An artificial intelligence-based data intelligent labeling method

Abstract

The application discloses a kind of data intelligent labeling methods based on artificial intelligence, it is related to intelligent medical technical field, including the following steps: obtaining the physiological parameter data of user, historical health parameter data and motion state data;According to the motion state data, output motion state label data by pre-constructed motion recognition model;The historical health parameter data is analyzed by sliding window, and the long-term change data of user health parameter is dynamically established and updated.This scheme can distinguish whether physiological parameter fluctuation is caused by motion through the correlation search of time window division and motion state label, for example, when detecting that the heart rate fluctuation of user during running exceeds threshold value, it can be judged as normal physiological response instead of abnormal by combining motion label, can effectively reduce the probability that normal physiological parameter fluctuation caused by motion is mislabeled as abnormal, while improving the pertinence of abnormal detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, specifically to an intelligent data annotation method based on artificial intelligence. Background Technology

[0002] With the rapid development of technologies such as artificial intelligence, big data analytics, and the Internet of Things, the healthcare field is gradually evolving towards intelligence and automation. Intelligent medical devices have been widely used in scenarios such as daily health management, chronic disease monitoring, and telemedicine. These devices can collect various physiological parameter data of users in real time, providing a rich data foundation for medical services and health management. At the same time, with the advancement of medical informatization, data analysis and judgment capabilities based on artificial intelligence have become an important trend and research hotspot in data processing in the healthcare field.

[0003] However, in the process of implementing the technical solution of the invention in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0004] Existing solutions often rely solely on single-type parameter data for anomaly detection, lacking the ability to collaboratively analyze multiple types of physiological parameters and motion state data. This can easily lead to insufficient accuracy and robustness in anomaly identification, failing to fully consider the long-term trends and individual differences in users' health parameters. Consequently, the sensitivity and specificity of anomaly detection are limited. Furthermore, the lack of deep integration between motion state recognition models and physiological parameter analysis results in normal parameter fluctuations caused by exercise being misjudged as anomalies, reducing the accuracy of anomaly data labeling. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent data annotation method based on artificial intelligence to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] This invention discloses an artificial intelligence-based intelligent data annotation method, applied to the intelligent annotation of abnormal physiological parameters of users in intelligent medical devices, comprising the following steps:

[0008] Acquire users' physiological parameters, historical health parameters, and exercise status data;

[0009] Based on the motion state data, motion state label data is output through a pre-built motion recognition model;

[0010] A sliding window analysis is performed on the historical health parameter data to dynamically establish and update long-term change data of user health parameters;

[0011] The physiological parameter fluctuation value is calculated based on the physiological parameter data. It is determined whether the physiological parameter fluctuation value is greater than the preset abnormal threshold. If it is, the motion state tag data is retrieved, and the preset abnormal threshold is adjusted according to the retrieval result.

[0012] The preset abnormal threshold is set based on long-term changes in the user's health parameters.

[0013] Next, determine whether the fluctuation value of the physiological parameter is greater than the adjusted preset abnormal threshold. If so, it is determined to be potential abnormal physiological parameter data.

[0014] The fluctuation values ​​of the physiological parameters and the long-term change data of the user's health parameters are input into a pre-built artificial intelligence discrimination model, and the confidence level of the potential abnormal physiological parameter data is output.

[0015] If the confidence level is greater than a set value, it is determined to be real abnormal physiological parameter data;

[0016] The real abnormal physiological parameter data are labeled using a structured data structure.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] 1. This solution can distinguish whether fluctuations in physiological parameters are caused by exercise by dividing time windows and associating them with exercise status tags. For example, when it is detected that a user's heart rate fluctuation exceeds a threshold during running, it can be determined as a normal physiological response rather than an abnormal one by combining the exercise tag. This can effectively reduce the probability of normal physiological parameter fluctuations caused by exercise being mislabeled as abnormal, and at the same time improve the targeting of abnormality detection.

[0019] 2. This solution uses a dynamic transition range and a gradual adjustment mechanism to keep the abnormal threshold synchronized with the user's actual movement state, thus solving the problem of misjudgment caused by threshold lag or sudden changes. It can effectively distinguish between normal physiological fluctuations and real abnormal data during movement state switching, reduce misjudgment caused by sudden changes in movement state, and improve the accuracy and reliability of abnormal data labeling.

[0020] 3. This solution effectively identifies such correlated fluctuations through a multivariate joint distribution model, avoiding misjudging normal physiological responses as abnormalities. It can establish adaptive anomaly detection standards based on the dynamic changes in the individual user's health status, effectively solving the false alarm problem caused by ignoring individual differences in traditional methods. At the same time, through multi-parameter joint analysis, it can accurately distinguish between normal fluctuations caused by exercise and pathological abnormal fluctuations, improving the accuracy of anomaly labeling. Attached Figure Description

[0021] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0022] Figure 1 This is a flowchart illustrating the steps of an artificial intelligence-based intelligent data annotation method according to the present invention.

[0023] Figure 2 A flowchart illustrating the process of adjusting a preset abnormal threshold provided by the present invention;

[0024] Figure 3 This is a flowchart illustrating the process of setting a preset abnormal threshold provided by the present invention.

[0025] Figure 4 This is a flowchart illustrating the data annotation module provided by the present invention. Detailed Implementation

[0026] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0027] In existing technologies, smart medical devices provide support for users' health management by collecting physiological parameter data in real time. However, existing abnormality detection methods usually rely on a single type of parameter and lack the ability to conduct multi-dimensional data collaborative analysis. For example, when changes in a user's exercise state cause fluctuations in heart rate, traditional methods may misjudge it as abnormal because they fail to identify the exercise state. At the same time, existing technologies fail to effectively track the long-term trends of users' health parameters, resulting in individual differences not being fully considered and abnormal threshold settings lacking dynamic adaptability.

[0028] To address the aforementioned issues, it was found that existing technologies suffer from shortcomings such as fragmented analysis of multi-source data, insufficient long-term trend modeling, and failure to eliminate motion interference factors. Analysis revealed that the accuracy of anomaly detection is limited by the single data dimension and poor adaptability to individual differences. Based on this, a proposal is made to integrate motion state recognition and physiological parameter analysis, establish a dynamic threshold adjustment mechanism, and introduce long-term health trend modeling. Furthermore, it is considered to introduce an artificial intelligence model to perform secondary verification of potential anomalies, forming a multi-stage collaborative anomaly labeling system.

[0029] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Example:

[0031] Please see Figures 1-4 An artificial intelligence-based intelligent data annotation method is applied to the intelligent annotation of abnormal physiological parameters of users in intelligent medical devices, including the following steps:

[0032] Acquire users' physiological parameters, historical health parameters, and exercise status data;

[0033] Based on motion state data, output motion state label data through a pre-built motion recognition model;

[0034] Sliding window analysis is performed on historical health parameter data to dynamically establish and update long-term change data of user health parameters;

[0035] The physiological parameter fluctuation value is calculated based on the physiological parameter data. It is determined whether the physiological parameter fluctuation value is greater than the preset abnormal threshold. If so, the motion state tag data is retrieved, and the preset abnormal threshold is adjusted according to the retrieval results.

[0036] The preset abnormal threshold is set based on long-term changes in the user's health parameters.

[0037] Next, determine whether the fluctuation value of the physiological parameter is greater than the adjusted preset abnormal threshold. If so, it is determined to be a potentially abnormal physiological parameter data.

[0038] The system inputs the fluctuation values ​​of physiological parameters and the long-term changes in users' health parameters into a pre-built artificial intelligence discrimination model, and outputs the confidence level of potentially abnormal physiological parameter data.

[0039] If the confidence level is greater than the set value, it is determined to be real abnormal physiological parameter data;

[0040] Real abnormal physiological parameter data are labeled using a structured data structure.

[0041] Physiological parameter data refers to vital signs indicators collected in real time through medical devices. Specifically, this can be achieved by using a smartwatch to monitor heart rate and blood oxygen saturation, or by using a blood pressure monitor to obtain blood pressure data, providing a basic data source for abnormality detection.

[0042] Motion status label data refers to the classification results of a user's activity status. Specifically, it can be achieved by collecting motion data using an accelerometer and then using a neural network model to identify states such as resting and walking, in order to distinguish the normal causes of fluctuations in physiological parameters.

[0043] Long-term changes in user health parameters refer to statistical characteristics that reflect individual health trends. Specifically, the mean, standard deviation, and other statistical quantities can be calculated using a sliding window, and the trend curve can be fitted through linear regression to provide a basis for personalized threshold settings.

[0044] Preset abnormal threshold adjustment refers to dynamically correcting the judgment criteria based on the motion state. Specifically, it can be achieved by establishing a threshold mapping table for different motion states, and automatically switching the corresponding threshold when a state change is detected, thus avoiding misjudgment caused by motion interference.

[0045] Artificial intelligence discrimination models refer to intelligent algorithms used for anomaly verification. Specifically, they can be implemented by using deep neural networks to jointly analyze fluctuation values ​​and long-term trends, and improve the reliability of anomaly judgment through confidence assessment.

[0046] The specific implementation process is as follows: After the physiological parameter data and motion state data are collected synchronously, the motion recognition model classifies and labels the motion state data; historical health data is continuously updated to the individual's health baseline through sliding window statistics, forming long-term change data of health parameters; when the detected physiological parameter fluctuation value exceeds the preset abnormal threshold, the current motion state label data is retrieved, and the preset abnormal threshold is adjusted according to the retrieval results; for potentially abnormal physiological parameter data that exceeds the adjusted preset abnormal threshold, the long-term change data of the user's health parameters is further combined with the input of artificial intelligence discrimination model for comprehensive discrimination, and finally only high-confidence abnormalities are structured and labeled; this process eliminates interference factors through motion state recognition, realizes threshold personalization by combining long-term trends, and improves the judgment accuracy by using artificial intelligence discrimination model.

[0047] Compared to existing technologies, current solutions typically use fixed thresholds and fail to differentiate between exercise states, such as directly classifying an increase in heart rate during running as abnormal. This solution dynamically adjusts the thresholds using exercise state tags, effectively distinguishing between physiological fluctuations and pathological abnormalities. Furthermore, existing technologies lack long-term trend tracking, while this solution establishes an individualized health baseline through sliding window analysis, making anomaly detection more aligned with the user's actual situation. Additionally, while existing methods often rely on a single threshold for judgment, this solution introduces an artificial intelligence discrimination model for multi-dimensional joint verification, significantly reducing the false alarm rate.

[0048] Through the above technical solutions, this application effectively solves the problem of misjudgment caused by motion interference, improves the specificity of anomaly detection through a dynamic threshold adjustment mechanism, and enhances detection sensitivity by making anomaly judgments more consistent with individual health change patterns through long-term trend modeling. The multi-stage verification process combined with motion state recognition ensures accurate judgment while achieving automated annotation of abnormal data, providing a reliable data foundation for subsequent medical diagnosis.

[0049] This application further proposes to obtain users' physiological parameter data, historical health parameter data, and exercise status data, specifically including:

[0050] The system collects users' physiological parameter data through smart medical devices (such as smartwatches, electrocardiogram monitors, pulse oximeters, etc.). The physiological parameter data includes heart rate, blood oxygen saturation, body temperature, and blood pressure.

[0051] Retrieve historical health parameter data of users over a period of time from cloud-based health records. Historical health parameter data includes past diseases, family history, and medication history.

[0052] The system collects users' motion status data by integrating sensors into smart medical devices. The motion status data includes steps, energy consumption, and speed.

[0053] Timestamps were added to physiological parameter data, historical health parameter data, and exercise status data.

[0054] Among them, historical health parameter data refers to the user's past health records, which can be retrieved through electronic health records stored in the cloud and used as benchmark data for individualized analysis;

[0055] Motion status data refers to information about a user's current or historical activity status. Specifically, it can be collected through embedded sensors such as accelerometers and gyroscopes to identify the impact of motion on physiological parameters.

[0056] Timestamps are precise time stamps added to each piece of data. This can be achieved by synchronizing the system clock to ensure the time alignment of data from multiple sources.

[0057] Specifically, physiological parameter data is collected in real time through smart medical devices. For example, when a user wears a smartwatch, their heart rate, blood oxygen saturation, and other data are continuously monitored. Historical health parameter data is retrieved from cloud-based health records. For example, information such as the user's disease records and medication history over the past three months is integrated into the current analysis. Exercise status data is acquired through sensors built into smart medical devices. For example, parameters such as the number of steps and speed recorded by accelerometers are used to determine the user's current activity intensity. All data is tagged with timestamps. For example, the collection time is automatically recorded when the data is generated, enabling subsequent analysis to match and correlate data from different sources according to a timeline.

[0058] Beneficial effects:

[0059] This application addresses the misjudgment problem caused by a single data source in existing technologies. By employing multi-dimensional data collection and time synchronization mechanisms, it improves the accuracy of anomaly detection. Simultaneously, the combination of historical health parameters and real-time motion status enables the system to dynamically adjust the anomaly judgment threshold, adapting to individual differences among different users. The addition of timestamp tags further ensures the consistency of analysis across data sources, providing a reliable data foundation for subsequent intelligent discrimination models.

[0060] This application further proposes a method for outputting motion state label data based on motion state data using a pre-built motion recognition model, specifically including:

[0061] The motion state data is preprocessed, including noise filtering and feature extraction (such as acceleration modulus, rate of change of angular velocity, etc.).

[0062] The preprocessed motion state data is then input into the pre-built motion recognition model.

[0063] The motion recognition model is built on neural networks, including multilayer perceptrons, recurrent neural networks, and long short-term memory network structures.

[0064] The motion state data and corresponding motion state label data for training are obtained through big data, and these are used as supervision signals to train the motion recognition model using supervised learning.

[0065] The exercise status label data includes resting, walking, running, and vigorous exercise;

[0066] The motion recognition model outputs motion state label data and associates the motion state label data with physiological parameter data at the corresponding time points.

[0067] Noise filtering refers to eliminating high-frequency interference signals in sensor-collected data through digital filtering algorithms. Specifically, Butterworth low-pass filters or moving average filters can be used to improve the signal-to-noise ratio of motion state data.

[0068] Feature extraction refers to extracting discriminative physical quantities from raw motion data. Specifically, the magnitude of the triaxial acceleration vector can be calculated using the acceleration modulus, or the rate of change of angular velocity can be calculated through differential operations, which are used to characterize the dynamic characteristics of different motion modes.

[0069] Neural network construction refers to the use of network structures with temporal processing capabilities. Specifically, long short-term memory networks can be used to capture the temporal correlation of motion data, or nonlinear feature mapping can be achieved through multilayer perceptrons to improve the accuracy of motion state classification.

[0070] Supervised learning refers to training a classification model using labeled samples. Specifically, it can use the cross-entropy loss function to optimize model parameters or update network weights through the backpropagation algorithm to establish a mapping relationship between motion data and state labels.

[0071] The specific implementation process is as follows: The motion state data is first filtered by a Butterworth low-pass filter to eliminate high-frequency noise. Then, the mean magnitude of the acceleration vector and the standard deviation of the angular velocity difference are extracted as feature vectors through a sliding window. This feature vector is input into a neural network model containing a bidirectional long short-term memory layer. The model is trained on a dataset containing labeled samples such as resting, walking, and running through supervised learning. After training, the model classifies the real-time collected motion data, outputs the corresponding motion state label data, and establishes a timestamp association between the motion state label data and the synchronously collected physiological parameter data. For example, when a vigorous exercise label is detected, the label will be associated with and stored with the heart rate and blood oxygen data within the corresponding time period.

[0072] Beneficial effects:

[0073] This application can accurately identify the dynamic characteristics under different motion states, avoiding the misjudgment of motion patterns that lead to the incorrect labeling of physiological parameter fluctuations as abnormal; the accurate correlation between motion state labels and physiological parameters provides a reliable basis for subsequent dynamic adjustment of thresholds, effectively reducing the false alarm rate and improving the accuracy of abnormal data labeling.

[0074] This application further proposes a sliding window analysis of historical health parameter data to dynamically establish and update long-term change data of user health parameters, specifically including:

[0075] Using a preset time window (e.g., 7 days, 30 days, etc.) as the unit, relevant data sequences within the time window are extracted from the collected historical health parameter data;

[0076] Within each sliding window, for each historical health parameter data, the mean, standard deviation, maximum value, minimum value, and quantile statistical characteristics are calculated respectively.

[0077] Based on the statistical results of each sliding window, the long-term trend of user health parameters is fitted using the linear regression method to obtain long-term change data of user health parameters.

[0078] As historical health parameter data is continuously collected, the sliding window moves forward in real time, automatically removing old data and adding new data, dynamically updating the long-term changes in the user's health parameters.

[0079] The time window refers to a fixed time interval used for segmented analysis of historical health parameter data. Specifically, it can be a unit such as 7 days or 30 days. The phased features of historical health parameters can be extracted by periodically dividing the data sequence.

[0080] Statistical characteristics refer to quantitative descriptive indicators of data distribution within a window. Specifically, the mean can be used to reflect the overall level, the standard deviation to measure the degree of fluctuation, and the quantiles to identify the range of extreme values. Through multi-dimensional calculations, a comprehensive characterization of the dynamic changes in health parameters can be achieved.

[0081] Linear regression is a mathematical model that fits the trend of parameter changes using the least squares method. Specifically, it can use time as the independent variable and health parameters as the dependent variable, and the regression coefficients reflect the rate and direction of parameter changes over time.

[0082] Dynamic updates refer to automatically adjusting the analysis window as new data is added. Specifically, a sliding window mechanism can be used to remove old data and include new data, and real-time iterative calculations can be used to ensure the timeliness of long-term changing data.

[0083] Specifically, historical health parameter data is divided into continuous time windows of fixed length, such as 30 days per analysis period. Within each window, statistical measures such as mean and standard deviation are calculated for parameters such as heart rate and blood pressure to quantify the data distribution characteristics within that period. Based on the statistical results of multiple windows, a linear regression model is used to fit the trend line of health parameters changing over time, such as the annual growth rate of heart rate or the monthly fluctuation range of blood pressure. As the smart device continues to collect new data, the window automatically slides forward. For example, the earliest day's data is removed and the latest day's data is added each day, the statistical characteristics are recalculated and the long-term change data of the user's health parameters are updated, so that the long-term change data can reflect the evolution of the user's health status in real time.

[0084] Through the above technical solution, this application can dynamically track the long-term evolution of user health parameters, provide personalized baseline reference for anomaly detection, effectively distinguish between normal physiological fluctuations and potential abnormal signals, reduce misjudgments caused by individual differences or the progression of chronic diseases, and improve the sensitivity and specificity of anomaly labeling.

[0085] This application further proposes a method for determining whether the fluctuation value of physiological parameters exceeds a preset abnormal threshold, and if so, retrieving motion state label data specifically includes:

[0086] For physiological parameter data, calculate the fluctuation value of physiological parameters within a set time window. The fluctuation value of physiological parameters is the difference between the maximum and minimum values ​​of physiological parameter data within that time window.

[0087] The fluctuation value of the physiological parameter is compared with the preset abnormal threshold. If the fluctuation value of the physiological parameter is greater than the preset abnormal threshold, the motion status label data corresponding to the time window is automatically retrieved within the time window when the fluctuation value of the physiological parameter is determined to be abnormal.

[0088] Among them, the physiological parameter fluctuation value refers to the dynamic change of the parameter calculated by the difference between the maximum and minimum values ​​within a time window. Specifically, it can be implemented using a sliding window statistical method, such as using a time window length of 5 minutes or 30 minutes to calculate the range of physiological parameters within the window in real time. This feature is used to dynamically reflect the real-time fluctuation status of the user's physiological parameters.

[0089] Among them, the preset abnormal threshold refers to the critical value of the fluctuation range dynamically set based on the long-term change data of the user's health parameters. Specifically, it can be achieved by using the quantiles or standard deviation multiples of the historical data statistical distribution. For example, the threshold can be set to the 95th quantile of the historical fluctuation value. This feature is used to make personalized judgments on whether the fluctuation of physiological parameters exceeds the normal range.

[0090] Specifically, physiological parameter data is divided into continuous time windows, such as every 5 minutes. Within each window, the difference between the maximum and minimum values ​​of parameters such as heart rate and blood pressure is calculated in real time as the physiological parameter fluctuation value. When the physiological parameter fluctuation value exceeds a preset abnormal threshold, such as a heart rate fluctuation exceeding 20 beats per minute, the system automatically triggers the retrieval of the exercise status tag data corresponding to that window time, such as whether the user was running or resting during this period. By associating and matching the time window of abnormal fluctuation with the exercise status tag, it is possible to identify whether the abnormal physiological parameter is caused by exercise, thereby avoiding misjudgment.

[0091] Beneficial effects:

[0092] This application can effectively reduce the probability of normal physiological parameter fluctuations caused by exercise being mislabeled as abnormal. For example, when a user is engaged in strenuous exercise, the system can dynamically adjust the judgment logic by retrieving exercise status tags to avoid incorrectly labeling exercise-induced heart rate increases as arrhythmias. At the same time, this method improves the targeting of anomaly detection, for example, by highlighting only abnormal fluctuations in non-exercise states, reducing the burden of redundant data processing.

[0093] This application further proposes adjusting the preset anomaly threshold based on the search results, specifically including:

[0094] Retrieve motion status label data for the current time and its immediate time window;

[0095] Preset the normal fluctuation range and abnormal fluctuation threshold of physiological parameter data under different motion state label data;

[0096] Based on the search results of motion state tag data, the corresponding physiological parameter data fluctuation abnormal threshold is matched;

[0097] When a change in motion status label data is detected in a short period of time (such as from rest to strenuous exercise), a dynamic transition range is set to gradually adjust the abnormal fluctuation threshold of physiological parameter data.

[0098] The preset abnormal threshold is adjusted to a dynamically adjusted threshold for fluctuations in physiological parameter data corresponding to the motion state label data.

[0099] The dynamic transition interval refers to the buffer zone set when the motion state changes rapidly. It can be implemented using time weighting or state probability models to avoid misjudgment caused by threshold abrupt changes.

[0100] Gradual adjustment refers to gradually changing the abnormal threshold based on the duration and intensity of the motion state switching. This can be achieved using linear interpolation or exponential smoothing algorithms, so that the threshold changes are synchronized with the user's actual physiological changes.

[0101] Specifically, when the system detects a user's activity level shifting from rest to vigorous activity, it recognizes the shift and triggers a dynamic transition interval. Within this interval, the abnormal fluctuation threshold for physiological parameters is not immediately set to the higher threshold corresponding to vigorous activity. Instead, it is gradually adjusted based on the duration of the shift. For example, in the first 30 seconds after the shift, the threshold increases by a fixed percentage per second until it reaches the target value. This gradual adjustment mechanism effectively prevents physiological parameter fluctuations caused by sudden changes in activity level from being misjudged as abnormal.

[0102] Through the above technical solution, this application can effectively distinguish between normal physiological fluctuations and real abnormal data during the transition of motion state, reduce misjudgment caused by sudden changes in motion state, and improve the accuracy and reliability of abnormal data labeling.

[0103] This application further proposes setting a preset abnormal threshold based on long-term changes in user health parameters, specifically including:

[0104] Model the long-term changes in various health parameters of users to obtain the current user's health parameter baseline and fluctuation range;

[0105] Based on the current user's baseline and fluctuation range of health parameters, dynamically set personalized abnormal thresholds for each health parameter and use them as the initial preset abnormal thresholds;

[0106] By combining the correlations among health parameters, a multivariate joint distribution is used to determine the initial preset abnormal threshold;

[0107] With real-time updates of long-term health parameter data, the preset abnormal thresholds are periodically and adaptively adjusted.

[0108] Among them, the health parameter baseline refers to the typical value of the user's health parameters calculated through statistical characteristics. Specifically, the mean or median within the sliding window can be used as the baseline value to establish a benchmark reference for personalized anomaly detection.

[0109] Multivariate joint distribution refers to considering the statistical correlation between different health parameters. Specifically, covariance matrix or principal component analysis methods can be used to construct a joint probability distribution model to avoid the risk of misjudgment caused by isolated judgment of a single parameter.

[0110] Specifically, during the operation of intelligent medical devices, the system first analyzes and updates long-term change data of parameters such as user heart rate and blood oxygen saturation through a sliding window. For example, every 30 days is an analysis cycle. The mean, standard deviation, and trend line slope of each parameter are calculated. Based on this data, a user-specific health parameter baseline is established. For example, the average heart rate of the past three months is used as the baseline value in the resting state. Combining the correlation between blood pressure and heart rate, a multivariate Gaussian distribution model is used to calculate the joint fluctuation range. Values ​​that deviate from this distribution by more than two standard deviations are set as the initial abnormal threshold. When the device detects a trend change in the user's recent health parameters, such as a decrease of 5 beats / minute in the average resting heart rate for two consecutive weeks, the system automatically triggers the threshold update mechanism, recalculates the multivariate joint distribution, and generates a new abnormal threshold.

[0111] Through the above technical solution, this application can establish an adaptive anomaly detection standard based on the dynamic changes in the individual user's health status, effectively solving the problem of false alarms caused by ignoring individual differences in traditional methods. For example, for users with sinus bradycardia, the system can automatically raise the lower limit of the abnormal heart rate threshold to avoid mislabeling normal physiological states as abnormal data. At the same time, through multi-parameter joint analysis, it can accurately distinguish between normal fluctuations caused by exercise and pathological abnormal fluctuations, improving the accuracy of anomaly labeling.

[0112] This application further proposes inputting physiological parameter fluctuation values ​​and long-term change data of user health parameters into a pre-built artificial intelligence discrimination model, and outputting the confidence level of potentially abnormal physiological parameter data, specifically including:

[0113] The fluctuation values ​​of physiological parameters and the long-term changes in user health parameters are used together as input features and combined to form a comprehensive feature vector;

[0114] The comprehensive feature vector is input into a pre-built artificial intelligence discrimination model, which is constructed through a neural network.

[0115] The confidence level of the potential abnormal physiological parameter data is output by the artificial intelligence discrimination model.

[0116] Among them, the comprehensive feature vector refers to a multi-dimensional data set formed by concatenating the fluctuation values ​​of physiological parameters and long-term change data after normalization. Specifically, it can be implemented using feature fusion technology to provide multi-dimensional input for the model.

[0117] Artificial intelligence discrimination models refer to classification or regression models built on neural networks. Specifically, they can be implemented using convolutional neural networks or deep neural network structures, and are used to perform nonlinear mapping on input features and output confidence scores.

[0118] The specific implementation process is as follows: physiological parameter fluctuation values ​​are calculated through a sliding window. For example, the difference between the maximum and minimum heart rate values ​​is calculated within a 30-second time window. This value reflects the instantaneous fluctuation of physiological parameters. Long-term changes in user health parameters are analyzed through a sliding window using historical health parameter data, such as the mean and standard deviation of heart rate over the past 30 days. Linear regression is used to fit the long-term trend, forming dynamically updated baseline data. The comprehensive feature vector is generated by standardizing and merging the above two types of data. For example, features such as heart rate fluctuation values, long-term mean, and standard deviation are concatenated into a vector and input into a pre-trained artificial intelligence discrimination model. The artificial intelligence discrimination model adopts a deep neural network structure, such as a multilayer perceptron containing an input layer, hidden layer, and output layer. The nonlinear combination of features is calculated through forward propagation, and finally, a confidence score between 0 and 1 is output, representing the credibility of potentially abnormal data.

[0119] Through the above technical solution, this application can combine the short-term fluctuation characteristics and long-term trends of physiological parameters, and achieve accurate identification of potential abnormal data through multi-dimensional feature fusion and deep neural network modeling. For example, after a user engages in strenuous exercise, even if the heart rate fluctuation value exceeds the conventional threshold, the system can still identify that the user is in the normal exercise recovery stage through long-term data, avoiding misjudgment. At the same time, when an abnormal shift in heart rate fluctuation accompanied by a long-term trend is detected, the model can output a high-confidence score to accurately identify real abnormal events. This technical solution effectively solves the problem of misjudgment caused by ignoring individual differences and dynamic changes in existing methods, and improves the accuracy of anomaly labeling.

[0120] 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.

[0121] 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. An artificial intelligence-based data intelligent labeling method applied to intelligent labeling of abnormal physiological parameters of users in intelligent medical devices, characterized in that, The method comprises the following steps: acquiring physiological parameter data, historical health parameter data and motion state data of a user; outputting motion state label data according to the motion state data through a pre-constructed motion recognition model; performing sliding window analysis on the historical health parameter data to dynamically establish and update long-term change data of the user's health parameters; calculating a physiological parameter fluctuation value according to the physiological parameter data, and determining whether the physiological parameter fluctuation value is greater than a preset abnormal threshold value, and if so, searching for motion state label data and adjusting the preset abnormal threshold value according to the search result, specifically comprising: searching for motion state label data in a current time and a nearby time window; presetting normal fluctuation ranges and fluctuation abnormal threshold values of physiological parameter data under different motion state label data; matching the fluctuation abnormal threshold value of the corresponding physiological parameter data according to the search result of the motion state label data; when it is detected that the motion state label data switches in a short time, setting a dynamic transition interval and gradually adjusting the fluctuation abnormal threshold value of the physiological parameter data; adjusting the preset abnormal threshold value to the fluctuation abnormal threshold value of the physiological parameter data corresponding to the motion state label data after dynamic adjustment; wherein the preset abnormal threshold value is set according to the long-term change data of the user's health parameters; determining whether the physiological parameter fluctuation value is greater than the adjusted preset abnormal threshold value, and if so, determining that it is potential abnormal physiological parameter data; inputting the physiological parameter fluctuation value and the long-term change data of the user's health parameters into a pre-constructed artificial intelligence discrimination model to output a confidence degree of the potential abnormal physiological parameter data; determining whether the confidence degree is greater than a set value, and if so, determining that it is real abnormal physiological parameter data; annotating the real abnormal physiological parameter data in a structured data structure.

2. The data intelligent labeling method based on artificial intelligence according to claim 1, characterized in that: The acquisition of the physiological parameter data, the historical health parameter data and the motion state data of the user specifically comprises: collecting the physiological parameter data of the user through intelligent medical equipment, wherein the physiological parameter data includes heart rate, blood oxygen saturation, body temperature and blood pressure; calling historical health parameter data of the user in the past period of time from a cloud health record, wherein the historical health parameter data includes past diseases, family history and drug use history; collecting the motion state data of the user through sensors integrated in the intelligent medical equipment, wherein the motion state data includes steps, activity energy consumption and speed; adding a time stamp label to the physiological parameter data, the historical health parameter data and the motion state data.

3. The data intelligent labeling method based on artificial intelligence according to claim 1, characterized in that: The outputting of the motion state label data according to the motion state data through the pre-constructed motion recognition model specifically comprises: preprocessing the motion state data, wherein the preprocessing includes noise filtering and feature extraction; and inputting the preprocessed motion state data into the pre-constructed motion recognition model; the motion recognition model is constructed based on a neural network, including a multilayer perceptron, a recurrent neural network and a long short-term memory network structure; training the motion recognition model in a supervised learning manner by acquiring training motion state data and corresponding motion state label data through big data and taking them as a supervision signal; The motion state label data includes rest, walking, running, and strenuous exercise. The motion state label data is output by a pre-constructed motion recognition model, and the motion state label data is associated with the physiological parameter data at the corresponding time point.

4. The data intelligent labeling method based on artificial intelligence according to claim 1, characterized in that: The long-term change data of the user's health parameters is dynamically established and updated by performing sliding window analysis on the historical health parameter data, and specifically includes the following steps: In a preset time window, relevant data sequences within the time window are extracted from the collected historical health parameter data. In each sliding window, the mean, standard deviation, maximum value, minimum value, and quantile statistical features are calculated for each historical health parameter data. Based on the statistical results of each sliding window, the long-term change trend of the user's health parameters is fitted using a linear regression method to obtain the long-term change data of the user's health parameters. As the historical health parameter data is continuously collected, the sliding window moves forward in real time, old data is automatically excluded, and new data is added, and the long-term change data of the user's health parameters is dynamically updated.

5. The data intelligent labeling method based on artificial intelligence according to claim 1, characterized in that: If the physiological parameter fluctuation value is greater than the preset abnormal threshold, the motion state label data corresponding to the time window is automatically searched within the time window in which the physiological parameter fluctuation value is determined to be abnormal. The physiological parameter fluctuation value is calculated within a set time window, and the physiological parameter fluctuation value is the difference between the maximum and minimum values of the physiological parameter data within the time window. The physiological parameter fluctuation value is compared with the preset abnormal threshold, and if the physiological parameter fluctuation value is greater than the preset abnormal threshold, the motion state label data corresponding to the time window is automatically searched within the time window in which the physiological parameter fluctuation value is determined to be abnormal.

6. The data intelligent labeling method based on artificial intelligence according to claim 1, characterized in that: The preset abnormal threshold is set according to the long-term change data of the user's health parameters, and specifically includes the following steps: The long-term change data of the user's health parameters is modeled to obtain the health parameter baseline and fluctuation range of the current user. The individualized abnormal threshold of each health parameter is dynamically set based on the health parameter baseline and fluctuation range of the current user, and is used as the initial preset abnormal threshold. The initial preset abnormal threshold is combined with the correlation between health parameters to determine the preset abnormal threshold using a multivariate joint distribution. The preset abnormal threshold is periodically self-adapted with the real-time update of the long-term change data of the health parameters.

7. The data intelligent labeling method based on artificial intelligence according to claim 1, characterized in that: The physiological parameter fluctuation value and the long-term change data of the user's health parameters are input into a pre-constructed artificial intelligence discrimination model to output the confidence of the potential abnormal physiological parameter data, and specifically includes the following steps: The physiological parameter fluctuation value and the long-term change data of the user's health parameters are jointly used as input features to form a comprehensive feature vector. The comprehensive feature vector is input into a pre-constructed artificial intelligence discrimination model, and the artificial intelligence discrimination model is constructed by a neural network. The confidence of the potential abnormal physiological parameter data is output by the artificial intelligence discrimination model.

Images