Internet hospital-based artificial intelligence auxiliary diagnosis and treatment system for endocrine diseases

Abstract

The invention relates to the technical field of medical data, in particular to an endocrine disease artificial intelligence auxiliary diagnosis and treatment system based on an internet hospital, and the system comprises a data collection module which is used for accessing a wearable device, a home detection device, a patient autonomous report, an offline medical record and multi-source data of a third-party platform; when the system is used, efficient fusion processing is performed on multi-channel, fragmented, high-real-time and cross-platform heterogeneous data, so that the limitation that the prior art depends on a static or single data source is broken through, the data timeliness is improved, accurate data support is provided for rapid judgment of an emergency situation, diagnosis and treatment lag is avoided, and the diagnosis and treatment efficiency is improved. By means of personalized treatment scheme adjustment based on real-time data dynamic diagnosis and treatment, a static scheme recommendation mode is convenient to change, the remote diagnosis and treatment response speed is improved, the problem of unexpected situation processing lag is convenient to solve, resource barriers are broken through online and offline resource linkage, online decisions are made to smoothly fall into the ground, and the auxiliary diagnosis and treatment effect is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data, in particular to an endocrine disease artificial intelligence assisted diagnosis and treatment system based on an Internet hospital. BACKGROUND

[0002] Endocrine diseases are a class of complex diseases caused by abnormal hormone secretion (too much, too little or resistance) in the body, covering diabetes, thyroid disease, pituitary tumor, adrenal disease, gonadal disease, etc. Its diagnosis and treatment need to combine multi-dimensional information such as symptoms, signs, laboratory tests (hormone levels), imaging data, etc. and individual differences, and the disease course changes significantly.

[0003] Patent No. CN111489820A, in the specification, records "The present application relates to an artificial intelligence-based auxiliary diagnosis and treatment system, at least comprising: a patient client (1), a database (3), a server (4) and a medical client (2), wherein the patient client (1) can record patient residence information, the database (3) can store the medication data of a specific patient in a manner related to medication, wherein the server (4) is configured to: count the medication data of the specific patient for the same disease type, to obtain the frequency and / or quantity of the same or similar drugs being repeatedly taken by the specific patient within a set time period; in the case where the server (4) determines that the frequency and / or quantity of the same or similar drugs being repeatedly taken by the specific patient is greater than a set threshold based on the analysis of its analysis module, the medical client (2) and / or the patient client (1) is prompted", the above technology, although it combines medication frequency analysis with regional patient data to solve the two major pain points of the decline in efficacy caused by the lack of medication data at the grassroots level and the difficulty in diagnosing regional diseases, but endocrine diseases are of various types, such as diabetes, thyroid disease, pituitary disease, etc. Its diagnosis and treatment are complex. Currently, the Internet hospital artificial intelligence assisted diagnosis and treatment system relies on static data or a single data source, which cannot effectively process the heterogeneous data from wearable devices, home testing devices, patient self-reporting, offline medical records and third-party platforms in the Internet hospital, which are multi-channel, fragmented, real-time and cross-platform, resulting in a lag in the judgment and handling of sudden conditions in assisted diagnosis and treatment. At the same time, the lack of linkage management of online and offline resources leads to general assisted diagnosis and treatment effect.

[0004] In summary, developing an endocrine disease artificial intelligence assisted diagnosis and treatment system based on an Internet hospital is still a key problem that needs to be solved in the field of medical data technology. SUMMARY

[0005] The present application aims to solve the problem that there are various endocrine diseases in the prior art, such as diabetes, thyroid disease, pituitary disease and the like, and the diagnosis and treatment thereof is complicated, at present, the artificial intelligence assisted diagnosis and treatment system of the Internet hospital mainly depends on static data or a single data source, and cannot effectively process the heterogeneous data from multiple channels, fragmentation, real-time and cross-platform in the Internet hospital, such as wearable devices, home detection devices, patient self-reporting, offline medical records and third-party platforms, resulting in the situation that the judgment and processing of the sudden situation are lagged in the assisted diagnosis and treatment, and at the same time, the linkage management of online and offline resources is lacked, resulting in the problem that the effect of the assisted diagnosis and treatment is general.

[0006] To achieve the above object, the present application provides an endocrine disease artificial intelligence assisted diagnosis and treatment system based on an Internet hospital, comprising: A data acquisition module is used for accessing multi-source data of wearable devices, home detection devices, patient self-reporting, offline medical records and third-party platforms; A time sequence alignment module is used for aligning the multi-source data according to timestamps by using a high-precision time sequence alignment algorithm, and outputting aligned data; An intelligent learning module is used for identifying endocrine disease data correlation in the aligned data by using a deep learning model, generating a physiological digital twin of a patient and updating it in real time; A dynamic diagnosis and treatment module is used for outputting an individualized treatment scheme according to the physiological digital twin combined with a knowledge graph; A doctor-patient linkage module is used for providing a data visualization dashboard for doctors and pushing interactive diagnosis and treatment suggestions for patients through an online and offline resource linkage database.

[0007] Further, the operation process of the data acquisition module for accessing multi-source data of wearable devices, home detection devices, patient self-reporting, offline medical records and third-party platforms comprises: The data acquisition module uses an SSL / TLS encryption transmission protocol, and establishes a local cache mechanism, wherein the local cache mechanism establishes a local cache pool Each data is cached separately, and the cache scheduling strategy uses a priority weighted sliding window mechanism, and the expression is as follows:

[0008] In the formula, represents the state of the local cache pool corresponding to the i-th data at time t, represents the size of the sliding window, represents the summation of all items in the interval from time t to t+i-1, represents the summation of all items in the interval from time t to t+i-1, represents the summation of all items in the interval from time t to t+i-1, represents the summation of all items in the interval from time t to t+i-1, represents the summation of all items in the interval from time t to t+i-1, represents the summation of all items in the interval from time t to t+i-1. weight coefficient of time data, represents the time the first original data content of road data, represents all time corresponding weight coefficient The sum is 1, is a hyperparameter for controlling the weight decay speed, is an exponential decay function, is a normalization factor.

[0009] Further, the time alignment module is used to adopt a high-precision time alignment algorithm to align the multi-source data according to the time stamp, and the operation process of the aligned data includes: Adopting a high-precision time alignment algorithm, for the multi-source data with heterogeneous data source devices, different device data collection frequency

[0010] In the formula, represents the original data sequence of the heterogeneous data source, represents the time stamp of the device at the sampling time, represents the observation value of the device at the sampling time, represents the total number of samples of the device, and a spline function interpolation method with disturbance compensation is adopted according to the global unified time stamp to reconstruct each device data, and the expression is:

[0011] In the formula, represents the interpolation reconstruction value of the device at the global unified time , represents the global unified time stamp, represents the contribution of the original sampling point to the reconstruction value at the global unified time , is a Gaussian kernel part, is a disturbance compensation coefficient, is a cubic spline function at the global unified time .

[0012] Further, the time sequence alignment module is used to adopt a high-precision time sequence alignment algorithm to perform alignment processing on the multi-source data according to the timestamps, and an operation process of outputting the aligned data includes: The alignment processing constructs a global alignment target function to perform a minimum asynchronous cost processing on the interpolation sequences on the global unified timestamps, and an expression is as follows:

[0013] In the formula, is a loss function of the global alignment, indicates traversing each global unified timestamp in a global unified timestamp set indicates traversing all data pairs, is a weight of a data source pair, indicating importance of the first data source and the second data source, indicates interpolation data of the first data source at time indicates interpolation data of the second data source at time is a square of the Euclidean distance, an expression of outputting the aligned data is as follows:

[0014] In the formula, indicates a data set after alignment, indicates a global unified timestamp, and indicates interpolation data of the first data source to the Nth data source at time indicates a total number of data sources, indicates traversing each time point in a global unified timestamp set .

[0015] Further, the intelligent learning module is used to identify endocrine disease data correlation in the aligned data through a deep learning model, generate a physiological digital twin of a patient, and update the physiological digital twin in real time, and an operation process includes: The deep learning model is based on an attention mechanism, and an expression is as follows:

[0016] In the formula, indicates a hidden state vector of the Nth channel / modality at the moment is an activation function, indicates a hidden state vector of the Nth channel / modality at the moment ​​​​​​​​​​​​The original input data of channels / modalities, represents the embedding weight matrix, represents the embedding bias vector, express Moment The attention output vector of each channel / modality, Indicates that the splicing operation will The outputs of the attention heads are stitched together. Indicates the The value weight matrix of the attention head will be hidden state Convert to a vector of values, Indicates the use of attention weight Weighted sum of value vectors, represents the total number of attention heads, express Moment Positions in the attention head and location The attention weights are set, and the deep learning model is optimized by training through Apache Hadoop and Spark frameworks, and a cross-validation method is used.

[0017] Furthermore, the intelligent learning module is used to identify the association of endocrine disease data in the aligned data through a deep learning model, generate a physiological digital twin of the patient, and update the operation process in real time, including: The physiological digital twin of the patient is generated and updated in real time. The state vector of the physiological digital twin includes but is not limited to core states such as hormone secretion level and metabolic rate, which is generated by hidden state mapping, and the expression is:

[0018] Where, express The state vector of the physiological digital twin at each moment, represents the state mapping weight matrix, express The hidden state vector at time t, Represents the state mapping bias vector. When new data is input, the historical state and the new observation value are fused through Kalman filtering. The expression is:

[0019] Where, express Time has come The state prediction value at the moment, represents the state transition matrix, express The state vector of the physiological digital twin at each moment, represents the control matrix, yes The control vector at time t, express Time has come The state prediction error covariance matrix at time , express The error covariance matrix of the state at the moment, Represents the state transition matrix The transposed matrix of represents the process noise covariance matrix, express The Kalman gain at time t, represents the observation matrix, is the inverse covariance matrix for calculating observation noise, is the observation noise covariance matrix, After integrating new observations The physiological digital twin state is corrected at all times. express The new observation value at time , According to the predicted state The derived theoretical observations are used to compare with the new observations Compare and calculate the correction amount, After integrating new observations The error covariance matrix after time correction, Represents the identity matrix used for identity transformation in matrix operations.

[0020] Furthermore, the dynamic diagnosis and treatment module is used to output a personalized treatment plan based on the physiological digital twin combined with the knowledge graph. The operation process includes: The dynamic diagnosis and treatment module uses a reinforcement learning model to load the quarterly updated endocrine disease diagnosis and treatment knowledge graph in the form of triples, and performs semantic parsing and feature mapping. Based on the physiological digital twin combined with the knowledge graph, a decision feature vector is generated. The expression is:

[0021] Where, express The decision feature vector generated at each moment, Indicates that the fusion weight matrix is ​​a learnable parameter used to perform a linear transformation on the concatenated vector. It is a knowledge graph feature fusion calculation based on the attention mechanism. is the total number of entities involved in the calculation in the knowledge graph, is the attention weight matrix, a learnable parameter used to calculate the physiological digital twin state vector and knowledge graph entity vectors The association weight between Indicates the The vector of entities, is the first entity vectors, Indicates that it is calculated by the exponential function and The relevance score after attention matrix transformation, Represents all entities with The sum of the correlation scores is used for normalization, is the attention weight representing the Knowledge graph entities relative to physiological digital twins the importance of The vector concatenation operation represents the concatenation of the physiological digital twin state vector and the feature vector obtained by fusing the knowledge graph through the attention mechanism. is the learnable parameter of the fusion bias vector and Cooperate and perform offset adjustment on the concatenated and linearly transformed vectors.

[0022] Furthermore, the dynamic diagnosis and treatment module is used to output a personalized treatment plan based on the physiological digital twin combined with the knowledge graph. The operation process includes: The reinforcement learning model adopts an Actor-Critic architecture and consists of a strategy function and a value function. The output is a personalized treatment plan. The knowledge graph is updated quarterly to incorporate the latest research results and clinical experience. The reward function of the reinforcement learning model is centered on treatment effect indicators, including blood sugar control target rate and complication rate. The expression is:

[0023] Where, Indicates the Instant rewards at all times, is the weight coefficient The reward weight for achieving blood sugar control targets Control the penalty weights for other adverse clinical events, It is an exponential function that makes the reward show a smooth decay / increase trend with blood sugar deviation. Indicates the The actual blood sugar value at the moment, Indicates target blood sugar value, Indicates the tolerance parameter for blood sugar fluctuations, represents the square of blood glucose deviation, Representative Other adverse clinical events at this time include but are not limited to the risk of hypoglycemia and ketosis.

[0024] Furthermore, the doctor-patient linkage module is used to link databases through online and offline resources, provide doctors with data visualization dashboards, and push interactive diagnosis and treatment recommendations to patients. The operation process includes: The doctor-patient linkage module includes a doctor side and a patient side. The doctor side provides a data visualization dashboard to display the physiological digital twin and treatment plan in real time and supports manual adjustment by the doctor. The expression is:

[0025] Where, Indicates the current patient portrait Hedi Historical patient portraits The similarity value between Represents the portrait vector of the current patient, Indicates the A portrait vector of a historical patient, Expressing arrive The items are accumulated, is the total number of features in the patient portrait, Indicates the The weight of the feature, Indicates the current patient Features With historical patients No. Features The numerical product of Indicates the current patient portrait The weighted norm of Indicates historical patients portrait The weighted norm of .

[0026] Furthermore, the doctor-patient linkage module is used to link databases through online and offline resources, provide doctors with data visualization dashboards, and push interactive diagnosis and treatment recommendations to patients. The operation process includes: The interactive diagnosis and treatment suggestions pushed by the patient end include but are not limited to animated demonstrations of insulin injection adjustment methods. The online and offline resource linkage database is used to calculate the adaptability with the patient portrait. The expression is:

[0027] Where, Indicates that for Treatment recommendations The fitness calculation results are: Indicates the Treatment recommendations awaiting evaluation. It is from arrive Accumulate the following formulas, is the total number of historical patient cases and other related data involved in the calculation, Indicates the current patient portrait Hedi Historical patient portraits The similarity value between Represents the portrait vector of the current patient, Indicates the A portrait vector of a historical patient, Is the evaluation function that measures the diagnosis and treatment recommendations Apply to Corresponding portraits of historical patients The effect produced when selecting The highest recommendation is pushed in the form of animation. The online and offline resource linkage database includes but is not limited to the diagnosis and treatment resources, equipment, expert information of offline medical institutions and the key information of doctor-patient communication recorded through natural language processing technology to feed back the deep learning model.

[0028] Beneficial effects Compared with the known public technology, the technical solution provided by the present invention has the following beneficial effects: When in use, the present invention efficiently integrates and processes multi-channel, fragmented, real-time and cross-platform heterogeneous data, which is conducive to breaking through the limitations of existing technologies that rely on static or single data sources, improving data timeliness, and providing accurate data support for rapid judgment of emergencies, avoiding delays in diagnosis and treatment. It is easy to adjust personalized treatment plans based on real-time data dynamic diagnosis and treatment, which is convenient for changing static plan recommendation modes, improving the response speed of remote diagnosis and treatment, and solving the problem of delays in emergency processing. By linking online and offline resources, it breaks down resource barriers, enables online decisions to be smoothly implemented, and is conducive to improving the effect of auxiliary diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a system diagram of an artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital in the present invention. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described 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 should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and in the above description of the drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0032] The present application will be further described in detail below with reference to the accompanying drawings: Embodiments: As Figure 1 shown, the present application provides an endocrine disease artificial intelligence auxiliary diagnosis and treatment system based on an Internet hospital, comprising: A data acquisition module for accessing multi-source data of wearable devices, home detection devices, patient self-reporting, offline medical records and third-party platforms; Further, the operation process of the data acquisition module for accessing multi-source data of wearable devices, home detection devices, patient self-reporting, offline medical records and third-party platforms includes: The data acquisition module uses an SSL / TLS encryption transmission protocol and establishes a local cache mechanism, and the local cache mechanism establishes a local cache pool Each data is cached separately, and the cache scheduling strategy uses a priority weighted sliding window mechanism, and the expression is:

[0033] In the formula, represents the state of the local cache pool corresponding to the i-th data at time t, represents the size of the sliding window, represents the sum of all items in the interval from time t to t+1, represents the weight coefficient of the data at time t, represents the original data content of the i-th data at time t, represents the sum of the weight coefficients corresponding to all times t, represents the sum of the weight coefficients corresponding to all times t, ​​​​​​​​is a hyperparameter for controlling the weight decay speed, is an exponential decay function, is a normalization factor.

[0034] Specifically, the data acquisition module securely accesses multi-source health data and efficiently manages data through a local caching mechanism, uses the SSL / TLS encryption protocol to ensure data transmission security, and establishes a local cache pool for each data channel. Based on the priority weighted sliding window mechanism, the cache is scheduled, the data range is determined through the sliding time window, the data at different times in the window is assigned a weight that decays over time (recent data has a higher weight), and after normalization, the weighted sum is determined to determine the cache pool state. Important data is given priority to retain, which is conducive to protecting the privacy and compliance of the entire data transmission, and at the same time, it can alleviate the problem of data loss caused by network instability or device disconnection.

[0035] The time alignment module is configured to align the multi-source data according to the timestamps using a high-precision time alignment algorithm, and output aligned data. Further, the time alignment module is configured to align the multi-source data according to the timestamps using a high-precision time alignment algorithm, and output aligned data. The operation process includes: Using a high-precision time alignment algorithm, for the multi-source data with heterogeneous data source devices, different device data collection frequencies,

[0036] In the formula, denotes the original data sequence of the heterogeneous data source, denotes the timestamp of the device at the sampling time, denotes the observation value of the device at the sampling time, denotes the total number of samples of the device, and a spline function interpolation method with disturbance compensation is used to reconstruct each device data according to the global unified timestamp. The expression is:

[0037] In the formula, denotes the interpolation reconstruction value of the device at the global unified time , denotes the global unified timestamp, denotes the measure of the original sampling point Global unified time The contribution of the reconstruction value at is the Gaussian kernel part, is the disturbance compensation coefficient, It is the global unified time The cubic spline function at .

[0038] Furthermore, the timing alignment module is used to align multi-source data based on timestamps using a high-precision timing alignment algorithm. The operation process of outputting aligned data includes: The alignment process constructs a global alignment objective function and aligns the Interpolation sequences are processed to minimize asynchronous cost, expression:

[0039] Where, is the global alignment loss function, Indicates traversing the global unified timestamp set Each global unified timestamp in , Indicates traversing all data pairs, is the weight of the data source pair. Hedi The importance of data sources, Indicates the Data sources at time The interpolated data at Is the square of the Euclidean distance, output alignment data, expression:

[0040] Where, represents the aligned data set, Represents a global unified timestamp, indicating the timestamps from 1 to Data sources at time The interpolated data at Indicates the total number of data sources, Indicates traversing the global unified timestamp set At each time point .

[0041] Specifically, the timing alignment module uses a high-precision timing alignment algorithm to uniformly process the time dimension of multi-source heterogeneous data, and outputs aligned data that can be directly used for subsequent analysis. For the original data sequences of multiple heterogeneous data sources, based on the global unified timestamp, the data is reconstructed using a combination of Gaussian kernel and cubic spline function with disturbance compensation. The asynchronous cost of multi-source data is then minimized through the global alignment objective function, and finally a set of aligned data with a unified time axis is generated to avoid information loss or distortion, so that multi-source data has optimal consistency and timing synchronization at the same time scale, which is convenient for flexible application in various medical wearables, sensors and smart devices, and supports the precise integration and processing of medical big data.

[0042] An intelligent learning module, configured to identify endocrine disease data associations in the aligned data using a deep learning model, generate a physiological digital twin of the patient, and update it in real time; Furthermore, the intelligent learning module is used to identify the association of endocrine disease data in the aligned data through a deep learning model, generate a physiological digital twin of the patient, and update the operation process in real time, including: The deep learning model is based on the attention mechanism, expressed as:

[0043] Where, express Moment The hidden state vector of each channel / modality, yes activation function, express Moment The original input data of channels / modalities, represents the embedding weight matrix, represents the embedding bias vector, express Moment The attention output vector of each channel / modality, Indicates that the splicing operation will The outputs of the attention heads are stitched together. Indicates the The value weight matrix of the attention head will be hidden state Convert to a vector of values, Indicates the use of attention weight Weighted sum of value vectors, represents the total number of attention heads, express Moment Positions in the attention head and location The attention weights are set, and the deep learning model is optimized by training through Apache Hadoop and Spark frameworks, and a cross-validation method is used.

[0044] Furthermore, the intelligent learning module is used to identify the association of endocrine disease data in the aligned data through a deep learning model, generate a physiological digital twin of the patient, and update the operation process in real time, including: The physiological digital twin of the patient is generated and updated in real time. The state vector of the physiological digital twin includes but is not limited to core states such as hormone secretion level and metabolic rate, which is generated by hidden state mapping, and the expression is:

[0045] Where, express The state vector of the physiological digital twin at each moment, represents the state mapping weight matrix, express The hidden state vector at time t, Represents the state mapping bias vector. When new data is input, the historical state and the new observation value are fused through Kalman filtering. The expression is:

[0046] Where, express Time has come The state prediction value at the moment, represents the state transition matrix, express The state vector of the physiological digital twin at each moment, represents the control matrix, yes The control vector at time t, express Time has come The state prediction error covariance matrix at time , express The error covariance matrix of the state at the moment, Represents the state transition matrix The transposed matrix of represents the process noise covariance matrix, express The Kalman gain at time t, represents the observation matrix, is the inverse covariance matrix for calculating observation noise, is the observation noise covariance matrix, After integrating new observations The physiological digital twin state is corrected at all times. denotes a new observation at time t, denotes a predicted state derived theoretical observation is used to compare with the new observation to calculate a correction, denotes the error covariance matrix after correction at time t, denotes the error covariance matrix after correction at time t, denotes an identity matrix used for identity transformation in matrix operation.

[0047] Specifically, the intelligent learning module identifies data correlations of endocrine diseases from multi-source aligned data through a deep learning model based on an attention mechanism, generates and updates a patient's physiological digital twin in real time, processes multi-modal data such as monitoring data of wearable devices and home detection instruments through an embedding layer and an attention mechanism, generates hidden states, and converts them into digital twin states through a mapping matrix. Distributed training is performed with the aid of Apache Hadoop and Spark frameworks, and the deep learning model is optimized through cross-validation to facilitate adaptation to massive medical data. When new data is input, Kalman filtering is used to fuse historical states and new observations, dynamically correct the twin, improve real-time performance, and help improve the monitoring and prediction accuracy of disease progression.

[0048] The dynamic diagnosis and treatment module is configured to output an individualized treatment plan based on the physiological digital twin and a knowledge graph. Further, the operation flow of the dynamic diagnosis and treatment module for outputting an individualized treatment plan based on the physiological digital twin and a knowledge graph comprises: The dynamic diagnosis and treatment module adopts a reinforcement learning model, loads the quarterly updated endocrine disease diagnosis and treatment knowledge graph in the form of triples, and performs semantic analysis and feature mapping. Based on the physiological digital twin and the knowledge graph, a decision feature vector is generated, and the expression is as follows:

[0049] In the formula, denotes a decision feature vector generated at time t, denotes that the fusion weight matrix is a learnable parameter used for linear transformation of the spliced vector, is a knowledge graph feature fusion calculation based on an attention mechanism, is the total number of entities participating in the calculation in the knowledge graph, is an attention weight matrix learnable parameter used for calculating the physiological digital twin state vector and the association weight between the knowledge graph entity vector , denotes the vector of entities, is the th entity vector, is calculated by exponential function is the relevance score of the th entity after attention matrix transformation, is the sum of relevance scores of all entities and for normalization, is the attention weight indicating the importance of the th knowledge graph entity relative to the physiological digital twin, is the vector concatenation operation indicating the concatenation of the physiological digital twin state vector and the feature vector obtained by fusing the knowledge graph through the attention mechanism, is the fusion bias vector which is a learnable parameter and works together to adjust the offset of the vector after concatenation and linear transformation.

[0050] Further, the dynamic diagnosis and treatment module, for outputting an individualized treatment scheme operation process according to the physiological digital twin combined with the knowledge graph, comprises: The reinforcement learning model adopts an Actor-Critic architecture composed of a policy function and a value function, the output individualized treatment scheme, the knowledge graph is updated once a quarter, and the latest research results and clinical experience are included, and the reward function of the reinforcement learning model takes the diagnosis and treatment effect index as the core, including the blood glucose control compliance rate and the complication incidence rate, the expression is:

[0051] In the formula, represents the instant reward at the th moment, is the weight coefficient controls the reward weight brought by blood glucose control controls the penalty weight of other adverse clinical events, is an exponential function that makes the reward present a smooth decay / growth trend with blood glucose deviation, represents the actual blood glucose value at the th moment, represents the target blood glucose value, represents the tolerance parameter of blood glucose fluctuation, represents the square of blood glucose deviation, represents the other adverse clinical events at the th moment, including but not limited to hypoglycemia risk, ketosis.

[0052] ​Specifically, by dynamically integrating the patient's digital twin with the knowledge graph, it is convenient to adjust the treatment plan in real time according to the patient's status. The reward mechanism of reinforcement learning is centered on the blood sugar target rate and the incidence of complications. The plan evolves dynamically, taking into account both control effect and risk prevention. The quarterly updated knowledge graph and deep learning model evolve together, so that personalized treatment plans are always based on the latest clinical and scientific research results, improving efficacy and cutting-edge nature.

[0053] The doctor-patient linkage module is used to link databases through online and offline resources, provide doctors with data visualization dashboards, and push interactive diagnosis and treatment recommendations to patients; Furthermore, the doctor-patient linkage module is used to link databases through online and offline resources, provide doctors with data visualization dashboards, and push interactive diagnosis and treatment recommendations to patients. The operation process includes: The doctor-patient linkage module includes a doctor side and a patient side. The doctor side provides a data visualization dashboard to display the physiological digital twin and treatment plan in real time and supports manual adjustment by the doctor. The expression is:

[0054] Where, Indicates the current patient portrait Hedi Historical patient portraits The similarity value between Represents the portrait vector of the current patient, Indicates the A portrait vector of a historical patient, Expressing arrive The items are accumulated, is the total number of features in the patient portrait, Indicates the The weight of the feature, Indicates the current patient Features With historical patients No. Features The numerical product of Indicates the current patient portrait The weighted norm of Indicates historical patients portrait The weighted norm of .

[0055] Furthermore, the doctor-patient linkage module is used to link databases through online and offline resources, provide doctors with data visualization dashboards, and push interactive diagnosis and treatment recommendations to patients. The operation process includes: The patient end push interactive diagnosis and treatment suggestion includes but is not limited to animation demonstration insulin injection adjustment method, the online and offline resource linkage database, the adaptation degree of calculation and patient image expression:

[0056] In the formula, The adaptation degree calculation result of the first Diagnosis and treatment suggestion , The first diagnosis and treatment suggestion to be evaluated , From To The formula is accumulated, The total number of related data such as historical patient cases participating in calculation, Indicates the similarity value between the current patient image And the first Historical patient image , Indicates the current patient image vector, Indicates the first Historical patient image vector, The evaluation function represents the effect of applying the diagnosis and treatment suggestion To the first Corresponding image of the historical patient , select The highest suggestion is pushed in the form of animation, the online and offline resource linkage database, including but not limited to offline medical institutions diagnosis and treatment resources, equipment, expert information and through natural language processing technology Record key information of doctor-patient communication, and return to deep learning model.

[0057] ​Specifically, the doctor-patient linkage module constructs an efficient collaboration mechanism between doctors and patients by integrating online and offline resource linkage databases. The doctor end uses a data visualization dashboard to view the patient's physiological digital twin in real time, such as dynamic blood glucose curves, thyroid hormone metabolism models, and artificial intelligence generated treatment plans, and can manually adjust them according to clinical experience. The patient end receives interactive diagnosis and treatment recommendations customized based on their profile, such as animated demonstrations of insulin injection site rotation and dose adjustment operations. The application effect of diagnosis and treatment recommendations in similar cases is combined to push the most suitable recommendations to patients. The database records key information in doctor-patient communication, such as patient adverse reactions to drugs and changes in lifestyle habits, which is fed back to the deep learning model for continuous optimization. Combined with doctor experience and machine intelligence, the system proposes treatment recommendation paths that better fit the patient's condition. Through animation demonstration and interactive recommendations, the system reduces the understanding threshold of complex treatment operations for patients, improves compliance and treatment effect, and records doctor-patient communication through natural language processing to form a diagnosis-feedback-learning-optimization closed loop, which is conducive to improving the intelligence level and clinical practicability of the system.

[0058] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements will not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital, characterized by: include: Data collection module, used to access multi-source data from wearable devices, home testing equipment, patient self-reporting, offline medical records, and third-party platforms; The timing alignment module is used to align multi-source data based on timestamps using a high-precision timing alignment algorithm and output aligned data; An intelligent learning module, configured to identify endocrine disease data associations in the aligned data using a deep learning model, generate a physiological digital twin of the patient, and update it in real time; A dynamic diagnosis and treatment module, which is used to output a personalized treatment plan based on the physiological digital twin combined with the knowledge graph; The doctor-patient linkage module is used to link databases through online and offline resources, provide doctors with data visualization dashboards, and push interactive diagnosis and treatment recommendations to patients.

2. The artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital according to claim 1 is characterized in that: The data collection module is used to access multi-source data from wearable devices, home testing equipment, patient self-reporting, offline medical records, and third-party platforms. The operation process includes: The data acquisition module uses SSL / TLS encryption transmission protocol and establishes a local cache mechanism. The local cache mechanism establishes a local cache pool Each data path is cached separately, and the cache scheduling strategy adopts a priority-weighted sliding window mechanism, expressed as: Where, Indicates at time Time The status of the local cache pool corresponding to the path data, represents the size of the sliding window, Indicates that from the moment arrive Sum all items in this interval. Indicates at time The weight coefficient of the data, Indicates at time Time The original data content of the road data, Indicates all moments The corresponding weight coefficient The sum is 1, is a hyperparameter used to control the speed of weight decay. is an exponential decay function, is the normalization factor.

3. The artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital according to claim 2 is characterized in that: The timing alignment module uses a high-precision timing alignment algorithm to align multi-source data based on timestamps. The process of outputting aligned data includes: Using high-precision time series alignment algorithm, Heterogeneous data source devices, different device data Acquisition frequency , generate data sequence, expression: Where, Indicates the The original data sequence of heterogeneous data sources, Indicates the The device in The timestamp of the sampling moment, Indicates the The device in The observation value at each sampling moment, Indicates the The total number of samples of each device is reconstructed based on the global unified timestamp using the spline function interpolation method with disturbance compensation. The expression is: Where, Indicates the Devices at the global unified time The interpolated reconstruction value at , Represents a globally unified timestamp, Represents the measurement of the original sampling point Global unified time The contribution of the reconstruction value at is the Gaussian kernel part, is the disturbance compensation coefficient, It is the global unified time The cubic spline function at .

4. The artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital according to claim 3 is characterized in that: The timing alignment module uses a high-precision timing alignment algorithm to align multi-source data based on timestamps. The process of outputting aligned data includes: The alignment process constructs a global alignment objective function and aligns the Interpolation sequences are processed to minimize asynchronous cost, expression: Where, is the global alignment loss function, Indicates traversing the global unified timestamp set Each global unified timestamp in , Indicates traversing all data pairs, is the weight of the data source pair. Hedi The importance of data sources, Indicates the Data sources at time The interpolated data at Is the square of the Euclidean distance, output alignment data, expression: Where, represents the aligned data set, Represents a global unified timestamp, indicating the timestamps from 1 to Data sources at time The interpolated data at Indicates the total number of data sources, Indicates traversing the global unified timestamp set At each time point .

5. The artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital according to claim 4 is characterized in that: The intelligent learning module is used to identify endocrine disease data associations in the aligned data through a deep learning model, generate a physiological digital twin of the patient, and update the operating procedures in real time, including: The deep learning model is based on the attention mechanism, expressed as: Where, express Moment The hidden state vector of each channel / modality, yes activation function, express Moment The original input data of channels / modalities, represents the embedding weight matrix, represents the embedding bias vector, express Moment The attention output vector of each channel / modality, Indicates that the splicing operation will The outputs of the attention heads are stitched together. Indicates the The value weight matrix of the attention head will be hidden state Convert to a vector of values, Indicates the use of attention weight Weighted sum of value vectors, represents the total number of attention heads, express Moment Positions in the attention head and location The attention weights are set, and the deep learning model is optimized by training through Apache Hadoop and Spark frameworks and adopting cross-validation method.

6. The artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital according to claim 5 is characterized in that: The intelligent learning module is used to identify endocrine disease data associations in the aligned data through a deep learning model, generate a physiological digital twin of the patient, and update the operating procedures in real time, including: The physiological digital twin of the patient is generated and updated in real time. The state vector of the physiological digital twin includes the core state of hormone secretion level and metabolic rate, which is generated by hidden state mapping, and the expression is: Where, express The state vector of the physiological digital twin at each moment, represents the state mapping weight matrix, express The hidden state vector at time t, Represents the state mapping bias vector. When new data is input, the historical state and the new observation value are fused through Kalman filtering. The expression is: Where, express Time has come The predicted state value at the moment, represents the state transition matrix, express The state vector of the physiological digital twin at each moment, represents the control matrix, yes The control vector at time t, express Time has come The state prediction error covariance matrix at time , express The error covariance matrix of the state at the moment, Represents the state transition matrix The transposed matrix of represents the process noise covariance matrix, express The Kalman gain at time t, represents the observation matrix, is the inverse covariance matrix for calculating the observation noise, is the observation noise covariance matrix, After integrating new observations The physiological digital twin state after constant correction, express The new observation value at time , According to the predicted state The derived theoretical observations are used to compare with the new observations Compare and calculate the correction amount, After integrating new observations The error covariance matrix after time correction, Represents the identity matrix used for identity transformation in matrix operations.

7. The artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital according to claim 6 is characterized in that: The dynamic diagnosis and treatment module is used to output a personalized treatment plan based on the physiological digital twin and the knowledge graph. The operation process includes: The dynamic diagnosis and treatment module uses a reinforcement learning model to load the quarterly updated endocrine disease diagnosis and treatment knowledge graph in the form of triples, and performs semantic parsing and feature mapping. Based on the physiological digital twin combined with the knowledge graph, a decision feature vector is generated. The expression is: Where, express The decision feature vector generated at each moment, Indicates that the fusion weight matrix is ​​a learnable parameter used to perform a linear transformation on the concatenated vector. It is a knowledge graph feature fusion calculation based on the attention mechanism. is the total number of entities involved in the calculation in the knowledge graph, is the attention weight matrix, a learnable parameter used to calculate the physiological digital twin state vector and knowledge graph entity vectors The association weight between Indicates the The vector of entities, is the first entity vectors, Indicates that it is calculated by the exponential function and The relevance score after attention matrix transformation, Represents all entities with The sum of the correlation scores is used for normalization, is the attention weight representing the Knowledge graph entities relative to physiological digital twins the importance of The vector concatenation operation represents the concatenation of the physiological digital twin state vector and the feature vector obtained by fusing the knowledge graph through the attention mechanism. is the learnable parameter of the fusion bias vector and Cooperate and perform offset adjustment on the concatenated and linearly transformed vectors.

8. The artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital according to claim 7 is characterized in that: The dynamic diagnosis and treatment module is used to output a personalized treatment plan based on the physiological digital twin and the knowledge graph. The operation process includes: The reinforcement learning model adopts an Actor-Critic architecture and consists of a strategy function and a value function. The output is a personalized treatment plan. The knowledge graph is updated quarterly to incorporate the latest research results and clinical experience. The reward function of the reinforcement learning model is centered on treatment effect indicators, including blood sugar control target rate and complication rate. The expression is: Where, Indicates the Instant rewards at all times, is the weight coefficient The reward weight for achieving blood sugar control targets Control the penalty weights for other adverse clinical events, It is an exponential function that makes the reward show a smooth decay / increase trend with blood sugar deviation. Indicates the The actual blood sugar value at the moment, Indicates target blood sugar value, Indicates the tolerance parameter for blood sugar fluctuations, represents the square of blood glucose deviation, Representative Other adverse clinical events at this time included the risk of hypoglycemia and ketosis.

9. The artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital according to claim 8 is characterized in that: The doctor-patient linkage module is used to link databases through online and offline resources, provide doctors with data visualization dashboards, and push interactive diagnosis and treatment recommendations to patients. The operation process includes: The doctor-patient linkage module includes a doctor side and a patient side. The doctor side provides a data visualization dashboard to display the physiological digital twin and treatment plan in real time and supports manual adjustment by the doctor. The expression is: Where, Indicates the current patient portrait Hedi Historical patient portraits The similarity value between Represents the portrait vector of the current patient, Indicates the A portrait vector of a historical patient, Expressing arrive The items are accumulated, is the total number of features in the patient portrait, Indicates the The weight of the feature, Indicates the current patient Features With historical patients No. Features The numerical product of Indicates the current patient portrait The weighted norm of Indicates historical patients portrait The weighted norm of .

10. The artificial intelligence-assisted diagnosis and treatment system for endocrine diseases based on an Internet hospital according to claim 9 is characterized in that: The doctor-patient linkage module is used to link databases through online and offline resources, provide doctors with data visualization dashboards, and push interactive diagnosis and treatment recommendations to patients. The operation process includes: The interactive diagnosis and treatment suggestions pushed by the patient end include an animated demonstration of the insulin injection adjustment method. The online and offline resource linkage database is used to calculate the compatibility with the patient portrait. The expression is: Where, Indicates that for Treatment recommendations The fitness calculation results are: Indicates the Treatment recommendations awaiting evaluation. It is from arrive Accumulate the following formulas, is the total number of historical patient case data involved in the calculation, Indicates the current patient portrait Hedi Historical patient portraits The similarity value between Represents the portrait vector of the current patient, Indicates the A portrait vector of a historical patient, Is the evaluation function that measures the diagnosis and treatment recommendations Apply to Corresponding portraits of historical patients The effect produced when selecting The highest recommendation is pushed in the form of animation. The online and offline resource linkage database includes the diagnosis and treatment resources, equipment, expert information of offline medical institutions, and the key information of doctor-patient communication recorded through natural language processing technology to feed back the deep learning model.

Images