Remote patient tracking management method based on big data

Abstract

The invention relates to the technical field of remote management, in particular to a remote patient tracking management method based on big data, which comprises the following steps: periodically acquiring physiological data submitted by a patient; encoding the time sequence characteristics of the physiological data by using an automatic encoder to generate continuous characteristic representation; constructing a global distribution model, calculating a probability value of newly submitted physiological data in global distribution, and marking low probability points as distribution abnormity; according to the method, data continuity features and distribution features are integrated, a comprehensive abnormal score of each data point is calculated through a weighted scoring method, score weights are dynamically adjusted in combination with personalized health portraits of patients, index deviations related to health risks of the patients are preferentially concerned, threshold values are set according to the comprehensive abnormal scores, and comprehensive abnormal data points are identified. According to the method, continuity features and distribution features are combined, an anomaly detection strategy is dynamically adjusted through a weighted scoring method, and significant anomaly points are marked.

Description

Technical Field

[0001] The present invention relates to the field of remote management technology, and in particular to a remote patient tracking management method based on big data. Background Art

[0002] With the rapid development of medical technology and information technology, telemedicine has gradually become an important way to manage chronic diseases, monitor the health of the elderly, and intervene in the health of sub-healthy people. In remote patient management, medical institutions usually rely on patients to submit physiological monitoring data through portable medical devices: sphygmomanometers, oximeters, and smart bracelets, including key health indicators such as heart rate, blood pressure, and blood oxygen saturation. These data are uploaded to the remote management system through the Internet of Things technology for doctors or systems to analyze and evaluate, so as to achieve real-time tracking and management of patients' health status. However, in the existing remote management methods, the physiological monitoring data submitted by patients often have the following technical problems:

[0003] Due to improper operation by patients (such as incorrect use of equipment or manual entry errors), the collected data may have large deviations, making it difficult to truly reflect the patient's actual health status.

[0004] Some current systems rely on simple static thresholds (such as an alarm when the heart rate is higher than a certain value) for anomaly detection, which are unable to identify complex dynamic change characteristics and long-term deviation characteristics. For potential health risks in the data (such as short-term drastic fluctuations or long-term deviations from the health baseline), existing detection methods are often not sensitive enough and are prone to missed reports or false alarms.

[0005] Traditional remote management systems are mostly based on universal health assessment standards (such as fixed normal ranges) and do not fully consider differences between individuals (such as the patient's medical history and long-term health status), resulting in strong generalization of test results but poor adaptability. Summary of the invention

[0006] The present invention provides a remote patient tracking management method based on big data.

[0007] The remote patient tracking management method based on big data includes the following steps:

[0008] S1, big data collection: regularly obtain physiological data submitted by patients, including heart rate, blood pressure, and blood oxygen saturation, and synchronize the timestamps of the collected data;

[0009] S2, data continuity feature extraction: extract short-term dynamic change features of physiological data based on time series analysis method, use sliding window technology to calculate the mean, standard deviation and change rate in continuous time period, use automatic encoder to encode the time series characteristics of physiological data, generate continuous feature representation, and capture abnormal short-term mutation signals;

[0010] S3, data distribution feature modeling: construct a global distribution model, use kernel density estimation to analyze the distribution characteristics of patients' historical physiological data, extract the global probability distribution of health indicators, calculate the probability value of newly submitted physiological data in the global distribution, and mark low-probability points as distribution anomalies;

[0011] S4, joint anomaly score calculation: Comprehensive data continuity characteristics and distribution characteristics are combined to calculate the comprehensive anomaly score of each data point through a weighted scoring method. The scoring weight is dynamically adjusted in combination with the patient's personalized health portrait, with priority given to indicator deviations related to the patient's health risks. Thresholds are set according to the comprehensive anomaly score to identify comprehensive abnormal data points.

[0012] Optionally, the S2 specifically includes:

[0013] S21, time series data segmentation: Use sliding window technology to segment the patient's physiological data, set the window length and sliding step length, ensure that each time period contains enough observation points, extract physiological data in chronological order within each sliding window, form time series segments within continuous time periods, and ensure the integrity and seamless connection of the sequence;

[0014] S22, short-term dynamic change feature calculation: For each window X i ={x i ,x i+1 ,...,x i+w-1}, in each sliding window, the statistical features of the physiological data are calculated, including:

[0015] i. Mean: reflects the overall level of data within this period of time. Among them, μ i is the mean of the data in the i-th window;

[0016] ii. Standard deviation: reflects the fluctuation range of data within the time period. Among them, σ i is the fluctuation range of the data in the i-th window;

[0017] iii. Rate of change: By differentiating adjacent data points, the rate and direction of data change are evaluated. Among them, r i is the rate of change of data in the i-th window.

[0018] S23, Autoencoder Modeling: Construct a time series feature extraction model based on autoencoder, and the dynamic feature of each sliding window is represented as F i =[μ i ,σ i ,r i ], where μ i is the mean, σi is the standard deviation, r i To change the rate, the autoencoder consists of an encoder and a decoder:

[0019] i. Encoder: maps the time series in the sliding window to a low-dimensional feature space to capture the core dynamic characteristics of the data;

[0020] ii. Decoder: reconstructs the time series from low-dimensional features and calculates the reconstruction error;

[0021] S24, continuous feature representation generation: Input the physiological data in each time window into the encoder to generate a low-dimensional continuous feature representation Z, retain the information of short-term dynamic changes, and perform time serialization on the generated feature representation to form a global feature trajectory for capturing abnormal short-term mutation signals;

[0022] S25, abnormal short-term mutation signal detection: analyze the trajectory changes of the continuous feature representation Z, identify the change amplitude of the feature value, including sudden increase or decrease, mark the detected abnormal signal, and combine it with the global distribution characteristics to verify whether it is a potential health risk.

[0023] Optionally, the encoder converts the dynamic feature F i Projection to low-dimensional feature space Z i middle:

[0024] Z i =f(W e ·F i +b e ), where Z i is a low-dimensional continuous feature representation, W e ,b e is the encoder’s weight matrix and bias vector, and f(·) is the ReLU activation function;

[0025] The decoder is constructed from the low-dimensional features Z i Reconstructing the original dynamic features in, is the dynamic feature reconstructed by the decoder, W d ,b d are the weight matrix and bias vector of the decoder, and g(·) is the activation function.

[0026] Optionally, the generation of the continuous feature representation in S24 includes:

[0027] The dynamic features F of each window are transformed into i Convert to low-dimensional feature Z i , and obtain the continuous feature sequence: Z = {Z1, Z2, ..., Z n1}, where n1 is the number of low-dimensional features, Z is a low-dimensional continuous feature representation sequence, and the feature representation Z provides a simplified dynamic description of the time series for subsequent detection of abnormal signals.

[0028] Optionally, the variation range of the identification feature value in S25 is expressed as: ΔZ i =‖Z i+1 -Z i ‖, where ΔZ i is the variation of the features of adjacent windows;

[0029] Abnormality score: Among them, A i is the anomaly score of the ith window, μ ΔZ ,σ ΔZ are the mean and standard deviation of the variation, A i When >τ, it is marked as a preliminary anomaly, and the anomaly score threshold when τ is determined based on the statistical distribution.

[0030] Optionally, the S3 specifically includes:

[0031] S31, Global distribution model construction: Collect historical physiological data of patients X old , using kernel density estimation analysis to construct a global probability distribution model of health indicators f(x z ):

[0032] S32, global probability distribution feature extraction: based on f(x z ) Extract the global distribution characteristics of health indicators and characterize the overall distribution pattern of patient health data;

[0033] S33, calculation of probability value of new data point: for the newly submitted physiological data point x new , using the global probability distribution model f(x z ) Calculate its probability value: P(x new )=f(x new ), judge P(x new ) is lower than the set probability interval τ p , if P(x new )<τ p , marked as abnormal, τ p is the decision threshold for low probability points (calculated based on the 5% percentile of historical data).

[0034] Optionally, S3 also includes bandwidth parameter optimization: selecting the optimal bandwidth h by a cross-validation method to balance the smoothness and fitting accuracy of the distribution model: h * =argmin h CV(h), where CV(h) is the cross-validation error under bandwidth h.

[0035] Optionally, the global probability distribution model f(x z ) is expressed as:

[0036] Among them, x z represents the data point to be calculated at position z (specific data value), the physiological data position for the current generation of probability density calculation, which is the newly submitted data or a specified position on the distribution, f(x z ) represents the physiological data location x z The probability density function of ), h represents the bandwidth parameter, which controls the smoothness of the kernel function, n2 represents the number of historical data points, and x g Represents a historical data point, which is the specific data value of the gth data point in the historical physiological data (the patient's historical heart rate, blood pressure or blood oxygen saturation value). g refers to the index of traversing historical data, which is used to participate in the estimation of probability density and represents known historical data. z -x g Indicates the distance between the data point to be calculated and the historical data point. is the bandwidth normalized distance, measuring x z ,x g similarity.

[0037] Optionally, the comprehensive abnormality score S in S4 total Calculated as:

[0038] S total =w A ·A i +w dist ·S dist , where w A With w dist They are the weights of the continuity anomaly score and the distribution characteristic anomaly score;

[0039] Distribution characteristic abnormality score S dist Calculated as:

[0040] Using the kernel density estimation model f(x z ) Calculate the new data point x new The probability density P(x new ), based on the threshold τ of low probability points p , defines the anomaly score S of the distribution characteristics dist :

[0041] Among them, S dist =0 means that the data point conforms to the global distribution and there is no distribution anomaly. Based on the comprehensive anomaly score of the normal distribution in the historical data, the comprehensive threshold τ is set. total(take 95% quantile), if S total >τ total , then mark the data point as a comprehensive outlier.

[0042] Optionally, S4 further includes adjusting the weight w according to the patient's personalized health profile A and w dist :

[0043] If the patient's health profile shows a high risk of short-term fluctuations, increase w A ;

[0044] If the patient's health profile shows a long-term abnormal distribution risk, increase w dist ;

[0045] w cont +w dist =1,w cont ,w dist ∈[0,1].

[0046] Beneficial effects of the present invention:

[0047] The present invention realizes accurate anomaly detection of patients' physiological data by performing multi-dimensional analysis on the continuity and distribution characteristics of comprehensive data. It uses sliding window technology and automatic encoder to capture short-term dynamic changes in patients' physiological data and discover short-term anomalies in real time, effectively solving the limitation of traditional methods that rely only on single-point data.

[0048] The present invention constructs a global probability distribution model of health data through kernel density estimation, accurately identifies abnormal points that deviate from the health baseline for a long time, combines continuity characteristics and distribution characteristics, and dynamically adjusts the abnormal detection strategy through a weighted scoring method. It can not only mark significant abnormal points, but also give priority to deviations in key indicators related to patient health risks, thereby improving detection accuracy and real-time response capabilities.

[0049] The present invention realizes personalized management of different patients' health status by constructing personalized health portraits of patients, dynamically adjusting detection models and weight parameters, and dynamically optimizing the weights of continuity features and distribution features in combination with health portrait information. For example, for patients with high risk of short-term fluctuations, short-term dynamic features are given priority; for patients with high risk of long-term deviations, distribution feature analysis is strengthened.

[0050] The present invention, automated data analysis and intelligent risk analysis mechanisms, significantly reduce the cost of remote patient management, while optimizing the efficiency of allocation of medical resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of data continuity feature extraction according to an embodiment of the present invention;

[0054] Figure 3 A schematic diagram of data distribution feature modeling according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0056] like Figure 1-Figure 3 As shown, the remote patient tracking management method based on big data includes the following steps:

[0057] S1, big data collection: regularly obtain physiological data submitted by patients, including heart rate, blood pressure, and blood oxygen saturation, and synchronize the timestamps of the collected data;

[0058] Use the unified time synchronization service to obtain the standard timestamp of the server, time-mark the data submitted by the patient, and perform time difference correction based on the time zone information of the patient's location for uploaded data with time zone differences or delays to ensure the temporal consistency of the data. After the timestamp synchronization is completed, the data is stored in the database in a time series format, providing an accurate data basis for subsequent continuous feature extraction and distribution feature modeling;

[0059] S2, data continuity feature extraction: extract short-term dynamic change features of physiological data based on time series analysis method, use sliding window technology to calculate the mean, standard deviation and change rate in continuous time period, use automatic encoder to encode the time series characteristics of physiological data, generate continuous feature representation, and capture abnormal short-term mutation signals;

[0060] S3, data distribution feature modeling: construct a global distribution model, use kernel density estimation to analyze the distribution characteristics of patients' historical physiological data, extract the global probability distribution of health indicators, calculate the probability value of newly submitted physiological data in the global distribution, and mark low-probability points as distribution anomalies;

[0061] S4, joint anomaly score calculation: Comprehensive data continuity characteristics and distribution characteristics are combined to calculate the comprehensive anomaly score of each data point through a weighted scoring method. The scoring weight is dynamically adjusted in combination with the patient's personalized health portrait, with priority given to indicator deviations related to the patient's health risks. Thresholds are set according to the comprehensive anomaly score to identify comprehensive abnormal data points.

[0062] S2 specifically includes:

[0063] S21, time series data segmentation: Use sliding window technology to segment the patient's physiological data, set the window length and sliding step length, ensure that each time period contains enough observation points, extract physiological data in chronological order within each sliding window, form time series segments within continuous time periods, and ensure the integrity and seamless connection of the sequence;

[0064] Time series data segmentation: Let X = {x1, x2, ..., x T} is the time series of physiological data, the total time length is T, and the sliding window technology is used to segment X, the window length is w, and the sliding step length is s;

[0065] Each piece of data X i Expressed as: X i ={x i ,x i+1 ,...,x i+w-1},i=1,s,s+1,…,T-w+1, where X is the original physiological time series data, x t is the physiological data (heart rate, blood pressure or blood oxygen saturation) at time t, w is the sliding window length, s is the sliding step length, X i is the time series segment within the i-th window;

[0066] The sequence {X1, X2, …} generated by the sliding window provides segmented input for subsequent calculations, which facilitates the analysis of dynamic change characteristics in a short period of time;

[0067] S22, short-term dynamic change feature calculation: For each window X i ={x i ,x i+1 ,...,x i+w-1}, in each sliding window, the statistical features of the physiological data are calculated, including:

[0068] i. Mean: reflects the overall level of data within this period of time. Among them, μ i is the mean of the data in the i-th window;

[0069] ii. Standard deviation: reflects the fluctuation range of data within the time period. Among them, σ i is the fluctuation range of the data in the i-th window;

[0070] iii. Rate of change: By differentiating adjacent data points, the rate and direction of data change are evaluated. Among them, r i is the rate of change of data in the i-th window;

[0071] The above characteristics (μ i ,σ i ,r i ) forms a dynamic feature description of each window, which is used as input feature for subsequent autoencoder modeling.

[0072] S23, Autoencoder Modeling: Construct a time series feature extraction model based on autoencoder, and the dynamic feature of each sliding window is represented as F i =[μ i ,σ i ,r i ], where μ i is the mean, σ i is the standard deviation, r i To change the rate, the autoencoder consists of an encoder and a decoder:

[0073] i. Encoder: maps the time series in the sliding window to a low-dimensional feature space to capture the core dynamic characteristics of the data;

[0074] ii. Decoder: reconstructs the time series from low-dimensional features and calculates the reconstruction error;

[0075] S24, continuous feature representation generation: Input the physiological data in each time window into the encoder to generate a low-dimensional continuous feature representation Z, retain the information of short-term dynamic changes, and perform time serialization on the generated feature representation to form a global feature trajectory for capturing abnormal short-term mutation signals;

[0076] S25, abnormal short-term mutation signal detection: analyze the trajectory changes of the continuous feature representation Z, identify the change amplitude of the feature value, including sudden increase or decrease, mark the detected abnormal signal, and combine it with the global distribution characteristics to verify whether it is a potential health risk.

[0077] The encoder transforms the dynamic features F i Projection to low-dimensional feature space Zi middle:

[0078] Z i =f(W e ·F i +b e ), where Z i is a low-dimensional continuous feature representation, W e ,b e is the encoder’s weight matrix and bias vector, and f(·) is the ReLU activation function;

[0079] The decoder is trained from the low-dimensional features Z i Reconstructing the original dynamic features in, is the dynamic feature reconstructed by the decoder, W d ,b d are the weight matrix and bias vector of the decoder, and g(·) is the activation function.

[0080] Optimize the autoencoder by minimizing the reconstruction error: The model is trained by minimizing the reconstruction error, ensuring that the encoder can extract effective features of the time series.

[0081] The continuous feature representation generation in S24 includes:

[0082] The dynamic features F of each window are transformed into i Convert to low-dimensional feature Z i , and obtain the continuous feature sequence: Z = {Z1, Z2, ..., Z n1}, where n1 is the number of low-dimensional features, Z is a low-dimensional continuous feature representation sequence, and the feature representation Z provides a simplified dynamic description of the time series for subsequent detection of abnormal signals.

[0083] The variation range of the identification feature value in S25 is expressed as: ΔZ i =‖Z i+1 -Z i ‖, where ΔZ i is the variation of the features of adjacent windows;

[0084] Abnormality score: Among them, A i is the anomaly score of the ith window, μ ΔZ ,σ ΔZ are the mean and standard deviation of the variation, A i>τ, it is marked as a preliminary anomaly. The anomaly score threshold at τ is determined according to the statistical distribution: τ = μ1 + k1·σ1, where μ1 is the mean of the anomaly score (calculated through historical data), σ1 is the standard deviation of the anomaly score, and k1 is an adjustment factor with a value of 2 or 3, corresponding to approximately 95% or 99.7% of the data being within the normal range.

[0085] S3 specifically includes:

[0086] S31, Global distribution model construction: Collect historical physiological data of patients X old , using kernel density estimation analysis to construct a global probability distribution model of health indicators f(x z ):

[0087] S32, global probability distribution feature extraction: based on f(x z ) Extract the global distribution characteristics of health indicators and characterize the overall distribution pattern of patient health data;

[0088] S33, calculation of probability value of new data point: for the newly submitted physiological data point x new , using the global probability distribution model f(x z ) Calculate its probability value: P(x new )=f(x new ), judge P(x new ) is lower than the set probability interval τ p , if P(x new )<τ p , marked as abnormal, τ p is the decision threshold for low probability points (calculated based on the 5% percentile of historical data).

[0089] Determine the 5% percentile of x using the cumulative distribution function F(x) 5％ :

[0090] x 5％ Substitute into the kernel density estimation model f(x) to calculate τ p =f(x 5％ ).

[0091] S3 also includes bandwidth parameter optimization: the optimal bandwidth h is selected by cross-validation method to balance the smoothness and fitting accuracy of the distribution model: h * =argmin h CV(h), where CV(h) is the cross-validation error under bandwidth h.

[0092] The global probability distribution model f(x z ) is expressed as:

[0093] Among them, xz represents the data point to be calculated at position z (specific data value), the physiological data position for the current generation of probability density calculation, which is the newly submitted data or a specified position on the distribution, f(x z ) represents the physiological data location x z The probability density function of ), h represents the bandwidth parameter, which controls the smoothness of the kernel function, n2 represents the number of historical data points, and x g Represents a historical data point, which is the specific data value of the gth data point in the historical physiological data (the patient's historical heart rate, blood pressure or blood oxygen saturation value). g refers to the index of traversing historical data, which is used to participate in the estimation of probability density and represents known historical data. z -x g Indicates the distance between the data point to be calculated and the historical data point. is the bandwidth normalized distance, measuring x z ,x g similarity.

[0094] The composite anomaly score S in S4 total Calculated as:

[0095] S total =w A ·A i +w dist ·S dist , where w A With w dist They are the weights of the continuity anomaly score and the distribution characteristic anomaly score;

[0096] Distribution characteristic abnormality score S dist Calculated as:

[0097] Using the kernel density estimation model f(x z ) Calculate the new data point x new The probability density P(x new ), based on the threshold τ of low probability points p , defines the anomaly score S of the distribution characteristics dist :

[0098] Among them, S di =0 means that the data point conforms to the global distribution and there is no distribution anomaly. Based on the comprehensive anomaly score of the normal distribution in the historical data, the comprehensive threshold τ is set. total (take 95% quantile), if S total >τ total , then mark the data point as a comprehensive outlier.

[0099] S4 also includes adjusting the weight w according to the patient's personalized health profile A and w dist :

[0100] If the patient's health profile shows a high risk of short-term fluctuations, increase w A ;

[0101] If the patient's health profile shows a long-term abnormal distribution risk, increase w dist ;

[0102] w cont +w dist =1,w cont ,w dist ∈[0,1].

[0103] Construction of personalized health profile of patients: The personalized health profile of patients is a dynamic model established based on historical data, real-time data and patient-specific characteristics (age, gender, health status), which is used to describe the baseline and abnormal preference characteristics of the patient's health status, as follows:

[0104] (1) Input data: patient historical health data: including time series data collected over a long period of time, such as heart rate, blood pressure, and blood oxygen saturation; patient basic information: age, gender, and weight; real-time health monitoring data: the latest data points collected from portable devices.

[0105] (2) Establishment of health baseline for portrait modeling:

[0106] Calculate the historical mean of each health indicator and standard deviation Develops a personalized healthy baseline range for the patient; normal ranges: k2 is the deviation multiple, which is 2, covering 95% of the normal data range.

[0107] Dynamic risk feature extraction: Analyze short-term data volatility (heart rate change rate) and long-term distribution deviation (long-term high blood pressure), and record high-risk areas;

[0108] Use statistical methods or machine learning algorithms to identify unusual patterns in health data, such as:

[0109] Short-term risk: continuous fluctuations in data (continuous high heart rate).

[0110] Long-term risk: A trend away from a healthy baseline (gradual increase in blood pressure).

[0111] Labeled health portrait:

[0112] Label the patient's health profile according to different risk sources:

[0113] Short-term volatility risk is high: it is suitable to increase the continuity weight;

[0114] The risk of long-term bias is significant: it is appropriate to increase the distribution weight.

[0115] Specific adjustment strategies:

[0116] Significant short-term volatility risk: Increase w A , reduce w dist , let w cont =0.7,w dist =0.3;

[0117] Significant risk of long-term distribution bias: Increase w dist, , reduce w A , let w cont =0.3,w dist =0.7;

[0118] No clear preference: weight equally.

[0119] Therefore, remote patient tracking management methods:

[0120] By collecting patients' physiological data (such as heart rate, blood pressure, and blood oxygen saturation) and performing real-time analysis, it helps doctors and management systems understand patients' health status at any time, provide personalized health portraits, and clarify patients' health baselines and risk characteristics.

[0121] Automatically identify potential abnormalities in the physiological data submitted by patients, mark health risks (abnormal data fluctuations, long-term deviations from the baseline), and promptly issue warnings to doctors and patients, thereby improving the accuracy of remote health management and reducing risks caused by inaccurate or delayed data.

[0122] By combining the continuity and distribution characteristics of the data, a comprehensive abnormality score is provided to prioritize patients with higher health risks, assist doctors in optimizing intervention decisions, dynamically adjust the analysis model, optimize the detection rules based on the patient's personalized health portrait, and adapt to the health characteristics of different patients.

[0123] Automated data processing and anomaly detection significantly reduce the workload of manual review, mark health risks in layers, and ensure that medical resources are focused on high-risk patients.

[0124] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A remote patient tracking management method based on big data, characterized in that: The following steps are involved: S1, big data collection: regularly obtain physiological data submitted by patients, including heart rate, blood pressure, and blood oxygen saturation, and synchronize the timestamps of the collected data; S2, data continuity feature extraction: extract short-term dynamic change features of physiological data based on time series analysis method, use sliding window technology to calculate the mean, standard deviation and change rate in continuous time period, use automatic encoder to encode the time series characteristics of physiological data, generate continuous feature representation, and capture abnormal short-term mutation signals; S3, data distribution feature modeling: construct a global distribution model, use kernel density estimation to analyze the distribution characteristics of patients' historical physiological data, extract the global probability distribution of health indicators, calculate the probability value of newly submitted physiological data in the global distribution, and mark low-probability points as distribution anomalies; S4, joint anomaly score calculation: Comprehensive data continuity characteristics and distribution characteristics are combined to calculate the comprehensive anomaly score of each data point through a weighted scoring method. The scoring weight is dynamically adjusted in combination with the patient's personalized health portrait, with priority given to indicator deviations related to the patient's health risks. Thresholds are set according to the comprehensive anomaly score to identify comprehensive abnormal data points.

2. The remote patient tracking management method based on big data according to claim 1 is characterized in that: The S2 specifically includes: S21, time series data segmentation: segment the patient's physiological data using sliding window technology, set the window length and sliding step length, extract physiological data in time sequence within each sliding window, and form time series segments within a continuous time period; S22, short-term dynamic change feature calculation: In each sliding window, the statistical features of the physiological data are calculated, including: i. Mean: reflects the overall level of data within the time period; ii. Standard deviation: reflects the fluctuation range of data within the time period; iii. Rate of change: Evaluate the rate and direction of data change by differentiating adjacent data points; S23, Autoencoder Modeling: Construct a time series feature extraction model based on autoencoder, and the dynamic feature of each sliding window is represented as F i =[μ i ,σ i ,r i ], where μ i is the mean, σ l is the standard deviation, r i To change the rate, the autoencoder consists of an encoder and a decoder: i. Encoder: maps the time series in the sliding window to a low-dimensional feature space to capture the core dynamic characteristics of the data; ii. Decoder: reconstructs the time series from low-dimensional features and calculates the reconstruction error; S24, continuous feature representation generation: Input the physiological data in each time window into the encoder to generate a low-dimensional continuous feature representation Z, retain the information of short-term dynamic changes, and perform time serialization on the generated feature representation to form a global feature trajectory for capturing abnormal short-term mutation signals; S25, abnormal short-term mutation signal detection: analyze the trajectory changes of the continuous feature representation Z, identify the change amplitude of the feature value, including sudden increase or decrease, and mark the detected abnormal signal.

3. The remote patient tracking management method based on big data according to claim 2 is characterized in that: The encoder transforms the dynamic feature F i Projection to low-dimensional feature space Z i In the decoder, the low-dimensional feature Z i Reconstructing the original dynamic features 4. The remote patient tracking management method based on big data according to claim 3 is characterized in that: The generation of the continuous feature representation in S24 includes: The dynamic features F of each window are transformed into i Convert to low-dimensional feature Z i , and obtain the continuous feature sequence: Z = {Z1, Z2, ..., Z n1 }, where n1 is the number of low-dimensional features and Z is the low-dimensional continuous feature representation sequence.

5. The remote patient tracking management method based on big data according to claim 4 is characterized in that: The variation range of the identification characteristic value in S25 is expressed as: ΔZ i =‖Z i+1 -Z i ‖, where ΔZ i is the variation of the features of adjacent windows; Abnormality score: Among them, A i is the anomaly score of the ith window, μ ΔZ ,σ ΔZ are the mean and standard deviation of the variation, A i When >τ, it is marked as a preliminary anomaly, and the anomaly score threshold when τ is determined based on the statistical distribution.

6. The remote patient tracking management method based on big data according to claim 5 is characterized in that: The S3 specifically includes: S31, Global distribution model construction: Collect historical physiological data of patients X old , using kernel density estimation analysis to construct a global probability distribution model of health indicators f(x z ): S32, global probability distribution feature extraction: based on f(x z ) Extract the global distribution characteristics of health indicators and characterize the overall distribution pattern of patient health data; S33, calculation of probability value of new data point: for the newly submitted physiological data point x new , using the global probability distribution model f(x z ) Calculate its probability value: P(x new )=f(x new ), judge P(x new ) is lower than the set probability interval τ p , if P(x new )<τ p , marked as abnormal, τ p is the decision threshold of low probability points.

7. The remote patient tracking management method based on big data according to claim 6 is characterized in that: S3 also includes bandwidth parameter optimization: the optimal bandwidth h is selected by cross-validation method to balance the smoothness and fitting accuracy of the distribution model: h * =argmin h CV(h), where CV(h) is the cross-validation error under bandwidth h.

8. The remote patient tracking management method based on big data according to claim 6 is characterized in that: The global probability distribution model f(x z ) is expressed as: Among them, x z represents the data point to be calculated at position z, the physiological data position of the current generation of probability density calculation, f(x z ) represents the physiological data location x z The probability density function of K(·) represents the kernel function, h represents the bandwidth parameter, which controls the smoothness of the kernel function, n2 represents the number of historical data points, and x g Represents a historical data point, which is the specific data value of the g-th data point in the historical physiological data. It is used to participate in the estimation of probability density and represents the known historical data. z -x g Indicates the distance between the data point to be calculated and the historical data point. is the bandwidth normalized distance.

9. The remote patient tracking management method based on big data according to claim 6, characterized in that: The comprehensive abnormality score S in S4 total Calculated as: S total =w A ·A i +w dist ·S dist , where w A With w dist They are the weights of the continuity anomaly score and the distribution characteristic anomaly score; Distribution characteristic abnormality score S dist Calculated as: Using the kernel density estimation model f(x z ) Calculate the new data point x new The probability density P(x new ), based on the threshold τ of low probability points p , defines the anomaly score S of the distribution characteristics dist : Among them, S dist =0 means that the data point conforms to the global distribution and there is no distribution anomaly. Based on the comprehensive anomaly score of the normal distribution in the historical data, the comprehensive threshold τ is set. total , if S total >τ total , then mark the data point as a comprehensive outlier.

10. The remote patient tracking management method based on big data according to claim 9, characterized in that: S4 also includes adjusting the weight w according to the patient's personalized health portrait A and w dist : If the patient's health profile shows a high risk of short-term fluctuations, increase w A ; If the patient's health profile shows a long-term abnormal distribution risk, increase w dist .

Images