Endocrine patient health management system based on big data

Abstract

The invention relates to the technical field of medical health management, and discloses an endocrine patient health management system based on big data. According to the system, a multi-source data acquisition module is adopted to acquire heterogeneous health data including physiological monitoring indexes, medication record data, medical image reports and the like, and living habit logs can also be included. The physiological feature noise reduction module is used for extracting steady-state physiological features, the multi-modal fusion module is used for generating a focus dynamic map, and the medication feedback analysis module is used for generating a drug response mode evolution sequence. The data are input into a multi-dimensional decision model, and a health management scheme is output by a personalized scheme generation module. And the living habit analysis module analyzes the living habit logs to generate influence factors, and participates in making a health management scheme. The system realizes multi-source data fusion analysis, can provide a personalized health management scheme for an endocrine patient, improves the disease diagnosis accuracy and treatment effect, and improves the life quality of the patient.

Description

Technical Field

[0001] The present invention relates to the field of medical health management technology, and specifically to an endocrine patient health management system based on big data. Background Art

[0002] Endocrine diseases are common and complex conditions that have a profound impact on patients' physical health and quality of life. Common endocrine diseases, such as diabetes and thyroid disease, require not only long-term treatment and management but also strict control of diet, sleep schedule, and exercise in daily life. However, traditional endocrine disease management models have numerous limitations and are unable to meet patients' growing health management needs.

[0003] In terms of data collection, the previous methods were relatively simple. Patients could usually only obtain physiological data during regular checkups in the hospital. This intermittent collection method could not fully and real-time reflect the patient's health status. For example, blood sugar levels can fluctuate greatly throughout the day due to factors such as diet and exercise. It is difficult to capture these subtle changes relying solely on occasional testing in the hospital, making it difficult for doctors to formulate accurate treatment plans based on incomplete data. At the same time, the traditional model lacks effective means of collecting patient lifestyle data, such as dietary preferences and exercise frequency, resulting in the neglect of this information, which is crucial for the management of endocrine diseases.

[0004] From the perspective of data analysis, traditional methods cannot fully tap the potential value of endocrine patient data. There are complex correlations between endocrine patients' physiological monitoring indicators, medication records, medical imaging and other data, but traditional analysis methods are often limited to simple analysis of a single data type and cannot comprehensively consider the interrelationships between multiple data types. For example, when evaluating the efficacy of a drug, only the relationship between drug dosage and blood sugar control effect is focused on, while ignoring the impact of factors such as the patient's lifestyle and hormone level changes on the efficacy of the drug. This leads to inaccurate evaluation of the effect of drug treatment and the inability to adjust the treatment plan in a timely manner.

[0005] Traditional health management plans lack personalization. Because each patient's condition, constitution, and lifestyle vary, a universal health management plan can't meet individual needs. For example, the same diet and exercise recommendations for obese patients with diabetes and those with normal body types may not achieve the desired therapeutic effect. Obese patients may require stricter dietary control and increased exercise, while those with normal body types may need a different plan based on their insulin secretion.

[0006] Furthermore, traditional endocrine disease management relies primarily on patients proactively seeking medical attention, lacking a long-term, ongoing follow-up management mechanism. This makes it difficult to receive timely attention and intervention for patients' health outside the hospital, which can easily lead to recurrence or worsening of their condition. For example, if patients fail to take medication or adjust their lifestyle according to doctor's orders after discharge, doctors will be unable to detect and provide timely guidance, potentially exacerbating their condition.

[0007] The rapid development of technologies such as big data and artificial intelligence has brought new opportunities for the health management of endocrine patients. However, current practical applications of these technologies still face numerous deficiencies, and a comprehensive and efficient big data-based health management system for endocrine patients has yet to be established. Therefore, developing a system that can integrate multi-source data, conduct in-depth analysis of data correlations, and provide personalized health management solutions is of great practical significance. Summary of the Invention

[0008] The purpose of the present invention is to provide an endocrine patient health management system based on big data to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an endocrine patient health management system based on big data, the system comprising:

[0010] Multi-source data acquisition module: used to obtain a heterogeneous set of health data from endocrine patients; the heterogeneous health data includes physiological monitoring indicators, medication record data, and medical imaging reports. The physiological monitoring indicators include blood glucose fluctuation curves and hormone level time series data, and the medication record data includes drug dosage association information and side effect feedback labels.

[0011] A physiological characteristic noise reduction module is configured to extract steady-state physiological characteristics based on the physiological monitoring indicators through an adaptive filtering algorithm, wherein the steady-state physiological characteristics include a baseline fluctuation threshold, an abnormal peak interval, and a metabolic rate parameter;

[0012] Multimodal fusion module: used to generate a dynamic map of lesions based on the medical imaging report through a cross-modal alignment algorithm, wherein the dynamic map includes organ tissue association paths and thermal weights of lesion areas;

[0013] Medication feedback analysis module: performs timestamp processing on the side effect feedback labels to generate a drug response pattern evolution sequence;

[0014] Personalized plan generation module: inputs the steady-state physiological characteristics, dynamic maps and drug response pattern evolution sequence into a multi-dimensional decision-making model to output a patient health management plan.

[0015] Preferably, the extracting of steady-state physiological characteristics by an adaptive filtering algorithm includes:

[0016] Performing baseline drift correction processing on the physiological monitoring indicators to generate denoised physiological signals;

[0017] Based on the preset steady-state reference model, the sliding window matching algorithm is used to identify physiological abnormal segments;

[0018] Wavelet transform algorithm is used to extract the frequency domain distribution characteristics of abnormal segments, and the metabolic deviation is generated by combining the baseline parameters in the steady-state reference model;

[0019] The metabolic deviation, baseline fluctuation threshold, and abnormal peak interval were coded as steady-state physiological characteristics.

[0020] Preferably, the generating of a dynamic lesion map by a cross-modal alignment algorithm comprises:

[0021] Performing feature segmentation processing on the medical imaging report to eliminate irrelevant background noise data;

[0022] The core lesion area of ​​the image is divided based on the region growing algorithm, and the spatial correlation between each area is calculated;

[0023] Generate tissue connection strength based on spatial correlation and preset thermal threshold;

[0024] The tissue connection strength is combined with the thermal weight of the lesion area to form a multidimensional dynamic map.

[0025] Preferably, the multi-dimensional decision model includes a feature fusion layer and a strategy optimization layer, and the feature fusion layer includes:

[0026] Normalizing the metabolic rate parameter in the steady-state physiological characteristics to obtain a first fusion vector;

[0027] Performing discrete Fourier transform encoding on the tissue connection strength in the dynamic atlas to generate a second fusion vector;

[0028] performing differential autoregressive calculation on the drug response pattern evolution sequence, extracting the pattern change gradient, and obtaining a third fusion vector;

[0029] The first fusion vector, the second fusion vector and the third fusion vector are combined into a low-dimensional decision sequence through weighted superposition.

[0030] Preferably, the strategy optimization layer includes:

[0031] Align the spatial dimensions of the low-dimensional decision sequence to generate a strategy weight matrix;

[0032] Extract regional priority features through the convolutional attention mechanism and generate a priority mapping matrix;

[0033] Perform matrix dot product operation on the policy weight matrix and the priority mapping matrix to generate policy optimization features;

[0034] The strategy optimization features are superimposed on the original low-dimensional decision sequence through residual connections to output a health management plan.

[0035] Preferably, the method for generating the drug response pattern evolution sequence includes:

[0036] Initialize the label weight according to the type of side effect feedback label and generate a temporal correlation matrix based on the timestamp;

[0037] The time correlation matrix is ​​used as a pattern evolution input to calculate the state transition probability through a hidden Markov model;

[0038] Update the pattern confidence of each time node through the Bayesian network and adjust the label weight;

[0039] Generate a sequence of drug response pattern evolution based on the adjusted marker weights.

[0040] Preferably, the method for constructing the steady-state reference model includes:

[0041] Collect standard physiological samples from patients with various typical endocrine diseases and extract baseline metabolic curves and frequency domain distribution characteristics;

[0042] Perform multi-scale decomposition on the baseline metabolic curve to generate a multi-resolution metabolic model;

[0043] Classify metabolic models according to disease types and associate them with a standard parameter database;

[0044] The classified metabolic model is stored as a steady-state reference model and is regularly updated based on newly collected samples.

[0045] Preferably, the parameter optimization method of the region growing algorithm includes:

[0046] Calculate the initial growth threshold and minimum regional density based on the distribution of historical image data;

[0047] The grid search algorithm is used to traverse the parameter combinations and select the parameters with the highest overlap between the segmentation results and the expert-annotated areas;

[0048] The growth threshold and minimum regional density are dynamically adjusted according to the degree of overlap to optimize the accuracy of tissue connection strength division.

[0049] Preferably, the heterogeneous health data also includes the user's life habit log.

[0050] Preferably, it also includes a lifestyle analysis module:

[0051] Performing time-series segmentation processing on the lifestyle log to extract behavior pattern labels and abnormal work and rest intervals;

[0052] Identify the correlation weights between lifestyle habits and physiological indicators based on association rule mining algorithms;

[0053] The correlation weight is integrated with the behavior pattern label to generate a lifestyle influence factor;

[0054] The multi-dimensional decision-making model is used to incorporate the factors affecting living habits into the process of generating health management plans.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] During the data collection phase, the multi-source data acquisition module can acquire a heterogeneous collection of health data from endocrine patients, covering physiological monitoring indicators, medication records, medical imaging reports, and even the user's lifestyle logs. This comprehensive data collection method overcomes the single-minded nature of traditional models. Taking blood glucose monitoring as an example, the system can collect the patient's blood glucose fluctuation curve in real time, recording the data at any time through wearable devices. Combined with time-series data on hormone levels, doctors can gain a more comprehensive understanding of the dynamic changes in the patient's endocrine system. The inclusion of lifestyle logs is also crucial, as they record the patient's daily behaviors, such as eating habits, exercise frequency, and sleep schedules, all of which are closely related to the development and control of endocrine diseases. The comprehensive collection of this data provides a solid data foundation for subsequent precise analysis and the development of personalized plans.

[0057] During the data processing and analysis phase, the physiological feature noise reduction module, the multimodal fusion module, and the medication feedback analysis module played an important role. The physiological feature noise reduction module uses an adaptive filtering algorithm to extract steady-state physiological features, such as baseline fluctuation thresholds, abnormal peak intervals, and metabolic rate parameters. This helps doctors accurately judge the patient's physiological state and promptly identify potential health issues. The multimodal fusion module generates a dynamic lesion map based on medical imaging reports, including organ-tissue association pathways and thermal weights of lesion areas. This intuitively presents the lesion status and its relationship with surrounding tissues, greatly improving the accuracy and efficiency of disease diagnosis. The medication feedback analysis module processes side effect feedback labels and generates a drug response pattern evolution sequence, allowing doctors to dynamically observe patients' reactions to drugs, evaluate drug efficacy and safety, and provide a strong basis for rationally adjusting medication regimens.

[0058] For the formulation of health management plans, the personalized plan generation module is the key. This module inputs steady-state physiological characteristics, dynamic maps and drug response pattern evolution sequences into a multi-dimensional decision-making model, and the output health management plan fully takes into account the individual differences of patients. For patients with different endocrine diseases, such as diabetes and thyroid disease, exclusive comprehensive management plans such as diet, exercise, and medication are formulated according to their respective physiological characteristics, disease development, and response to drugs. For diabetic patients with poor blood sugar control and exercise habits, the system may adjust the exercise plan and drug dosage according to their metabolic rate parameters and drug response to achieve better blood sugar control. This personalized management plan can significantly improve the treatment effect and improve the quality of life of patients.

[0059] Furthermore, the system's lifestyle analysis module further enhances the comprehensiveness of health management. By analyzing lifestyle logs, it extracts behavioral pattern labels and abnormal sleep and rest intervals, identifies correlation weights between lifestyle habits and physiological indicators, and generates lifestyle influencing factors, which are incorporated into the health management plan generation process. This allows doctors to fully consider the impact of lifestyle habits on disease when formulating plans, guiding patients to develop healthy lifestyles and assisting in the treatment and prevention of diseases at their root.

[0060] The system of this invention enables lifelong health management for endocrine patients, breaking the time and space constraints of traditional management models. Patients can record and upload their health data anytime and anywhere using various devices. Doctors can access and analyze this data in real time, adjust treatment plans promptly, and continuously track and intervene in patients' health status, effectively reducing disease risks and improving their health. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a working principle diagram of the endocrine patient health management system based on big data according to the present invention;

[0062] Figure 2 Flowchart for extracting steady-state physiological features for the adaptive filtering algorithm;

[0063] Figure 3 This is the workflow diagram of the multi-dimensional decision model-feature fusion layer;

[0064] Figure 4 Flowchart generated for the drug response pattern evolution sequence. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0066] See also Figure 1-Figure 4 The present invention provides a big data-based endocrine patient health management system, which aims to achieve precise health management for endocrine patients. The specific implementation method is described in detail below.

[0067] The multi-source data acquisition module is responsible for acquiring a heterogeneous collection of health data from endocrine patients. This heterogeneous health data includes physiological monitoring indicators, medication records, medical imaging reports, and, in some embodiments, the user's lifestyle log. Physiological monitoring indicators include blood glucose fluctuation curves and hormone level time series data. These data reflect the patient's real-time physiological state. For example, blood glucose fluctuation curves can visually display the changes in a patient's blood glucose at different time points, while hormone level time series data can reflect the dynamic changes of various hormones in the endocrine system. Medication record data includes medication dosage information and side effect feedback tags. The medication dosage information records the specific dosage and frequency of medication used by the patient, while the side effect feedback tags record various side effects experienced by the patient after medication use. Medical imaging reports are used for subsequent analysis of the patient's internal organs and tissues to determine whether there are pathological conditions. The lifestyle log records the patient's daily habits, such as diet, exercise, and daily routine.

[0068] The physiological feature denoising module extracts steady-state physiological features based on physiological monitoring indicators using an adaptive filtering algorithm. Specifically, the physiological monitoring indicators are first subjected to baseline drift correction to remove baseline drift caused by equipment errors or other factors, generating a denoised physiological signal, making the signal for subsequent analysis more accurate and reliable. Then, based on a preset steady-state reference model, a sliding window matching algorithm is used to identify abnormal physiological segments. This algorithm slides a fixed-size window across the denoised physiological signal and matches it with the steady-state reference model to identify segments that do not conform to normal physiological laws. A wavelet transform algorithm is then used to extract the frequency domain distribution characteristics of the abnormal segments, and the metabolic deviation is generated by combining the baseline parameters in the steady-state reference model. Finally, the metabolic deviation, baseline fluctuation threshold, and abnormal peak interval are encoded as steady-state physiological features. These steady-state physiological features can effectively reflect the stable state of the patient's body metabolism and provide an important basis for the formulation of subsequent health management plans.

[0069] The multimodal fusion module generates a dynamic map of lesions based on the medical imaging report through a cross-modal alignment algorithm. In this process, the medical imaging report is first subjected to feature segmentation processing to remove irrelevant background noise data to avoid interference with the subsequent analysis of the lesion area. The core lesion area of ​​the image is then divided based on the region growing algorithm, and the spatial correlation between the regions is calculated to determine the relationship between the lesion areas. The tissue connection strength is then generated based on the spatial correlation and the preset thermal threshold. This tissue connection strength can reflect the degree of connection between different tissue areas. Finally, the tissue connection strength is combined with the thermal weight of the lesion area to form a multidimensional dynamic map. The map can intuitively display the location, range and relationship with the surrounding tissues of the lesion area, providing doctors with more comprehensive information for judging the condition.

[0070] The medication feedback analysis module timestamps side effect feedback tags to generate a drug response pattern evolution sequence. By adding time information to the side effect feedback tags and constructing a temporal correlation matrix, the module then uses a hidden Markov model to calculate state transition probabilities. Furthermore, the module uses a Bayesian network to update the pattern confidence at each time node and adjust the tag weights. Ultimately, this generates a drug response pattern evolution sequence that reflects how a patient's body reacts to medication over time.

[0071] The personalized plan generation module inputs steady-state physiological characteristics, dynamic maps, and drug response pattern evolution sequences into a multidimensional decision model to output a patient health management plan. The multidimensional decision model consists of a feature fusion layer and a strategy optimization layer. In the feature fusion layer, the metabolic rate parameters in the steady-state physiological characteristics are normalized to generate a first fusion vector; the tissue connectivity strength in the dynamic map is encoded using a discrete Fourier transform to generate a second fusion vector; the drug response pattern evolution sequence is subjected to differential autoregression to extract the pattern change gradient to generate a third fusion vector; and the first, second, and third fusion vectors are combined into a low-dimensional decision sequence through weighted superposition. In the strategy optimization layer, the low-dimensional decision sequence is spatially aligned to generate a policy weight matrix. Regional priority features are extracted using a convolutional attention mechanism to generate a priority mapping matrix. A matrix dot product operation is performed between the policy weight matrix and the priority mapping matrix to generate a policy optimization feature. The policy optimization feature is then superimposed with the original low-dimensional decision sequence through residual connections to output a health management plan that fully considers all patient data characteristics and is highly personalized and targeted.

[0072] The present invention is further described in detail below with reference to specific embodiments:

[0073] Example 1:

[0074] This embodiment describes in detail the specific process of extracting steady-state physiological features through the adaptive filtering algorithm in the physiological feature noise reduction module.

[0075] In a practical application scenario, consider patient data from a hospital's endocrinology department. Assume that the collected physiological monitoring indicator data includes blood glucose fluctuation curves collected every 15 minutes for a week and time series data on hormone levels collected every two hours. First, baseline drift correction is performed on these physiological monitoring indicator data. Because baseline drift may occur during long-term operation of the data acquisition equipment, the collected data may contain certain errors. A baseline drift correction method based on polynomial fitting is used to fit a baseline curve based on the overall trend of the data over a period of time. This baseline curve is then subtracted from the original data to obtain a de-noised physiological signal.

[0076] Based on the preset steady-state reference model, physiological abnormality segments are identified through a sliding window matching algorithm. The steady-state reference model is pre-built and contains standard physiological sample characteristics of patients with a variety of typical endocrine diseases. The size of the sliding window is set here to 10 data points (i.e. corresponding to 150 minutes of blood glucose data or 20 hours of hormone data). The sliding window starts from the starting point on the denoised physiological signal, and the data in the window is compared with the normal data pattern in the steady-state reference model. If the change trend of the data in the window differs from the normal pattern by more than a certain threshold, the data segment in the window is considered to be a physiological abnormality segment. For example, in blood glucose data, under normal circumstances, blood glucose will rise for a period of time after a meal, and then gradually return to normal. If blood glucose continues to rise or fall abnormally in a certain window and exceeds the normal fluctuation range specified in the steady-state reference model, it is identified as an abnormal segment.

[0077] The wavelet transform algorithm is used to extract the frequency domain distribution characteristics of the abnormal fragments. Wavelet transform is a time-frequency analysis method that can decompose the signal at different time and frequency scales. The identified abnormal fragments are subjected to wavelet transform to obtain their energy distribution in different frequency bands. The metabolic deviation is generated by combining the benchmark parameters in the steady-state reference model. Assuming that the normal energy distribution range in a certain frequency band is specified as [a, b] in the steady-state reference model, the energy value of the abnormal fragment obtained by wavelet transform in this frequency band is E, then the metabolic deviation D can be calculated by the formula The calculation results show that D represents the metabolic deviation, E represents the energy value of the abnormal segment in a specific frequency range, and a and b represent the lower and upper limits of the normal energy distribution range in this frequency range in the steady-state reference model. The greater the metabolic deviation, the greater the difference between the metabolic status reflected by the abnormal segment and the normal state.

[0078] Metabolic deviation, baseline fluctuation threshold, and abnormal peak interval are encoded as steady-state physiological characteristics. The baseline fluctuation threshold is a threshold set during the denoising process based on the overall signal fluctuations and is used to measure whether the signal fluctuations are normal. The abnormal peak interval refers to the time interval between two adjacent abnormal peaks. These parameters are encoded according to specific coding rules, such as binary encoding, to convert the metabolic deviation, baseline fluctuation threshold, and abnormal peak interval into binary digital strings to form steady-state physiological characteristics. These steady-state physiological characteristics can more accurately reflect the patient's metabolic stability and provide an important basis for subsequent health management.

[0079] Example 2:

[0080] This embodiment describes the specific implementation process of generating a dynamic lesion map through a cross-modal alignment algorithm in a multimodal fusion module.

[0081] Take, for example, the thyroid ultrasound image report of an endocrine patient. Feature segmentation processing is performed on the medical image report. Because thyroid ultrasound images may contain irrelevant background information, such as images of skin and muscle tissue, this information can interfere with the analysis of thyroid lesions. A deep learning-based image segmentation algorithm is employed. This algorithm is trained on a large number of labeled thyroid ultrasound images to learn the characteristic differences between thyroid tissue and other tissues. This allows the algorithm to accurately segment thyroid tissue from the entire image and eliminate irrelevant background noise data.

[0082] The core lesion area of ​​the image is divided based on the region growing algorithm, and the spatial correlation between each area is calculated. The region growing algorithm uses some pixels with specific characteristics in the image as seed points, and then merges adjacent pixels similar to the seed points into the growing area according to certain growth criteria. The growth criteria are set here as the grayscale value similarity and spatial adjacency of the pixels. Assuming that a pixel with a low grayscale value and considered to be a lesion area is initially selected as the seed point, the algorithm will continuously search for pixels adjacent to it with grayscale values ​​within a certain range, and merge them into the growing area until the stopping condition is met (such as the size of the growing area no longer changes or changes very little). After dividing multiple core lesion areas, the spatial correlation between each area is calculated. The spatial correlation is determined by calculating the distance and relative position relationship between the two areas. For example, for area A and area B, their spatial correlation S can be calculated by the formula Calculate, where d represents the Euclidean distance between the centroids of region A and region B, and θ represents the angle between the line connecting the centroids of region A and region B and the horizontal. A larger value of S indicates a higher spatial correlation between the two regions, meaning they are closer in spatial location and have a certain correlation in relative direction.

[0083] The tissue connection strength is generated based on the spatial correlation and a preset thermal threshold. The preset thermal threshold is a value determined based on extensive historical image data and expert experience. When the spatial correlation is greater than the thermal threshold, a strong tissue connection is considered to exist between the two regions. Assuming the thermal threshold is T, if S > T, the tissue connection strength between the two regions is set to a higher value, such as 1; otherwise, it is set to 0.

[0084] The tissue connection strength is combined with the thermal weight of the lesion area to form a multidimensional dynamic map. The thermal weight of the lesion area is calculated based on the size, grayscale value and other characteristics of the lesion area, and is used to indicate the severity or importance of the lesion area. For example, the larger the lesion area and the greater the difference in grayscale value from normal tissue, the higher its thermal weight. The tissue connection strength and the thermal weight of the lesion area are combined in the form of a matrix to form a multidimensional dynamic map. Each element in the map represents the connection between the two areas and the importance of the lesion area. Doctors can use this map to intuitively understand the location, range and relationship between the thyroid lesion areas, providing more comprehensive information for diagnosis and treatment.

[0085] Example 3:

[0086] This article uses a diabetic endocrinologist as an example to illustrate the specific operations of the feature fusion layer in a multidimensional decision-making model. Assume that the patient's steady-state physiological characteristics, dynamic profile, and drug response pattern evolution sequence have been obtained.

[0087] For the metabolic rate parameters in the steady-state physiological characteristics, there are large differences between different patients. In order to better fuse these data in subsequent calculations, they need to be normalized to obtain the first fusion vector. For example, a reasonable range is determined through the metabolic rate parameter data of a large number of diabetic patients. For the metabolic rate parameter of the current patient, it is compared with this range. If it is near the lower limit of the range, a smaller value is assigned; if it is close to the upper limit, a larger value is assigned. In this way, the metabolic rate parameter is converted to a unified scale to form the various elements of the first fusion vector. For example, in the data collected from 1,000 diabetic patients, the metabolic rate parameter range is between 30-80, and the parameter of this patient is 50. After this relative position calculation, it is converted to 0.4 (assuming it is obtained based on a certain ratio conversion), becoming a value in the first fusion vector.

[0088] The tissue connectivity strengths in the dynamic atlas are then processed and encoded using a discrete Fourier transform (DFT) to generate a second fusion vector. The tissue connectivity strengths in the dynamic atlas are presented in matrix form, reflecting the relationship between the lesion and surrounding tissue. The DFT allows for a different perspective on this data. Imagine each row of the tissue connectivity strength matrix as a signal that varies over time (time is merely an analogy, not actual time). Using the DFT, these "signals" can be converted from the original "time domain" to the "frequency domain," yielding characteristic information at different frequencies. These eigenvalues ​​at different frequencies become the elements of the second fusion vector. For example, after a DFT of a row of tissue connectivity strength data, eigenvalues ​​at five different frequencies are obtained: 0.1, 0.3, 0.2, 0.4, and 0.6. These values ​​form part of the second fusion vector.

[0089] For the drug response pattern evolution sequence, a differential autoregression calculation is performed to extract the pattern change gradient, thereby obtaining the third fusion vector. The drug response pattern evolution sequence records the changes in the patient's physical response over time during medication. First, the changes in this sequence over time are observed, and the differences between adjacent time points are calculated to obtain a new change sequence. Then, based on the characteristics of this change sequence, a fitting method (similar to finding a curve that best represents the trend of these changes) is used to determine parameters that reflect the change pattern. These parameters are combined to represent the change gradient of the drug response pattern, forming the third fusion vector. For example, over a period of time, a patient's drug side effect feedback changes from occasional to increasingly frequent. By analyzing this change process, a numerical sequence representing the severity of this change is obtained, such as 0.2, 0.3, 0.5, etc. These values ​​constitute the third fusion vector.

[0090] Finally, the first fusion vector, the second fusion vector, and the third fusion vector are combined into a low-dimensional decision sequence through weighted superposition. Different weights are assigned to each fusion vector. Assume that the weights of the first fusion vector V1, the second fusion vector V2, and the third fusion vector V3 are w1, w2, and w3 respectively, and w1+w2+w3=1. The kth element d of the low-dimensional decision sequence D is k The formula d k =w1v 1k +w2v 2k +w3v 3k Calculated, where v 1k 、v 2k 、v 3kThe kth elements of the first, second, and third fusion vectors are respectively. Through weighted superposition, the characteristics of different types of data are comprehensively considered to generate a low-dimensional decision sequence containing multiple aspects of patient information, which provides a basis for subsequent strategy optimization and health management plan generation.

[0091] Example 4:

[0092] This embodiment focuses on the specific implementation steps of the strategy optimization layer in the multi-dimensional decision model. After obtaining the low-dimensional decision sequence, the strategy optimization layer starts working. The low-dimensional decision sequence is spatially aligned to generate a strategy weight matrix. Since the elements in the low-dimensional decision sequence may have different meanings and importance, it is necessary to align their spatial dimensions for subsequent unified processing. Assume that the low-dimensional decision sequence is a vector D = [d1, d2, ..., d n ], by normalizing each element to map it to the same scale range. For example, using the z-score normalization method, the element d i The result after normalization d′ i The formula Calculated, where is the mean of the low-dimensional decision sequence, and σ is the standard deviation. The standardized elements are combined into a new matrix, namely the policy weight matrix W, where each row corresponds to a standardized element and each column represents a different dimension (in this example, since we are processing a one-dimensional vector, the number of columns is 1).

[0093] The convolutional attention mechanism is used to extract regional priority features and generate a priority mapping matrix. The convolutional attention mechanism is a method that can automatically learn important areas in the data. The policy weight matrix W is taken as input and processed through a series of convolutional layers and pooling layers. Assume that the size of the convolution kernel is k×k and the step size is s. During the convolution process, the convolution kernel slides on the policy weight matrix to extract features from the local area of ​​each position. Through the convolution operation, a feature map F is obtained. Then the attention mechanism is used to calculate the importance score of each position in the feature map. Assume that the element in the i-th row and j-th column of the feature map F is f ij , its importance score is a ij The formula Calculated, where m and n are the number of rows and columns of the feature map respectively. The importance scores are combined into a matrix, namely the priority mapping matrix P, which reflects the importance of different elements in the low-dimensional decision sequence.

[0094] Perform matrix dot product operation on the policy weight matrix W and the priority mapping matrix P to generate policy optimization features. The matrix dot product operation can combine the weight information in the policy weight matrix with the importance information in the priority mapping matrix. Assume that the policy weight matrix W is an m×n matrix and the priority mapping matrix P is an n×p matrix (in this embodiment, since it is a matrix operation after processing a one-dimensional vector, n=1, p=1), their dot product result O is an m×p matrix, whose element o ij The formula The strategy optimization feature O comprehensively considers the weights and importance of elements in the low-dimensional decision sequence, further optimizing the decision information.

[0095] The policy optimization features are superimposed on the original low-dimensional decision sequence through residual connections to output the health management plan. Residual connections can effectively avoid the gradient vanishing problem during neural network training while retaining the information of the original data. The policy optimization features O are superimposed on the original low-dimensional decision sequence D to obtain the final output result H. The i-th element h of H is i The formula h i =d i +o i1 (Because in this embodiment, due to the particularity of the dimension of the correlation matrix, here o i1 The output H contains the optimized decision information. Based on H, the system generates a personalized health management plan for the patient, such as adjusting medication dosages and providing dietary and exercise recommendations, making the health management plan more accurate and effective.

[0096] Example 5:

[0097] This example describes in detail the method for generating a drug response pattern evolution sequence.

[0098] Take an endocrine patient who is taking hypoglycemic drugs as an example. The tag weight is initialized according to the type of side effect feedback label, and a time association matrix is ​​generated based on the timestamp. Side effect feedback labels may include hypoglycemia, gastrointestinal discomfort, and other types. An initial tag weight is assigned to each side effect feedback label. Assuming there are n side effect feedback labels, the initial tag weight vector W0 = [w 01 ,w 02 ,…,w 0n ], these initial weight values ​​can be set based on experience or historical data. For example, for hypoglycemia reactions that are more common and have a greater impact on patients, a relatively high initial weight can be given. At the same time, based on the timestamp information of the side effect feedback, a time correlation matrix T is constructed. The time correlation matrix T is an m×m matrix (m is the number of time points for recording side effect feedback), and its element tij Represents the time interval between the i-th time point and the j-th time point. If i=j, then t ij =0; if i≠j, t ij Available through |t i -t j |Calculate, where t i and t j are the time values ​​corresponding to the i-th and j-th time points respectively. In this way, the time correlation matrix reflects the closeness between the time points of each side effect feedback.

[0099] The time correlation matrix is ​​used as the input of pattern evolution, and the state transition probability is calculated through the hidden Markov model. The hidden Markov model is a statistical model used to describe the process of generating an observable sequence from a hidden Markov process. In this scenario, the hidden state can be regarded as the different response states of the patient to the drug, and the observable sequence is the side effect feedback label. Let the state set of the hidden Markov model be S = {s1, s2, ..., s k} (k is the number of states, for example, it can be divided into mild reaction, moderate reaction, severe reaction and other states according to different drug reaction degrees), the initial state probability vector π=[π1,π2,…,π k ], state transition probability matrix A=[a ij ] k×k , where a ij Indicates that from state s i Transfer to state s j The probability of . Using the time correlation matrix T and the observed side effect feedback label sequence, the state transition probability matrix A is calculated by the forward-backward algorithm and other methods. For example, in calculating a ij When considering the time relationship reflected by the time correlation matrix, i Transfer to state s j And the possibility of generating corresponding side effect feedback labels can be expressed as Where I is the indicator function, when the condition in the brackets is met, the value of I is 1, otherwise it is 0, s t represents the hidden state of the patient at time point t.

[0100] The pattern confidence of each time node is updated through the Bayesian network, and the label weight is adjusted. The Bayesian network is a graphical network based on probabilistic reasoning, which can use new evidence to update the estimate of the probability of an event. The state transition probability calculated by the hidden Markov model and the newly appeared side effect feedback label are input into the Bayesian network as evidence. The Bayesian network recalculates the pattern confidence corresponding to different side effect feedback labels at each time node based on this information. For example, for a certain time node t and side effect feedback label l, its pattern confidence P(l|E) (E represents all known evidence) can be calculated by the Bayesian formula Calculate, where P(E|l) is the probability of observing evidence E given the presence of side effect feedback label l, P(l) is the prior probability of the presence of side effect feedback label l, and l′ traverses all possible side effect feedback labels. Based on the updated pattern confidence, adjust the label weight. If the pattern confidence of a side effect feedback label increases at a certain time point, it indicates that the side effect feedback is more likely to be related to the drug response pattern at that time point, and its label weight is increased accordingly; otherwise, the weight is decreased.

[0101] A drug response pattern evolution sequence is generated based on the adjusted marker weights. Arranging the adjusted marker weights in chronological order creates a drug response pattern evolution sequence. This sequence clearly demonstrates how the patient's body's response to the drug changes at different time points during medication use, providing important insights for doctors to understand drug efficacy and patient responses, and facilitating timely adjustments to treatment plans.

[0102] Example 6:

[0103] This example illustrates the specific workflow of the lifestyle analysis module and how to incorporate lifestyle factors into the health management plan generation process. For example, let's take an endocrine patient with thyroid disease as an example. Assume that a detailed lifestyle log has been collected for this patient over a period of time, which records information such as the type and amount of food consumed daily, the duration and intensity of exercise, and their sleep and rest schedules.

[0104] Perform time-series segmentation on the lifestyle log to extract behavioral pattern labels and abnormal work and rest intervals. Divide the lifestyle log into different time periods in chronological order, such as dividing it by day. For daily dietary information, determine whether there are unhealthy eating habits, such as high-sugar, high-fat diets, etc., based on the type and intake of food, and assign corresponding behavioral pattern labels, such as "high-sugar diet" and "high-fat diet". For exercise time and intensity, if the patient exercises less than 3 times a week or the exercise time is less than 30 minutes each time, it can be marked as "insufficient exercise". For work and rest time, by setting the normal sleeping and waking time range (such as 10 pm to 6 am as the normal work and rest range), determine whether there is an abnormal work and rest interval. If the patient often falls asleep after 1 am and gets up before 8 am, it will be recorded as an abnormal work and rest interval.

[0105] Association rule mining algorithms can be used to identify correlation weights between lifestyle habits and physiological indicators. Association rule mining algorithms can discover associations between different variables in large amounts of data. They combine lifestyle data (such as behavioral pattern labels like diet, exercise, and sleep schedules) with physiological indicator data (such as hormone levels and blood sugar fluctuations) for analysis. For example, the Apriori algorithm can be used to mine association rules between lifestyle habits and physiological indicators by setting support and confidence thresholds. Suppose a correlation is found between "high-sugar diet" and "abnormal blood sugar fluctuations." The support is calculated by calculating the proportion of records meeting the "high-sugar diet and abnormal blood sugar fluctuation" condition to the total number of records. The confidence is calculated by calculating the probability of "abnormal blood sugar fluctuations" occurring under the "high-sugar diet" condition. The correlation weight is determined based on the support and confidence values. If both the support and confidence values ​​are high, the correlation between "high-sugar diet" and "abnormal blood sugar fluctuations" is strong, and a higher correlation weight is assigned; otherwise, a lower correlation weight is assigned.

[0106] The relevance weights are combined with the behavior pattern labels to generate a lifestyle influence factor. Each behavior pattern label is combined with its corresponding relevance weight. For example, the relevance weight of "high-sugar diet" is c1, and the relevance weight of "lack of exercise" is c2. The lifestyle influence factor I can be expressed as I = c1 × "high-sugar diet" + c2 × "lack of exercise" + .... Here, the behavior pattern labels such as "high-sugar diet" and "lack of exercise" can be represented by binary code, 1 indicates the presence of the behavior pattern, and 0 indicates its absence. In this way, the lifestyle information is quantified into a comprehensive influence factor.

[0107] Lifestyle influencing factors are incorporated into the health management plan generation process through a multidimensional decision-making model. The lifestyle influencing factors are used as a new input feature and input into the multidimensional decision-making model together with steady-state physiological characteristics, dynamic maps, and drug response pattern evolution sequences. In the feature fusion layer of the multidimensional decision-making model, the lifestyle influencing factors are processed accordingly, such as normalization, so that they can be fused and calculated with other features at the same scale. In the subsequent strategy optimization layer, the information brought by the lifestyle influencing factors is comprehensively considered to generate a more comprehensive health management plan. For example, if the lifestyle influencing factors indicate that the patient has problems with a high-sugar diet and insufficient exercise, the health management plan may recommend that the patient adjust the diet structure, reduce the intake of high-sugar foods, and increase the amount of exercise. The drug treatment plan will be optimized accordingly based on these adjustments to better control the disease.

[0108] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0109] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A big data-based endocrine patient health management system, characterized by: include: Multi-source data acquisition module: used to obtain heterogeneous health data sets of endocrine patients; The heterogeneous health data includes physiological monitoring indicators, medication record data and medical imaging reports; wherein the physiological monitoring indicators include blood sugar fluctuation curves and hormone level time series data, and the medication record data includes drug dosage association information and side effect feedback labels; A physiological characteristic noise reduction module is configured to extract steady-state physiological characteristics based on the physiological monitoring indicators through an adaptive filtering algorithm, wherein the steady-state physiological characteristics include a baseline fluctuation threshold, an abnormal peak interval, and a metabolic rate parameter; Multimodal fusion module: used to generate a dynamic map of lesions based on the medical imaging report through a cross-modal alignment algorithm, wherein the dynamic map includes organ tissue association paths and thermal weights of lesion areas; Medication feedback analysis module: performs timestamp processing on the side effect feedback labels to generate a drug response pattern evolution sequence; Personalized plan generation module: inputs the steady-state physiological characteristics, dynamic maps and drug response pattern evolution sequence into a multi-dimensional decision-making model to output a patient health management plan.

2. The endocrine patient health management system based on big data according to claim 1 is characterized in that: The method of extracting steady-state physiological characteristics by using an adaptive filtering algorithm includes: Performing baseline drift correction processing on the physiological monitoring indicators to generate denoised physiological signals; Based on the preset steady-state reference model, the sliding window matching algorithm is used to identify physiological abnormal segments; Wavelet transform algorithm is used to extract the frequency domain distribution characteristics of abnormal segments, and the metabolic deviation is generated by combining the baseline parameters in the steady-state reference model; The metabolic deviation, baseline fluctuation threshold, and abnormal peak interval were coded as steady-state physiological characteristics.

3. The endocrine patient health management system based on big data according to claim 1, characterized in that: The generating of the dynamic lesion map by the cross-modal alignment algorithm includes: Performing feature segmentation processing on the medical imaging report to eliminate irrelevant background noise data; The core lesion area of ​​the image is divided based on the region growing algorithm, and the spatial correlation between each area is calculated; Generate tissue connection strength based on spatial correlation and preset thermal threshold; The tissue connection strength is combined with the thermal weight of the lesion area to form a multidimensional dynamic map.

4. The endocrine patient health management system based on big data according to claim 1 is characterized in that: The multi-dimensional decision model includes a feature fusion layer and a strategy optimization layer. The feature fusion layer includes: Normalizing the metabolic rate parameter in the steady-state physiological characteristics to obtain a first fusion vector; Performing discrete Fourier transform encoding on the tissue connection strength in the dynamic atlas to generate a second fusion vector; performing differential autoregressive calculation on the drug response pattern evolution sequence, extracting the pattern change gradient, and obtaining a third fusion vector; The first fusion vector, the second fusion vector and the third fusion vector are combined into a low-dimensional decision sequence through weighted superposition.

5. The endocrine patient health management system based on big data according to claim 4 is characterized in that: The strategy optimization layer includes: Align the spatial dimensions of the low-dimensional decision sequence to generate a strategy weight matrix; Extract regional priority features through the convolutional attention mechanism and generate a priority mapping matrix; Perform matrix dot product operation on the policy weight matrix and the priority mapping matrix to generate policy optimization features; The strategy optimization features are superimposed on the original low-dimensional decision sequence through residual connections to output a health management plan.

6. The endocrine patient health management system based on big data according to claim 1 is characterized in that: The method for generating the drug response pattern evolution sequence comprises: Initialize the label weight according to the type of side effect feedback label and generate a temporal correlation matrix based on the timestamp; The time correlation matrix is ​​used as a pattern evolution input to calculate the state transition probability through a hidden Markov model; Update the pattern confidence of each time node through the Bayesian network and adjust the label weight; Generate a sequence of drug response pattern evolution based on the adjusted marker weights.

7. The endocrine patient health management system based on big data according to claim 2 is characterized in that: The method for constructing the steady-state reference model includes: Collect standard physiological samples from patients with various typical endocrine diseases and extract baseline metabolic curves and frequency domain distribution characteristics; Perform multi-scale decomposition on the baseline metabolic curve to generate a multi-resolution metabolic model; Classify metabolic models according to disease types and associate them with a standard parameter database; The classified metabolic model is stored as a steady-state reference model and is regularly updated based on newly collected samples.

8. The endocrine patient health management system based on big data according to claim 3 is characterized in that: The parameter optimization method of the region growing algorithm includes: Calculate the initial growth threshold and minimum regional density based on the distribution of historical image data; The grid search algorithm is used to traverse the parameter combinations and select the parameters with the highest overlap between the segmentation results and the expert-annotated areas; The growth threshold and minimum regional density are dynamically adjusted according to the degree of overlap to optimize the accuracy of tissue connection strength division.

9. The endocrine patient health management system based on big data according to claim 1, characterized in that: The heterogeneous health data also includes the user's life habit log.

10. The endocrine patient health management system based on big data according to claim 9 is characterized in that: Also includes lifestyle analysis module: Performing time-series segmentation processing on the lifestyle log to extract behavior pattern labels and abnormal work and rest intervals; Identify the correlation weights between lifestyle habits and physiological indicators based on association rule mining algorithms; The correlation weight is integrated with the behavior pattern label to generate a lifestyle influence factor; The multi-dimensional decision-making model is used to incorporate the factors affecting living habits into the process of generating health management plans.

Images