Disease prevention system and method based on deep learning and knowledge graph

Abstract

The application relates to the technical field of medical care information science, and discloses a disease prevention system and method based on deep learning and a knowledge graph, which comprises the following steps: converting physiological index data of a user at different time points into a physiological state trajectory vector in a multidimensional space, performing weighted transformation by using a risk weight tensor driven by the knowledge graph, introducing a historical trajectory, performing geometric relationship verification, and finally performing risk classification based on the transformation and verification results; the application reconstructs discrete static physical examination data into a dynamic trajectory vector capable of representing a health state evolution process, thereby changing a risk identification mode, and instead of being limited to an absolute value of a current index, the direction and rate of a change trend are quantified, and dynamic prediction of a chronic disease risk is realized.

Description

Technical Field

[0001] This invention relates to a disease prevention system and method based on deep learning and knowledge graphs, belonging to the field of healthcare informatics technology. Background Technology

[0002] Currently, using an individual's past health check data for early identification of chronic disease risks is a direction of technological development. The technical approach used in this field is to compare the values ​​of key physiological indicators in the user's physical examination report with risk thresholds defined by large-scale population statistics or clinical guidelines. When the indicator exceeds the preset threshold, the system generates a risk alert. This method of static risk assessment based on absolute values ​​has been applied in existing systems because of its clear logic and ease of implementation.

[0003] However, the essential characteristic of chronic diseases lies in their continuous and slow evolutionary process, rather than a series of isolated events. When the aforementioned static assessment methods are placed under the long-term and dynamic reality of full-cycle chronic disease management, an inherent limitation in information processing is exposed. Specifically, when processing users' health data over several years, existing methods focus on the indicator values ​​at each independent point in time, i.e., the user's health status, but discard the dynamic information connecting these status points, such as the rate and direction of indicator changes. This processing method results in a reduction of the dimensionality of time-series data, which means that an individual whose physiological indicator is stable at a critical high value for a long time may be assigned a similar risk level to an individual whose indicator deteriorates steadily from an ideal value to the same critical high value, even though the two individuals have different risk orientations in their physiological evolution process.

[0004] To address this issue, a direct approach is to employ time-series data prediction models to process existing low-frequency health checkup data points and predict future indicator values. However, when applied to annual health checkup data with long intervals and sparse data points, the reliability of these models fails to meet application requirements. More importantly, their design goal remains predicting future points, which does not solve a more fundamental problem: how to extract and quantify the user's overall evolution over a considerable period into a stable risk characteristic for deeper risk pattern identification. Therefore, existing technologies suffer from the following inherent technical problems: 1. When processing low-frequency standardized health data, there is a lack of a data representation method that can effectively preserve and quantify its time-series dynamic characteristics, leading to the loss of trend information; 2. It cannot effectively distinguish individuals with the same current physiological indicator values ​​but different historical evolution paths and rates, limiting the early detection of risks. Therefore, how to effectively extract and quantify a feature that characterizes the dynamics of a user's physiological state evolution from standard low-frequency health data records, thereby avoiding the inherent limitations of existing static assessment methods, becomes the technical problem this invention aims to solve. Summary of the Invention

[0005] This invention provides a disease prevention system and method based on deep learning and knowledge graphs. Its main purpose is to solve the problem of effectively extracting and quantifying a characteristic that can represent the dynamics of the user's physiological state evolution from standard, low-frequency health data records.

[0006] To achieve the above objectives, this invention provides a disease prevention system based on deep learning and knowledge graphs, the system comprising:

[0007] A state space construction module is configured to establish a standardized multidimensional physiological state space based on multiple preset physiological indicators, and to map the user's physiological indicator data at a first time point and a second time point at an interval of 9 to 15 months to the first physiological state point and the second physiological state point in the space, respectively.

[0008] A trajectory vector generation module, connected to the state space construction module, is configured to calculate a physiological state trajectory vector based on the first physiological state point and the second physiological state point.

[0009] A risk weight transformation module, connected to the trajectory vector generation module, is configured to use a risk weight tensor corresponding to each dimension of the multidimensional physiological state space and updated by a knowledge graph to perform a tensor product transformation on the physiological state trajectory vector to generate a risk-weighted trajectory vector.

[0010] A trajectory consistency verification module is configured to calculate the geometric relationship between the historical trajectory vector and the physiological state trajectory vector after obtaining the user's third physiological state point at a third time point earlier than the first time point, so as to determine a trajectory consistency metric.

[0011] A risk assessment module, connected to the risk weight transformation module and the trajectory consistency verification module, is configured to use the risk-weighted trajectory vector as the sole input for risk classification, and the output mode of the risk assessment module is gated by the trajectory consistency metric.

[0012] Preferably, the state space construction module is configured to perform linear mapping processing on the value range of each physiological indicator to serve as a measure of the corresponding dimension of the multidimensional physiological state space; and the direction of the physiological state trajectory vector represents the coordinated change trend of the user's physiological state in multiple indicator dimensions, while its magnitude quantifies the overall rate of change of the user's physiological state between the first time point and the second time point.

[0013] Preferably, the trajectory consistency verification module is configured to calculate historical trajectory vectors. With physiological state trajectory vector The cosine similarity is used to determine the trajectory consistency metric, and its calculation follows certain rules. Among them, the As a measure of trajectory consistency, this Let the vector be the one pointing from the third physiological state point to the first physiological state point. This is a physiological state trajectory vector. The Euclidean norm of a vector is denoted by .

[0014] Preferably, in the risk weight transformation module, the risk weight tensor is a diagonal tensor, and each diagonal element of the diagonal tensor is a risk weight value corresponding to a single dimension of the multidimensional physiological state space; the risk weight transformation module is also configured to obtain update instructions from an external medical knowledge base or clinical guideline knowledge graph to dynamically adjust each of the risk weight values ​​of the diagonal tensor.

[0015] Preferably, the output mode of the risk assessment module is gated by the trajectory consistency metric, specifically configured as follows: when the trajectory consistency metric is not lower than a preset verification threshold, the risk assessment module outputs an exact risk classification result; when the trajectory consistency metric is lower than the verification threshold, the risk assessment module stops outputting the exact risk classification result and generates a specific marker indicating a sudden change in trend.

[0016] Preferably, the system further includes an intelligent annotation module connected to the risk assessment module. The intelligent annotation module is configured to be triggered when the risk assessment module generates the specific marker, to perform a targeted query on the user's structured medical records between the first time point and the second time point to determine whether there is a preset key intervention event code; and based on the query results, to perform clinically-related intelligent annotation on the specific marker.

[0017] Preferably, when the intelligent annotation module is configured to perform clinical intelligent annotation, it follows these rules: if a key intervention event code is found and the physiological state trajectory vector points to the direction of health improvement, then the specific marker is annotated as an interventional positive inflection point; if no key intervention event code is found, then the specific marker is annotated as a pathological mutation or data anomaly of unknown origin.

[0018] Preferably, the preset verification threshold is set within the range of 0.3 to 0.7.

[0019] Preferably, the system further includes a time saliency attribution module, configured to acquire time series of at least one high-frequency behavioral agent data associated with the user between the first time point and the second time point, and based on the time series, determine one or more key time periods with the highest weight value in the attribution analysis for the formation of the physiological state trajectory vector, and provide the determined one or more key time periods together with the output of the risk assessment module. The physiological indicator data processed by the state space construction module in the system comes from an electronic health record database, the knowledge graph applied by the risk weight transformation module is a medical domain knowledge graph, and the structured medical records queried by the intelligent annotation module include medication history records and surgical history records.

[0020] A disease prevention method based on deep learning and knowledge graphs includes the following steps:

[0021] Step a: Establish a standardized multidimensional physiological state space based on multiple preset physiological indicators, and map the physiological indicator data of the user at the first time point and the second time point at an interval of 9 to 15 months to the first physiological state point and the second physiological state point in the space, respectively.

[0022] Step b: Calculate a physiological state trajectory vector based on the first physiological state point and the second physiological state point;

[0023] Step c: Using a risk weight tensor that corresponds to each dimension of the multidimensional physiological state space and is updated by the knowledge graph, perform a tensor product transformation on the physiological state trajectory vector to generate a risk-weighted trajectory vector.

[0024] Step d: After obtaining the user's third physiological state point at a third time point earlier than the first time point, calculate the geometric relationship between the historical trajectory vector and the physiological state trajectory vector to determine a trajectory consistency metric.

[0025] Step e involves using the risk-weighted trajectory vector as the sole input for risk classification, and the output pattern of this risk classification is gated by the trajectory consistency metric.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] 1. A standardized multidimensional physiological state space is defined by pre-defined multiple physiological indicators, and user data at two different time points is converted into a first physiological state point and a second physiological state point within this space. Then, a physiological state trajectory vector is calculated based on these two state points. This series of operations reconstructs the originally discrete and static physical examination data into a single data object that can comprehensively represent the evolution of health status within a specific time window. The directional attribute of this trajectory vector objectively presents the synergistic trend of changes in multiple physiological indicators, while its magnitude attribute directly quantifies the overall rate of this evolution process. This processing method, using conventional low-frequency health records, obtains the dynamic characteristics of individual health status changes without relying on complex prediction models, providing a new information dimension for subsequent risk analysis that existing technologies cannot provide.

[0028] 2. After obtaining the user's physiological state at the third time point, the historical trajectory vector is calculated and compared geometrically with the currently calculated physiological state trajectory vector to obtain a trajectory consistency measure. This process introduces an internal verification step based on the historical continuity of the data itself for the system's risk assessment. When the direction and rate of the trajectory vectors within two consecutive time windows show high continuity, the confidence level of the current risk trend is confirmed. When there is a significant deviation between the two, the system marks it as a trend change. This approach avoids erroneous risk assessments caused by accidental data fluctuations at a single time point, such as detection errors or short-term physical discomfort, and makes the entire analysis process more tolerant of unavoidable data noise in the real world.

[0029] 3. This invention further uses the case where the trajectory consistency measurement is lower than a preset threshold as an internal trigger condition to query whether there are key intervention events in the user's medical records within the corresponding time period. This step transforms the trend change verification from a simple data quality control step into an analysis step with preliminary cause indication. When a significant trajectory deviation corresponds to an effective medical intervention in time, the system can classify this deviation as a positive improvement in health status; conversely, in the absence of a corresponding intervention event record, the deviation is pointed to an unexplained status change that requires more attention. Thus, when the system faces data anomalies, it no longer suspends or prompts for re-examination, but rather performs a preliminary nature distinction.

[0030] 4. This invention also introduces a risk weight tensor corresponding to each dimension of the physiological state space. Before risk analysis, this tensor is used to perform a weighted transformation on the physiological state trajectory vector. This risk weight tensor itself can be updated based on external medical knowledge. This mechanism ensures that all subsequent analyses, including trajectory vector clustering and consistency verification, are performed in a weighted space that reflects the differences in the risk contribution of different physiological indicators. If a vector with a small amplitude in the original space changes mainly in the high-risk weight dimension defined by current medical consensus, its risk level will be increased accordingly. This allows the judgment benchmark of the entire assessment system to keep pace with continuously developing medical knowledge and avoids misjudgments caused by outdated knowledge. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall system architecture and dynamic risk assessment process of the present invention.

[0032] Figure 2 This is a schematic diagram illustrating the dynamic update of risk weights driven by the knowledge graph of this invention.

[0033] Figure 3 This is a closed-loop monitoring and state transition diagram of user health trends in this invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] The disease prevention system based on deep learning and knowledge graph provided by this invention has an overall architecture that mainly includes a state space construction module, a trajectory vector generation module, a risk weight transformation module, a trajectory consistency verification module, a risk assessment module, an intelligent annotation module, and a time saliency attribution module. These modules work together to transform discrete, multi-source electronic health record data into structured information capable of dynamic trend analysis and risk identification.

[0036] In a typical application scenario—long-term monitoring and early warning of chronic disease risks for a specific user group—a common challenge for healthcare information systems is handling annual health check data with long time spans, low sampling frequencies, and heterogeneous indicator dimensions. To address this challenge, the state space construction module in this invention is configured to first establish a standardized multidimensional physiological state space based on multiple preset physiological indicators. This procedure unifies raw physiological data with different physical units and numerical ranges into a dimensionless logical framework capable of geometric analysis. Specifically, the module first selects a set of physiological indicators closely related to a specific chronic disease, such as type 2 diabetes, from a medical knowledge graph or authoritative clinical guidelines as N orthogonal dimensions of the space. Examples include fasting blood glucose, glycated hemoglobin, total cholesterol, low-density lipoprotein cholesterol, body mass index (BMI), systolic blood pressure, and diastolic blood pressure. Subsequently, the state space construction module performs linear mapping on the value range of each dimension to serve as the metric for the corresponding dimension of the multidimensional physiological state space. This processing follows a deterministic normalization procedure. ,in, It is the first Normalized coordinate values ​​of each indicator This is the original measured value of the indicator. It is the ideal healthy value for this indicator as defined by clinical guidelines. This refers to the risk diagnostic threshold for that indicator; for example, for the fasting blood glucose indicator, if its ideal value is... The risk threshold is Then the measurement value of a user at the first time point It will be converted into coordinate values ​​in this dimension. Thus, any user's health check report at a specific point in time, after being processed by the state space construction module, will be mapped to a specific physiological state point (PSP) in this N-dimensional space. This physiological state point, as a structured data object, provides a unified data foundation for subsequent dynamic trend analysis.

[0037] Furthermore, in order to transform the evolution of a user's health status from a series of static snapshots at specific time points into a quantifiable dynamic feature, the trajectory vector generation module is configured to be based on two physiological state points of the user at a first time point and a second time point at an interval of 9 to 15 months. and The core of this step is to calculate a Physiological State Trajectory Vector (PSTV). This involves representing changes in health status as a geometric object with direction and length within the constructed multidimensional physiological state space. The calculation process is a vector subtraction: ,in, For the user within the corresponding time window The change in the normalized value of a physiological indicator; for example, if a user at the first time point The physiological state point is At the second time point The physiological state point is Then the physiological state trajectory vector calculated by the trajectory vector generation module is The direction of this physiological state trajectory vector represents the coordinated change trend of the user's physiological state across all N indicator dimensions, and its magnitude... This quantifies the overall rate of change in the user's physiological state during that time period. By generating this physiological state trajectory vector, the system extracts the two dynamic information of the speed and direction of health evolution from the original time series data and solidifies them into a single data object. It should be noted that since this physiological state trajectory vector only contains the amount of change rather than the absolute value of the individual's physiological indicators, the user's privacy information is desensitized during its generation and subsequent processing.

[0038] However, different physiological indicators do not contribute equally to the final health risk, and this contribution weight changes dynamically with updates in medical knowledge. To address this issue, the risk weight transformation module is configured to use a risk weighting tensor (RWT) corresponding to each dimension of the multidimensional physiological state space and updated by a knowledge graph to perform a tensor product transformation on the physiological state trajectory vector generated in the preceding steps, thereby generating a risk-weighted trajectory vector (RW-PSTV) that better reflects the risk evolution trend. In one implementation, the risk weighting tensor is a... The diagonal tensor, whose diagonal elements That is, the first The risk weights are set to zero for each physiological indicator dimension. This risk weight transformation module is also configured to periodically retrieve update instructions from an external medical knowledge base or clinical guideline knowledge graph to dynamically adjust the risk weight values. For example, when a new large-scale cohort study confirms that, for a specific population, the risk contribution of changes in glycated hemoglobin is 1.5 times that of changes in body mass index, the system can update the corresponding weights. , The transformation process specifically involves converting the physiological state trajectory vector... With this risk weight tensor Perform dot product operation to generate risk-weighted trajectory vector. Continuing from the previous example, if the set weights... Then the original vector Will be transformed into Through this weighted transformation, the system incorporates external, updated medical consensus into the quantitative assessment of individual health evolution trends in a computable manner, enabling subsequent risk assessments to dynamically keep pace with new medical knowledge.

[0039] Considering that single health check data may be affected by short-term, non-persistent factors, to improve the robustness of risk assessment, the trajectory consistency verification module is configured to calculate the geometric relationship between the historical trajectory vector and the current physiological state trajectory vector after obtaining the user's third physiological state point at a third time point earlier than the first time point. This determines a trajectory consistency metric, which is used to judge whether the latest health evolution trend is a continuation of a long-term stable trend or an artifact that may be caused by random fluctuations in the data. Specifically, this module first calculates the historical trajectory vector... Subsequently, by calculating the historical trajectory vector Trajectory vector of current physiological state (i.e., the aforementioned) The cosine similarity of the trajectory is used to determine the consistency measure of the trajectory, and its calculation follows the rules: ,in, As a measure of trajectory consistency, Let be the vector pointing from the third physiological state point to the first physiological state point. This is the trajectory vector of the current physiological state. The Euclidean norm of a vector; The range of values ​​is When the value is close to 1, it indicates that the current trend is highly consistent with the historical trend and has high credibility. When the value deviates significantly from 1, it suggests that the current state point may be an anomaly or that the user's health status has indeed undergone a real and drastic change. This trajectory consistency measure is passed to the subsequent risk assessment stage as an internal verification parameter based on the historical continuity of the data itself.

[0040] Ultimately, the risk assessment module is configured to use the risk-weighted trajectory vector as the sole input for risk classification, and its output mode is gated by the trajectory consistency metric calculated by the aforementioned trajectory consistency verification module. This design aims to couple the quantitative analysis of trends with the reliability verification of trends to output more contextualized risk information. The module's internal operating procedure is set as follows: when the trajectory consistency metric... Not lower than a preset verification threshold, for example (The value of this threshold is set at) to Within this range, its specific calibration is based on receiver operating characteristic (ROC) curve analysis on a reference dataset to find the operating point that balances the sensitivity and specificity of trend mutation detection. The risk assessment module then considers the current trend reliable and outputs a precise risk classification result. For example, it might combine this risk-weighted trajectory vector with a curve obtained from a large-scale population analysis. The risk trajectory pattern library, which is pre-built for cluster analysis, is compared and classified into clusters such as rapid deterioration, hidden one-dimensional deterioration, or health improvement. Conversely, when the trajectory consistency metric is lower than the verification threshold, the risk assessment module stops outputting the exact risk classification result and generates a specific marker indicating a sudden change in trend. In this way, the system avoids generating false risk alerts due to accidental data fluctuations at a single point in time, and improves the stability of the entire analysis process.

[0041] To further interpret the clinical significance of trend changes, the system also includes an intelligent annotation module. This module is configured to be triggered when the risk assessment module generates the aforementioned specific markers. It performs a targeted query on the user's structured medical records (e.g., medication and surgical history) between the first and second time points to determine if a pre-defined key intervention event code exists. Based on the query results, it provides clinically relevant intelligent annotations to the specific markers. The annotation rules are as follows: if a key intervention event code is found, such as the ATC code for the first use of a certain type of highly effective hypoglycemic drug, and the current physiological state trajectory vector points towards health improvement (e.g., negative blood glucose and lipid components), then the specific marker is annotated as an interventional positive inflection point. If no relevant key intervention event code is found, then the specific marker is annotated as a pathological mutation or data anomaly of unknown origin. This intelligent annotation... The system elevates a simple data quality control signal into an insight with preliminary causal indications that can guide subsequent clinical decisions. Furthermore, to explore key contributing periods leading to changes in annual health trajectories, the system includes a temporal saliency attribution module. This module is configured to acquire time series of at least one high-frequency behavioral proxy data associated with the user between a first and a second time point, such as the average daily steps recorded by a smart bracelet. Based on this time series, it identifies one or more key time periods with high weight in the attribution analysis of the formation of the physiological state trajectory vector. This identified key time periods are provided together with the output of the risk assessment module. This module uses low-precision, high-frequency behavioral data to explain the causes of evolutionary trends calculated from high-precision, low-frequency physiological data, thereby providing users and doctors with more refined attribution information that can guide behavioral interventions.

[0042] Example 1: This example demonstrates the operation of the general technical solution described in a specific scenario. In an application for early identification of chronic disease risk based on annual electronic health records for a large population, there are two representative user data processing scenarios. User A, a 45-year-old male, has the following raw measurements of fasting blood glucose levels in his annual health check reports for the past five years: , , , and Under the information processing method based on a single threshold comparison, since the annual measurement values ​​have not exceeded the diagnostic criteria for prediabetes, This user was consistently flagged as low-risk for the past five years; User B, a 48-year-old male with a similar profile, had stable fasting blood glucose levels for the previous four years. Around [value missing], but due to a short-term stress event before the fifth health check, the measurement value accidentally increased to [value missing]. Similarly, the aforementioned processing method also marks it as low risk. This method cannot distinguish between a continuous deterioration trend and an occasional data fluctuation, thus causing the omission of potential risks. When the system of this invention processes user A's data, the state space construction module first converts the fasting blood glucose value for each year into a series of coordinates in the corresponding dimension of the multidimensional physiological state space, according to the normalization procedure explained in the specific implementation. These coordinates are respectively , , , and The trajectory vector generation module then calculates the physiological state trajectory vectors for the four years. Its component in the blood glucose dimension is approximately: These vectors exhibit high stability and continuity in direction and magnitude; correspondingly, when the trajectory consistency verification module calculates the trajectory vectors of user A for the most recent two years... and When considering the geometric relationship between them, since their directions are nearly identical, the calculated trajectory consistency measure is... The value is This value is higher than the system's preset verification threshold. The risk assessment module therefore confirms the current situation. The deteriorating trend is reliable and persistent; meanwhile, the risk weight transformation module, based on update instructions obtained from an external medical knowledge graph, sets higher risk weight values ​​for blood glucose-related dimensions, for example... The generated risk-weighted trajectory vector Compared to the original vector, it has a larger magnitude. Ultimately, the risk assessment module uses this high-confidence, high-weight vector. By comparing with the risk trajectory pattern library, user A is classified into a high-risk trend cluster of continuous one-dimensional deterioration, and corresponding early risk warning information is generated.

[0043] When processing user B's data, the system executed the same state point and trajectory vector generation process. The physiological state trajectory vectors for the first three years had a component close to zero in the blood glucose dimension, while the trajectory vector for the last year... The component in this dimension exhibits a significant jump in value when the trajectory consistency verification module calculates its historical trajectory vector. It is approximately a zero vector, and is related to the current trajectory vector. When considering the geometric relationship between two objects, since their directions are not continuous, the calculated trajectory consistency measure is... The value is This value is lower than the preset verification threshold. This result, acting as a gating signal, causes the risk assessment module to halt its precise risk classification of the user and instead generate a specific marker indicating a sudden trend change. Upon receiving this specific marker, the intelligent annotation module is triggered and performs a targeted query on user B's structured medical records for that time period. After finding no key intervention events related to the blood glucose mutation, the module ultimately annotates the specific marker as a pathological mutation or data anomaly of unknown origin and recommends a re-examination of this physiological indicator in the short term. For two user data sets that yield no difference under traditional information processing methods, this invention's system outputs two completely different types of information. For user A, the system, based on the continuity analysis of physiological state trajectory vectors, identified a continuously developing risk process that was masked by static numerical values. For user B, the system, through the synergistic effect of trajectory consistency verification and intelligent annotation, separated a false trace caused by random data fluctuations from the real risk trend, thus avoiding a false risk alert. By shifting the focus of risk assessment from isolated absolute values ​​of physiological indicators to the quantification and verification of the dynamic characteristics of multidimensional physiological state evolution processes, the system balances the ability to detect slowly evolving risks and the ability to suppress random noise in data when processing low-frequency health data.

[0044] Example 2: To objectively verify the effectiveness of the technical solution of this invention in distinguishing between continuous risk evolution and accidental data fluctuations, this example conducted a comparative experiment based on a publicly available dataset. The purpose of the experiment was to quantitatively evaluate the performance difference of the complete technical solution of this invention in early risk identification tasks compared to traditional threshold methods and some technical solutions lacking key modules. The experimental data came from a publicly available, anonymized longitudinal cohort study health dataset, which contained a large number of users' physiological indicator records over several consecutive years. Two representative subsets of test data were selected for this experiment. The first subset contained data from 1000 users whose key physiological indicators showed a stable, unidirectional evolutionary trend over a five-year observation period, but never exceeded the clinical risk threshold; this group was defined as the continuous trend group. The second subset also contained data from 1000 users whose physiological indicators remained stable in the first four years of the observation period. However, a significant, non-persistent data jump occurred in the fifth year, and this group was defined as the transient disturbance group. The experiment set up three treatment groups: a control group, a partial functional group, and the sample group of the present invention. The control group adopted a risk assessment method based on absolute value thresholds. When the measured value of any physiological indicator exceeded 95% of its corresponding risk threshold, a risk alarm was generated. The partial functional group adopted the state space construction and physiological state trajectory vector generation method of the present invention, but its risk assessment logic was simplified to judge only based on the magnitude of the physiological state trajectory vector of the most recent year. When the magnitude exceeded the preset value, an alarm was generated. The trajectory consistency verification module was not enabled in this group. The sample group of the present invention adopted the complete technical solution in the specific implementation, including the generation of physiological state trajectory vectors, the transformation of risk weight tensors, and a risk assessment mechanism gated by trajectory consistency measurement. The verification threshold of trajectory consistency measurement was set to 0.5.

[0045] During the experiment, all data from the aforementioned persistent trend group and transient disturbance group were input into the systems of the three treatment groups for analysis. The system output was recorded as either a risk warning generated or no risk warning generated. For the persistent trend group, a risk warning generated once was counted as a true positive (TP), and for the transient disturbance group, a risk warning generated once was counted as a false positive (FP). By counting the total number of TP and FP for each treatment group on the two test subsets, the true positive rate (TPR) and false positive rate (FPR) were calculated to evaluate the accuracy and robustness of risk identification. Table 1 shows an example of the processing results of 10 user data randomly selected from the two test subsets.

[0046] Table 1: Comparison of processing results for typical user data by different processing groups.

[0047] User ID Data subset Real trend type control group output partial function group output Sample output of this invention U001 Continuing trend group Continued deterioration No warning Risk warning Risk warning U002 Continuing trend group Continued deterioration No warning Risk warning Risk warning U003 Continuing trend group Continued deterioration No warning Risk warning Risk warning U004 Continuing trend group Continued deterioration No warning Risk warning Risk warning U005 Continuing trend group Continued deterioration No warning Risk warning Risk warning U006 Transient disturbance group Accidental jump No warning Risk warning Unannounced (sudden trend change) U007 Transient disturbance group Accidental jump No warning Risk warning Unannounced (sudden trend change) U008 Transient disturbance group Accidental jump No warning Risk warning Unannounced (sudden trend change) U009 Transient disturbance group Accidental jump No warning Risk warning Unannounced (sudden trend change) U010 Transient disturbance group Accidental jump No warning Risk warning Unannounced (sudden trend change)

[0048] Based on the statistical results of all 2000 user data, the control group, due to the limitations of its assessment method, had a true positive rate of 0% for the persistent trend group, failing to identify any potential risks below the threshold, and its false positive rate was also 0%. Some functional groups, by introducing physiological state trajectory vectors, improved their ability to identify the persistent trend group, achieving a true positive rate of 91.3%. However, because they could not distinguish between true and false trends, their false positive rate for the transient disturbance group was as high as 88.7%, generating numerous false alarms. The sample group of this invention, after applying the complete technical solution, achieved a true positive rate of 0% for the persistent trend group. The positive rate was 89.5%, comparable to the recognition ability of some functional groups. However, its false positive rate for the transient disturbance group was reduced to 5.2% (see Table 1). Its output results were mostly trend change markers rather than direct risk warnings. The experimental data showed that the introduction of the trajectory consistency verification module was the key factor that led to a significant reduction in the false positive rate of the sample group of this invention. This module filters out artifacts caused by single data jumps by performing geometric verification on the historical continuity of the trajectory vector. Thus, while maintaining the ability to detect the continuous risk evolution process, it enhances the system's tolerance to data noise.

[0049] Example 3: This example combines Figures 1 to 3 This section describes disease prevention systems and methods based on deep learning and knowledge graphs, such as... Figure 1 As shown, the process begins by acquiring user physiological indicator data from the electronic health record database. The state space construction module processes this data into physiological state points (PSPs) in a multi-dimensional space. Subsequently, the trajectory vector generation module calculates physiological state trajectory vectors (PSTVs) representing the direction and rate of health evolution based on the PSPs at different time points. Simultaneously, the trajectory consistency verification module calculates the geometric relationship between historical and current trajectories to determine a trajectory consistency metric. Before risk assessment, the risk weight transformation module utilizes authoritative medical knowledge and clinical guidelines provided by a medical knowledge graph to weight the PSTVs. The process generates a risk-weighted trajectory vector (RW-PSTV). The core of this process lies in a decision point controlled by a consistency metric. If the trajectory consistency is high, the risk assessment module compares the RW-PSTV with the pattern library and outputs a precise risk classification result, such as persistent unidimensional deterioration or health improvement. If the consistency is below the threshold, it indicates a trend abrupt change. At this time, the intelligent annotation module is activated. It queries structured medical records, including medication history and surgical history, to perform clinical intelligent annotation on the abrupt change and finally outputs a specific marker indicating the trend abrupt change. This marker is annotated as an interventional inflection point or a pathological mutation of unknown origin.

[0050] like Figure 2As shown, the horizontal axis represents time, extending from Q1 2023 to Q2 2024, and the vertical axis represents the risk weight values. The four lines in the figure represent blood glucose (FBG), glycated hemoglobin (HbA1c), low-density lipoprotein (LDL-C), and body mass index (BMI), respectively. The figure clearly shows that the risk weights of different indicators are not static. For example, the weight of HbA1c was increased from 1.1 to 1.2 in Q3 2023, while the weights of FBG and LDL-C were increased from 1.2 to 1.3 and from 1.3 to 1.5, respectively, in Q1 2024. The weight of BMI remained at the baseline value of 1.0 throughout this period. This dynamic adjustment process intuitively demonstrates that the system can internalize updates in external medical knowledge into adaptive adjustments of risk assessment model parameters, thereby ensuring that the risk assessment benchmark is always synchronized with the latest medical consensus.

[0051] like Figure 3 As shown, after system initialization, the user's physiological trajectory is usually in a stable monitoring state. When a new data analysis event is received, if the conditions of high trajectory consistency and high risk level are met, the system will switch to a high-risk trend warning state and generate a warning. If subsequent analysis finds that the data is abnormal or fluctuates in a short period of time, it can return to the stable monitoring state after review and confirmation. If the condition of the new data analysis is low trajectory consistency, the system enters a trend change pending investigation state and stops risk classification. Then, the intelligent annotation analysis event is triggered. If it is identified as an intervention-oriented improvement inflection point, the state is changed to post-intervention improvement confirmation. If it is identified as a pathological mutation of unknown origin, the system will prompt that manual intervention or clinical review is required. When a user in the improvement confirmation or stable monitoring state has a stable health status according to subsequent analysis, the system will continue to maintain the stable monitoring state. This figure fully shows how the system achieves dynamic tracking, classification and feedback of the user's health trend through a series of preset events and conditions, forming an intelligent management closed loop.

[0052] Example 4: In a specific deployment scenario, the system of this invention needs to be integrated into a regional health information platform. This platform aggregates historical electronic health records from multiple medical institutions. Due to differences in the physiological baselines and data recording habits of people in different regions, directly using a general verification threshold or weight tensor may result in suboptimal system performance. Therefore, before the system is officially launched, a localized calibration for this specific dataset needs to be performed; to determine the trajectory consistency metric. To determine the verification threshold, the system first extracts an anonymized dataset from the regional health information platform, containing 20,000 users, each with at least three consecutive years of health records, as a calibration set. This calibration set, retrospectively annotated by clinical experts, is divided into 10,000 confirmed cases of continuous trend and 10,000 confirmed cases of random data fluctuation. The system's trajectory consistency verification module is configured to process each sample in this calibration set and calculate its corresponding... Value, then, with The value is the classifier output, with expert annotation results as the gold standard, through a... to Within the range, with To iterate through all possible validation thresholds, the true positive rate and false positive rate corresponding to each threshold point were calculated, and the receiver operating characteristic (ROC) curve was plotted. To achieve a balance between sensitivity and specificity, the calibration procedure used the maximization of the Youden exponent (the sum of the true positive rate and specificity minus one) as the criterion for determining the threshold. In this calibration, when the validation threshold was set to... At that time, the Yoden index reached its maximum value. ,therefore, This was determined as the initial working threshold for the trajectory consistency verification module in the deployment instance of this region.

[0053] Regarding the risk weight tensor The mechanism for ensuring timeliness and restructuring assumes that after the system has been operating stably for a period of time, an authoritative medical association releases new cardiovascular disease prevention guidelines. These guidelines explicitly state that for men over 50 years of age, a small annual increase in low-density lipoprotein cholesterol (LDL-C), for example, a change greater than [missing information], is acceptable. Its contribution to the risk of long-term cardiovascular events, compared to changes in body mass index (BMI) with an equivalent degree of normalization, should be given greater attention, approximately 1.5 times; the system's maintenance module allows authorized administrators to translate this new medical knowledge into... For the update operation, the administrator first filters the system to identify a specific user group of men over 50 years old, and then targets this group... The template will contain the weights corresponding to the LDL-C dimensions. From the initial value Adjusted to ,in Maintain as the baseline value After this update, when processing new user data for this specific group, the system will automatically apply this knowledge-updated, non-isotropic weighted space to calculate risk-weighted trajectory vectors. This procedure ensures that the system's risk assessment logic keeps pace with the evolution of external medical knowledge. In addition, the system has set a performance monitoring trigger. When the overall prediction accuracy of the risk assessment module drops by more than 5% compared to the historical best level in three consecutive months of validation set testing, the system will automatically trigger a re-clustering analysis of all newly accumulated anonymous physiological state trajectory vectors in the past year to generate a new risk trajectory pattern library, replacing the old pattern library, thereby ensuring the adaptability of risk pattern recognition.

[0054] Example 5: In a specific application scenario, to ensure the accuracy and reproducibility of the system's annotation of physiological state trajectory vector trend changes and the attribution analysis of annual trajectory causes, an offline knowledge base construction and model parameter calibration are required before the system goes live, targeting the application domain, such as diabetes management. To build the key intervention event knowledge base required for the intelligent annotation module, the system executes a standardized data filling procedure. This procedure first extracts all drug and treatment procedure codes related to glucose metabolism regulation from internationally recognized medical coding systems, including the World Health Organization's Anatomical Therapeutic and Chemical Classification (ATC) and the International Classification of Diseases and Related Health Problems (ICD-9) surgical codes. Then, a clinical endocrinology expert group reviews the extracted code list. The process involves nuclei analysis and screening, retaining only events recognized as having a significant and rapid impact on core physiological indicators such as fasting blood glucose or glycated hemoglobin. For each selected event code, the expert group assigns an expected direction vector of impact. For example, for the first use of SGLT-2 inhibitors, the component of the direction vector in the blood glucose dimension is labeled as negative; for weight loss surgery, the components in both the body mass index (BMI) and blood glucose dimensions are labeled as negative. Ultimately, this structured information—intervention event code, applicable disease domain, and expected direction vector of impact—is stored in a targeted database in the system backend. This provides the intelligent annotation module with a comparable and objective attribution benchmark when facing trend changes.

[0055] To calibrate the parameters of the weighted model in the time-significant attribution module, the system used an anonymized retrospective dataset containing 5,000 users with daily average step count records for more than one year and annual start and end point health check reports. This module first performed individualized normalization on the daily average step count time series for each user, calculating the standard deviation multiple of the weekly average step count to their individual annual average step count, using this as a behavioral proxy indicator reflecting the volatility of their lifestyle. The weighted model was set as a non-linear function, representing the contribution weight of a given week's behavioral state to the total annual trajectory. The weighting is positively correlated with the degree of negative deviation of the behavioral agent metric for that week. Specifically, when the behavioral agent metric is at or above the user's personal baseline, the weighting is increased. The weighting approaches a baseline value, such as 0.1, and when the indicator deviates significantly negatively, such as when the number of steps is below two standard deviations of the individual average for a consecutive week, the weighting is adjusted. The value tends to an upper limit, such as 1.0. The key parameters in the model, namely the inflection point and slope of the nonlinear function, are calibrated through an optimization algorithm. The goal of this algorithm is to maximize the overlap between the key contribution period of the model output and the time period of known major negative life events retrospectively labeled by clinicians in the dataset. Through this calibration process, the time significance attribution module has the ability to transform non-standardized consumer-grade health data into quantitative analysis that can be used to explain the causes of changes in clinical-grade physiological data.

[0056] Example 6: In a specific system initialization scenario, to construct a risk trajectory pattern library that reflects the main health evolution paths of the target population, unsupervised learning and clustering analysis using large-scale historical data is required before system deployment. To construct the risk trajectory pattern library, the system first extracts all eligible annual physiological state trajectory vectors from a retrospective cohort dataset containing 500,000 users, each with more than five consecutive years of anonymized electronic health records. After processing by the risk weight transformation module, a risk-weighted trajectory vector containing millions of vectors is formed. The system uses an improved K-means clustering algorithm to process the high-dimensional dataset. To address the sensitivity of the traditional K-means algorithm to the selection of initial centroids, this procedure uses the K-means++ algorithm for initialization to make the distribution of initial cluster centers more reasonable. The core objective of clustering is to group vectors with similar orientations and magnitudes in N-dimensional space into the same cluster. The number of clusters, K, which represents the total number of risk patterns, is determined based on a process aimed at balancing model complexity and explanatory power. This involves calculating the silhouette coefficient and Calinski-Harabasz index under different K values ​​and selecting the K value that makes both indices reach a higher value as the number of clusters, for example, K=8.

[0057] After clustering, each resulting cluster represents a typical health status evolution pattern. To assign a clear risk level to these patterns, the system executes a retrospective association analysis procedure. For each cluster, the system calculates the actual incidence rate of a specific chronic disease, such as type 2 diabetes, for all users belonging to that cluster within the next five years. Based on this incidence rate, and combined with the centroid vector of each cluster—which represents the most typical evolutionary direction and rate of the pattern—a risk level is assigned to each cluster by a clinical expert group. Examples of risk levels include rapid multidimensional deterioration, concealed single-dimensional deterioration of blood glucose, stable maintenance, or intervention improvement. The collection of these risk-level-labeled clusters, including their centroid vectors, intra-cluster sample variance, and associated long-term incidence rate data, constitutes a risk trajectory pattern library, which is stored in the system backend. When the risk assessment module receives a new user risk-weighted trajectory vector... The algorithm implementation path is as follows: first, calculate the Euclidean distance between the vector and the centroid vector of each cluster in the pattern library; then, classify the user into the cluster with the closest distance; and output the risk level and associated statistical data corresponding to the cluster as the risk classification result.

[0058] To further verify the indispensable role of each core technical component in the technical solution of this invention, especially the trajectory consistency verification module, in achieving accurate and robust risk identification, the following comparative examples are provided.

[0059] Comparative Example 1: This comparative example aims to illustrate the crucial role of the trajectory consistency verification module and its gating mechanism in suppressing data noise and avoiding false warnings. Except for the absence of the trajectory consistency verification module, which simplifies the risk assessment logic to directly based on the magnitude of the risk-weighted trajectory vector, all other technical conditions, data sources, and processing steps used in this comparative example, including state space construction, trajectory vector generation, and risk weight transformation, are completely identical to the samples described in Examples 1 and 2. The risk assessment method used in this comparative example follows these rules: after calculating the user's risk-weighted trajectory vector (RW-PSTV), its Euclidean norm, i.e., the magnitude, is directly calculated. If the modulus exceeds a threshold determined based on large-scale historical data statistics, which is equivalent to the boundary of the cluster used by the sample group of this invention to distinguish between stable maintenance type and hidden one-dimensional deterioration type, the system directly outputs a high-risk trend warning. This method represents a simplified technical path based on this invention, omitting the core verification link, and is easier to think of.

[0060] To verify its ability to process complex real-world data, a similar experimental setup to that in Example 2 was used. Two test subsets were constructed from a retrospective dataset containing 50,000 users, each with five consecutive years of health records: a persistent trend group (Group A) containing data from 2,500 users whose key physiological indicators (such as fasting blood glucose and low-density lipoprotein) showed a stable, unidirectional, and significant annual variation over the five-year observation period, but never exceeded a single clinical risk threshold. This group represents the early, persistent risks that this invention aims to effectively identify. The transient disturbance group (Group B) also contained data from 2,500 users whose physiological indicators... For the first four years, the data remained near ideal health values. However, a non-persistent data jump occurred in a single record in the fifth year. Retrospective medical records confirmed that this was caused by short-term stress, detection error, or acute self-healing illness. This group represents data noise that is common in the real world and should not be classified as chronic disease risk. All data from the above groups A and B were input into the simplified system described in this comparative example and the sample group of the present invention (using the complete technical solutions of Examples 2 and 6) for analysis. For group A, a successful risk warning output was counted as a true positive (TP); for group B, a risk warning output was counted as a false positive (FP). The experimental results are shown in Table 2.

[0061] Table 2: Performance comparison table between the simplified comparative system and the sample group of the present invention.

[0062] Test metrics Scaled Simplification System Sample of the present invention True Positive Rate (TPR) in Group A (Continuing Trend Group) 93.5% 91.8% False positive rate (FPR) in Group B (transient disturbance group) 92.1% 4.7%

[0063] The experimental results of Comparative Example 1 show that, in the absence of a mechanism to verify the historical continuity of physiological state trajectories, the risk assessment method based solely on the current state of the trajectory vector (e.g., magnitude) can identify most persistent trends (true positive rate 93.5%), but at the cost of failing to effectively distinguish between genuine, continuous physiological evolution and artifacts caused by random fluctuations in single data points. Consequently, it generates a false positive warning rate as high as 92.1% for the transient disturbance group. In practical applications, this high false positive rate will lead to a large number of unnecessary subsequent clinical re-examinations, significantly increasing the burden on the medical system and potentially causing unnecessary psychological anxiety for users. In contrast, the sample group of this invention, by introducing a trajectory consistency verification module, maintains an extremely high identification ability for persistent trend groups (true positive rate 91.8%, with no statistical difference from the simplified system) while successfully suppressing the false positive rate of the transient disturbance group to an extremely low level of 4.7%. The vast majority of its output to Group B users is a trend change to be investigated marker, rather than a direct risk warning.

[0064] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A disease prevention system based on deep learning and a knowledge graph, characterized by, The system comprises: a state space construction module configured to establish a standardized multi-dimensional physiological state space according to a plurality of preset physiological indicators, and map physiological indicator data of a user at a first time point and a second time point, which are spaced apart by 9 to 15 months, into a first physiological state point and a second physiological state point in the space, respectively; a trajectory vector generation module connected to the state space construction module and configured to calculate a physiological state trajectory vector based on the first physiological state point and the second physiological state point; a risk weight transformation module connected to the trajectory vector generation module and configured to perform tensor product transformation on the physiological state trajectory vector by using a risk weight tensor corresponding to each dimension of the multi-dimensional physiological state space and updated by a knowledge graph, to generate a risk-weighted trajectory vector; a trajectory consistency verification module configured to calculate a geometric relationship between a historical trajectory vector and the physiological state trajectory vector after obtaining a third physiological state point of the user at a third time point earlier than the first time point, to determine a trajectory consistency measure; a risk assessment module connected to the risk weight transformation module and the trajectory consistency verification module, configured to perform risk classification by taking the risk-weighted trajectory vector as the only input, and an output mode of the risk assessment module is gated by the trajectory consistency measure; wherein the output mode of the risk assessment module is gated by the trajectory consistency measure, specifically configured as follows: when the trajectory consistency measure is not lower than a preset verification threshold, the risk assessment module outputs an exact risk classification result; when the trajectory consistency measure is lower than the verification threshold, the risk assessment module stops outputting the exact risk classification result and generates a specific marker indicating that a trend has mutated; The system further comprises an intelligent annotation module connected to the risk assessment module, which is configured to be triggered when the risk assessment module generates the specific marker, to perform a targeted query on structured medical records of the user between the first time point and the second time point, to determine whether there is a preset key intervention event code; and based on the result of the query, the specific marker is intelligently annotated with clinical properties; When the intelligent annotation module is configured to perform intelligent annotation with clinical properties, the following rules are followed: if the key intervention event code is found, and the physiological state trajectory vector points to a healthy improvement direction, the specific marker is annotated as an interventionally good inflection point; if the key intervention event code is not found, the specific marker is annotated as a pathologically mutated or data abnormality of unknown origin. 2.The disease prevention system based on deep learning and knowledge graph of claim 1, wherein, The state space construction module is configured to perform linear mapping processing on the value range of each physiological indicator, as the measurement of the corresponding dimension of the multi-dimensional physiological state space; Moreover, the direction of the physiological state trajectory vector represents the synergistic change trend of the user's physiological state in multiple indicator dimensions, and the modulus quantifies the overall rate of change of the user's physiological state between the first time point and the second time point. 3.The disease prevention system based on deep learning and knowledge graph of claim 1, wherein, The trajectory consistency check module is configured to determine a trajectory consistency measure by computing a cosine similarity of a history trajectory vector to a physiological state trajectory vector , which computation follows the rule wherein the trajectory consistency measure, the is a vector pointing from the third physiological state point to the first physiological state point, the is a physiological state trajectory vector, and the denotes the Euclidean norm of a vector. 4.The disease prevention system based on deep learning and knowledge graph of claim 1, wherein, In the risk weight transformation module, the risk weight tensor is a diagonal tensor, each diagonal element of the diagonal tensor is a risk weight value corresponding to a single dimension of the multi-dimensional physiological state space; the risk weight transformation module is further configured to obtain an update instruction from an external medical knowledge base or a clinical guideline knowledge graph. 5.The disease prevention system based on deep learning and knowledge graph of claim 1, wherein, The preset check threshold is set to a value in a range of 0.3 to 0.

7. 6.The disease prevention system based on deep learning and knowledge graph of claim 1, wherein, The system further comprises a time saliency attribution module configured to obtain a time series of at least one high-frequency behavior proxy data associated with the user between the first time point and the second time point, and based on the time series, determine one or more key time periods having the highest weight value in the attribution analysis on the physiological state trajectory vector, and provide the determined one or more key time periods together with the output of the risk assessment module, the physiological indicator data processed by the state space construction module in the system is derived from an electronic health record database, the knowledge graph applied by the risk weight transformation module is a medical field knowledge graph, and the structured medical record queried by the intelligent annotation module includes medication history records and surgery history records. 7.A disease prevention method based on deep learning and a knowledge graph, applied to the disease prevention system of claim 1, characterized in that, The method comprises the following steps: Step a, establishing a standardized multi-dimensional physiological state space according to a plurality of preset physiological indicators, and mapping physiological indicator data of a user at a first time point and a second time point with an interval of 9 to 15 months into a first physiological state point and a second physiological state point in the space respectively; Step b, calculating a physiological state trajectory vector based on the first physiological state point and the second physiological state point; Step c, performing tensor product transformation on the physiological state trajectory vector by using a risk weight tensor corresponding to each dimension of the multi-dimensional physiological state space and updated by a knowledge graph, to generate a risk weighted trajectory vector; Step d, after obtaining a third physiological state point of the user at a third time point earlier than the first time point, calculating the geometric relationship between the historical trajectory vector and the physiological state trajectory vector to determine a trajectory consistency measure; Step e, using the risk weighted trajectory vector as the only input for risk classification, and the output mode of the risk classification is gated by the trajectory consistency measure.

Images