation of appropriate medical interventions. This method fully leverages the multi-node data acquisition advantages of chronic disease surveillance networks. By integrating physiological indicator data from multiple monitoring nodes, it avoids the impact of errors or data gaps that may occur in single-node data acquisition on anomaly identification results. The fusion of multi-node data enables the construction of a more complete and comprehensive physiological indicator data system for target patients, providing a more solid data foundation for the division of sub-time series data segments and the correlation of physiological indicator characteristics among patients with the same disease features, further improving the accuracy of anomaly identification. Simultaneously, the entire operational process can be implemented using existing hardware facilities of chronic disease surveillance networks, eliminating the need for additional complex equipment systems. This demonstrates strong feasibility in practical application and can be widely applied to chronic disease surveillance scenarios of different scales and types, providing more scientific and effective technical support for the health management of chronic disease patients. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the working principle of the data anomaly identification method for the automatic chronic disease monitoring network described in this invention. Figure 2 A flowchart for dividing time intervals and handling time gaps; Figure 3 Flowchart for filling blank time intervals; Figure 4 A flowchart for linking patient characteristics with the same disease features; Figure 5 This is a flowchart for adjusting the granularity of the time interval. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a method for identifying outliers in data from an automated chronic disease monitoring network, the method comprising: At multiple monitoring nodes in the chronic disease surveillance network, wearable devices or implantable sensors acquire time-series data sets of physiological indicators corresponding to the target patient, including multiple physiological parameters such as blood pressure, blood glucose, and heart rate. The time-series data sets are stored in chronological order of collection, with each data point containing a timestamp and measurement value. Based on the collection time order and the number of data points, a sliding window algorithm is used to divide the time-series data sets into multiple consecutive sub-time-series data segments, each corresponding to a preset time interval. Data mining algorithms are used to retrieve the first patient node in the chronic disease surveillance network that shares the same disease characteristics as the target patient, and its corresponding first physiological indicator feature is extracted. The multiple sub-time-series data segments of the target patient and the first physiological indicator feature are input into an anomaly detection model to calculate the anomaly probability of the physiological indicators within each time interval. Finally, the anomaly identification data is output in the form of visual charts on a preset monitoring interface, including anomaly probability curves on the time axis and anomaly level markers.

[0018] Example 1: See Figure 2 In automated chronic disease monitoring network systems, the integrity of time-series data directly impacts the accuracy of anomaly detection. This embodiment addresses the potential collection discontinuity issue in physiological indicator time-series data sets by proposing a time gap detection and repair method based on feature comparison. The system first receives raw physiological indicator data streams from multiple monitoring nodes. These data are stored in a circular buffer in timestamp order, forming a continuous time-series data set. Each data point contains a timestamp accurate to milliseconds and a calibrated physiological indicator value.

[0019] In the initial time interval segmentation phase, the system employs a sliding segmentation algorithm with an adaptive window size. The algorithm calculates the standard number of data points required for each time interval based on a preset base time unit (e.g., 5 minutes) and data acquisition frequency (e.g., once per minute). For example, a 5-minute interval should contain 5 data points at the standard acquisition frequency. The system traverses the entire time-series dataset in chronological order, performing cumulative time calculations for each consecutive data segment. When the actual number of data points within a time interval reaches more than 90% of the standard value, the interval is marked as a complete data segment, and its start and end timestamps are recorded. When the system detects that the number of data points in a sub-time-series data segment is significantly insufficient (below 80% of the standard value), a time gap detection process is initiated. The system extracts statistical characteristics of the data segment and its two preceding and following adjacent data segments, including key indicators such as moving average, standard deviation, and autocorrelation coefficient. These characteristics reflect the basic changing patterns and short-term fluctuations of physiological indicators. By comparing the characteristic differences between adjacent data segments, the system can determine whether there are any interruptions in data acquisition.

[0020] The system calculates the relative differences between adjacent data segments at the mean level. If the difference exceeds twice the standard deviation of the historical fluctuation range, further analysis is triggered. The system then examines the rate of change of variance between data segments; a sudden change in variance exceeding a preset threshold indicates potential data gaps. The comparison of autocorrelation coefficients is used to assess the consistency of data patterns; a significantly decreased autocorrelation coefficient suggests a possible time gap. For abnormal data segment groups marked as potentially having time gaps, the system initiates a predictive repair mechanism. This mechanism analyzes the overall trend of data changes before and after the gap, considering factors such as circadian rhythms and individual patient characteristics. By establishing regression models for the data segments before and after the gap, the system predicts possible data change paths during the missing period. Based on the prediction results, the system inserts virtual data points at corresponding positions on the timeline; these data points are specially marked to distinguish them from actual measurements.

[0021] The prediction of the time gap length employs an estimation algorithm based on feature changes to analyze characteristics such as slope changes and extreme point distributions in the data before and after the gap. Combined with historical data patterns of this type of patient, the system estimates the most likely duration of data loss. For example, when a jump in the mean level of the preceding and following data segments is detected, but the trend remains continuous, the system determines that there has been a short-term data acquisition interruption; conversely, when the preceding and following data segments exhibit completely different fluctuation patterns, it may correspond to a longer period of device offline.

[0022] After predicting the time gap, the system recalculates the time interval boundaries for each sub-time series data segment. The adjustment process maintains the original temporal order of the data, only marking and compensating for gaps detected. The new time interval division ensures complete temporal coverage for each segment while preserving the true acquisition time information of the original data. This approach guarantees the temporal continuity of subsequent analysis while avoiding distortion caused by data padding. The system also employs a multi-level verification mechanism to ensure the accuracy of time gap detection. Primary verification confirms the specific time of acquisition interruption by checking device logs; intermediate verification compares the data integrity of multiple relevant physiological indicators; and advanced verification references data patterns from other patients with similar disease characteristics during the same period. When verification results at each level contradict each other, the system automatically selects the most conservative estimation scheme to avoid error accumulation caused by over-padding.

[0023] In terms of implementation details, the system adopts a layered processing architecture. The bottom-level data processing module is responsible for receiving and buffering raw data; the intermediate analysis module performs time gap detection and feature comparison; and the upper-level decision module completes the final time interval adjustment. The modules communicate asynchronously through message queues to ensure the system's responsiveness when processing large-scale data. The time gap marking information is stored in a dedicated metadata area, maintaining association with the raw data but stored independently for easy subsequent tracing and correction. Different processing parameters are set for different types of physiological indicators. For indicators with large fluctuations, such as instantaneous heart rate, a shorter time window and a more lenient gap detection threshold are used; while for relatively stable indicators, such as body temperature, a longer time window and stricter judgment conditions are used. This differentiated processing can adapt to the unique variation patterns of different physiological indicators and improve the accuracy of gap detection.

[0024] During system operation, the time gap detection model is continuously updated. By recording the deviation between each gap prediction and the actual equipment logs, the weight parameters of feature comparison are automatically adjusted. This self-learning mechanism enables the system to gradually adapt to the individual characteristics of specific patients and the characteristics of different monitoring devices, continuously improving the accuracy of time gap identification. All adjustment records are meticulously recorded in the system log for quality control and algorithm optimization.

[0025] At the user interface level, the system provides a time gap visualization tool. Healthcare professionals can view a complete timeline display, with actual measurement data and predicted filling data highlighted in different colors. The interface supports drill-down functionality, allowing users to view the detailed rationale and adjustment process for each gap detection. This transparent presentation helps healthcare professionals understand the system's decision-making process and allows for manual intervention when necessary.

[0026] Example 2: See Figure 3In the actual operation of chronic disease monitoring networks, waveform feature analysis of physiological indicator data is a crucial step in identifying data anomalies and time gaps. This system employs a multi-level waveform feature comparison method to achieve accurate detection and intelligent repair of data integrity. When a monitoring node collects a patient's blood pressure data, the raw signal first undergoes filtering and noise reduction by a preprocessing module to eliminate noise components caused by motion artifacts or equipment interference.

[0027] The system scans the entire time-series data using a sliding window approach, identifying potential peaks and troughs within each window. The window size is dynamically adjusted based on the signal sampling rate; for example, a short window of 1 second is used for ECG data sampled 60 times per minute, while a longer window of 15 minutes is used for blood glucose data sampled once per hour. Each extreme point must meet amplitude and duration conditions to avoid misclassifying short-term noise as valid extreme values. Identified peaks and troughs are recorded in a feature point list, along with their time location and amplitude value.

[0028] Waveform comparison between adjacent data segments is performed in three dimensions: the number of extreme points, the distribution of extreme points, and the waveform morphology. The system first counts the number of peaks and troughs in two data segments and calculates their absolute difference. Considering normal fluctuations under different physiological conditions, the system does not simply compare the numerical difference, but rather standardizes it against the typical fluctuation range of similar patients. For example, the allowable difference in the number of peaks is smaller for heart rate data at rest; while during exercise recovery, the criteria for determining the numerical difference are relaxed. This dynamic threshold mechanism avoids false alarms caused by oversensitivity.

[0029] When the difference in the number of extreme points does not exceed a threshold, the system transitions to a more refined waveform matching analysis. The peaks and troughs that are closest in time from the two data segments are selected to form matching point pairs. The composite distance between each pair of extreme points is calculated in both time and amplitude, taking into account the combined effects of time offset and amplitude variation. The system performs a weighted summation of the distance values ​​for all matching point pairs, with the weighting coefficients dynamically adjusted based on the significance of the extreme points, giving higher weight to major peaks than minor fluctuations. The summation result reflects the overall fit between the two waveforms; values ​​below the threshold are considered continuous data, while values ​​above the threshold indicate a potential time gap.

[0030] The system analyzes the waveform changes before and after the gap, including periodic characteristics, amplitude variation rate, and phase information. For indicators with obvious periodicity, such as respiratory rate, a periodic extension method is used to predict the missing waveform; for trend indicators, such as blood glucose changes, linear prediction is performed based on historical slopes. The prediction process considers the influence of circadian rhythms; for example, physiological parameter fluctuations are usually more gradual at night than during the day. The generated predicted waveforms are smoothly connected to the actual data before and after, maintaining a natural waveform transition.

[0031] For completely blank time intervals in the monitoring network, the system initiates a data reconstruction mechanism. This reconstruction process is not a simple mean-filling, but rather a comprehensive extrapolation based on multi-source information. First, the system retrieves concurrent data for the same patient at other monitoring nodes, identifying relevant indicators for reference. For example, when blood pressure data is missing, heart rate and activity levels from the same period are referenced. Second, it queries monitoring data from other patients with similar disease characteristics during the same time period to establish a population reference model. Finally, it combines the patient's individual historical data patterns to generate the most probable physiological parameter change curves.

[0032] Data reconstruction employs a layered generation strategy. The base layer establishes the overall trend of change, reflecting the macroscopic direction of physiological indicators; the intermediate layer adds periodic fluctuations to simulate normal physiological rhythms; and the detail layer introduces reasonable random variations to maintain the natural characteristics of the data. The reconstruction results undergo multiple verifications: time-series verification checks the rationality of the time distribution of data points, range verification ensures that the values ​​are within the physiologically possible range, and trend verification confirms that the rate of change conforms to medical common sense. Data that fails verification is regenerated until all constraints are met. When extreme outliers are detected, they are not immediately excluded but their context is analyzed. Certain real clinical events, such as arrhythmias, do produce abnormal waveforms, and these medically significant abnormalities need to be retained. The system distinguishes between real pathological events and acquisition anomalies by comparing changes in multiple related indicators. For example, a sudden drop in heart rate alone may be due to equipment malfunction, but if it is accompanied by a drop in blood pressure and a decrease in blood oxygen, it is more likely a real clinical event.

[0033] The system continuously monitors and analyzes the reliability indicators of the results. When a large number of suspected gaps are found to be false alarms after manual review, the system automatically raises the judgment threshold; conversely, it lowers the threshold to improve sensitivity. This adaptive mechanism allows the system to adapt to different monitoring environments and patient groups. All parameter adjustments are recorded in the configuration log, forming a complete traceability chain. Medical staff can view waveform overlays of adjacent data segments side-by-side to intuitively understand the basis for the system's gap judgment. The interface supports annotation and commentary on automatically filled data, allowing medical personnel to add clinical background information. This human feedback is collected by the system and used to optimize subsequent analysis algorithms.

[0034] The data storage employs an architecture that separates raw and derived data. Raw measurement data is stored in read-only mode, while all filled and repaired data is stored in a separate derived area, with detailed records of the reasons and methods used for processing. This design ensures the integrity of the raw data while providing appropriate data versions for different application scenarios. The data access interface provides corresponding data views based on the user's role. Researchers can obtain complete data containing all filled markers, while the clinical alarm system uses a validated continuous data stream.

[0035] During system operation, the waveform feature analysis module undergoes continuous performance monitoring, recording operational metrics such as processing time, resource consumption, and result consistency for each gap detection. When performance degradation or abnormal results are detected, a diagnostic process is automatically triggered to check hardware resources, data quality, and algorithm status. The diagnostic results generate maintenance suggestions, indicating possible manual intervention measures.

[0036] Diabetes monitoring focuses on trend analysis of blood glucose changes, using longer time windows and trend-sensitive feature extraction parameters; cardiovascular disease monitoring focuses on instantaneous heart rate variability, employing short time windows and high temporal resolution analysis methods. These specialized configuration templates can be combined and applied according to the patient's specific diagnosis to provide personalized monitoring solutions. The entire waveform feature analysis process emphasizes a balance between medical rationality and engineering practicality. The system does not pursue a perfect mathematical fit, but rather ensures that the generated filled data conforms to actual clinical conditions. All automated processing retains the option for manual review and correction, forming a human-machine collaborative workflow. This design philosophy allows the system to possess both the efficiency advantages of intelligent processing and maintain necessary professional medical supervision.

[0037] Example 3: See Figure 4 In chronic disease surveillance networks, the construction and association of feature maps are the core technical path to achieve accurate anomaly detection. During system initialization, a multidimensional feature map is created for the target patient, represented using an attribute graph structure. Nodes in the graph are divided into two categories: entity nodes and feature nodes. Entity nodes contain patient identification information (medical record number, age group, gender identifier, confirmed disease code, etc.), while feature nodes correspond to the time-series characteristics of physiological indicators (systolic blood pressure trend vector, blood glucose change matrix, etc.). Semantic associations are established between nodes through directed edges, such as "patient-suffers from-diabetes" or "blood glucose characteristic-impact-systolic blood pressure characteristic". Each edge is accompanied by a time decay weight factor, reflecting the time-dependent strength of the association.

[0038] When the system retrieves patient nodes with the same disease characteristics, it employs a semantic matching algorithm based on medical ontology. The algorithm first parses the disease characteristic description of the target patient and maps it to node positions in a standard medical terminology tree. While traversing the monitoring network database, it calculates the semantic distance between candidate patients and the ontology, selecting patients within a distance threshold as similar nodes. For each selected patient with the same disease characteristics, the system extracts their three-month time-series physiological indicator data, converting it into a standardized feature vector set through feature engineering.

[0039] The graph merging process employs a layered fusion strategy. The first layer performs node alignment, matching nodes representing the same semantic meaning from different graphs (e.g., the "morning fasting blood glucose" node). The second layer processes unmatched nodes, determining whether to create a new node or merge it into an existing one based on node attribute similarity. The third layer performs relationship fusion, weighting multiple edges between identical nodes. The resulting comprehensive feature path graph contains a cross-patient feature association network, where the edge weights are calculated using the following formula: in: This represents the fusion weight from node i to j; This represents the number of associated paths; It is the confidence coefficient of the k-th original path (0.8~1.2); The time decay function (the highest value is taken for the nearest correlation); Minimum time factor (default 30 days); The time span (in days) is used for association. This formula ensures that recent high-frequency associations receive higher weights.

[0040] Establish the transition probability matrix, matrix elements The transition probability from feature node u to v is determined by three factors: temporal proximity (the reciprocal of the time interval between the occurrences of the two features), numerical correlation (Pearson correlation coefficient), and semantic similarity (path distance in the medical ontology). Next, a random walk with restart is performed, starting from the target patient's current feature node and jumping back to the starting point with a probability of 0.15 to avoid deviating from the core region. During the walk, the frequency of node visits is recorded to form a ranking of feature node importance.

[0041] Primary screening retains nodes with a metastasis probability exceeding 0.6, while secondary screening requires a temporal correlation (normalized value of dynamic time-normalized distance) of less than 0.3. For nodes that pass screening, the system calculates their mutual information value with core features, removing weakly correlated nodes with mutual information below 0.05. The final target node set must meet the coverage criterion: that is, it must be able to explain more than 80% of the physiological variation in the target patient.

[0042] The feature aggregation stage employs attention-weighted graph convolution operations. For each target node, the system calculates a weighted sum of the feature vectors of its neighboring nodes, with the weights determined by the attention score. The attention score is calculated using a three-layer perceptron, with input including node feature differences, time intervals, and relation type encodings. The aggregated feature vectors are transformed by a non-linear activation function to form a 128-dimensional feature representation. The feature representations of all target nodes constitute a key physiological indicator feature set, which preserves the spatiotemporal correlation characteristics of the original features.

[0043] The system implements a dynamic atlas update mechanism during operation, automatically performing incremental updates to the atlas every morning at midnight: newly added monitoring data triggers feature node updates, and expired associations decay over time. When a change in a patient's clinical status is detected (such as the onset of new complications), the system initiates an atlas reconstruction process, recalculating the set of patients with the same disease characteristics and feature association paths. Atlas version management uses snapshot technology, retaining historical versions from the most recent seven days, supporting retrospective analysis of abnormal detection results.

[0044] In terms of computational optimization, the system employs a partitioned parallel processing strategy, dividing the comprehensive feature path map into multiple subgraphs according to disease type and distributing them across different computing nodes. For large-scale maps, sparsification is implemented, pruning edges with weights below 0.1 and isolated nodes with degrees less than 3. Frequently accessed subgraphs are cached in an in-memory database, with cache replacement managed using the LRU algorithm.

[0045] The system incorporates a medical rule engine that periodically scans the atlas for contradictory relationships (such as conflicting nodes between "diabetic patients" and "normal insulin secretion"). Detected contradictory relationships trigger a manual review process, where clinicians confirm or correct the relationships through a review interface. All corrections are recorded in the audit log for subsequent optimization of the automated algorithm.

[0046] Node positions are automatically arranged based on association strength, with different types of nodes using differentiated color coding (red for disease nodes, blue for physiological indicator nodes). The interface supports timeline filtering, allowing users to view the feature association status within a specific time period. Healthcare professionals can explore implicit paths between feature nodes by dragging and dropping, and the system displays the confidence score and clinical interpretation of the paths in real time.

[0047] The system monitors real-time changes in key features. When three or more key features simultaneously deviate from the normal range, a multi-feature collaborative early warning is triggered. The warning level is dynamically calculated based on the degree of feature deviation and clinical importance, prioritizing notifications for feature abnormalities directly related to the core disease. All warning events are displayed in conjunction with the associated paths in the feature map, helping medical staff quickly pinpoint the root cause of the abnormality.

[0048] The construction of the feature map follows the disease relationships defined by medical ontology while incorporating statistical associations from actual monitoring data. Key feature selection balances temporal relevance and clinical interpretability, ensuring that the generated anomaly markers are both consistent with data patterns and medically meaningful. The system continuously learns from feedback from medical personnel, gradually optimizing the feature association model to form a closed-loop knowledge evolution mechanism.

[0049] Example 4: In the anomaly identification generation stage of the chronic disease monitoring system, multidimensional data transformation and visualization constitute the core technology module. The system receives three anomaly probability dimensions from the preprocessing flow: time offset dimension data reflects the periodic anomalies in physiological indicator fluctuations, indicator deviation dimension data characterizes the degree to which measured values ​​deviate from the baseline, and trend deviation dimension data describes the unconventionality of the rate of change. These raw dimensional data are transformed into operational dimensional data through specific transformation rules, ultimately forming an anomaly identification that can be intuitively interpreted.

[0050] The system calculates the cross-correlation function between the target patient's physiological indicator sequence and a reference sequence from similar patients, determining the time lag value corresponding to the maximum correlation coefficient as the original time offset. This offset is input into a nonlinear converter, with the conversion rules set as follows: when the lag value is in the 0-5 minute range, a high-frequency trend is output (fluctuation once per minute); the 5-15 minute range is converted to a medium-frequency trend (fluctuation once every 3 minutes); and values ​​exceeding 15 minutes are mapped to a low-frequency trend (fluctuation once every 10 minutes). The converted target time dimension indicates that the data directly reflects the frequency characteristics of abnormal fluctuations, with high-frequency anomalies requiring urgent attention.

[0051] The processing of deviation dimension data involves two stages: standardization and normalization. The original deviation value is calculated using Mahalanobis distance, representing the degree to which the current measurement deviates from the patient's individual baseline distribution. The system normalizes this distance value by dividing it by the historical maximum deviation value, obtaining the relative deviation coefficient in the 0-1 interval. The converter employs a piecewise linear interpolation strategy: coefficients 0-0.3 map to intensity values ​​0-30; 0.3-0.6 correspond to 30-70; and values ​​above 0.6 map to 70-100. The generated target intensity dimension represents a unified scale for quantifying the severity of abnormalities in the data.

[0052] The system performs a linear fit on the physiological indicator sequence within the current time window, using the absolute value of the slope as the original trend rate of change. A conversion rule establishes an inverse relationship between rate and duration: a rate of 0-0.1 units / minute corresponds to a duration of 120 minutes; 0.1-0.2 corresponds to 60 minutes; 0.2-0.5 corresponds to 30 minutes; and values ​​exceeding 0.5 are mapped to 15 minutes. This non-linear mapping ensures that rapidly deteriorating trends receive more urgent short-term warnings (see Table 1).

[0053] Table 1: Data Transformation Mapping Table for Anomaly Dimensions The system automatically adjusts the weighting coefficients of each dimension based on the patient's current clinical status: for patients in the stable phase, the time dimension is emphasized (weight 0.5); for patients in the acute phase, the intensity dimension is emphasized (weight 0.6); and for patients in the recovery phase, the duration dimension is emphasized (weight 0.4). The weighted calculation results generate a comprehensive abnormal score from 0 to 100. A score exceeding 70 triggers a yellow warning, and a score exceeding 85 is upgraded to a red alert.

[0054] The user interface establishes a spherical coordinate system: the radial axis represents the target intensity dimension data, the angle axis corresponds to the target time dimension data, and the height axis maps to the target duration dimension data. Each anomalous event is represented as a colored sphere in the coordinate system, with the sphere's diameter expanding as the anomalous score increases. The system has a built-in automatic view cruise function; when a high-risk anomaly is detected, it automatically focuses on the corresponding sphere and enhances the warning with a pulsating light effect.

[0055] When a user selects a specific abnormal sphere, the system automatically expands the associated panel: the left side displays the original physiological indicator curves, marking the location of the abnormality; the middle section presents the reverse mapping path from the three-dimensional coordinates to the original dimensions; and the right side lists the historical treatment plans for similar patients. A timeline zoom control is provided at the bottom of the panel, allowing users to view changes in associated indicators before and after the abnormal event.

[0056] Anomalies with a comprehensive score of 70-85 are recorded internally by the system and displayed as an amber marker on the monitoring interface; anomalies with a score of 85-95 are pushed to the responsible nurse via instant message; anomalies with a score of 95 or above simultaneously activate a voice alarm and connect to the on-duty doctor's terminal. All warning events generate structured logs, including detailed values ​​before and after 3D data conversion and decision path traceability codes.

[0057] The system projects abnormal events from a patient's past three months onto a spherical coordinate system, creating a spatiotemporal distribution cloud map. The cloud map density coloring function identifies high-frequency abnormal areas, while the heatmap layer displays periods of abnormal clustering. Medical staff can select areas in the cloud map to access analysis reports for similar abnormalities in batches.

[0058] The medical team can adjust mapping parameters based on clinical experience, such as adjusting the blood glucose trend deviation threshold for diabetic patients from 0.2 to 0.15 units / minute. All custom rules must be reviewed by the medical safety verification engine to avoid misjudgments due to conflicting settings. Modified rules automatically generate version branches for differentiated application in different wards. It is compatible with various devices from mobile terminals to the command center's large screen. The touch interface supports gesture operations: two-finger rotation adjusts the viewing angle, pinching zooms to switch display granularity, and long-pressing a sphere brings up a details card. The command center's large screen mode additionally provides a holographic projection option, projecting a 3D model of an abnormal event into physical space for multi-angle consultation.

[0059] The raw 3D data is stored in a time-series database, the transformation rules are stored in a configuration library, and the visualization status information is cached in an in-memory database. This architecture ensures that core data processing performance is not affected during frequent interface interactions. The system automatically generates an interface snapshot every five minutes and stores it as an audit trail, supporting post-event review and analysis.

[0060] Example 5: See Figure 5The dynamic optimization mechanism of the chronic disease surveillance network improves monitoring accuracy by continuously adapting to individual patient characteristics. During system initialization, a basic monitoring frequency is established, and data is segmented using a uniform time window length for all physiological indicators. During operation, a background analysis thread is initiated to extract fluctuation characteristics from the historical physiological indicator data of the target patients. This analysis employs a sliding window statistical method, with a configurable window length from 24 hours to 30 days, calculating the coefficient of variation and range ratio for each physiological indicator within the corresponding time period. The coefficient of variation reflects the degree of data dispersion, and the range ratio characterizes the amplitude of fluctuation; the combination of these two factors forms a data stability profile for each individual patient.

[0061] The system sets volatility level thresholds: when the coefficient of variation is below 0.05 and the range ratio is less than 0.1, it is judged as a low volatility state, and the time interval of the sub-time series data segment is automatically extended to 150% of the base value; when the coefficient of variation is between 0.05 and 0.15 or the range ratio is between 0.1 and 0.3, the base time window is maintained; when any indicator exceeds the upper limit threshold, the time window is shortened to 80% of the base value. The adjustment process adopts a gradual change strategy, with each change not exceeding 20% ​​of the current window, to avoid sudden changes in the monitoring rhythm affecting trend analysis.

[0062] The system maintains a global time window configuration table, recording the optimal window parameters for each patient's physiological indicators. When the configuration is updated, relevant monitoring nodes are notified via a publish-subscribe mechanism. After receiving the new configuration, the nodes apply the new window settings at the start of the next complete monitoring cycle. During the transition period, dual-window parallel processing is used: the old window completes the current data analysis, and the new window begins collecting data for the next cycle, ensuring data continuity is not affected by configuration switching.

[0063] For monitoring nodes distributed across different body parts, the system implements a distributed anomaly detection architecture. Each monitoring node deploys a lightweight analysis engine to independently process the physiological indicator data collected from its own location. The analysis engine comprises three processing units: a signal preprocessing unit for real-time filtering and noise reduction, a feature extraction unit for calculating statistical indicators and waveform features, and an anomaly scoring unit for generating independent anomaly probability values ​​based on site characteristics. Each node outputs an independent anomaly identification data stream once per second, with the data packet containing a timestamp, node number, anomaly score, and confidence index.

[0064] The system monitors the deployment of a time synchronization server on the network, coordinating the clocks of each node through a precise time protocol. Each abnormal data packet is appended with a high-precision timestamp, with errors controlled to the millisecond level. The central processing unit establishes a data receiving buffer, aligning and matching data packets from multiple nodes within the same time window. The time alignment algorithm allows for a time tolerance of ±100 milliseconds; data packets exceeding this tolerance are entered into an asynchronous processing queue.

[0065] The central processing unit performs spatiotemporal correlation analysis on the independent abnormal data streams reported by each node: First, it detects the temporal correlation of abnormal events in different parts and marks multiple abnormalities in different parts that occur within 5 consecutive seconds as collaborative events; second, it analyzes the spatial distribution pattern of abnormal intensity and identifies abnormal hot spots; finally, it calculates a comprehensive index of cross-part abnormalities, which is composed of the weighted sum of abnormal scores in each part, with the weight coefficients set according to the clinical importance of the part.

[0066] The main view displays a full-body schematic, with each monitoring node marked by a color-coded dot; the color depth reflects the real-time anomaly score. Clicking on a node enters a partial view, displaying detailed physiological parameter curves and anomaly probability time-series graphs for that area. When a collaborative event is triggered, a panoramic analysis panel automatically pops up, displaying the anomaly data streams of related areas side-by-side, with spatiotemporal correlation paths marked by connecting lines. A timeline zoom control is provided at the bottom of the interface, supporting adjustment of observation granularity from seconds to days.

[0067] The system determines the alarm level based on the comprehensive abnormality index and duration: an index of 70-85 lasting for 5 minutes triggers a primary alert, displaying an amber warning bar on the interface; an index of 85-95 lasting for 3 minutes activates an intermediate alarm, pushing a message to the mobile terminal; an index of 95 or above, or the simultaneous occurrence of high-score abnormalities in two or more areas, immediately activates an emergency alarm, triggering audible and visual alarms and establishing a remote consultation channel. All alarm events generate a comprehensive report containing multi-node data, automatically linking to the patient's electronic medical record.

[0068] The system tracks the computation latency, data quality, and resource utilization of each node in real time, automatically switching to backup analysis strategies in case of anomalies. Weekly network health reports are generated, statistically analyzing key indicators such as the frequency of time window adjustments and the success rate of multi-node synchronization. The configuration management interface allows medical personnel to manually adjust window parameters for specific patients; all manual operations are recorded in the audit log with the operator's identity clearly indicated.

[0069] Raw sensor data is stored locally on edge nodes for 7 days, feature data is uploaded to regional servers for 6 months, and anomalous event data is permanently stored in the central database. The archiving system establishes cross-layer indexes, supporting complete path queries from anomaly identifiers back to the original waveforms. Data access interfaces are configured with role-based access control, allowing researchers to access anonymized datasets for algorithm optimization.

[0070] The dynamic adjustment of the time window respects individual patient differences while maintaining the consistency of monitoring standards. Multi-node collaborative analysis takes into account both the sensitivity of local abnormalities and the overall state of the patient, transforming complex data relationships into intuitive clinical decision support through a visual interface. By continuously recording the effects of configuration adjustments and alarm response results, the system forms a closed-loop optimization mechanism, enabling the monitoring network to continuously adapt to the evolving needs of chronic disease management.

[0071] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying outlier values ​​in an automated chronic disease monitoring network, characterized in that, include: At multiple monitoring nodes in the chronic disease surveillance network, time-series data sets of physiological indicators corresponding to target patients are acquired; Based on the collection time sequence and number of data points of the physiological indicator time series data set, the time interval of the physiological indicator time series data set is divided to obtain multiple sub-time series data segments, including: Based on the acquisition time sequence and the number of data points, the cumulative time of each sub-time series data segment is recorded based on the time interval to determine the initial time interval when there is no time gap in each sub-time series data segment. When the total number of data points in each of the sub-time series data segments is less than the expected number of data points in the time interval, the physiological indicator characteristics of adjacent sub-time series data segments are compared, and abnormal sub-data segment groups with time gaps are marked. The predicted time gap value of the abnormal sub-data segment group is determined based on the characteristic difference between the physiological indicator features; Insert the predicted time gap value into the initial time interval corresponding to the abnormal sub-data segment group, and recalculate the cumulative time of each sub-time series data segment to obtain the time interval of each sub-time series data segment. The first physiological indicator feature associated with the first patient node that has the same disease characteristics as the target patient includes: Construct a first feature map corresponding to the target patient, wherein the first feature map includes the identification features of the target patient and the physiological indicator features corresponding to the multiple sub-time series data segments; Obtain a second feature map corresponding to the first patient node, wherein the second feature map includes the identification features of the first patient node and the first physiological indicator features; A hierarchical fusion strategy is adopted to merge the first feature map and the second feature map to generate a comprehensive feature path map. According to a preset temporal association strategy, physiological indicator features are selected in the comprehensive feature path map to obtain a set of key physiological indicator features. The abnormal identification data generated based on the multiple sub-time series data segments and the first physiological indicator features is output on the preset monitoring interface; Wherein, the abnormality identification data represents the probability of abnormality in the target patient's physiological indicators; The step of comparing the physiological indicator characteristics of adjacent sub-time series data segments and marking abnormal sub-data segment groups with time gaps includes: Extract the physiological indicator change trend characteristics of adjacent sub-time series data segments; Obtain the peak points and trough points in the trend characteristics of each physiological indicator, and calculate the difference in the number of peak points and the difference in the number of trough points in the trend characteristics of each physiological indicator. When the difference in the number of peak points is greater than the first quantity threshold or the difference in the number of trough points is greater than the second quantity threshold, it is determined that there is a time gap in the abnormal sub-data segment group and it is marked. When the difference in the number of peak points is less than the first quantity threshold and the difference in the number of trough points is less than the second quantity threshold, obtain the set of matching points of the nearest peak points and trough points in the trend characteristics of each physiological indicator; The total distance value of each set of matching points is calculated. When the total distance value is greater than the trend fit threshold, it is determined that there is a time gap in the abnormal sub-data segment group and it is marked. Identify the blank time intervals in the chronic disease surveillance network where data is missing; Select the second sub-time series data segment corresponding to the second target patient node adjacent to the blank time interval as the reference data segment; Calculate the mean value of the data corresponding to the same physiological index type in each of the aforementioned reference data segments; The mean value of the data is determined to be the fill value of the corresponding physiological indicator type within the blank time interval, and the filled time series data of the blank time interval is generated based on the fill value.

2. The method for identifying outlier data in an automated chronic disease monitoring network according to claim 1, characterized in that, The step involves selecting physiological indicator features from the comprehensive feature path map according to a preset temporal association strategy to obtain a set of key physiological indicator features, including: Based on the feature node transition probability defined by the preset temporal association strategy, the connected physiological indicator feature nodes are traversed in the comprehensive feature path map. Based on the temporal correlation between the physiological indicator feature nodes, target physiological indicator feature nodes that meet the preset correlation strength threshold are sampled. The physiological indicator features corresponding to the target physiological indicator feature nodes are aggregated to form a key physiological indicator feature set.

3. The method for identifying outlier data in an automated chronic disease monitoring network according to claim 2, characterized in that, The step of outputting abnormal identification data generated based on the multiple sub-time series data segments and the first physiological indicator features on the preset monitoring interface includes: The three anomaly probability dimension data corresponding to the multiple sub-time series data segments are respectively converted into operation dimension data; The three anomaly probability dimensions include time offset dimension data, indicator deviation dimension data, and trend deviation dimension data. The operational dimension data includes time-dimensional data, intensity-dimensional data, and duration-dimensional data. Target time dimension representation data is generated based on the time offset dimension data transformation; Based on the deviation dimension data of the aforementioned indicators, target intensity dimension representation data is generated; Based on the aforementioned trend deviation dimension data transformation, target persistence dimension representation data is generated; The anomaly identification data is generated based on the target time dimension representation data, target intensity dimension representation data, and target duration dimension representation data.

4. The method according to claim 3, characterized in that: The larger the time offset value in the time offset dimension data, the lower the trend frequency in the target time dimension data after conversion; The larger the deviation value of the indicator in the indicator deviation dimension data, the higher the intensity value in the corresponding transformed target intensity dimension data. The faster the rate of decline in the trend deviation dimension data, the shorter the duration in the corresponding target persistence dimension data.

5. The method for identifying outlier data in an automated chronic disease monitoring network according to claim 4, characterized in that, The method further includes: The historical physiological index fluctuation range of the target patient was tested to obtain the patient data fluctuation range test results. Based on the test results of the patient data fluctuation range, the time interval length of the multiple sub-time series data segments is adjusted to obtain the time interval granularity adjustment result; The time intervals of the physiological indicator time series data set are redefined based on the time interval granularity adjustment results.

6. The method for identifying outlier data in an automated chronic disease monitoring network according to claim 5, characterized in that, The chronic disease surveillance network includes multiple monitoring nodes distributed across different parts of the body; The step of generating the anomaly identification data based on the target time dimension representation data, the target intensity dimension representation data, and the target duration dimension representation data includes: Generate independent anomaly identification data streams for each of the multiple monitoring nodes; The independent anomaly identification data streams corresponding to the multiple monitoring nodes are output synchronously to form comprehensive anomaly identification data.

Images