An internet-based remote medical consultation system

Abstract

The application relates to the technical field of medical information, and discloses a remote medical inquiry system based on the Internet, which comprises the following modules: a data acquisition and quality guarantee module, which is used for collecting multi-source heterogeneous medical data, performing time sequence alignment, quality monitoring and intelligent repair interpolation; a feature extraction and fusion module, which is used for performing deep feature extraction, cross-modal semantic alignment and hierarchical attention fusion; a complication correlation reasoning module, which adopts a graph attention network and multi-hop reasoning to obtain a deep complication correlation reasoning result; a complication progression prediction module, which is used for constructing a complication progression prediction model, performing meta-learning enhancement and uncertainty quantification; an intelligent medication decision module, which is used for generating candidate schemes and screening a Pareto optimal scheme; a compliance management module, which is used for performing compliance causal inference and closed-loop optimization; and an effect evaluation module, which is used for performing effect evaluation and dynamic optimization; the application establishes a compliance improvement mechanism through causal inference and reinforcement learning, and realizes interpretable human-computer collaborative decision and closed-loop optimization management.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and more specifically, to an Internet-based remote medical consultation system. Background Technology

[0002] With the continued rise in global diabetes prevalence and the accelerating aging of the population, the number of patients with diabetes and multiple chronic complications is constantly expanding. Managing these patients involves multiple organ systems, various monitoring indicators, and the synergy of multiple treatment drugs, requiring frequent medical follow-ups, assessment of complication progression, and dynamic adjustments to medication regimens. However, due to factors such as patients' physical condition, geographical location, and uneven distribution of medical resources, many patients struggle to access timely, continuous, and professional medical services, leading to increasingly prominent problems such as accelerated complication progression, poor treatment outcomes, and increased medical costs. Internet-based telemedicine consultation systems offer a new technological path to address this dilemma, enabling remote doctor-patient interaction beyond geographical limitations. However, existing telemedicine systems still have significant shortcomings in their intelligent capabilities for managing complex chronic diseases.

[0003] Existing telemedicine consultation systems suffer from the following technical problems: a lack of effective fusion and analysis capabilities for time-series multimodal medical data; the inability to fully integrate and utilize heterogeneous data from multiple sources, such as continuous blood glucose monitoring, dynamic renal function indicators, fundus imaging sequences, nerve conduction test results, and lifestyle behavior data, leading to incomplete and inaccurate disease status assessments; a lack of intelligent identification and prediction capabilities for the evolutionary patterns of multiple complications, making it difficult to accurately predict the progression of complications and the risk of acute exacerbations, hindering proactive medical intervention; a lack of dynamic and personalized multi-drug synergistic optimization mechanisms, resulting in untimely or inappropriate adjustments to medication regimens and insufficient consideration of interactions between different drugs, impacting treatment effectiveness; and a lack of intervention strategies based on precise modeling of patient adherence behavior, leading to crude adherence management methods that fail to provide personalized solutions for different adherence barriers, significantly reducing the effectiveness of remote management. These technical problems severely restrict the application of telemedicine consultation systems in complex chronic disease management scenarios.

[0004] To address the aforementioned technical challenges, it is necessary to construct a novel telemedicine consultation system that possesses deep multimodal data fusion capabilities, intelligent disease evolution prediction capabilities, personalized medication optimization capabilities, and precise adherence management capabilities. Summary of the Invention

[0005] This invention provides an Internet-based remote medical consultation system that addresses the technical problems in related technologies, such as the lack of effective fusion and analysis capabilities for time-series multimodal medical data, the lack of intelligent identification and prediction capabilities for the evolution patterns of multiple complications, and the lack of dynamic and personalized multi-drug synergistic optimization mechanisms.

[0006] This invention provides an Internet-based remote medical consultation system, comprising:

[0007] The data acquisition and quality assurance module collects multi-source heterogeneous medical data, performs time-series alignment, quality monitoring, and intelligent repair and interpolation, and obtains a medical dataset with enhanced integrity.

[0008] The feature extraction and fusion module performs deep feature extraction, cross-modal semantic alignment, and hierarchical attention fusion on the complete medical dataset to obtain a hierarchically fused patient health status representation vector.

[0009] The complication association reasoning module, based on the hierarchically fused patient health status representation vector, executes graph attention network and multi-hop reasoning to obtain deep complication association reasoning results;

[0010] The complication progression prediction module constructs a complication progression prediction model based on the deep complication association reasoning results, performs meta-learning enhancement and uncertainty quantification, and obtains complication progression prediction results.

[0011] The intelligent medication decision-making module generates candidate solutions and selects the Pareto optimal solution based on the complication progression prediction results, thus obtaining a multi-level personalized medication recommendation plan;

[0012] The adherence management module, based on a multi-level personalized medication recommendation scheme, performs causal inference and closed-loop optimization of adherence to obtain an adaptive adherence improvement scheme.

[0013] The efficacy evaluation module, based on the adaptive compliance improvement plan, performs efficacy evaluation and dynamic optimization, and generates a treatment efficacy evaluation report.

[0014] In a preferred embodiment, the data acquisition and quality assurance module includes:

[0015] A data transmission channel is established with the patient's physiological signal monitoring equipment through the Internet of Things (IoT) device access protocol, and a local data caching mechanism is set up to achieve network fault tolerance.

[0016] A unified time reference axis is established using a timestamp alignment algorithm;

[0017] A real-time data quality monitoring algorithm was used to detect missing data, outliers, and noise. Missing data was processed using a temporal imputation method based on long short-term memory networks and a collaborative imputation method based on patient characteristics.

[0018] In a preferred embodiment, the feature extraction and fusion module includes:

[0019] A Transformer encoder architecture is used to extract deep features from time-series physiological signal data, and the time-series signal feature vector is obtained through position encoding and self-attention mechanism.

[0020] The Vision Transformer model was used to extract features from fundus image data to obtain image feature vectors.

[0021] The TabNet deep learning network, specifically designed for tabular data, is used to extract features from structured inspection report data, resulting in structured data feature vectors.

[0022] A cross-modal semantic alignment method based on contrastive learning is used to map the features of each modality to a common semantic space; a hierarchical cross-modal attention fusion mechanism is used to perform deep feature fusion to obtain a hierarchically fused patient health status representation vector.

[0023] In a preferred embodiment, the complication association reasoning module includes:

[0024] A pre-trained biomedical language model is used for knowledge extraction to identify disease entities, symptom entities, drug entities, and examination indicator entities. The causal relationships, co-occurrence relationships, and treatment relationships between entities are extracted to construct a knowledge graph of diabetic complications.

[0025] Incremental learning and knowledge fusion methods are used to dynamically update the knowledge graph, resulting in a dynamically evolving disease knowledge graph;

[0026] An entity linking algorithm is used to map the patient's symptoms and test indicators to entity nodes in the knowledge graph, thereby obtaining the representation of the patient's disease state in the knowledge graph.

[0027] A graph attention network is used to perform message passing and feature aggregation on a knowledge graph, and the association paths between complications and the risk of potential complications are identified through multi-hop information propagation.

[0028] In a preferred embodiment, the complication progression prediction module includes:

[0029] A self-supervised learning method was used to pre-train a multi-task complication prediction model on an electronic medical record database of diabetic patients. The disease evolution pattern was learned through masked prediction task and contrastive prediction task. A multi-task learning framework was constructed to simultaneously predict the occurrence and progression of diabetic nephropathy, retinopathy, peripheral neuropathy and cardiovascular complications.

[0030] The pre-trained model is adapted to a specific patient population in the target medical institution using transfer learning techniques to obtain a transfer-optimized prediction model.

[0031] By learning model initialization parameters on the complication combination prediction task, the data is divided into a support set and a query set for each task. Gradient updates are performed on the support set to obtain task-specific parameters, performance is evaluated and meta-loss is calculated on the query set, and the model initialization parameters are updated according to the meta-loss to obtain a meta-learning-enhanced prediction model.

[0032] Complication progression is predicted based on patient health status representation vectors, deep complication association inference results, and patient historical time series data. Bayesian deep learning method is used to quantify uncertainty, calculate prediction mean and prediction variance, and obtain complication progression prediction results with confidence.

[0033] In a preferred embodiment, the intelligent medication decision-making module includes:

[0034] A multi-objective optimization function was constructed using the analytic hierarchy process, and a scoring function for blood glucose control effect, complication delay effect, side effect risk, and quality of life was defined.

[0035] We employ deep reinforcement learning to train an intelligent medication decision agent, construct a Markov decision process for medication decisions, and incorporate drug interaction constraints and rule constraints from clinical practice guidelines.

[0036] Based on the patient's metabolic enzyme genotype, liver and kidney function indicators, drug sensitivity, and allergy history, a personalized drug combination regimen is generated. The synergistic therapeutic effect is calculated using a drug synergistic effect assessment model, and the personalized dosing dose is determined through a pharmacokinetic model.

[0037] The Pareto optimal selection algorithm is used to find the optimal set of solutions, resulting in a multi-level personalized medication recommendation scheme.

[0038] In a preferred embodiment, the compliance management module includes:

[0039] We construct a causal relationship network for compliance behavior using causal inference methods, and use a structural causal model to construct a directed acyclic graph to represent causal relationships, thereby identifying factors that have a causal effect on compliance.

[0040] Counterfactual reasoning techniques were used to evaluate the causal effects of different interventions, resulting in an assessment of the effectiveness of the interventions.

[0041] Patients are categorized into cognitive impairment type, time management difficulty type, side effect sensitivity type, and social support deficiency type. Corresponding intervention combinations are matched for different types of patients, including health education, intelligent reminder optimization, medication regimen adjustment, and social support network construction.

[0042] Online reinforcement learning algorithms are used to continuously optimize compliance prediction models and intervention strategies, resulting in an adaptive compliance improvement scheme.

[0043] In a preferred embodiment, the effect evaluation module includes:

[0044] The system uses explainable artificial intelligence technology to generate visual explanation reports of the decision-making basis, and employs attention weight visualization method, counterfactual explanation method and rule extraction technology to display the decision-making basis;

[0045] Enables collaborative decision-making between doctors and patients using remote audio and video communication technology, and provides interactive solution adjustment functions;

[0046] Multimedia health education content generation technology is used to convey treatment plans to patients and generate personalized health management plans.

[0047] The efficacy evaluation model and deviation analysis method are used to assess the difference between the actual effect and the expected effect of the treatment plan, and the treatment effect data is fed back to the machine learning module for continuous model training and optimization.

[0048] In a preferred embodiment, the data acquisition and quality assurance module further includes:

[0049] Set up a local data caching mechanism on the device side. When a network interruption is detected, the data will be automatically stored in the device's local cache. Once the network is restored, the cached data will be automatically synchronized and uploaded in timestamp order.

[0050] A real-time data quality monitoring algorithm is used to detect missing data and calculate the missing rate of each data source, identify outliers in the data that exceed the physiologically possible range, and detect glitches caused by equipment failure.

[0051] For short-term missing time-series data, an intelligent imputation model based on long short-term memory network is used; for long-term missing data, an estimation imputation method based on similar patient matching is used.

[0052] For outlier data, determine whether it is a real pathological abnormality or a measurement error based on the degree of deviation. If it is determined to be a measurement error, smooth it or delete it. If it is determined to be a real abnormality, retain it and mark it.

[0053] Confidence scores are labeled on all imputed and repaired data to obtain a medical dataset with enhanced integrity.

[0054] A computer-readable storage medium for storing computer-readable instructions that, when read by a computer, enable the execution of the aforementioned Internet-based remote medical consultation system.

[0055] The beneficial effects of this invention are as follows:

[0056] This system employs a Transformer architecture combined with contrastive learning for cross-modal semantic alignment, effectively integrating heterogeneous data such as temporal physiological signals, medical images, and structured diagnostic indicators. It captures collaborative and complementary information across multiple abstraction levels, constructing a comprehensive and accurate representation of patient health status. Deep reasoning techniques based on dynamically evolving disease knowledge graphs and graph attention networks can uncover complex correlation patterns and pathological mechanisms among various complications, identifying potential disease evolution paths. A complication progression prediction model built using transfer learning and meta-learning techniques maintains high prediction accuracy even in scenarios with small sample sizes and rare complication combinations, providing quantitative assessments of uncertainty and offering reliable quantitative evidence for clinical decision-making. These technological innovations enable the system to accurately predict complication progression, identify the risk of acute exacerbations in advance, support proactive medical interventions, effectively delay disease progression, and improve patient prognosis.

[0057] Intelligent medication decision-making models trained using deep reinforcement learning can achieve an optimal balance among multiple treatment goals, including glycemic control, delaying complications, minimizing side effects, and improving quality of life. They comprehensively consider individual factors such as patient metabolic characteristics, genotype, liver and kidney function, as well as drug synergies and interactions, generating scientifically sound personalized medication plans. Pareto optimality screening provides physicians with multi-level options, supporting flexible clinical decision-making. Adherence behavior models constructed using causal inference techniques can identify causal barriers that truly affect patient adherence. Counterfactual reasoning assesses the causal effects of different interventions, matching the most effective personalized intervention combinations to different types of patients. Combined with online reinforcement learning, intervention strategies are continuously optimized. These technological innovations make treatment plans more precise and effective, adherence management more scientific and efficient, ensure the long-term success of remote management, and ultimately achieve comprehensive benefits such as improving the quality of medical services, enhancing patients' quality of life, and reducing medical expenses. Attached Figure Description

[0058] Figure 1 This is a block diagram of an Internet-based remote medical consultation system according to the present invention. Detailed Implementation

[0059] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0060] At least one embodiment of the present invention discloses an Internet-based remote medical consultation system, such as... Figure 1 As shown, it includes:

[0061] The data acquisition and quality assurance module collects multi-source heterogeneous medical data, performs time-series alignment, quality monitoring, and intelligent repair and interpolation, and obtains a medical dataset with enhanced integrity.

[0062] Specifically, the following steps are included:

[0063] Step 1.1: Establish a data transmission channel for the physiological signal monitoring equipment;

[0064] Based on the patient's unique identifier in the telemedicine system, an IoT device access protocol is used to establish a secure data transmission channel with physiological signal monitoring devices worn or used by the patient, such as continuous glucose monitors, ambulatory blood pressure monitors, smart bracelets, and pulse oximeters. A local data caching mechanism is set up on the device side. When the network connection is normal, monitoring data is uploaded to the cloud server in real time; when a network interruption is detected, the data is automatically stored in the device's local cache; after the network is restored, the cached data is automatically synchronized and uploaded in timestamp order. This results in a real-time physiological signal data stream with network fault tolerance capabilities, containing physiological parameters such as blood glucose levels, blood pressure levels, heart rate, and blood oxygen saturation, along with their corresponding timestamp information.

[0065] Step 1.2: Obtain electronic medical record data from medical institutions;

[0066] Based on the data interface between the Hospital Information System (HIS) and the Laboratory Information System (LIS), and using the HL7 FHIR international medical information exchange standard protocol, the system allows patients to access their electronic medical records at medical institutions with patient authorization. It acquires renal function indicators such as serum creatinine, blood urea nitrogen, and urinary microalbumin from regular renal function test reports; retinal photographs from fundus imaging examinations; nerve conduction velocity test reports; and structured and unstructured medical documents such as historical inpatient medical records, outpatient medical records, medication records, and surgical records. Simultaneously, it acquires basic patient information including age, gender, height, weight, duration of diabetes, past medical history, family history, and allergy history. This results in a multi-source, heterogeneous medical data set from medical institutions, encompassing comprehensive medical information required for patient disease management.

[0067] Step 1.3: Perform multimodal data timing alignment;

[0068] Based on real-time physiological signal data streams and multi-source heterogeneous medical datasets, a timestamp alignment algorithm is employed. Due to differences in sampling frequencies across different data sources—with continuous blood glucose monitoring data sampling intervals of several minutes, blood pressure monitoring data sampling intervals of tens of minutes, and renal function test data sampling intervals of several weeks or months—a unified time reference axis needs to be established to map all data onto this standard time axis according to their acquisition time. For high-frequency sampled data, downsampling techniques are used to extract key time point values; for low-frequency sampled data, the original sampling frequency is maintained. A time window association method is used to group and associate different modalities within the same time window, resulting in time-aligned multimodal medical data that is comparable and correlated in the time dimension.

[0069] Step 1.4: Implement data quality monitoring and assessment;

[0070] Based on time-series aligned multimodal medical data, a real-time data quality monitoring algorithm is employed to automatically check the completeness, consistency, and reasonableness of the data. This includes detecting missing data and calculating the missing rate for each data source; identifying outliers such as blood glucose levels exceeding physiological limits or blood pressure values ​​significantly deviating from the patient's historical normal range; detecting data noise such as spikes caused by equipment malfunctions; and generating a data quality assessment report when data quality issues are detected. This report includes information such as data integrity score, outlier location markers, and data reliability level.

[0071] Step 1.5: Perform intelligent data repair and interpolation;

[0072] Based on the data quality assessment report, a multi-strategy data repair algorithm was employed to address the identified data quality issues. For short-term missing time-series data, such as missing blood glucose values ​​at several time points, a time-series imputation model based on Long Short-Term Memory (LSTM) networks was used to intelligently imputate the data using the patient's historical time-series patterns and data values ​​from previous and subsequent time points. For long-term missing data, such as data from tests that have not been performed for a long time, a collaborative imputation method based on similar patient matching was used. This method retrieved patient groups in the database with similar characteristics to the patient in terms of age, gender, disease duration, and complication types, and used the mean of the corresponding indicator from similar patients for estimation and imputation. For outlier data, the degree of deviation was used to determine whether it was a true pathological abnormality or a measurement error. If it was determined to be a measurement error, it was smoothed or deleted; if it was determined to be a true abnormality, it was retained and marked as data of key concern. Confidence scores were labeled on all imputed and repaired data to obtain a medical dataset with enhanced integrity. This dataset, while maintaining the original true data, improved the integrity and usability of the data through intelligent imputation and repair, providing a high-quality data foundation for subsequent analysis.

[0073] The feature extraction and fusion module performs deep feature extraction, cross-modal semantic alignment, and hierarchical attention fusion on the complete medical dataset to obtain a hierarchically fused patient health status representation vector.

[0074] Specifically, the following steps are included:

[0075] Step 2.1: Extract deep features of time-series physiological signals;

[0076] Based on time-series physiological signal data from a medical dataset enhanced for integrity, including continuous blood glucose monitoring sequences, blood pressure monitoring sequences, and heart rate sequences, a Transformer encoder architecture is employed for deep feature extraction. The Transformer encoder consists of multiple stacked self-attention layers and feedforward neural network layers. For time-series signals containing multiple time points, positional encoding embeds temporal location information into the input, enabling the model to perceive the temporal order of the data, resulting in a position-embedded input representation, which is then fed into the Transformer encoder. The self-attention mechanism calculates the association weights between any two time points in the sequence through the following steps: linearly transforming the input sequence to obtain three different representations: query matrix, key matrix, and value matrix; calculating the product of the query matrix and the transpose of the key matrix to measure the similarity between time points; scaling the similarity using the dimension of the key vector to prevent the vanishing gradient problem; converting the similarity into a probability distribution form of attention weights using a softmax normalization function; and weighted summing of the value matrix using these attention weights to obtain the final attention output. A multi-head attention mechanism captures temporal dependencies from different representation subspaces, resulting in a context-aware feature representation for each time point after multi-layer encoding. Global average pooling is performed on the feature representations at all time points to obtain a fixed-dimensional time series signal feature vector. This feature vector comprehensively encodes the local fluctuation patterns and long-term evolution trends of the time series signal.

[0077] Step 2.2: Extract depth features from fundus images;

[0078] Based on fundus image data from a medical dataset with enhanced integrity, the Vision Transformer model was used for image depth feature extraction. The Vision Transformer divides the 2D image into multiple image patches, flattens each patch into a vector, and obtains the patch embedding representation through linear projection. This embedding representation is then fed into the Transformer encoder after adding positional encoding. The Transformer encoder learns the spatial relationships and semantic associations between different image patches through a self-attention mechanism, capturing the spatial distribution patterns and interrelationships of pathological features such as retinal vessel morphology, microaneurysms, hard exudates, and hemorrhages in fundus images. The input fundus image contains three dimensions: image height, image width, and number of image channels. After encoding by the Vision Transformer, the feature representations corresponding to the classification labels are extracted as the global feature vector for the entire image, resulting in a fixed-dimensional image feature vector. This feature vector contains the visual semantic information of fundus lesions.

[0079] Step 2.3: Extract deep features from structured data;

[0080] Based on structured laboratory reports and patient basic information data from a comprehensive medical dataset, including numerical and categorical data such as renal function indicators, liver function indicators, blood lipid indicators, glycated hemoglobin, patient age, and disease course, TabNet, a deep learning network specifically designed for tabular data, is used for feature extraction. TabNet employs a sequential attention mechanism to progressively select the most relevant features for processing, learning the non-linear interactions between features through feature selection masks and feature transformers. For an input data vector containing multiple structured features, TabNet progressively aggregates feature information through a multi-step decision-making process, focusing on different feature subsets at each step, ultimately outputting a fixed-dimensional comprehensive structured data feature vector. This feature vector captures the correlation patterns between various biochemical indicators and provides a comprehensive representation of the patient's metabolic state, resulting in a structured data feature vector.

[0081] Step 2.4, achieve cross-modal semantic alignment;

[0082] Based on temporal signal feature vectors, image feature vectors, and structured data feature vectors, a cross-modal semantic alignment method based on contrastive learning is adopted to learn the alignment mapping between modalities in a common semantic space. The goal of contrastive learning is to make features from different modalities within the same patient and time window closer in semantic space, while features from different patients or different health states are farther apart. Each modal feature vector is mapped to a common semantic space of a unified dimension through a modality-specific projection network, resulting in aligned temporal feature representations, image feature representations, and structured data feature representations. The contrastive learning loss function adopts an information noise contrastive estimation form, maximizing the similarity of positive sample pairs (features from different modalities within the same patient) and minimizing the similarity of negative sample pairs (features from different patients). The projection network is trained using this loss function, aligning features from different modalities but with the same semantics in the common space, resulting in semantically aligned cross-modal feature representations. These features are comparable and fusionable at the semantic level.

[0083] Step 2.5: Perform hierarchical cross-modal feature fusion;

[0084] Semantic alignment-based cross-modal feature representation includes temporal feature representation, image feature representation, and structured data feature representation. A hierarchical cross-modal attention fusion mechanism is employed for deep feature fusion. Hierarchical fusion models the interaction relationships between modalities at multiple abstract levels. The first layer of attention fusion focuses on the low-level detailed interactions of different modal features, calculating mutual attention weights between modalities. For temporal and image modal features, it calculates the attention between temporal features as queries and image features as keys and values, obtaining an enhanced representation of temporal features by image information. Similarly, it calculates mutual attention enhancement between other modal pairs. The second layer of attention fusion focuses on mid-level semantic interactions, further calculating the three-way attention between the three modalities based on the enhanced features of the first layer, modeling the collaborative mode of the three modalities. The third layer of attention fusion focuses on high-level abstract semantic integration, performing adaptive weighted clustering on the fusion results of the first two layers. The system learns the importance weights of each modality in different patients and disease states. For some patients, imaging data may be more important, while for others, time-series data may be more critical. The system adaptively adjusts the fusion weights through a gating mechanism. The feature vectors of the multi-level fusion are concatenated and mapped to a fixed dimension through a fully connected network to obtain a hierarchical fusion patient health status representation vector. This representation vector comprehensively integrates the patient's physiological signal time-series features, imaging pathological features, and biochemical index features. It captures cross-modal collaborative and complementary information at multiple abstract levels, providing a high-quality comprehensive representation foundation for subsequent disease analysis and prediction.

[0085] In some embodiments, due to significant differences in the information richness of modal data across different patients and disease stages, a dynamic modality importance assessment and adaptive fusion mechanism can be employed. The aim is to dynamically adjust the fusion weights of each modality based on specific circumstances to achieve the optimal fusion effect. Specifically, a modality quality assessment module is introduced during the fusion process. This module calculates a quality score for each modality based on indicators such as the integrity score, signal-to-noise ratio, and feature discriminancy of the data. Modalities with higher quality scores are assigned greater weights during fusion, while modalities with lower quality scores are assigned smaller weights or, in extreme cases, may be excluded. An adaptive weight allocation strategy is trained using reinforcement learning. The agent selects the optimal combination of fusion weights based on the current patient's modality data quality and disease characteristics. The strategy is optimized through feedback from downstream tasks such as complication prediction, learning the optimal fusion strategy for different scenarios. This method can improve the robustness and effectiveness of the fusion representation under conditions of varying data quality.

[0086] The complication association reasoning module, based on the hierarchically fused patient health status representation vector, executes graph attention network and multi-hop reasoning to obtain deep complication association reasoning results;

[0087] Specifically, the following steps are included:

[0088] Step 3.1: Construct an initial disease knowledge graph;

[0089] Based on large-scale medical literature databases and clinical practice guideline databases, a pre-trained biomedical language model, BioBERT, is used for knowledge extraction. BioBERT is pre-trained on large-scale biomedical texts to learn language representations in the medical field. For the input medical literature text, BioBERT is used for named entity recognition to identify disease entities such as diabetic nephropathy and diabetic retinopathy, symptom entities such as proteinuria and vision loss, drug entities such as metformin and ACEI antihypertensive drugs, and test indicator entities such as serum creatinine and urinary microalbumin, resulting in an entity list. A relation extraction model is used to identify the relationships between entities, including causal relationships such as hyperglycemia leading to glomerular damage, co-occurrence relationships such as diabetic nephropathy often accompanied by hypertension, and treatment relationships such as ACEI drugs being used to treat diabetic nephropathy. The relation extraction model is based on BERT's sentence pair classification architecture to determine whether a specific relationship exists between two entities and the type of relationship. Through automated knowledge extraction from massive amounts of medical literature, a knowledge set containing a large number of entity and relation triples is obtained. This knowledge set covers various aspects of diabetes and its complications, including pathological mechanisms, clinical manifestations, diagnostic criteria, and treatment plans. The extracted entities are used as nodes in the knowledge graph, and the relations are used as edges to construct an initial knowledge graph of diabetes complications, resulting in an automatically constructed disease knowledge graph that organizes domain knowledge in a structured form.

[0090] Step 3.2: Implement dynamic updates to the knowledge graph;

[0091] Based on an automatically constructed disease knowledge graph and feedback from clinical experts, incremental learning and knowledge fusion methods are used to dynamically update and improve the knowledge graph. Medical knowledge is constantly being updated, new research findings are continuously published, and clinical guidelines are regularly revised; therefore, the knowledge graph needs to have dynamic evolution capabilities. New knowledge is periodically extracted from the latest published medical literature, new entities and relationships are identified, and the new knowledge is integrated with the existing knowledge graph. The fusion process requires knowledge conflict detection. When new knowledge contradicts existing knowledge, the credibility of the knowledge is judged based on factors such as the authority of the knowledge source, publication time, and level of evidence, retaining the knowledge with higher credibility. For domain knowledge annotated by clinical experts or corrected knowledge extraction errors, expert knowledge is directly integrated into the graph as high-confidence knowledge. Vector representations of entities and relationships are learned through knowledge graph embedding technology, making semantically similar entities close in distance in the vector space. Vector representations can be used for knowledge reasoning and completion, inferring possible but not yet explicitly labeled relationships. Through continuous incremental updates and knowledge fusion, a dynamically evolving disease knowledge graph is obtained. This graph maintains the timeliness and accuracy of knowledge and can reflect the latest medical understanding.

[0092] Step 3.3: Map patient status to knowledge graph;

[0093] Based on a hierarchically fused patient health status representation vector and a dynamically evolving disease knowledge graph, an entity linking algorithm is used to map patient symptoms and detection indicators to entity nodes in the knowledge graph. For the patient health status representation vector, a decoding network maps it to an interpretable symptom and indicator space, obtaining a list of the patient's current significant symptoms and abnormal indicators, such as proteinuria, elevated serum creatinine, and decreased vision. For each symptom and indicator, the corresponding entity node is retrieved in the knowledge graph, and entity disambiguation is used to handle polysemy, ensuring correct mapping to the correct entity. All patient-related entity nodes are marked as active nodes, forming a subgraph on the knowledge graph. This subgraph represents the projection of the patient's current disease state into the knowledge space, resulting in a representation of the patient's disease state in the knowledge graph. This representation establishes a connection between the patient's data features and medical knowledge.

[0094] Step 3.4: Perform graph attention network inference;

[0095] Based on the representation of patient disease states in a knowledge graph and the dynamic evolution of the disease knowledge graph, a graph attention network (GAN) is employed for message passing and feature aggregation on the knowledge graph. The GAN can perform deep learning on graph-structured data, learning the embedded representation of nodes. For each node in the knowledge graph, the GAN calculates the attention weights between the node and its neighbors to reflect the importance of the neighbors to the central node through the following steps: linear transformation of the weight matrices is performed on the feature vectors of the central node and its neighbors; the transformed feature vectors are concatenated; the concatenated result is linearly combined using the attention parameter vector; a nonlinear transformation is performed using a linear rectified activation function with leakage; and the probability distribution of the attention weights is obtained by calculating the exponential normalization of all neighbors of the central node. The representation of the central node is updated by aggregating the features of the neighbors through attention weighting. Specifically, the update process involves: linear transformation of the weight matrices is performed on the feature vectors of all neighbors of the central node; the calculated attention weights are used to weight and sum these transformed features; and a nonlinear transformation is performed using an activation function to obtain the updated node representation. Multi-layered graph attention networks enable multi-hop information propagation. Information spreads along the relational edges of the knowledge graph from the activated node. After multi-hop propagation, related nodes that are close to the activated node on the graph are also activated or receive high responses. These nodes represent potential complications, comorbidities, and pathological mechanisms closely related to the patient's current disease state. By analyzing activation patterns and node response strength, the association paths between various complications in the patient are identified. For example, diabetes leads to glomerulosclerosis through high blood sugar, resulting in nephropathy, which in turn exacerbates hypertension and affects retinal vessels, forming an association network between complications. Simultaneously, potential complications that are currently undiagnosed but at high risk are identified—that is, complication nodes on the knowledge graph that are highly associated with diagnosed complication nodes but have not yet manifested in the patient. Deep complication association inference results are obtained, including causal relationships, comorbid relationships, explanations of pathological mechanisms, and early warnings of potential complication risks, providing doctors with comprehensive support for disease association analysis.

[0096] In some embodiments, due to the temporal dynamics of the evolutionary path of diabetic complications, temporal knowledge graphs and time-aware reasoning methods can be employed to more accurately model the disease's evolution over time. Specifically, temporal information is added to the edges of the knowledge graph to represent the time or time period in which the relationship occurs, such as the average number of years after the onset of diabetes before kidney disease develops. A temporal knowledge graph is constructed, and during the reasoning process, not only the semantic information of entities and relationships is considered, but also the constraints of the time dimension. A time-aware graph neural network is used, incorporating a time decay mechanism during message passing. Knowledge closer to the current time is given higher weight, while knowledge further away has its weight decayed. Through temporal reasoning, the future evolutionary trajectory of a patient's disease state can be predicted, and the complications that a patient may experience at different points in the future can be inferred. This method can provide more accurate risk prediction in the time dimension.

[0097] The complication progression prediction module constructs a complication progression prediction model based on the deep complication association reasoning results, performs meta-learning enhancement and uncertainty quantification, and obtains complication progression prediction results.

[0098] Specifically, the following steps are included:

[0099] Step 4.1: Pre-train the multi-task complication prediction base model;

[0100] Based on a large-scale electronic medical record database of diabetic patients, a self-supervised learning method was used to pre-train a multi-task complication prediction model. The electronic medical record database contains longitudinal medical records of tens of thousands to hundreds of thousands of patients, recording complete time-series data from the diagnosis of diabetes to the development of complications. Self-supervised learning does not require manual annotation; it learns useful representations from the data itself by designing proxy tasks. A masking prediction task is used, randomly masking observations at certain time points in the patient's time-series data. The model is trained to predict the masked observations based on data from before and after these time points, learning the time-dependent patterns of disease evolution. A contrastive prediction task is used, with patient pairs with similar disease progression trajectories as positive samples and patient pairs with different disease progression trajectories as negative samples. The model is trained to make the representations of positive sample pairs similar and the representations of negative sample pairs dissimilar, learning discriminative features to distinguish different disease evolution patterns. A multi-task learning framework is constructed to simultaneously predict the occurrence and progression of various complications, including diabetic nephropathy, retinopathy, peripheral neuropathy, and cardiovascular complications. Each complication prediction task shares a bottom-level feature extraction network to learn general disease representations. Each task has an independent task-specific layer to learn the specific patterns of that complication. Connections are established between tasks through shared parameters and task association regularization to capture the interactive effects between complications; for example, the progression of nephropathy can affect the risk of cardiovascular complications. The model is pre-trained on large-scale data, and the model parameters are optimized through backpropagation and gradient descent to obtain a pre-trained base prediction model. This model learns the general knowledge and patterns of diabetic complication evolution and possesses strong feature representation capabilities.

[0101] Step 4.2: Perform transfer learning model adaptation;

[0102] Based on patient data from the target medical institution and a pre-trained basic prediction model, transfer learning is used to adapt the model to a specific patient population. Patient populations in different medical institutions differ in demographic characteristics, geographical environment, lifestyle, and medical conditions, and directly applying the pre-trained model may result in domain bias. Transfer learning fine-tunes the model on target domain data to adapt it to the characteristics of the target domain. The low-level parameters of the pre-trained model are used as initialization; these parameters have already learned general disease representations. The model is then further trained on patient data from the target medical institution, with a small learning rate for fine-tuning to avoid destroying the useful features learned during pre-training. During fine-tuning, the parameters of task-specific layers are optimized to adapt to the disease characteristic distribution of the target population. Through transfer learning, even with relatively small amounts of data in the target domain, the model can achieve good predictive performance by leveraging pre-trained knowledge, resulting in a transfer-optimized prediction model with higher predictive accuracy on the target patient population.

[0103] Step 4.3: Apply meta-learning to enhance few-shot generalization ability;

[0104] To address the extremely small sample size problem of rare complication combinations, Model-Agnostic Meta-Learning (MAML) is employed to enable the model to rapidly adapt to new tasks. Diabetic patients may simultaneously experience various combinations of complications, and certain specific complication combinations have very small sample sizes in the dataset. Traditional machine learning methods are prone to overfitting and poor generalization in small sample situations. Meta-learning learns quickly on multiple related tasks, allowing the model to adapt rapidly to new tasks with a small sample size. The core idea of ​​the MAML algorithm is to learn a good model initialization parameter, so that starting from this initialization, good performance can be achieved on new tasks with a small number of gradient updates. Meta-learning training... During the process, multiple tasks are sampled from the dataset, each task corresponding to the prediction of a combination of complications. For each task, the data is divided into a support set and a query set. A small number of gradient updates are performed on the support set to obtain task-specific parameters. Performance is evaluated and meta-loss is calculated on the query set. The meta-optimizer updates the model's initialization parameters based on the meta-loss of all tasks, so that the initialization can quickly adapt to various tasks. After training, when faced with a new rare complication combination prediction task, only a few gradient updates are needed on a small number of samples of that task, and the model can quickly adapt and give an accurate prediction, resulting in a meta-learning-enhanced prediction model. This model improves the generalization ability in small samples and rare scenarios, and can provide accurate prediction support for patients with various complication combinations.

[0105] Step 4.4: Generate a confidence-based prediction of complication progression;

[0106] Based on a hierarchically fused patient health status representation vector, deep complication association inference results, and patient historical time-series data, a meta-learning-enhanced prediction model is used to predict the progression of complications. The patient health status representation vector and association inference results are used as model inputs, and the model outputs the probability of progression and severity scores for each complication at different future time periods. For example, it predicts the probability of diabetic nephropathy progressing from its current stage to the next stage within the next three months, six months, and one year, and predicts the trend of retinopathy severity scores. Simultaneously, an uncertainty quantification method is used to assess the reliability of the predictions. Predictive uncertainty in the medical field stems from data noise, model limitations, and the inherent randomness of diseases. Accurately quantifying uncertainty is crucial for clinical decision-making. A Bayesian deep learning method is employed. By applying a probability distribution to the model parameters and performing multiple forward propagation samplings, the distribution of the predicted results is obtained. The predicted mean and predicted variance are calculated. The predicted variance reflects the uncertainty of the model regarding the prediction; a large variance indicates high uncertainty, and a small variance indicates low uncertainty. When the predicted uncertainty of a complication is high, the system marks it as requiring the doctor's close attention and suggests more detailed examination and evaluation. When a complication is predicted to have a high probability of progression and a high confidence level, the system generates an acute exacerbation risk warning to remind the doctor to intervene in a timely manner, obtaining a complication progression prediction result with confidence. This result includes the progression probability, severity score, prediction time range, uncertainty assessment, and risk level classification of each complication, providing a quantitative reference for the doctor's clinical decision-making.

[0107] In some embodiments, since the progression of complications is significantly influenced by individual patient differences and external factors, a personalized dynamic prediction model can be employed. The aim is to dynamically adjust the prediction results based on the patient's real-time feedback. Specifically, an online learning mechanism is constructed. As new monitoring and follow-up data are subsequently generated by the patient, the model continuously updates its predictions for that patient. An incremental learning algorithm is used, eliminating the need to retrain the model from scratch when new data arrives. Instead, parameters are updated based on the existing model, rapidly absorbing new information. By tracking the deviation between the patient's actual disease progression and the predicted results, the prediction model is dynamically corrected. When a significant prediction deviation is detected, the cause of the deviation is analyzed. This could be due to rare pathological changes in the patient or external intervention factors such as a change in treatment plan. The model is adjusted based on the deviation feedback. This personalized dynamic prediction mechanism makes the predictions closer to the patient's actual disease progression trajectory, improving the long-term accuracy of the predictions.

[0108] The intelligent medication decision-making module generates candidate solutions and selects the Pareto optimal solution based on the complication progression prediction results, thus obtaining a multi-level personalized medication recommendation plan;

[0109] Specifically, the following steps are included:

[0110] Step 5.1: Construct a multi-objective optimization function;

[0111] Based on a hierarchical fusion of patient health status representation vectors, deep complication association inference results, and confidence-based complication progression prediction results, a multi-objective optimization function is constructed using the analytic hierarchy process (AHP). Treatment goals for diabetic patients with complications are multi-dimensional, including glycemic control (stabilizing blood glucose levels within a target range), complication progression slowing (decelerating the rate of deterioration of complications such as nephropathy and retinopathy), minimizing drug side effects (reducing the risk of adverse drug reactions), and improving patient quality of life (reducing medication complexity and impact on daily life). These goals may conflict; for example, intensive glucose control may increase the risk of hypoglycemia and affect quality of life. The AHP is used to construct a hierarchical structure of goals, with the top level representing the overall treatment effect and the lower levels decomposing into specific goals. The relative importance weights of each goal are determined through expert evaluation or clinical data statistics. The process of constructing a multi-objective optimization function is as follows: Define four sub-objective functions: a glycemic control effect scoring function, a complication delay effect scoring function, a side effect risk scoring function, and a quality of life scoring function; assign corresponding weight coefficients to each sub-objective function, and the weights can be adjusted according to the patient's individual preferences; multiply each sub-objective function with its corresponding weight coefficient and then perform a weighted sum to obtain a comprehensive multi-objective optimization function, where the side effect risk score is represented by a negative number to achieve the minimization objective; obtain a quantified multi-objective function, which transforms the complex treatment objectives into an optimizable mathematical form.

[0112] Step 5.2: Train the intelligent medication decision-making agent;

[0113] Based on a drug knowledge base and pharmacogenomics database, a deep reinforcement learning method is used to train an intelligent medication decision-making agent. Medication decision-making is a sequential decision problem that requires optimization in multiple decision-making steps, such as drug selection, dosage determination, and timing of administration. Reinforcement learning learns the optimal decision strategy through agent-environment interaction, constructing a Markov decision process for medication decision-making. The state space includes the patient's health status, current medication status, and treatment history, while the action space includes available drug types, dosage levels, and administration regimens. The reward function is defined according to a multi-objective function, with high rewards for treatments with good therapeutic effects and few side effects, and low rewards or penalties for treatments with poor side effects. A deep Q-network (Deep Q-network) is employed. The decision-making agent is trained using Q-Network (DQN) or policy gradient methods. The agent undergoes extensive trials in a clinical simulation environment modeled based on real clinical data and drug action mechanisms, simulating patient responses to different medication regimens. The agent tries different medication strategies in the simulation, observes treatment effects and side effects, and updates its decision-making strategy based on reward signals, gradually learning the optimal medication decisions for different patient states. Drug interaction constraints are incorporated into the training process; large negative rewards are given when the selected drug combinations have serious interactions, forcing the agent to learn to avoid irrational drug use. Rules from clinical practice guidelines are also incorporated, using the recommended first-line and second-line drug sequences as prior knowledge to guide the agent's learning. After extensive training and iterations, an intelligent medication decision-making model is obtained, which can intelligently recommend the optimal drug combinations and medication regimens based on the patient's state.

[0114] Step 5.3: Generate personalized multi-drug combination regimens;

[0115] Based on patients' personalized characteristics and an intelligent medication decision-making model, personalized multi-drug combination regimens are generated. Patient-specific characteristics include metabolic enzyme genotypes (e.g., CYP2C9 genotype affects the metabolic rate of certain hypoglycemic drugs), liver function indicators (e.g., transaminase levels affect the liver's ability to metabolize drugs), kidney function indicators (e.g., glomerular filtration rate affects the kidney's ability to excrete drugs), drug sensitivity (i.e., the patient's history of drug reactions, allergy history, and contraindications), etc. These patient characteristics are input into the intelligent medication decision-making model. The model selects appropriate drug types based on the patient's metabolic characteristics and complication types. For patients with renal insufficiency, nephrotoxic drugs are avoided; for patients with abnormal liver function, the dosage of drugs metabolized by the liver is adjusted. The model predicts drug efficacy and side effect risks based on the patient's drug genotype. For patients with certain genotypes, a drug with better efficacy or fewer side effects is preferentially recommended. Considering the synergistic effects and interactions of multiple drugs, a drug synergy assessment model is used to calculate the synergistic therapeutic effect of different drug combinations. The calculation process for the synergistic effect is as follows: first, the efficacy scores of each drug used alone are evaluated separately; then, the efficacy scores of the two drugs are evaluated separately. The combined efficacy score for combined use is calculated, followed by the difference between the combined efficacy score and the sum of the two monotherapy efficacy scores to measure synergistic gain. This difference is then normalized by dividing by the square root of the product of the two monotherapy efficacy scores to obtain the synergistic effect index. A synergistic effect index greater than zero indicates positive synergy, meaning the combined drug effect is better than the sum of the monotherapy scores. A synergistic effect index less than zero indicates an antagonistic effect. Drug combinations with positive synergistic effects are selected, and a dosage optimization algorithm is used to determine personalized dosages for each drug. Dosage optimization comprehensively considers factors such as patient weight, age, liver and kidney function, and the severity of complications. Pharmacokinetic models are used to predict blood drug concentrations at different doses, selecting doses that achieve the therapeutic concentration window and have controllable side effect risks. Monte Carlo simulations are used to evaluate the efficacy and risk distribution of the regimen, considering individual differences and randomness in patient responses. A large number of simulation trials are used to generate probability distributions of the regimen's effects, resulting in multiple candidate drug regimens and their risk-benefit assessments. Each candidate regimen includes detailed information such as specific drug combinations, drug dosages, administration times, expected efficacy distribution, and side effect risk distribution.

[0116] Step 5.4: Select the Pareto optimal medication regimen;

[0117] Based on multiple candidate medication regimens and their risk-benefit assessments, a Pareto optimal selection algorithm is used to find the optimal set of regimens in a multi-objective space. Different medication regimens may have varying performance on different treatment objectives; one regimen may have good glycemic control but slightly more side effects, while another may have fewer side effects but slightly weaker glycemic control. Pareto optimality is defined as the absence of any other regimen that is no worse than the current regimen in all objectives and is better than it in at least one objective. The Pareto optimal set of regimens represents the optimal balance in a multi-objective trade-off. The score of each candidate regimen on each objective dimension is calculated and represented as a point in the multi-objective space. By comparing all candidate regimens, the Pareto optimal regimen is selected. The Pareto optimal regimens are then categorized according to different objective preferences: efficacy-priority regimens are recommended for patients who are more concerned about glycemic control; safety and comfort regimens are recommended for patients who are more concerned about side effects and quality of life; and comprehensive optimal regimens are recommended for patients who need to balance all aspects. This presents doctors with multiple Pareto optimal regimens and their characteristics, allowing doctors to choose the most suitable regimen based on the patient's specific situation and preferences, resulting in a multi-level personalized medication recommendation plan. This set of regimens provides doctors with flexible decision-making options and supports personalized clinical decision-making.

[0118] In some embodiments, since the patient's condition is dynamic and the medication regimen needs to be adjusted promptly according to changes in the condition, a dynamic regimen monitoring and adaptive adjustment mechanism can be adopted to achieve continuous optimization of the medication regimen. Specifically, during the patient's medication regimen, the system continuously monitors the patient's blood glucose data, renal function indicators, symptom changes, etc. When a high-risk warning is triggered in the prediction results of complication progression with confidence, or when the actual changes in the patient's condition exceed the expected range, the system automatically triggers the regimen adjustment process, re-enters the patient's latest health status data, and calls the intelligent medication decision model to generate adjusted medication recommendations. The adjustment recommendations include specific operations such as adding a certain drug, reducing the dosage of a certain drug, or changing the type of drug. The system calculates the difference in efficacy and risk before and after the adjustment, pushes the adjustment recommendations and their basis to the doctor, and implements the adjusted regimen after the doctor reviews and confirms it. This dynamic adjustment mechanism enables the treatment regimen to respond promptly to changes in the patient's condition and maintain optimal treatment results.

[0119] The adherence management module, based on a multi-level personalized medication recommendation scheme, performs causal inference and closed-loop optimization of adherence to obtain an adaptive adherence improvement scheme.

[0120] Specifically, the following steps are included:

[0121] Step 6.1: Construct a compliance causal network;

[0122] Based on patient history medication records, lifestyle data, and environmental data, a causal inference method was used to construct a causal network for adherence behavior. Patient adherence is influenced by multiple factors, including personal factors such as health awareness, self-management ability, and psychological state; medication factors such as medication complexity and side effect experience; social environmental factors such as family support and work schedule conflicts; and healthcare service factors such as the quality of doctor-patient communication and follow-up frequency. These factors have complex causal relationships and interactions with adherence. Traditional association analysis can only identify related factors but cannot distinguish causal relationships, while causal inference, through the construction of causal graphs and counterfactual reasoning, can identify the true causal factors. A structural causal model (SCM) was used, with adherence behavior as the outcome variable and each influencing factor as the cause variable, constructing a directed acyclic graph (DAG) to represent the causal relationships between the variables. Using causal inference techniques such as do-calculus and backdoor adjustment, causal effects are estimated from observational data. For each potential factor, its mean causal effect (ATE) on adherence is calculated; that is, the magnitude of the causal effect of a change in this factor on adherence after controlling for other confounding factors. Key factors with positive or negative causal effects on adherence are identified. For example, causal analysis reveals that work schedule conflicts are the main causal factor leading to missed medication doses, and side effect experiences are the main causal factor leading to patients discontinuing medication on their own. An adherence causal network is obtained, which clearly shows the causal pathways and key barriers affecting patient adherence, providing a scientific basis for precise intervention.

[0123] Step 6.2: Assess the causal effects of the intervention;

[0124] Based on adherence causal networks, counterfactual reasoning techniques are employed to assess the causal effects of different interventions. Counterfactual reasoning answers the question of how patient adherence will change if a certain intervention is implemented. For each identified key barrier factor, corresponding interventions are designed. For example, flexible medication reminders are designed to address work-related time conflicts; personalized health education interventions are designed to address insufficient health awareness; and medication regimen optimization interventions are designed to address side effect experiences. Using counterfactual reasoning, based on patients' historical data and causal models, the expected changes in patient adherence under counterfactual scenarios of implementing a particular intervention are simulated, and the causal effect of the intervention is calculated—that is, the improvement in adherence after intervention compared to no intervention. By comparing the magnitude of the causal effects of different interventions, the most effective intervention for the patient is identified. Because different patients have different barrier factors, the effects of the same intervention are heterogeneous. Counterfactual reasoning can identify personalized optimal intervention combinations for each patient, yielding intervention effect evaluation results. These results quantify the causal improvement effect of each intervention on patient adherence, providing evidence to support intervention strategy selection.

[0125] Step 6.3: Generate personalized intervention strategies;

[0126] Based on the evaluation results of intervention effectiveness and the individualized characteristics of patients, a personalized intervention strategy generation algorithm is used to match the most effective intervention combination for patients with different adherence types. Patients are classified according to their main adherence impairment types. Cognitive impairment patients lack sufficient understanding of the severity of their illness and the importance of treatment. The intervention focus for these patients is health education, providing disease information, complication risk warnings, and successful treatment cases to improve their health awareness and treatment motivation. Time management difficulties patients forget to take their medication due to busy work and life schedules. The intervention focus for these patients is intelligent reminder optimization, intelligently selecting the most suitable medication reminder time based on their daily routines and schedules, and using multiple reminder methods such as mobile push notifications, SMS, and voice reminders to improve the reach and effectiveness of reminders. Side effect sensitivity patients resist medication due to the impact of drug side effects on their lives. The intervention focus for these patients is adjusting their medication regimen and managing side effects, communicating with doctors to adjust the type or dosage of medication to reduce side effects, and providing guidance on side effect coping strategies. Lack of social support patients lack family and social support, leading to difficulties in self-management. The intervention focus for these patients is building a social support network, involving family members in patient management, providing community support, and establishing a regular doctor-patient communication mechanism. A comprehensive intervention strategy package containing multiple intervention measures is generated for each patient. The combination and intensity of each intervention measure in the strategy package are personalized according to the patient's characteristics, resulting in a causal-driven personalized intervention strategy. This strategy provides precise intervention for the patient's actual obstacle factors, resulting in more significant intervention effects.

[0127] Step 6.4: Achieve closed-loop optimization management of compliance;

[0128] Based on the implementation of personalized intervention strategies and feedback data on patient adherence behavior, an online reinforcement learning algorithm is used to continuously optimize the adherence prediction model and intervention strategies, forming a closed-loop adherence management system. During the patient's execution of the treatment plan and acceptance of intervention, the system continuously collects actual medication adherence data, such as on-time medication rate and missed dose frequency; lifestyle adherence data, such as dietary control and exercise adherence; and adherence change trends. The actual adherence data is compared with the predicted adherence to evaluate the accuracy of the model's predictions. When prediction deviations are found, the adherence prediction model is updated, and new data is incorporated into the model training to improve its predictive ability. The actual effects of the implemented intervention measures are evaluated, and changes in patient adherence before and after the intervention are compared to verify whether the counterfactual reasoning prediction of the intervention effect is accurate. When certain interventions fail to achieve the expected results, the system analyzes the reasons and adjusts the intervention strategy. It employs a Multi-ArmedBandit algorithm to dynamically adjust the selection of interventions, treating different interventions as separate arms. Based on the historical benefits (intervention effects) and exploration needs of each arm, the system balances utilizing known effective interventions with exploring new ones. Through continuous feedback learning and strategy optimization, the system gradually learns the optimal intervention strategies for different patients and situations, resulting in an adaptive adherence improvement plan. This plan can dynamically adjust the content and intensity of interventions based on real-time patient feedback, achieving continuous improvement and personalized optimization of adherence management. This ensures patients maintain good treatment adherence in the long term, guaranteeing effective implementation of treatment plans and long-term success in disease management.

[0129] In some embodiments, since adherence behavior is significantly influenced by the patient's psychological state, a combined mechanism of psychological state assessment and psychological intervention can be employed to improve patient adherence from a psychological perspective. Specifically, the patient's psychological state, including depression, anxiety, and confidence in treatment, is assessed through self-report questionnaires, behavioral data analysis, and voice emotion analysis. When negative psychological states are identified, psychological support interventions are introduced, including online psychological counseling services, the dissemination of positive psychological content, and the establishment of patient support communities. By improving the patient's psychological state, adherence is indirectly improved. The combination of psychological and behavioral interventions forms a comprehensive adherence improvement system. This method can address adherence issues at a deeper level and improve long-term adherence maintenance rates.

[0130] The efficacy evaluation module, based on the adaptive compliance improvement plan, evaluates and dynamically optimizes the efficacy, and generates a treatment efficacy evaluation report.

[0131] Specifically, the following steps are included:

[0132] Step 7.1: Generate an interpretable intelligent recommendation report;

[0133] Based on multi-level personalized medication recommendation schemes, adaptive adherence improvement schemes, and confidence-based complication progression prediction results, Explainable Artificial Intelligence (XAI) technology is used to generate visual explanation reports of the decision-making basis. When referring to AI system recommendations, clinicians not only need to know the content of the recommendations, but also need to understand the basis and reasoning process behind them. This is crucial for building doctors' trust in AI and ensuring medical safety. Attention weight visualization is employed to demonstrate which patient characteristics and data the model prioritizes when making predictions and recommendations. For example, in predicting kidney disease progression, the model primarily relies on key characteristics such as rising serum creatinine, persistent positive urine protein, and poor blood pressure control. These key characteristics and their importance weights are displayed through heatmaps or highlighting. Counterfactual interpretation is used to answer how the prediction result would change if a certain characteristic value of the patient were different. For example, if the patient's glycated hemoglobin could be reduced to the target range, the risk of complication progression would decrease. This counterfactual analysis helps doctors understand the impact of various factors on the disease. Rule extraction technology is used to extract interpretable decision rules from the complex deep learning model. For example, when a patient's serum creatinine is greater than a certain threshold and urine protein quantification exceeds a certain threshold, the probability of kidney disease progression to a certain stage is greater than a certain threshold. The extracted rules are expressed in natural language, conforming to doctors' clinical thinking habits. The rationale for medication recommendations is clearly presented, explaining why the drug combination is recommended, which treatment goal each drug targets, how the drugs synergize, and which personalized characteristics of the patient led to dosage adjustments. All analysis results, recommendations, and their explanations are integrated into a structured intelligent recommendation report. The report includes modules such as a summary of the patient's health status, complication risk assessment, medication recommendations and their basis, adherence analysis, and intervention suggestions. It uses a combination of charts, text, and data to produce an interpretable intelligent recommendation report. This report improves doctors' understanding and trust in AI recommendations, supporting them in making more informed clinical decisions.

[0134] Step 7.2: Achieve collaborative decision-making between doctors, patients, and machines;

[0135] Based on interpretable intelligent recommendation reports and remote audio-visual communication technology, human-machine collaborative decision-making is achieved during remote doctor-patient consultations. According to the patient's appointment request and early warning information on the risk of acute exacerbation of complications, an intelligent scheduling algorithm matches endocrinologists with matching specialties, moderate current workload, and good historical evaluations. A real-time video communication connection is established between the doctor and patient, using WebRTC technology to achieve high-definition audio-visual transmission and screen sharing. Doctors can view the patient's electronic medical records, historical examination reports, recent monitoring data curves, and the intelligent recommendation report generated by the system in the consultation interface. Through video communication, doctors can understand the patient's subjective symptoms, living conditions, medication experiences, and other information that cannot be directly obtained from data. Combining the patient's complaints and the system's analysis results, doctors use their professional knowledge and clinical experience to adapt the system's recommended treatment plan. The system allows doctors to evaluate and directly adopt the recommended treatment plan, or adjust it based on the patient's specific circumstances. It offers an interactive adjustment function, allowing doctors to modify medication types, adjust dosages, and add or remove interventions. The system calculates the expected effects and risk changes of the adjusted plan in real time, displaying the impact of the adjustment on each treatment goal and assisting doctors in assessing the rationality of the adjustment. Through human-machine collaboration, it fully leverages the advantages of AI in data analysis and pattern recognition with the advantages of doctors in experience-based judgment and case handling, achieving a synergistic effect greater than the sum of its parts. The doctor ultimately confirms the treatment plan, and the system records the doctor's decisions and adjustment operations. This data will be used for subsequent model optimization and knowledge accumulation, resulting in a revised treatment plan reviewed by the doctor. This plan combines AI's intelligent analysis with the doctor's professional judgment, possessing both scientific validity and clinical applicability.

[0136] Step 7.3: Generate a patient-friendly treatment plan description;

[0137] Based on the doctor's reviewed and adjusted treatment plan, the system uses multimedia health education content generation technology and a patient-friendly presentation method to convey the treatment plan to the patient. Patient understanding and acceptance of the treatment plan are crucial for ensuring adherence. Professional medical terminology and complex treatment content may be difficult for patients to understand. Therefore, the system automatically generates personalized health education content based on the patient's education level and health literacy level. It explains the principles of the treatment plan in easy-to-understand language, such as using vivid analogies to illustrate how medications act on the body and help control the condition. Visual charts and graphs display the key points of the treatment plan, including the dosing schedule for each medication, dosage instructions, and specific dietary and exercise recommendations. Short videos or animations are created to illustrate correct medication usage, precautions, possible side effects and coping strategies, and visualizations of expected treatment effects are presented. For example, if the treatment plan is strictly followed, the expected improvement curve of the patient's blood glucose control and the expected reduction in the risk of complications can be visually demonstrated to enhance the patient's confidence and motivation in treatment. Doctors explain the treatment plan to patients during video consultations, answer their questions, and emphasize the importance of treatment and the necessity of adherence. The system simultaneously pushes treatment plan documents, health education materials, and medication reminder settings to the patient's mobile application, making it convenient for the patient to view and follow them at any time. Electronic prescriptions are generated and pushed to partner pharmacies, and drug delivery or medication pickup instructions are arranged. Personalized health management plans are generated, including comprehensive guidance such as medication plans, dietary recommendations, exercise recommendations, monitoring plans, and follow-up arrangements. Patients receive a friendly explanation of the treatment plan, which ensures that they fully understand the treatment plan and are willing to actively implement it, thereby improving the actual effectiveness of the treatment plan.

[0138] Step 7.4: Perform efficacy evaluation and system optimization;

[0139] Based on data from the treatment regimen's execution process and subsequent patient follow-up, an efficacy evaluation model and deviation analysis method are used to assess the difference between the actual and expected effects of the treatment regimen, and to generate system optimization suggestions. After the patient begins the treatment regimen, the system continuously tracks the patient's monitoring data, medication records, adherence behavior, symptom changes, etc. At set follow-up time points, such as one month and three months later, the system automatically analyzes the patient's treatment effect, assessing whether blood glucose control has been achieved, whether the progression of complications has been delayed, and whether quality of life has improved. The actual treatment effect is compared with the expected effect when the treatment regimen was formulated, and the deviation is calculated. When the actual effect is better than expected, success factors are analyzed, which may be due to particularly good patient adherence or a particularly good response to the treatment regimen. The characteristics of successful cases are recorded for use in optimizing the model and guiding the treatment of similar patients. When the actual effect is worse than expected, the reasons are analyzed, which may be due to poor patient adherence. In cases where adverse events occur, complications progress beyond prediction, medication regimens are unsuitable for the patient, or unexpected adverse reactions occur, the system generates treatment adjustment suggestions. These suggestions include strengthening adherence interventions, adjusting medication regimens, increasing monitoring frequency, and recommending the patient undergo detailed examinations at the hospital. These suggestions are then sent to the doctor for evaluation, who decides whether to implement the adjustments, thus forming a dynamic optimization loop for the treatment plan. Simultaneously, treatment effect data, deviation analysis results, and treatment plan adjustment records are fed back to the system's machine learning module for continuous model training and optimization. Through feedback learning from numerous real clinical cases, the system's predictive models, medication decision-making models, and adherence prediction models are continuously improved, resulting in continuously enhanced predictive accuracy and recommendation rationality. This leads to treatment effect evaluation reports and system optimization suggestions. This closed-loop mechanism ensures the continuous evolution of the system and the continuous optimization of treatment plans, providing patients with increasingly precise and effective remote medical management services.

[0140] In some embodiments, due to the unique and complex nature of the medical field, a continuous learning mechanism based on a medical expert knowledge base and a collective intelligence fusion mechanism can be adopted to continuously enrich the system's medical knowledge and improve its decision-making capabilities. Specifically, a knowledge review and feedback mechanism involving medical experts is established. Endocrinologists are regularly invited to review the system's recommended cases, assessing their rationality and safety. Expert suggestions and supplementary knowledge are integrated into the system's knowledge base and rule base. Federated learning technology is used to collaboratively train models across multiple medical institutions. While protecting patient privacy in each institution, model parameter updates are shared, and data characteristics and clinical experience from different institutions are integrated to improve the model's generalization ability and robustness. Through collective intelligence fusion, the system can aggregate multi-party expert knowledge and multi-center clinical data to build a more comprehensive and reliable intelligent medical decision-making system.

[0141] A computer-readable storage medium for storing computer-readable instructions that, when read by a computer, enable the execution of the aforementioned Internet-based remote medical consultation system.

[0142] In one embodiment of the present invention, a specific example is provided:

[0143] Patient Zhang, 60 years old, has been diagnosed with type 2 diabetes for 8 years. He currently suffers from diabetic nephropathy, diabetic retinopathy and hypertension. Because he lives in the suburbs, it takes him most of the day to go to the hospital for a follow-up visit. In addition, he has difficulty walking. He hopes to manage his condition in the long term through a telemedicine system.

[0144] In the data acquisition and quality assurance module, the system underwent a 14-day field test in the endocrinology department of a tertiary hospital. During the test, IoT devices such as continuous glucose monitors and smart blood pressure monitors were deployed, and data interfaces were established with the hospital information system and the laboratory information system (LIS). The test covered patient Zhang's daily life and medical examination scenarios. The system collected 2016 data points of blood glucose data from the past two weeks through the continuous glucose monitor worn by the patient, and collected blood pressure data three times a day through the smart blood pressure monitor. At the same time, it obtained the patient's most recent kidney function test report and fundus examination report from the hospital information system.

[0145] Table 1 shows examples of data collected by the physiological signal monitoring device:

[0146] Table 1: Examples of data collected by physiological signal monitoring equipment;

[0147]

[0148] Table 2 shows examples of data obtained from the electronic medical record system of medical institutions:

[0149] Table 2: Examples of data obtained from electronic medical record systems in medical institutions;

[0150]

[0151] The system uses an LSTM model to intelligently imput the few missing data points, resulting in a more complete medical dataset.

[0152] In the feature extraction and fusion module, the system uses a Transformer encoder to extract time-series features from blood glucose time-series data, identifying characteristic patterns such as large fluctuations in blood glucose and high postprandial blood glucose peaks. It uses VisionTransformer to extract visual features from fundus images, identifying pathological features such as retinal microaneurysms and a small number of hard exudates. It uses TabNet to extract features from structured data such as renal function, comprehensively reflecting the patient's metabolic disorders and declining renal function. Through cross-modal contrastive learning and hierarchical attention fusion, the system obtains a comprehensive health status representation vector for the patient.

[0153] In the complication association reasoning module, the system uses a disease knowledge graph to infer that there is a mutually aggravating relationship between diabetic nephropathy and hypertension in patients. Hyperglycemia and hypertension work together to accelerate glomerular sclerosis, and the progression of kidney disease leads to water and sodium retention, which further increases blood pressure, forming a vicious cycle. At the same time, the system identifies the potential risk of cardiovascular complications in patients and obtains the complication association analysis results.

[0154] In the complication progression prediction module, the system uses a meta-learning-enhanced prediction model to predict that the probability of a patient's kidney disease progressing from the current stage to the next stage within the next six months is moderate, and the probability of retinopathy progression is low, but continuous monitoring is required. At the same time, it provides a prediction confidence assessment to obtain a complication progression prediction result with confidence.

[0155] In the intelligent medication decision-making module, the system generates personalized medication plans for patients based on multi-objective optimization. It recommends a combination of hypoglycemic drugs including metformin and DPP-4 inhibitors, and ACEI drugs for kidney protection and blood pressure reduction. The dosage is adjusted based on the patient's age and renal function. Through Pareto optimal screening, three candidate options are provided to the doctor for selection, resulting in a multi-level personalized medication recommendation plan.

[0156] In the adherence management module, the system analyzed the patient's historical data and found that the patient had adherence disorder due to time management difficulties. The patient often forgot to take medication because he was rushing in the morning. The system generated a personalized intervention strategy for the patient, adjusted the morning medication reminder time to the time when the patient washes up after getting up, increased the family participation reminder mechanism, and improved the patient's adherence through multiple methods, resulting in an adaptive adherence improvement plan.

[0157] In the efficacy evaluation module, doctors conduct remote video consultations with patients, review the interpretable intelligent recommendation reports generated by the system, fine-tune the medication plan after understanding the patient's condition, and finally confirm the treatment plan and explain it to the patient in detail. The system generates patient-friendly treatment instructions and health management plans and pushes them to the patient. After the patient starts implementing the plan, the system continuously tracks the effect and evaluates it regularly, forming a closed-loop management system for continuous optimization.

[0158] Through the application of this system, after three months of remote management, patient Zhang's blood sugar control rate improved, his renal function indicators tended to stabilize, and his quality of life improved. He also avoided the inconvenience of frequent trips to the hospital, demonstrating the practical application value and clinical benefits of this invention.

[0159] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An internet-based remote medical consultation system, characterized in that, include: The data acquisition and quality assurance module collects multi-source heterogeneous medical data, performs time-series alignment, quality monitoring, and intelligent repair and interpolation, and obtains a medical dataset with enhanced integrity. The feature extraction and fusion module performs deep feature extraction, cross-modal semantic alignment, and hierarchical attention fusion on the integrity-enhanced medical dataset to obtain a hierarchically fused patient health status representation vector. Specifically, this includes: using a Transformer encoder architecture to extract deep features from time-series physiological signal data, obtaining time-series signal feature vectors through positional encoding and self-attention mechanisms; using a Vision Transformer model to extract features from fundus image data, obtaining image feature vectors; using a TabNet deep learning network specifically for tabular data to extract features from structured test report data, obtaining structured data feature vectors; using a contrastive learning-based cross-modal semantic alignment method to map features from each modality to a common semantic space; and using a hierarchical cross-modal attention fusion mechanism to perform deep feature fusion, obtaining a hierarchically fused patient health status representation vector. The complication association reasoning module, based on a hierarchically fused patient health status representation vector, executes graph attention networks and multi-hop reasoning to obtain deep complication association reasoning results. Specifically, it includes: using a pre-trained biomedical domain language model for knowledge extraction, identifying disease entities, symptom entities, drug entities, and examination indicator entities; extracting causal relationships, co-occurrence relationships, and treatment relationships between entities to construct a diabetes complication knowledge graph; dynamically updating the knowledge graph using incremental learning and knowledge fusion methods to obtain a dynamically evolving disease knowledge graph; using an entity linking algorithm to map patient symptoms and examination indicators to entity nodes in the knowledge graph to obtain the patient's disease status representation in the knowledge graph; and using a graph attention network for message passing and feature aggregation on the knowledge graph, identifying association paths between complications and potential complication risks through multi-hop information propagation. The complication progression prediction module, based on deep complication association inference results, constructs a complication progression prediction model, performs meta-learning enhancement and uncertainty quantification, and obtains complication progression prediction results. Specifically, it includes: pre-training a multi-task complication prediction base model on a diabetic patient electronic medical record database using a self-supervised learning method; learning disease evolution patterns through masked prediction tasks and contrastive prediction tasks; constructing a multi-task learning framework to simultaneously predict the occurrence and progression of diabetic nephropathy, retinopathy, peripheral neuropathy, and cardiovascular complications; adapting the pre-trained model to a specific patient population in a target medical institution using transfer learning techniques to obtain a transfer-optimized prediction model; learning model initialization parameters on a complication combination prediction task; dividing data into support and query sets for each task; performing gradient updates on the support set to obtain task-specific parameters; evaluating performance and calculating meta-loss on the query set; updating model initialization parameters based on meta-loss to obtain a meta-learning enhanced prediction model; and predicting complication progression trends based on patient health status representation vectors, deep complication association inference results, and patient historical time-series data, using Bayesian deep learning methods for uncertainty quantification, calculating the prediction mean and prediction variance, and obtaining complication progression prediction results with confidence. The intelligent medication decision-making module generates candidate treatments and selects the Pareto optimal treatment based on complication progression prediction results, resulting in a multi-level personalized medication recommendation scheme. Specifically, this includes: constructing a multi-objective optimization function using the analytic hierarchy process (AHP), defining scoring functions for glycemic control effectiveness, complication delay effectiveness, side effect risk, and quality of life; training an intelligent medication decision-making agent using deep reinforcement learning to construct a Markov decision process for medication decisions, incorporating drug interaction constraints and rules from clinical guidelines; generating drug combination schemes based on the patient's metabolic enzyme genotype, liver and kidney function indicators, drug sensitivity, and allergy history; calculating the synergistic therapeutic effect using a drug synergy evaluation model; determining personalized dosage through a pharmacokinetic model; and using a Pareto optimal selection algorithm to find the optimal set of schemes, resulting in a multi-level personalized medication recommendation scheme. The adherence management module, based on a multi-level personalized medication recommendation scheme, performs causal inference and closed-loop optimization of adherence to obtain an adaptive adherence improvement scheme. Specifically, it includes: constructing a causal relationship network of adherence behavior using causal inference methods; constructing a directed acyclic graph to represent causal relationships using a structural causal model to identify factors with causal effects on adherence; evaluating the causal effects of different interventions using counterfactual reasoning techniques to obtain intervention effect evaluation results; classifying patients into cognitive impairment type, time management difficulty type, side effect sensitivity type, and social support lack type, and matching corresponding health education, intelligent reminder optimization, medication scheme adjustment, and social support network construction intervention combinations for different types of patients; and continuously optimizing the adherence prediction model and intervention strategies using online reinforcement learning algorithms to obtain an adaptive adherence improvement scheme. The efficacy evaluation module, based on the adaptive compliance improvement plan, performs efficacy evaluation and dynamic optimization, and generates a treatment efficacy evaluation report.

2. The Internet-based remote medical consultation system according to claim 1, characterized in that, The data acquisition and quality assurance module includes: A data transmission channel is established with the patient's physiological signal monitoring equipment through the Internet of Things (IoT) device access protocol, and a local data caching mechanism is set up to achieve network fault tolerance. A unified time reference axis is established using a timestamp alignment algorithm; A real-time data quality monitoring algorithm is used to detect missing data, outliers, and noise. Missing data were processed using a temporal imputation method based on long short-term memory networks and a collaborative imputation method based on patient characteristics.

3. The Internet-based remote medical consultation system according to claim 1, characterized in that, The performance evaluation module includes: The system uses explainable artificial intelligence technology to generate visual explanation reports of the decision-making basis, and employs attention weight visualization method, counterfactual explanation method and rule extraction technology to display the decision-making basis; Enables collaborative decision-making between doctors and patients using remote audio and video communication technology, and provides interactive solution adjustment functions; Multimedia health education content generation technology is used to convey treatment plans to patients and generate personalized health management plans. The efficacy evaluation model and deviation analysis method are used to assess the difference between the actual effect and the expected effect of the treatment plan, and the treatment effect data is fed back to the machine learning module for continuous model training and optimization.

4. The Internet-based remote medical consultation system according to claim 1, characterized in that, The data acquisition and quality assurance module also includes: Set up a local data caching mechanism on the device side. When a network interruption is detected, the data will be automatically stored in the device's local cache. Once the network is restored, the cached data will be automatically synchronized and uploaded in timestamp order. A real-time data quality monitoring algorithm is used to detect missing data and calculate the missing rate of each data source, identify outliers in the data that exceed the physiologically possible range, and detect glitches caused by equipment failure. For short-term missing time-series data, an intelligent imputation model based on long short-term memory network is used; for long-term missing data, an estimation imputation method based on similar patient matching is used. For outlier data, determine whether it is a real pathological abnormality or a measurement error based on the degree of deviation. If it is determined to be a measurement error, smooth it or delete it. If it is determined to be a real abnormality, retain it and mark it. Confidence scores are labeled on all imputed and repaired data to obtain a medical dataset with enhanced integrity.

5. A computer-readable storage medium, characterized in that, It is used to store computer-readable instructions, which, when read by a computer, enable the operation of an Internet-based telemedicine consultation system as described in any one of claims 1-4.

Images