5G-based chronic disease follow-up data AI analysis method and system

Abstract

The invention relates to the technical field of data analysis, in particular to a chronic disease follow-up visit data AI analysis method and system based on 5G. The method comprises the following steps: acquiring a follow-up report uploaded by a patient through a low-delay data transmission channel provided by a 5G network; analyzing the follow-up visit report, and determining a self-evaluation scale, physiological index detection data and a medication record; analyzing the self-evaluation scale, and determining the lifestyle of the patient; historical complication data of the patient is called from the database, and according to the physiological index detection data, the lifestyle, the historical complication data and the medication record, multi-modal data integration is carried out through a 5G network to obtain a structured data set; based on the dynamic knowledge graph, analyzing the structured data set, and determining the risk evolution condition of the patient; and acquiring real-time behavior information of the patient, and generating a dynamic intervention scheme according to the risk evolution condition of the patient and the real-time behavior information of the patient. The emergency event processing efficiency is improved, and medical advice discarding caused by scheme rigidity is reduced.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of data analysis, in particular to a chronic disease follow-up data AI analysis method and system based on 5G. BACKGROUND

[0002] Chronic disease management relies on long-term, dynamic follow-up monitoring and personalized intervention to delay disease progression and reduce the risk of complications. With the deep integration of 5G communication technology and medical artificial intelligence, the chronic disease intelligent follow-up system has gradually realized the transformation of data collection and analysis, achieving the goal of automation instead of manual work.

[0003] However, the traditional follow-up system usually generates static health recommendations by analyzing single collected data. This method often deviates from the prediction due to various complex real situations, ultimately leading to the attenuation of health management effect. SUMMARY

[0004] The application provides a chronic disease follow-up data AI analysis method and system based on 5G to solve the above problems.

[0005] In a first aspect, the application provides a chronic disease follow-up data AI analysis method based on 5G, which comprises: acquiring a follow-up report uploaded by a patient through a low-latency data transmission channel provided by a 5G network; analyzing the follow-up report to determine a self-assessment scale, physiological index detection data, and medication record; analyzing the self-assessment scale to determine the patient's lifestyle; retrieving the patient's historical complication data from a database, and integrating multi-modal data through a 5G network based on the physiological index detection data, the lifestyle, the historical complication data, and the medication record to obtain a structured data set; analyzing the structured data set based on a dynamic knowledge graph to determine the patient's risk evolution; acquiring real-time behavior information of the patient, and generating a dynamic intervention plan based on the patient's risk evolution and the real-time behavior information of the patient.

[0006] Through the scheme, the follow-up report uploaded by the patient is acquired through a low-delay data transmission channel provided by a 5G network, avoiding waveform segmentation misplacement caused by data packet disorder or delay. The follow-up report is analyzed to determine a self-evaluation form, physiological index detection data, and medication records, avoiding time-frequency resolution mismatch caused by sampling rate difference. The self-evaluation form is analyzed to determine the lifestyle of the patient, making lifestyle quantification features more consistent with real behavior deviation, and avoiding subjective errors of artificial interpretation. Historical complication data of the patient is retrieved from a database, and based on the physiological index detection data, lifestyle, historical complication data, and medication records, multi-modal data integration is performed through the 5G network to obtain a structured data set, which helps to eliminate feature extraction redundancy caused by scattered multi-modal data storage. Based on a dynamic knowledge graph, the structured data set is analyzed to determine the risk evolution of the patient, ensuring that the model responds to the dynamic evolution of the disease course. Real-time behavior information of the patient is acquired, and based on the risk evolution of the patient and the real-time behavior information of the patient, a dynamic intervention scheme is generated to improve the efficiency of critical event handling and reduce the abandonment of medical orders caused by rigid schemes.

[0007] Optionally, based on the physiological index detection data, the lifestyle, the historical complication data, and the medication records, multi-modal data integration is performed through the 5G network to obtain a structured data set, including: The physiological index detection data is analyzed to determine an oxygen saturation waveform; The oxygen saturation waveform is subjected to wavelet denoising processing to obtain oxygen saturation data; The medication records are analyzed to determine a drug name; Based on the drug name, a mapping relationship between the drug and the ATC classification is established; Based on the oxygen saturation data, the mapping relationship, the lifestyle, the historical complication data, multi-modal data integration is performed through the 5G network to obtain a structured data set containing time series features.

[0008] Optionally, the oxygen saturation waveform is subjected to wavelet denoising processing to obtain oxygen saturation data, including: Basic information of the wearable device is acquired, and based on the basic information, a wearable device sampling rate is determined; Based on the wearable device sampling rate, a maximum decomposition level constraint of the oxygen saturation waveform is determined; Through inverse wavelet transform, the oxygen saturation waveform is reconstructed based on the maximum decomposition level constraint to obtain a final waveform; The final waveform is analyzed to determine a frequency spectrum distribution feature; Based on the frequency spectrum distribution feature, a wavelet basis function is determined; Based on the wavelet basis function, wavelet denoising processing is performed to obtain oxygen saturation data.

[0009] Optionally, the wavelet base function is determined according to the frequency spectrum distribution feature, including: The physiological index detection data is analyzed to determine a patient breathing frequency range; The optimal time-frequency resolution is determined according to the patient breathing frequency range; The mutation point density of blood oxygen saturation is determined according to the frequency spectrum distribution feature; The base function support length is determined according to the mutation point density; The discretization parameter is determined according to the sampling rate of the wearable device; The wavelet base function is matched according to the optimal time-frequency resolution, the base function support length, and the discretization parameter.

[0010] Optionally, the wavelet denoising processing is performed according to the wavelet base function to obtain blood oxygen saturation data, including: The final waveform is analyzed based on the wavelet base function to determine a noise distribution feature; The high-frequency components in each layer coefficient that exceed a preset threshold are compressed by a soft threshold according to the noise distribution feature to obtain a denoised waveform; The sampling points are determined according to the sampling rate of the wearable device; The denoised waveform is analyzed based on the sampling points to determine a waveform gradient change; The artifact data segment is determined according to the waveform gradient change; The artifact data segment is subjected to local interpolation processing to obtain blood oxygen saturation data.

[0011] Optionally, the structured data set is analyzed based on the dynamic knowledge graph to determine a patient risk evolution situation, including: Disease-complication association rules in clinical guidelines are obtained; A patient case is obtained, and an actual decision graph is constructed according to diagnosis and treatment path data in the patient case; The time decay factor is determined according to the structured data set; The association strength weight is calculated according to the time decay factor; The patient risk evolution model containing the time decay factor and the association strength weight is constructed by a graph neural network based on the actual decision graph and the disease-complication association rules, and the patient risk evolution situation is determined.

[0012] Optionally, the time decay factor is determined according to the structured data set, including: The complication occurrence time is determined according to the disease-complication association rules; The current time interval from the occurrence of the complication of the patient to the current time is determined according to the patient case; According to the time interval of the complication and the current time interval, a time decay factor is determined.

[0013] Optionally, the calculation of the correlation strength weight according to the time decay factor comprises: The medication record, the mapping relationship, and the lifestyle are analyzed to determine drug adherence data. The patient case is analyzed to determine the patient's age. The drug adherence data, the patient's age, and the lifestyle are used to adjust the time decay factor. The correlation degree weight is calculated according to the adjusted time decay factor.

[0014] Optionally, a dynamic intervention scheme is generated according to the patient risk evolution and the patient real-time behavior information, comprising: The physiological index detection data is analyzed to determine the physiological index deviation. The physiological index deviation is used to determine the intervention priority. The medication record is analyzed to determine the actual medication amount. The patient case is analyzed to determine the theoretical medication amount. The actual medication amount and the theoretical medication amount are used to determine the dose deviation. The dose deviation and the patient risk evolution are used to adjust the drug recommendation. The patient real-time behavior information and the lifestyle are analyzed to determine the lifestyle change trend. The lifestyle change trend is used to match a behavior intervention template, and based on the intervention template, a dynamic intervention scheme is generated according to the lifestyle change trend and the drug recommendation.

[0015] In a second aspect, the present application provides a 5G-based chronic disease follow-up data AI analysis system, comprising: A report acquisition module is configured to acquire a follow-up report uploaded by a patient through a low-latency data transmission channel provided by a 5G network, analyze the follow-up report, and determine a self-rating scale, physiological index detection data, and medication record. A form analysis module is configured to analyze the self-rating scale and determine the patient's lifestyle. A data integration module is configured to retrieve historical complication data of the patient from a database, perform multi-modal data integration through a 5G network based on the physiological index detection data, the lifestyle, the historical complication data, and the medication record, and obtain a structured data set. A data analysis module is configured to analyze the structured data set based on a dynamic knowledge graph and determine the patient risk evolution. A scheme generation module is configured to acquire real-time behavior information of a patient, and generate a dynamic intervention scheme according to a risk evolution condition of the patient and the real-time behavior information of the patient.

[0016] Optionally, when the data integration module integrates the physiological index detection data, the lifestyle, the historical complication data and the medication record through the 5G network to obtain a structured data set, the data integration module is configured to: analyze the physiological index detection data to determine an oxygen saturation waveform; perform wavelet denoising processing on the oxygen saturation waveform to obtain oxygen saturation data; analyze the medication record to determine a drug name; establish a mapping relationship between the drug and an ATC classification according to the drug name; integrate the oxygen saturation data, the mapping relationship, the lifestyle and the historical complication data through the 5G network to obtain a structured data set containing time series features.

[0017] Optionally, when the data integration module performs wavelet denoising processing on the oxygen saturation waveform to obtain oxygen saturation data, the data integration module is configured to: acquire basic information of a wearable device, and determine a sampling rate of the wearable device according to the basic information; determine a maximum decomposition layer constraint of the oxygen saturation waveform according to the sampling rate of the wearable device; reconstruct the oxygen saturation waveform according to the maximum decomposition layer constraint through inverse wavelet transform to obtain a final waveform; analyze the final waveform to determine a frequency spectrum distribution feature; determine a wavelet basis function according to the frequency spectrum distribution feature; perform wavelet denoising processing according to the wavelet basis function to obtain oxygen saturation data.

[0018] Optionally, when the data integration module determines a wavelet basis function according to the frequency spectrum distribution feature, the data integration module is configured to: analyze the physiological index detection data to determine a patient breathing frequency range; determine an optimal time-frequency resolution according to the patient breathing frequency range; determine a mutation point density of oxygen saturation according to the frequency spectrum distribution feature; determine a basis function support length according to the mutation point density; determine a discretization parameter according to the sampling rate of the wearable device; match a wavelet basis function according to the optimal time-frequency resolution, the basis function support length and the discretization parameter.

[0019] Optionally, the data integration module, according to the wavelet basis function, performs wavelet denoising processing to obtain the blood oxygen saturation data, and is configured to: analyze the final waveform based on the wavelet basis function to determine noise distribution characteristics; perform soft threshold compression on high-frequency components exceeding a preset threshold in each layer coefficient according to the noise distribution characteristics to obtain a denoised waveform; determine a sampling point according to the sampling rate of the wearable device; analyze the denoised waveform based on the sampling point to determine a waveform gradient change; determine an artifact data segment according to the waveform gradient change; perform local interpolation processing on the artifact data segment to obtain the blood oxygen saturation data.

[0020] Optionally, the data analysis module, based on a dynamic knowledge graph, analyzes the structured data set to determine the patient risk evolution, and is configured to: obtain disease-complication association rules in a clinical guideline; obtain a patient case and construct an actual decision graph based on diagnosis and treatment path data in the patient case; determine a time decay factor according to the structured data set; calculate an association strength weight according to the time decay factor; based on the actual decision graph and the disease-complication association rules, construct a patient risk evolution model containing the time decay factor and the association strength weight through a graph neural network to determine the patient risk evolution.

[0021] Optionally, the data analysis module, according to the structured data set, determines a time decay factor, and is configured to: determine a complication occurrence time according to the disease-complication association rules; determine a current time interval from when the patient developed the complication to the current time according to the patient case; determine the time decay factor according to the complication occurrence time and the current time interval.

[0022] Optionally, the data analysis module, according to a time decay factor, calculates an association strength weight, and is configured to: analyze the medication record, the mapping relationship, and the lifestyle to determine drug adherence data; analyze the patient case to determine the patient's age; adjust the time decay factor according to the drug adherence data, the patient's age, and the lifestyle; Calculate the correlation weight according to the adjusted time decay factor.

[0023] Optionally, when the scheme generation module generates a dynamic intervention scheme according to the patient risk evolution and the patient real-time behavior information, it is used for: Analyzing physiological index detection data to determine physiological index deviation; According to the physiological index deviation, determine the intervention priority; Analyzing the medication record to determine the actual medication amount; According to the patient case, determine the theoretical medication amount; According to the actual medication amount and the theoretical medication amount, determine the dose deviation; According to the dose deviation and the patient risk evolution, adjust the drug suggestion; Analyzing the patient real-time behavior information and the lifestyle to determine the lifestyle change trend; According to the lifestyle change trend, match the behavior intervention template, and based on the intervention template, generate a dynamic intervention scheme according to the lifestyle change trend and the drug suggestion. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0025] Figure 1 An application scenario schematic diagram is provided for an embodiment of the present application; Figure 2 A flowchart of a chronic disease follow-up data AI analysis method based on 5G is provided for an embodiment of the present application; Figure 3 A structure schematic diagram of a chronic disease follow-up data AI analysis system based on 5G is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical scheme in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0027] In addition, the term "and / or" in this document merely describes an associated relationship between associated objects, which means that there can be three relationships, for example, A and / or B can represent three cases of A alone, A and B together, and B alone. In addition, the character " / " in this document generally represents an "or" relationship between the front and rear associated objects unless otherwise specified.

[0028] The embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0029] Traditional follow-up systems usually generate static health recommendations by analyzing single collected data. This method often deviates from the prediction due to various complex real situations, resulting in the attenuation of health management effect.

[0030] Based on this, the present application provides a chronic disease follow-up data AI analysis method and system based on 5G. The method obtains the follow-up report uploaded by the patient through the low-delay data transmission channel provided by the 5G network, avoids waveform segmentation misplacement caused by data packet disorder or delay, analyzes the follow-up report to determine the self-evaluation scale, physiological index detection data and medication record, avoids time-frequency resolution mismatch caused by sampling rate difference, analyzes the self-evaluation scale to determine the patient's lifestyle, makes the lifestyle quantitative characteristics more consistent with the real behavior deviation, avoids subjective errors of artificial interpretation, retrieves the patient's historical complication data from the database, integrates the multi-modal data through the 5G network according to the physiological index detection data, lifestyle, historical complication data and medication record, obtains the structured data set, which helps to eliminate the feature extraction redundancy caused by the storage dispersion of multi-modal data, analyzes the structured data set based on the dynamic knowledge graph, determines the patient's risk evolution, ensures the model to respond to the dynamic evolution of the disease course, obtains the real-time behavior information of the patient, generates the dynamic intervention scheme according to the patient's risk evolution and the real-time behavior information of the patient, improves the efficiency of critical event handling, and reduces the abandonment of medical advice caused by the rigidity of the scheme.

[0031] Figure 1 An application scenario provided by the present application is provided. When analyzing chronic disease follow-up data, the method provided by the present application is applied.

[0032] Specifically, the method provided by the application is applied to any server, the server interacts with a patient intelligent terminal, and a low-delay data transmission channel provided by a 5G network arranged by the patient intelligent terminal is used to obtain a follow-up report uploaded by the patient, the follow-up report is analyzed, a self-evaluation scale, physiological index detection data and medication records are determined, the self-evaluation scale is analyzed, the lifestyle of the patient is determined, the lifestyle quantification characteristics are more in line with real behavior deviation, historical complication data of the patient is called from a database, the physiological index detection data, the lifestyle, the historical complication data and the medication records are used to perform multi-modal data integration through the 5G network, a structured data set is obtained, and feature extraction redundancy caused by scattered multi-modal data storage is eliminated. Based on a dynamic knowledge graph, the structured data set is analyzed, the risk evolution of the patient is determined, and it is ensured that the model responds to the dynamic evolution of the disease course. Real-time behavior information of the patient is obtained, a dynamic intervention scheme is generated according to the risk evolution of the patient and the real-time behavior information of the patient, the efficiency of critical event processing is improved, and the abandonment of medical advice caused by rigid schemes is reduced.

[0033] The specific implementation can refer to the following embodiments.

[0034] Figure 2 A flowchart of a chronic disease follow-up data AI analysis method based on 5G provided by an embodiment of the application, the method of the embodiment can be applied to the server in the above scene. As shown in the figure, the method comprises the following steps. Figure 2 S201, a follow-up report uploaded by a patient is obtained through a low-delay data transmission channel provided by a 5G network; the follow-up report is analyzed, a self-evaluation scale, physiological index detection data and medication records are determined; The low-delay data transmission channel can be a communication link established based on the 5G network.

[0035] The follow-up report can be an electronic document containing a self-evaluation scale of the patient, physiological index detection data collected by a wearable device and medication records.

[0036] The self-evaluation scale can be a standardized questionnaire filled out by the patient.

[0037] The physiological index detection data can be time series signals such as blood oxygen saturation waveform, blood pressure and blood glucose continuously collected by the wearable device.

[0038] The medication records can be code information containing drug names, dosages, timestamps of taking and mapping to the ATC classification system.

[0039] ​Specifically, an edge gateway supporting a 5G low latency communication protocol is deployed on the side of the patient intelligent terminal, an end-to-end low latency data transmission channel is established according to a 5G network slicing technology, and a follow-up report uploaded by the patient intelligent terminal is received in real time. Then, the follow-up report is parsed by using multi-threading, and a self-rating scale, physiological index monitoring data and medication records are separated.

[0040] In S202, the lifestyle of the patient is determined by analyzing the self-rating scale. The lifestyle can be a set of behavior tags extracted from the self-rating scale.

[0041] Specifically, the self-rating scale is filtered by using a regular expression to extract fields, and then based on a transfer learning theory framework, a pre-trained natural language processing model is used to capture lifestyle entities by using a named entity recognition layer. Finally, a rule engine is applied to map the recognition results to an ICD-10-Z coding system to generate the lifestyle of the patient.

[0042] In S203, the historical complication data of the patient is retrieved from the database, and the physiological index detection data, lifestyle, historical complication data and medication records are integrated by using a 5G network to obtain a structured data set. The historical complication data can be a structured record of the type, occurrence time and severity of the patient's previously diagnosed complications.

[0043] The structured data set can be a timestamp-aligned multi-modal data set.

[0044] Specifically, the patient's historical complication events are retrieved from the database, sorted by occurrence time and labeled with duration and severity. Then, the drug names in the medication records are mapped to ATC classification codes, associated to the target metabolic pathway according to the ATC hierarchy (anatomy-therapy-chemistry), and the percentage deviation of the current dose from the prescribed dose is calculated. Further, based on the 5G network, the physiological index detection data, lifestyle and drug deviation are aligned on a unified time axis to generate a structured data set containing a time series feature matrix and a static feature vector.

[0045] In S204, the structured data set is analyzed based on a dynamic knowledge graph to determine the patient's risk evolution. The dynamic knowledge graph can be a graph structure model that integrates clinical guideline association rules, time decay factors and individual parameters.

[0046] The patient's risk evolution can be the current complication risk level and the future risk trend curve.

[0047] Specifically, disease-complication association rules are extracted from a clinical guideline library to initialize the node and edge weights of a dynamic knowledge graph. Then, a time decay factor is introduced to dynamically adjust the edge weights. Next, the structured data set is encoded into node feature vectors and edge feature matrices. Finally, the neighbor node information is aggregated through a graph attention mechanism to output the patient risk evolution.

[0048] In S205, real-time behavior information of the patient is acquired, and a dynamic intervention scheme is generated according to the patient risk evolution and the real-time behavior information of the patient.

[0049] The real-time behavior information of the patient can be the current blood oxygen saturation, exercise state, and recent smoking record of the patient.

[0050] The dynamic intervention scheme can be a step-by-step response strategy matching the risk level and behavior deviation degree.

[0051] Specifically, the real-time behavior information of the patient is acquired by continuously collecting the number of steps, GPS location, and microphone ambient sound through a wearable device. Then, according to the complication risk probability output by the dynamic knowledge graph, the intervention priority is divided according to the clinical guideline threshold of the risk stratification standard quantified in the authoritative medical guideline. Further, the intervention measure sequence is output according to the patient risk evolution. Finally, the intervention measures are matched with the historical behavior data of the patient, thereby generating a dynamic intervention scheme.

[0052] Through this scheme, the patient uploaded follow-up reports are acquired through the low-latency data transmission channel provided by the 5G network, avoiding waveform segmentation misplacement caused by data packet out-of-order or delay. The follow-up reports are analyzed to determine the self-rating scale, physiological index detection data, and medication record, avoiding time-frequency resolution mismatch caused by sampling rate difference. The self-rating scale is analyzed to determine the lifestyle of the patient, making the lifestyle quantification features more consistent with the real behavior deviation, and avoiding subjective errors of artificial interpretation. The historical complication data of the patient are retrieved from the database, and the physiological index detection data, lifestyle, historical complication data, and medication record are integrated through the 5G network to obtain a structured data set, which helps to eliminate feature extraction redundancy caused by scattered multi-modal data storage. Based on the dynamic knowledge graph, the structured data set is analyzed to determine the patient risk evolution, ensuring that the model responds to the dynamic evolution of the disease course. The real-time behavior information of the patient is acquired, and a dynamic intervention scheme is generated according to the patient risk evolution and the real-time behavior information of the patient, improving the efficiency of critical event handling and reducing the abandonment of medical orders caused by rigid schemes.

[0053] In some embodiments, the physiological index detection data is parsed to determine a blood oxygen saturation waveform; the blood oxygen saturation waveform is subjected to wavelet denoising processing to obtain blood oxygen saturation data; the medication record is parsed to determine a drug name; a mapping relationship between the drug and ATC classification is established according to the drug name; and multi-modal data integration is performed through a 5G network according to the blood oxygen saturation data, the mapping relationship, lifestyle, and historical complication data to obtain a structured data set containing time series features.

[0054] The blood oxygen saturation waveform can be continuous time series data reflecting the proportion change of oxygenated hemoglobin and reduced hemoglobin in blood.

[0055] The blood oxygen saturation data can be a structured numerical set obtained by wavelet denoising processing of the blood oxygen saturation waveform.

[0056] The drug name can be a generic name or chemical name of the drug.

[0057] The ATC classification can be an anatomic, therapeutic, and chemical classification.

[0058] The mapping relationship can be an association rule between the drug name and the corresponding ATC code.

[0059] The 5G network can be a fifth-generation mobile communication technology.

[0060] The time series feature can be a data attribute set arranged in chronological order.

[0061] Specifically, physiological index detection data is received in real time through a 5G network, and a sliding window segmentation algorithm is used to extract a blood oxygen saturation waveform. Further, based on the blood oxygen saturation waveform, the optimal wavelet basis function based on the wavelet adaptive denoising theory is dynamically selected according to the sampling rate of the wearable device, and the time-frequency resolution parameter is adjusted in combination with the real-time respiratory frequency. Then, the motion artifact interference segment is identified, and only the distorted segment is filled with the missing data segment using the local cubic spline interpolation method. Subsequently, the structured / unstructured medication record is subjected to natural language processing to extract the drug name. The parsed drug name is matched with the WHO ATC classification database to construct a mapping relationship table of drug-ATC classification. The blood oxygen saturation data, the mapping relationship, the lifestyle, and the historical complication data are synchronously transmitted through a 5G network low-latency channel; then, each modal data stream is aligned with a millisecond-level timestamp to generate a structured data set containing time series features.

[0062] By the scheme, the physiological index detection data is analyzed, and the blood oxygen saturation waveform is determined, which helps to eliminate the problem of losing physiological rhythm characteristics due to dependence on discrete sampling points, and provides an original signal basis for dynamic noise reduction and feature extraction. The blood oxygen saturation waveform is subjected to wavelet noise reduction processing to obtain blood oxygen saturation data, avoiding signal distortion or high-frequency noise residue caused by fixed wavelet basis. The drug record is analyzed to determine the drug name, which helps to eliminate the problem of isolated storage of drug information, and provides structured input for ATC classification mapping. According to the drug name, the mapping relationship between the drug and the ATC classification is established to support accurate evaluation of the influence of dose deviation on pathology. According to the blood oxygen saturation data, the mapping relationship, the lifestyle, and the historical complication data, multi-modal data integration is performed through the 5G network to obtain a structured data set containing time series features, which helps to eliminate the problem of asynchronous multi-modal data caused by transmission delay, and provides real-time data basis for dynamic risk assessment.

[0063] In some embodiments, the basic information of the wearable device is obtained, and the sampling rate of the wearable device is determined according to the basic information; the maximum decomposition layer constraint of the blood oxygen saturation waveform is determined according to the sampling rate of the wearable device; the blood oxygen saturation waveform is reconstructed according to the maximum decomposition layer constraint through inverse wavelet transform to obtain a final waveform; the final waveform is analyzed to determine the spectral distribution feature; the wavelet basis function is determined according to the spectral distribution feature; and the wavelet noise reduction processing is performed according to the wavelet basis function to obtain the blood oxygen saturation data.

[0064] The wearable device can be a portable monitoring device for collecting physiological signals of a patient.

[0065] The basic information can be a set of device parameters attached by the wearable device during data collection.

[0066] The sampling rate of the wearable device can be the number of times the wearable device collects blood oxygen saturation waveforms per second.

[0067] The maximum decomposition layer constraint can be an upper limit of the wavelet decomposition layer number.

[0068] The inverse wavelet transform can be a mathematical process of reconstructing a time domain waveform through a wavelet basis function.

[0069] The final waveform can be a blood oxygen saturation waveform reconstructed by inverse wavelet transform.

[0070] The spectral distribution feature can be the frequency domain energy distribution characteristic of the final waveform.

[0071] The wavelet basis function can be a mother wavelet function type used for wavelet decomposition and reconstruction.

[0072] Specifically, the basic information uploaded by the wearable device is parsed through the device communication protocol to extract the device identifier and hardware configuration parameters; then, the hardware technical specification document disclosed by the wearable device manufacturer is matched with the industry standard preset sampling rate database according to the device identifier to determine the sampling rate of the current blood oxygen saturation waveform of the wearable device. Based on the Nyquist sampling theorem, the maximum decomposition layer number is calculated according to the wearable device sampling rate and the lower limit of the effective frequency band of the blood oxygen saturation waveform; then, the upper limit of the wavelet decomposition layer number is set according to the maximum decomposition layer number. The discrete wavelet transform is performed on the blood oxygen saturation waveform, and the multi-scale decomposition is performed under the constraint of the maximum decomposition layer number; then, the blood oxygen saturation waveform is reconstructed through the inverse wavelet transform, the low-frequency approximation coefficients and the thresholded high-frequency detail coefficients are retained, and the final waveform is generated. Subsequently, the time domain signal of the final waveform is analyzed for power spectral density, the spectral leakage is reduced through segmented windowing and overlapping averaging, and the spectral distribution characteristics are determined. If the spectral distribution characteristics are concentrated in the corresponding respiratory frequency band, the db4 wavelet basis function is selected to enhance the low-frequency resolution; at the same time, if the spectral distribution characteristics are extended to cover the pulse wave and harmonics, the sym6 wavelet basis function is selected to balance the time-frequency localization characteristics. The blood oxygen saturation waveform is decomposed by N layers using the selected wavelet basis function, and the threshold value is adaptively set according to the layer number to remove high-frequency motion noise or smooth low-frequency noise; finally, the inverse transform is performed on the threshold-processed coefficients to output the final blood oxygen saturation data.

[0073] Through the scheme, the basic information of the wearable device is acquired, the wearable device sampling rate is determined according to the basic information, the mismatch of the noise reduction parameters caused by the difference between the device models is eliminated, and the performance degradation of the fixed wavelet basis function in the high / low frequency scene is avoided. According to the wearable device sampling rate, the maximum decomposition layer number constraint of the blood oxygen saturation waveform is determined to prevent loss of low-frequency characteristics caused by excessive decomposition. The blood oxygen saturation waveform is reconstructed according to the maximum decomposition layer number constraint through the inverse wavelet transform to obtain the final waveform, which filters out the high-frequency noise in the original waveform while maintaining the time domain integrity of the pulse waveform. The spectral distribution characteristics are determined by analyzing the final waveform to eliminate the problem of incorrect selection of the basis function caused by ignoring the spectral characteristics. According to the spectral distribution characteristics, the wavelet basis function is determined to improve the signal-to-noise ratio in the low frequency band and avoid signal distortion of the fixed wavelet basis under strong interference. The blood oxygen saturation data is obtained through wavelet denoising processing based on the wavelet basis function, which helps to meet the input quality requirements of the complication prediction model.

[0074] In some embodiments, the physiological indicator detection data is parsed to determine the patient's respiratory frequency range; the optimal time-frequency resolution is determined according to the patient's respiratory frequency range; the mutation point density of the blood oxygen saturation is determined according to the spectral distribution characteristics; the basis function support length is determined according to the mutation point density; the discretization parameters are determined according to the wearable device sampling rate; and the wavelet basis function is matched according to the optimal time-frequency resolution, the basis function support length and the discretization parameters.

[0075] The patient breathing frequency range can be a frequency value range corresponding to the patient breathing related frequency band.

[0076] The optimal time-frequency resolution can be a balance parameter of the wavelet transform time window length and the frequency bandwidth.

[0077] The blood oxygen saturation can be a percentage of oxygenated hemoglobin in blood relative to the total amount of combined hemoglobin.

[0078] The mutation point density can be the number of spectral energy mutation points per unit bandwidth in the high-frequency noise frequency band.

[0079] The basis function support length can be the effective duration of the wavelet basis function in the time domain.

[0080] The discretization parameter can be a wavelet transform discretization scale and a translation step.

[0081] Specifically, based on the continuous blood oxygen saturation waveform collected by the wearable device, the physiological index detection data is analyzed, the main frequency band energy peak value is extracted through fast Fourier transform, the breathing related frequency band is located, and the patient breathing frequency range corresponding to the frequency domain gravity center is calculated. Then, taking the upper limit of the patient breathing frequency range as the benchmark, based on the constant principle of time-frequency resolution product, the optimal time-frequency resolution is determined. Subsequently, based on the spectral distribution characteristics, the number of energy mutation points of the high-frequency noise frequency band in the final waveform power spectral density is counted, and the mutation point density of the blood oxygen saturation is calculated. Further, according to the mutation point density, the basis function support length is dynamically adjusted, if it is a high mutation point density scene, a short support length wavelet basis is selected, and the high-frequency transient noise generated by the motion artifact is suppressed through the time domain fast decay characteristic; at the same time, if it is a low mutation point density scene, a long support length wavelet basis is used. Based on the sampling rate of the wearable device, according to the scale-frequency relationship formula of wavelet transform, the discretization parameter is calculated. The optimal time-frequency resolution, the basis function support length, and the discretization parameter are input into the wavelet basis function selector, and the wavelet basis function that meets the parameter constraint is matched through the scale-frequency mapping formula of wavelet transform.

[0082] By the scheme, physiological index detection data is analyzed, and a patient respiratory frequency range is determined, avoiding signal oversmoothing or noise residual problems caused by ignoring respiratory dynamic characteristics. According to the patient respiratory frequency range, an optimal time-frequency resolution is determined, ensuring that small changes in respiratory rhythm can be accurately captured while suppressing high-frequency noise interference. According to the frequency spectrum distribution characteristics, the mutation point density of blood oxygen saturation is determined, avoiding high-frequency noise residual or excessive truncation of effective signals caused by basis functions. According to the mutation point density, the basis function support length is determined, retaining the rapid fluctuation characteristics of blood oxygen saturation, and avoiding misjudgment of medium and low frequency physiological rhythms as noise. According to the sampling rate of the wearable device, the discretization parameters are determined, avoiding frequency aliasing in the wavelet transform process, ensuring that the time-frequency decomposition result is consistent with the time resolution of the original signal, and improving the compatibility of the denoised waveform with the sampling characteristics of the wearable device. According to the optimal time-frequency resolution, the basis function support length, and the discretization parameters, the wavelet basis function is matched, realizing the rapid filtering of motion artifacts and the accurate separation of respiratory and pulse rhythms, and ensuring the universality of the denoising algorithm on different sampling rate devices.

[0083] In some embodiments, based on the wavelet basis function, the final waveform is analyzed to determine noise distribution characteristics; according to the noise distribution characteristics, the high-frequency components in each layer of coefficients that exceed a preset threshold are soft threshold compressed to obtain a denoised waveform; according to the sampling rate of the wearable device, sampling points are determined; based on the sampling points, the denoised waveform is analyzed to determine waveform gradient changes; according to the waveform gradient changes, artifact data segments are determined; the artifact data segments are subjected to local interpolation processing to obtain blood oxygen saturation data.

[0084] The noise distribution characteristics can be the energy intensity and frequency band concentration characteristics of the noise signal in the frequency domain.

[0085] The preset threshold can be a critical value preset according to the noise distribution characteristics. It is pre-stored in a server and called when used.

[0086] The high-frequency component can be a coefficient component corresponding to a high-frequency subband after wavelet decomposition.

[0087] The denoised waveform can be a blood oxygen saturation waveform after soft threshold compression processing.

[0088] The sampling points can be equally spaced time points generated according to the nominal sampling rate of the wearable device.

[0089] The waveform gradient change can be the instantaneous change rate of the blood oxygen saturation amplitude between adjacent sampling points.

[0090] The artifact data segment can be a time period composed of consecutive abnormal gradient change points.

[0091] Specifically, based on the wavelet basis function, the final waveform is subjected to multi-scale wavelet decomposition, and through multi-scale decomposition, the original signal is separated layer by layer to obtain high-frequency coefficients and low-frequency coefficients of each layer; then, the amplitude distribution standard deviation of the high-frequency coefficients is counted, and a preset noise energy threshold is established combined with the noise standard deviation estimation method in the wavelet denoising theory, so as to determine the noise distribution characteristics. Further, according to the noise distribution characteristics, the soft threshold of each decomposition level is dynamically set; then, the components with an amplitude exceeding the preset threshold in the high-frequency coefficients of each layer are subjected to nonlinear compression to suppress the high-frequency noise energy; further, the low-frequency coefficients are kept unchanged, and the wavelet coefficients are reconstructed to obtain the denoised waveform. Subsequently, according to the sampling rate of the wearable device, the time domain sequence of the denoised waveform is divided at equal interval time points to generate sampling points that are time-aligned with the original signal. Further, the amplitude change rate of the denoised waveform between adjacent sampling points is calculated to generate a waveform gradient sequence; then, the sampling points with an absolute gradient value exceeding a preset physiological fluctuation range determined according to the clinical statistical results of the physiological change rate of blood oxygen saturation are marked as waveform gradient changes. Subsequently, the waveform gradient changes are subjected to continuity detection, and if the number of continuous abnormal points exceeds a preset threshold, it is determined as an artifact data segment. Finally, the front and rear effective data boundaries of the artifact data segment are subjected to linear interpolation to replace the abnormal sampling point values in the artifact segment; thus, the interpolated and corrected continuous waveform is output as the final usable blood oxygen saturation data.

[0092] Through the scheme, based on the wavelet basis function, the final waveform is analyzed to determine the noise distribution characteristics, which helps to eliminate the noise residue or signal oversmoothing problem caused by the fact that the wavelet basis function cannot adapt to different device sampling rates. According to the noise distribution characteristics, the high-frequency components exceeding the preset threshold in the coefficients of each layer are subjected to soft threshold compression to obtain the denoised waveform, which overcomes the physiological rhythm feature blurring defect caused by the moving average method. According to the sampling rate of the wearable device, the sampling points are determined to ensure that the denoised waveform is time-aligned with the original signal, providing an accurate time reference for gradient analysis. Based on the sampling points, the denoised waveform is analyzed to determine the waveform gradient changes, quantify the instantaneous fluctuation intensity of the blood oxygen saturation, and effectively capture abnormal mutations caused by motion interference. According to the waveform gradient changes, the artifact data segment is determined to avoid misjudging temporary physiological fluctuations as noise, improving the specificity of artifact identification. The artifact data segment is subjected to local interpolation processing to obtain the blood oxygen saturation data, which helps to eliminate the effective signal distortion problem caused by global filtering and ensures the input data quality for subsequent complication prediction.

[0093] In some embodiments, a disease-complication association rule in a clinical guideline is obtained; a patient case is obtained, and an actual decision graph is constructed according to diagnosis and treatment path data in the patient case; a time decay factor is determined according to a structured data set; an association strength weight is calculated according to the time decay factor; and a patient risk evolution model containing the time decay factor and the association strength weight is constructed based on the actual decision graph and the disease-complication association rule through a graph neural network, to determine a patient risk evolution condition.

[0094] The clinical guideline can be a standardized diagnosis and treatment specification issued by an authoritative medical institution.

[0095] The disease-complication association rule can be a causal logic relationship extracted from the clinical guideline.

[0096] The patient case can be an electronic health record data of a patient individual.

[0097] The diagnosis and treatment path data can be a medical event sequence with a time stamp in the patient case.

[0098] The actual decision graph can be a directed graph with disease nodes as starting points, and complication nodes and diagnosis and treatment operation nodes connected through time-stamped edges.

[0099] The time decay factor can be a decay coefficient calculated based on a complication occurrence time interval.

[0100] The association strength weight can be an edge weight from a disease node to a complication node.

[0101] The graph neural network can be a deep learning model for processing graph structure.

[0102] The patient risk evolution model can be a prediction model that inputs the actual decision graph and the association rule, and outputs a risk level of each complication and a warning time window in a future time period.

[0103] Specifically, disease-complication association rules in clinical guidelines are extracted from authoritative medical databases. Medication records, examination timelines, and complication occurrence time nodes in patient medical records are extracted to generate timestamped diagnosis and treatment path data. Then, taking the patient as the graph center node, diagnosis and treatment events as attribute nodes, and complications as association nodes, the event-complication edges are connected in chronological order to construct an actual decision graph. Based on the time interval of complication occurrence recorded in the structured data set, a preset time decay function is established according to the time proximity principle in clinical medicine to calculate the time decay factor. The time decay factor and the patient's current age, drug adherence score are input into the weight distribution function to dynamically adjust the association strength weight of the disease node to the complication node. The actual decision graph and the disease-complication association rules are input into the graph neural network, and the time decay factor and the association strength weight are injected into the node embedding process to model the complication risk propagation path through the graph attention mechanism. Then, based on the node state output by the graph neural network, the risk level and warning time window of each complication in the future time period are determined. Then, the high-probability complication type and the timing risk distribution under the current disease state are identified, and the patient's risk evolution is output.

[0104] By the scheme, the disease-complication association rules in clinical guidelines are obtained, the clinical knowledge disconnection problem caused by relying on static rules is eliminated, and global constraint conditions are provided for the risk propagation path. The patient case is obtained, the actual decision graph is constructed according to the diagnosis and treatment path data in the patient case, and the defect of the complication association edge without time sequence attribute is eliminated. According to the structured data set, the time decay factor is determined, the dynamic contribution degree of the historical complication to the current risk is quantified, and the importance of the recent event caused by the weight is avoided. According to the time decay factor, the association strength weight is calculated, and the adaptability of the patient risk evolution model to individualized disease course characteristics is enhanced. Based on the actual decision graph and the disease-complication association rules, the patient risk evolution model containing the time decay factor and the association strength weight is constructed through the graph neural network, the patient risk evolution is determined, which helps to eliminate the defect of ignoring the chain evolution of the disease, and avoids the intervention failure caused by the risk lag evaluation.

[0105] In some embodiments, according to the disease-complication association rules, the complication occurrence time is determined; according to the patient case, the current time interval from the occurrence of the complication of the patient to the current time is determined; and according to the complication occurrence time and the current time interval, the time decay factor is determined.

[0106] The complication occurrence time can be the exact time point when the complication is determined to occur.

[0107] The occurrence of the complication of the patient can be that the patient actually occurs the complication event meeting the clinical diagnostic criteria.

[0108] The current time interval can be a time length difference from a complication occurrence time stamp to a current time.

[0109] Specifically, according to a complication trigger condition defined in the disease-complication association rule, a complication occurrence time is extracted from a complication record table of the structured data set. Then, based on the patient case, a time interval from a complication occurrence to a current time is calculated for each complication event. Further, the complication occurrence time and the time interval are input into a preset time decay function, and a decay rate parameter set according to a complication recurrence cycle in the clinical guideline is combined to determine a time decay factor.

[0110] Through the scheme, the complication occurrence time is determined according to the disease-complication association rule, and the interval calculation error caused by ambiguous time records is avoided. According to the patient case, the current time interval from the complication occurrence to the current time is determined, and a calculable discrete input is provided for the time decay function to replace the qualitative description of the recent / future. According to the complication occurrence time and the current time interval, the time decay factor is determined, which is helpful to intuitively reflect the medical prior knowledge that the recent complication has a greater impact on the current risk prediction.

[0111] In some embodiments, the medication record, the mapping relationship, and the lifestyle are analyzed to determine drug compliance data; the patient case is analyzed to determine the patient age; the drug compliance data, the patient age, and the lifestyle are used to adjust the time decay factor; and the association degree weight is calculated according to the adjusted time decay factor.

[0112] The drug compliance data can be a numerical indicator quantitatively reflecting the standard deviation of the actual medication behavior of the patient from the medical order.

[0113] The patient age can be a physiological age value of the patient accurate to a year.

[0114] The association degree weight can be a quantitative parameter representing the association strength of the disease and the complication at the current time point.

[0115] Specifically, the drug name, the dose, and the taking time sequence are extracted from the medication record, the overlap degree of the actual medication time window and the medical order time window is calculated based on the mapping relationship of the ATC classification, and the dose deviation value is quantified; then, the compliance score is corrected in combination with the abnormal event in the lifestyle, and the drug compliance data is output. The birth date field is extracted from the basic information of the patient case, so that the patient age is calculated according to the current date. Subsequently, the time decay factor is linearly weighted with the drug compliance data, the patient age, and the lifestyle to calculate the adjusted time decay factor. Further, the association weight is read from the disease-complication association rule library, the association weight is normalized with the time decay factor, and the final association degree weight is determined.

[0116] By the scheme, the medication record, the mapping relationship and the lifestyle are analyzed, the drug compliance data is determined, the hidden dose reduction behavior is identified, and the problem of ignoring the metabolic interaction is eliminated. The patient case is analyzed, the patient age is determined, and the metabolic rate estimation deviation caused by rounding is avoided. According to the drug compliance data, the patient age and the lifestyle, the time attenuation factor is adjusted, which helps to break through the limitation of the attenuation rate. According to the adjusted time attenuation factor, the correlation degree weight is calculated, and the problem that a single factor excessively affects the overall risk assessment is avoided.

[0117] In some embodiments, the physiological index detection data is analyzed to determine a physiological index deviation degree; the intervention priority is determined according to the physiological index deviation degree; the medication record is analyzed to determine an actual medication amount; the patient case is analyzed to determine a theoretical medication amount; the dose deviation is determined according to the actual medication amount and the theoretical medication amount; the drug suggestion is adjusted according to the dose deviation and the patient risk evolution; the real-time behavior information and the lifestyle of the patient are analyzed to determine a lifestyle change trend; the behavior intervention template is matched according to the lifestyle change trend; and the dynamic intervention scheme is generated based on the intervention template, the lifestyle change trend and the drug suggestion.

[0118] The physiological index deviation degree can be a quantitative index reflecting the severity of the deviation of the current physiological state from the normal range.

[0119] The intervention priority can be a step response level for determining the disposal order and resource allocation weight of different health events.

[0120] The actual medication amount can be the total amount of the drug actually taken by the patient.

[0121] The theoretical medication amount can be the total amount of the drug expected to be taken.

[0122] The dose deviation can be the percentage difference between the actual medication amount and the theoretical medication amount.

[0123] The drug suggestion can be the drug risk evolution.

[0124] The lifestyle change trend can be a parameter quantitatively describing the change direction and rate of the behavior pattern.

[0125] The behavior intervention template can be a pre-set step intervention scheme library.

[0126] Specifically, the blood oxygen saturation waveform of the physiological index detection data is obtained, time domain features and frequency domain features are calculated by using a sliding window algorithm, then the time domain features and the frequency domain features are compared with a dynamic threshold interval set by a clinical guideline to generate a physiological index deviation degree, a weight value with a time attenuation factor in a disease-complication association rule library is called, then the physiological index deviation degree and a real-time complication risk level are input into a preset priority matrix to generate an intervention priority. Based on a mapping relationship of ATC classification, a trade name in a medication record is converted into a standard dosage unit to count an actual medication amount. Then, prescription data in a patient case is analyzed, a liver and kidney function adjustment coefficient is combined to output a theoretical medication amount. Then, based on the actual medication amount and the theoretical medication amount, a medication deviation is calculated according to a drug sensitivity coefficient defined in the ATC classification. Then, the weight value in the disease-complication association graph is dynamically adjusted based on the time attenuation factor, and the drug suggestion is adjusted according to the complication probability promotion predicted based on the dose deviation and the patient risk evolution. Subsequently, a sliding window is used to count real-time behavior information and lifestyle of the patient, a linear regression is used to fit a slope value of each behavior index to determine a lifestyle change trend. An intervention level is selected according to the direction and intensity of the lifestyle change trend. Finally, the adjusted drug suggestion is logically coupled with a matched behavior intervention template, and a clinical rationality checking rule is added to determine a dynamic intervention scheme.

[0127] By the scheme, physiological index detection data is analyzed to determine a physiological index deviation degree, accurate quantification of physiological state abnormalities is realized, and the misjudgment problem caused by noise residues is eliminated. According to the physiological index deviation degree, an intervention priority is determined to avoid acute event response lag. Medication records are analyzed to determine an actual medication amount, and the missing problem of pharmacokinetic analysis caused by isolated storage is eliminated. According to a patient case, a theoretical medication amount is determined to avoid dose recommendation deviation caused by neglecting physiological state changes of the patient. According to the actual medication amount and the theoretical medication amount, a dose deviation is determined, which is helpful to identify high-risk deviation types and enhance the targeting of intervention suggestions. According to the dose deviation and the patient risk evolution, the drug suggestion is adjusted to eliminate the distortion problem of drug efficacy evaluation caused by static weights. Real-time behavior information and lifestyle of the patient are analyzed to determine a lifestyle change trend, and the response blind area to sudden behavior changes is broken through. According to the lifestyle change trend, a behavior intervention template is matched, and based on the intervention template, a dynamic intervention scheme is generated according to the lifestyle change trend and the drug suggestion to ensure the executability and clinical safety of the scheme.

[0128] Figure 3 A structural schematic diagram of a chronic disease follow-up data AI analysis system based on 5G provided by an embodiment of the present application is shown in Figure 3As shown, the 5G-based chronic disease follow-up data AI analysis system 300 of the embodiment comprises a report acquisition module 301, a form analysis module 302, a data integration module 303, a data analysis module 304, and a scheme generation module 305.

[0129] The report acquisition module 301 is configured to acquire a follow-up report uploaded by a patient through a low-latency data transmission channel provided by a 5G network, and parse the follow-up report to determine a self-rating scale, physiological index detection data, and medication record. The form analysis module 302 is configured to analyze the self-rating scale to determine a lifestyle of the patient. The data integration module 303 is configured to retrieve historical complication data of the patient from a database, and perform multi-modal data integration on the physiological index detection data, the lifestyle, the historical complication data, and the medication record through the 5G network to obtain a structured data set. The data analysis module 304 is configured to analyze the structured data set based on a dynamic knowledge graph to determine a patient risk evolution condition. The scheme generation module 305 is configured to acquire real-time behavior information of the patient, and generate a dynamic intervention scheme based on the patient risk evolution condition and the real-time behavior information of the patient.

[0130] Optionally, when the data integration module 303 performs multi-modal data integration on the physiological index detection data, the lifestyle, the historical complication data, and the medication record through the 5G network to obtain a structured data set, the data integration module 303 is configured to: parse the physiological index detection data to determine a blood oxygen saturation waveform; perform wavelet denoising processing on the blood oxygen saturation waveform to obtain blood oxygen saturation data; parse the medication record to determine a drug name; establish a mapping relationship between the drug and an ATC classification based on the drug name; perform multi-modal data integration on the blood oxygen saturation data, the mapping relationship, the lifestyle, and the historical complication data through the 5G network to obtain a structured data set containing time series features.

[0131] Optionally, when the data integration module 303 performs wavelet denoising processing on the blood oxygen saturation waveform to obtain blood oxygen saturation data, the data integration module 303 is configured to: acquire basic information of a wearable device, and determine a sampling rate of the wearable device based on the basic information; determine a maximum decomposition level constraint of the blood oxygen saturation waveform based on the sampling rate of the wearable device. Reconstruct the blood oxygen saturation waveform according to the maximum decomposition layer constraint through inverse wavelet transform to obtain a final waveform; Analyze the final waveform to determine a spectral distribution feature; Determine a wavelet base function according to the spectral distribution feature; Perform wavelet denoising processing according to the wavelet base function to obtain blood oxygen saturation data.

[0132] Optionally, when the data integration module 303 determines the wavelet base function according to the spectral distribution feature, it is used for: Analyzing the physiological index detection data to determine a patient breathing frequency range; Determining an optimal time-frequency resolution according to the patient breathing frequency range; Determining a mutation point density of blood oxygen saturation according to the spectral distribution feature; Determining a base function support length according to the mutation point density; Determining a discretization parameter according to the wearable device sampling rate; Matching a wavelet base function according to the optimal time-frequency resolution, the base function support length, and the discretization parameter.

[0133] Optionally, when the data integration module 303 performs wavelet denoising processing according to the wavelet base function to obtain blood oxygen saturation data, it is used for: Analyzing the final waveform based on the wavelet base function to determine a noise distribution feature; Performing soft threshold compression on high-frequency components exceeding a preset threshold in each layer coefficient according to the noise distribution feature to obtain a denoised waveform; Determining a sampling point according to the wearable device sampling rate; Analyzing the denoised waveform based on the sampling point to determine a waveform gradient change; Determining an artifact data segment according to the waveform gradient change; Performing local interpolation processing on the artifact data segment to obtain blood oxygen saturation data.

[0134] Optionally, when the data analysis module 304 analyzes the structured data set based on a dynamic knowledge graph to determine a patient risk evolution, it is used for: Obtaining disease-complication association rules in a clinical guideline; Obtaining a patient case and constructing an actual decision graph according to diagnosis and treatment path data in the patient case; Determining a time decay factor according to the structured data set; Calculating an association strength weight according to the time decay factor; Based on the actual decision graph, the disease-complication association rule, and through a graph neural network, a patient risk evolution model containing the time decay factor and the association strength weight is constructed to determine the patient risk evolution condition.

[0135] Optionally, when the data analysis module 304 determines the time decay factor according to the structured data set, it is used for: determining the complication occurrence time according to the disease-complication association rule; determining a current time interval from the complication occurrence time to the current time according to the patient case; determining the time decay factor according to the complication occurrence time and the current time interval.

[0136] Optionally, when the data analysis module 304 calculates the association strength weight according to the time decay factor, it is used for: analyzing the medication record, the mapping relationship, and the lifestyle to determine drug adherence data; analyzing the patient case to determine the patient age; adjusting the time decay factor according to the drug adherence data, the patient age, and the lifestyle; calculating the association strength weight according to the adjusted time decay factor.

[0137] Optionally, when the scheme generation module 305 generates a dynamic intervention scheme according to the patient risk evolution condition and the patient real-time behavior information, it is used for: analyzing the physiological index detection data to determine a physiological index deviation; determining an intervention priority according to the physiological index deviation; analyzing the medication record to determine an actual medication amount; determining a theoretical medication amount according to the patient case; determining a dose deviation according to the actual medication amount and the theoretical medication amount; adjusting a drug suggestion according to the dose deviation and the patient risk evolution condition; analyzing the patient real-time behavior information and the lifestyle to determine a lifestyle change trend; matching a behavior intervention template according to the lifestyle change trend, and generating a dynamic intervention scheme based on the intervention template, the lifestyle change trend, and the drug suggestion.

[0138] The system of the embodiment can be used to execute the method of any of the above embodiments, and has similar implementation principles and technical effects, which will not be described here again.

Claims

1. A 5G-based chronic disease follow-up data AI analysis method, characterized in that, The method comprises the following steps: acquiring a follow-up report uploaded by a patient through a low-delay data transmission channel provided by a 5G network; analyzing the follow-up report to determine a self-assessment scale, physiological index detection data, and medication record; analyzing the self-assessment scale to determine the patient's lifestyle; calling historical complication data of the patient from a database, and integrating multi-modal data through the 5G network based on the physiological index detection data, the lifestyle, the historical complication data, and the medication record to obtain a structured data set; analyzing the structured data set based on a dynamic knowledge graph to determine the patient's risk evolution; 2. The method of claim 1, wherein, acquiring real-time behavior information of the patient, and generating a dynamic intervention scheme based on the patient's risk evolution and the real-time behavior information of the patient. The method for integrating multi-modal data through the 5G network based on the physiological index detection data, the lifestyle, the historical complication data, and the medication record to obtain a structured data set comprises the following steps: analyzing the physiological index detection data to determine an oxygen saturation waveform; performing wavelet denoising processing on the oxygen saturation waveform to obtain oxygen saturation data; analyzing the medication record to determine a drug name; establishing a mapping relationship between the drug and the ATC classification based on the drug name; 3. The method of claim 2, wherein, integrating multi-modal data through the 5G network based on the oxygen saturation data, the mapping relationship, the lifestyle, and the historical complication data to obtain a structured data set containing time series features. The method for performing wavelet denoising processing on the oxygen saturation waveform to obtain oxygen saturation data comprises the following steps: acquiring basic information of a wearable device, and determining a wearable device sampling rate based on the basic information; determining a maximum decomposition layer constraint of the oxygen saturation waveform based on the wearable device sampling rate; reconstructing the oxygen saturation waveform based on the maximum decomposition layer constraint through inverse wavelet transform to obtain a final waveform; analyzing the final waveform to determine a frequency spectrum distribution feature; determining a wavelet basis function based on the frequency spectrum distribution feature; 4. The method of claim 3, wherein, performing wavelet denoising processing based on the wavelet basis function to obtain oxygen saturation data. The method for determining a wavelet basis function based on the frequency spectrum distribution feature comprises the following steps: analyzing the physiological index detection data to determine a patient breathing frequency range; determining an optimal time-frequency resolution based on the patient breathing frequency range; determining a mutation point density of oxygen saturation based on the frequency spectrum distribution feature; determining a basis function support length based on the mutation point density; determining a discretization parameter based on the wearable device sampling rate; 5. The method of claim 3, wherein, matching a wavelet basis function based on the optimal time-frequency resolution, the basis function support length, and the discretization parameter. The method for performing wavelet denoising processing based on the wavelet basis function to obtain oxygen saturation data comprises the following steps: analyzing the final waveform based on the wavelet basis function to determine a noise distribution feature; performing soft threshold compression on high-frequency components exceeding a preset threshold in each layer coefficient based on the noise distribution feature to obtain a denoised waveform; determining a sampling point based on the wearable device sampling rate; analyzing the denoised waveform based on the sampling point to determine a waveform gradient change; determining an artifact data segment based on the waveform gradient change; The artifact data segment is locally interpolated to obtain blood oxygen saturation data.

6. The method of claim 2, wherein, The structured data set is analyzed based on the dynamic knowledge graph to determine the patient risk evolution condition, including: Obtain disease-complication association rules in the clinical guidelines; Obtain patient cases, and construct an actual decision graph based on the diagnosis and treatment path data in the patient cases; Determine a time decay factor based on the structured data set; Calculate an association strength weight based on the time decay factor; Based on the actual decision graph and the disease-complication association rules, a patient risk evolution model containing the time decay factor and the association strength weight is constructed through a graph neural network to determine the patient risk evolution condition.

7. The method of claim 6, wherein, Determine a time decay factor based on the structured data set, including: Determine the complication occurrence time based on the disease-complication association rules; Determine the current time interval from the occurrence of the complication to the current time based on the patient case; Determine the time decay factor based on the complication occurrence time and the current time interval.

8. The method of claim 6, wherein, The association strength weight is calculated based on the time decay factor, including: Analyze the medication record, the mapping relationship, and the lifestyle to determine drug adherence data; Analyze the patient case to determine the patient's age; Adjust the time decay factor based on the drug adherence data, the patient's age, and the lifestyle; Calculate the association degree weight based on the adjusted time decay factor.

9. The method of claim 6, wherein, Generate a dynamic intervention scheme based on the patient risk evolution condition and the patient real-time behavior information, including: Analyze the physiological index detection data to determine the physiological index deviation; Determine the intervention priority based on the physiological index deviation; Analyze the medication record to determine the actual medication amount; Determine the theoretical medication amount based on the patient case; Determine the dose deviation based on the actual medication amount and the theoretical medication amount; Adjust the drug recommendation based on the dose deviation and the patient risk evolution condition; Analyze the patient real-time behavior information and the lifestyle to determine the lifestyle change trend; Match the behavior intervention template based on the lifestyle change trend, and generate a dynamic intervention scheme based on the lifestyle change trend and the drug recommendation based on the intervention template. 10.A 5G-based chronic disease follow-up data AI analysis system applied to the method of any one of claims 1-9, characterized in that, Including: The report acquisition module is used to acquire the follow-up report uploaded by the patient through the low-delay data transmission channel provided by the 5G network; analyze the follow-up report to determine the self-evaluation scale, physiological index detection data, and medication record; The form analysis module is used to analyze the self-evaluation scale to determine the patient's lifestyle; The data integration module is used to retrieve the patient's historical complication data from the database, and integrate the physiological index detection data, the lifestyle, the historical complication data, and the medication record through the 5G network to obtain a structured data set; The data analysis module is used to analyze the structured data set based on the dynamic knowledge graph to determine the patient risk evolution condition; The scheme generation module is used to obtain patient real-time behavior information, and generate a dynamic intervention scheme based on the patient risk evolution condition and the patient real-time behavior information.

Images