A knowledge graph construction method based on chronic disease management

Abstract

The application relates to the technical field of big data processing, and discloses a knowledge graph building method based on chronic disease management. The present application uses deduction logic based on biological steady-state game, introduces thermodynamic dissipation and entropy increase constraint, accurately quantifies the irreversible loss of living beings in the disease course, overcomes the defect that the existing algorithm cannot simulate physiological function decline, at the same time, the present application constructs a physical filter based on the principle of biological impedance mismatch, uses mass transport limitation to eliminate sensor high-frequency artifact noise, significantly reduces the false positive rate while ensuring the real lesion recall rate, further, the present application converts the medical contraindication rule into a rigid topological potential constraint through contact mechanics and gibbs free energy functional, forcibly corrects the data violating common sense, ensures the high unity of the output graph in the physical observation reality and the medical logic self-consistency, and greatly improves the clinical explainability and decision safety of the auxiliary diagnosis and treatment system.

Description

Technical Field

[0001] This application relates to the field of big data processing technology, and in particular to a method for building a knowledge graph based on chronic disease management. Background Technology

[0002] With the deepening application of medical informatics and the Internet of Things (IoT) technologies, the full-cycle management of chronic diseases has generated massive amounts of multi-source heterogeneous data. On the one hand, clinical treatment guidelines and pharmacological literature constitute a relatively static medical theoretical foundation; on the other hand, hospital records and wearable devices generate a large amount of dynamic, time-varying vital sign monitoring data in real time. This time-series data records in detail the disease progression, the delayed response of drug efficacy, and the fluctuation patterns of physiological indicators. Constructing an intelligent knowledge system that can deeply integrate static pathological mechanisms and dynamic disease states is of great significance for assisting clinicians in formulating precise treatment plans and realizing patient risk warnings.

[0003] However, existing knowledge graph construction techniques are mostly based on static triple structures, lacking endogenous support for the time dimension, making it difficult to represent constant pathological mechanisms and fluid disease states within the same graph. Current technologies often use time information as node attribute labels, requiring the system to traverse numerous non-indexed fields when performing cross-time period searches, resulting in low data query efficiency and an inability to support complex temporal causal reasoning. Another common discrete-time slice subgraph scheme not only artificially severs the inherent logical links across time periods, making it difficult to capture long-term evolutionary patterns, but also causes significant computational resource consumption and substantial storage redundancy when dealing with high-frequency monitoring data due to frequent full-scale graph reconstruction. Therefore, how to achieve incremental fusion of temporal features and continuous modeling of disease evolution while maintaining topological stability is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0004] This application proposes a method for building a knowledge graph based on chronic disease management to address the problems mentioned in the background.

[0005] To achieve the above objectives, this application adopts the following technical solution: a method for constructing a knowledge graph based on chronic disease management, comprising the following steps:

[0006] Step S1: In response to the received non-time-varying domain knowledge data and time-varying domain physiological monitoring data, perform time reference normalization processing, map the non-time-varying domain knowledge data to a background potential field existing in the entire time domain, and map the time-varying domain physiological monitoring data to instantaneous particle states under normalized time coordinates. Perform full data traversal on the mapped background potential field and instantaneous particle states, and extract the coordinate tuples corresponding to the effective values ​​to generate a heterogeneous spatiotemporal tensor index set, wherein the heterogeneous spatiotemporal tensor index set contains quadruples of coordinate information that clearly distinguish between time invariants and time covariates;

[0007] Step S2: Obtain the heterogeneous spatiotemporal tensor index set generated in step S1, assign complex embedding vectors to entities and relations in the heterogeneous spatiotemporal tensor index set, introduce pathological angular velocity vectors that characterize the physical properties of disease change rate, perform trajectory deduction according to the complex domain rotation embedding algorithm, calculate the phase offset of entities under the passage of time, and generate a temporal prediction entity vector matrix with complex phase information.

[0008] Step S3: Obtain the time-series predicted entity vector matrix generated in step S2 and the real-time observation data at the current moment. Calculate the state deviation norm between the time-series predicted entity vector matrix and the real-time observation data. Activate the damped oscillation gated update mechanism based on the state deviation norm. Use the biological damping function in the damped oscillation gated update mechanism to limit the state change rate and remove high-frequency noise interference. Output the dynamic node state vector after physical filtering correction.

[0009] Step S4: Obtain the dynamic node state vector after physical filtering correction generated in step S3, construct a Gibbs free energy functional containing data fitting terms and logical penalty terms, model the medical taboo rules as an infinitely high potential barrier in the Gibbs free energy functional, perform reverse gradient calibration on the dynamic node state vector after physical filtering correction by minimizing the Gibbs free energy functional, and output a dynamic chronic disease knowledge graph under medical logical constraints.

[0010] Furthermore, in step S1, time reference normalization processing is performed to map non-time-varying domain knowledge data to a background potential field existing in the entire time domain, and to map time-varying domain physiological monitoring data to instantaneous particle states under normalized time coordinates. The specific operations are as follows:

[0011] A10 acquires electronic medical record data and device log data throughout the patient's life cycle, performs a minimization extreme value search operation to determine the start timestamp of a valid medical event, and sets the start timestamp as the physical zero-time reference point;

[0012] A11: Acquire time-varying domain physiological monitoring data with absolute timestamps, perform differential subtraction between the absolute timestamps and the physical zero-time reference point, and generate normalized time manifold coordinates.

[0013] A12 performs a non-negative numerical domain determination operation on the normalized time manifold coordinates. If the determination result is negative, the data is truncated and discarded. If the determination result is non-negative, the data is retained and verified time-varying domain physiological monitoring data is generated.

[0014] A13, acquire non-time-varying domain knowledge data, construct a full-time-domain dimension mask matrix, perform broadcast copy operation on the non-time-varying domain knowledge data along the time axis dimension of the normalized time manifold coordinates, and generate a background potential field data structure with constant values ​​on any time slice.

[0015] A14 maps the verified time-varying domain physiological monitoring data to discrete node positions in the sparse tensor space corresponding to the normalized time manifold coordinates to generate instantaneous particle state data, and performs isomorphic alignment storage of the instantaneous particle state data and the background potential field data structure based on the same coordinate system.

[0016] Furthermore, in step S1, the specific operation of performing a full data traversal on the mapped background potential field and instantaneous particle states to extract the coordinate tuples corresponding to the effective values ​​to generate a heterogeneous spatiotemporal tensor index set is as follows:

[0017] A20 constructs a fourth-order tensor space container and obtains the metadata of the hardware device to extract the physical measurement accuracy value, and generates the signal confidence factor based on the physical measurement accuracy value;

[0018] A21, construct a domain orthogonality discriminant operator to distinguish data types, and configure the spatiotemporal existence density calculation logic through the domain orthogonality discriminant operator;

[0019] A22 performs a full traversal of the background potential field data structure and instantaneous particle state data. For non-time-varying domain knowledge data, it calls the all-pass unit potential field function through the domain orthogonality discriminant operator and calculates the first spatiotemporal existence density value in combination with the signal confidence factor.

[0020] A23, for time-varying domain physiological monitoring data, the second spatiotemporal existence density value is obtained by calling the discrete impulse function through the domain orthogonal discriminant operator and combining it with the signal confidence factor;

[0021] A24, obtain the preset minimum observable threshold, perform numerical comparison operations between the first spatiotemporal existence density value and the second spatiotemporal existence density value and the minimum observable threshold respectively, and retain valid data nodes whose values ​​are greater than the minimum observable threshold;

[0022] A25 extracts the four-dimensional coordinate information and corresponding existence density values ​​of valid data nodes, constructs coordinate tuples containing time invariant and time covariant labels, and aggregates all coordinate tuples to output a heterogeneous spatiotemporal tensor index set.

[0023] Furthermore, in step S2, the specific operation of assigning complex embedding vectors to entities and relations in the heterogeneous spatiotemporal tensor index set and introducing pathological angular velocity vectors representing the physical properties of disease change rate is as follows:

[0024] B10, traverse the heterogeneous spatiotemporal tensor index set generated in step S1, and obtain the background potential field data structure marked as time invariant and the instantaneous particle state data marked as time covariant.

[0025] B11 performs a complex field spatial mapping operation, mapping the background potential field data structure to the real part of a complex vector, and simultaneously mapping the instantaneous particle state data to the imaginary part of a complex vector, thereby constructing an initial complex state vector composed of the real and imaginary parts.

[0026] B12, perform a modulus calculation operation on the initial complex state vector to obtain the modulus value of the initial complex state vector, and perform a numerical comparison operation between the modulus value and the preset biological survival domain limit threshold. If the modulus value is greater than the biological survival domain limit threshold, then perform truncation and normalization processing on the initial complex state vector.

[0027] B13, Obtain the signal confidence factor generated in step S1, and perform a self-square operation on the signal confidence factor to generate confidence weight coefficients;

[0028] B14 performs first-order differential gradient calculation on the instantaneous particle state data about the time axis of the normalized time manifold coordinates to generate the gradient change rate, and performs tensor product operation on the gradient change rate and confidence weight coefficients to generate the pathological driving torque characterizing the intensity of external disturbance.

[0029] B15 performs a weighted harmonic average operation based on a diagonal matrix on the background potential field data structure to generate a steady-state stiffness tensor characterizing the organism's ability to resist changes in state.

[0030] B16 performs a tensor division operation on the pathological driving torque and the steady-state stiffness tensor, and outputs a pathological angular velocity vector, which represents the rate at which the phase deflection of the complex domain entity state occurs within a unit time step.

[0031] Furthermore, in step S2, the specific operation of performing trajectory deduction based on the complex domain rotation embedding algorithm, calculating the phase shift of the entity over time, and generating a time-series predicted entity vector matrix with complex phase information is as follows:

[0032] B20, obtain the normalized time manifold coordinates generated in step S1, and obtain the initial complex state vector and pathological angular velocity vector calculated by the preceding operations in this step;

[0033] B21. Construct an overdamped dynamic model based on non-equilibrium thermodynamics. First, define the thermodynamic dissipation coefficient to characterize the irreversible loss of the system, and define the relaxation time constant to characterize the steady-state recovery hysteresis of the system.

[0034] B22, calculate the ratio of the normalized time manifold coordinates to the relaxation time constant, and perform a natural constant exponential function operation on the negative value of the ratio to generate the time inertia decay factor;

[0035] B23, perform a multiplication operation between the time inertia decay factor and the pathological angular velocity vector to calculate the effective rotational phase angle after damping correction;

[0036] B24. Establish a polar coordinate system in the complex plane, perform a vector rotation operation on the initial complex state vector according to the effective rotation phase angle, and subtract the product of the thermodynamic dissipation coefficient and the normalized time manifold coordinate from the magnitude value of the initial complex state vector to generate a predicted complex state vector containing the updated phase and the updated magnitude.

[0037] B25 performs orthogonal feature decoupling operation on the predicted complex state vector, extracts its phase angle value as disease progression stage prediction data, extracts its magnitude value as residual physiological function index data, and performs stacking and splicing operation on the phase angle value and residual physiological function index data along the feature channel dimension to output a time-series predicted entity vector matrix with complex phase information.

[0038] Further, in step S3, the time-series predicted entity vector matrix generated in step S2 and the real-time observation data at the current moment are obtained. The state deviation norm between the time-series predicted entity vector matrix and the real-time observation data is calculated. Based on the state deviation norm, the damped oscillation gated update mechanism is activated. The biological damping function in the damped oscillation gated update mechanism is used to limit the state change rate and remove high-frequency noise interference. The specific operation of outputting the physically filtered corrected dynamic node state vector is as follows:

[0039] C10, obtain the time-series prediction entity vector matrix generated in step S2, and extract the components along the real part dimension of the time-series prediction entity vector matrix as physiological anchor point data.

[0040] C11, simultaneously acquire the background potential field data structure generated in step S1, and perform a normalization mapping operation based on the background potential field data structure on the real-time observation data at the current moment to generate normalized real-time observation values.

[0041] C12 constructs a complex domain holographic projection space, maps physiological anchor point data to the real part coordinates of the holographic observation complex vector, and maps normalized real-time observation values ​​to the imaginary part coordinates of the holographic observation complex vector, thereby generating the holographic observation complex vector;

[0042] C13, obtain the dynamic node state vector after physical filtering correction output from the previous time step, calculate the absolute value of the difference between the imaginary coordinates of the holographic observation complex vector and the imaginary coordinates of the dynamic node state vector after physical filtering correction, and divide the absolute value of the difference by the time step of the normalized time manifold coordinates generated in step S1 to calculate the instantaneous metabolic change rate.

[0043] C14 retrieves the preset physiological rate limit threshold, performs a subtraction difference operation on the instantaneous metabolic change rate and the physiological rate limit threshold, and inputs the operation result into a nonlinear inverse logistic decay function constructed based on the natural constant exponent to perform a mapping operation, outputting a physical reliability factor. This physical reliability factor exhibits a nonlinear approach to zero numerical characteristic when the instantaneous metabolic change rate is greater than the physiological rate limit threshold.

[0044] Furthermore, in step S3, the specific operation of using the biological damping function in the damped oscillation gating update mechanism to limit the state change rate and eliminate high-frequency noise interference, and outputting the physically filtered corrected dynamic node state vector is as follows:

[0045] C20, construct a biological admittance tensor, which is configured as the reciprocal of the physical impedance characterizing the biological system's acceptance of external state changes;

[0046] C21, Perform a vector subtraction operation on the holographic observation complex vector generated by the preceding operation in step S3 and the time-series prediction entity vector matrix generated by step S2 to generate an impedance mismatch deviation vector.

[0047] C22, obtain the physical confidence factor generated by the preceding operation in step S3, and perform a ternary tensor product operation on the impedance mismatch deviation vector, the physical confidence factor and the biological admittance tensor to generate an effective corrected flux vector.

[0048] C23 will perform vector addition on the flux vector and the time-series prediction entity vector matrix to generate a preliminary updated state vector.

[0049] C24, calculate the complex modulus of the initial updated state vector and the complex modulus of the time-series predicted entity vector matrix, and perform a numerical comparison operation between the complex modulus of the initial updated state vector and the complex modulus of the time-series predicted entity vector matrix.

[0050] C25. If the complex modulus of the initial updated state vector is greater than the complex modulus of the time-series predicted entity vector matrix, the system performs a forced entropy increase normalization operation to scale the complex modulus of the initial updated state vector to be equal to the complex modulus of the time-series predicted entity vector matrix, while keeping the phase angle of the initial updated state vector unchanged. Finally, the system outputs the dynamic node state vector after physical filtering correction following energy conservation constraint verification.

[0051] Furthermore, in step S4, the specific operation of obtaining the dynamic node state vector after physical filtering correction generated in step S3 and constructing the Gibbs free energy functional containing data fitting terms and logical penalty terms is as follows:

[0052] D10, obtain the dynamic node state vector after physical filtering correction output by step S3, and define the dynamic node state vector after physical filtering correction as physical anchor point data.

[0053] D11 performs a vector expansion operation on the physical anchor point data based on the concatenation of the real and imaginary channels to generate a generalized coordinate state vector in the real number domain.

[0054] D12, call the biological admittance tensor generated in step S3, and perform matrix inversion operation on the biological admittance tensor to generate the biological stiffness tensor;

[0055] D13 calculates the difference vector between the generalized coordinate state vector and the physical anchor point data, and performs a quadratic matrix multiplication operation on the difference vector and the biological stiffness tensor to generate the generalized elastic potential energy value.

[0056] D14 defines the generalized elastic potential energy as a data fitting term in the Gibbs free energy functional. The data fitting term is used to constrain the generalized coordinate state vector to maintain topological proximity to the physical anchor data in subsequent iterations.

[0057] Furthermore, in step S4, the specific operation of modeling the medical contraindication rule as an infinitely high potential barrier in the Gibbs free energy functional is as follows:

[0058] D20, obtain the preset medical logic geometric constraint set, which contains a set of geometric inequalities that define mutually exclusive boundaries in the multidimensional state space;

[0059] D21 calculates the shortest Euclidean distance from the generalized coordinate state vector to the compliance boundary defined in the medical logic geometric constraint set, and assigns positive or negative signs based on the spatial positional relationship of the generalized coordinate state vector relative to the compliance boundary to generate a logic penetration depth value.

[0060] D22 performs a numerical truncation operation on the logical penetration depth value, retaining only the negative part of the logical penetration depth value and performing an absolute value square operation on it to generate the square value of the penetration amount.

[0061] D23, construct the logical wall stiffness coefficient, multiply the logical wall stiffness coefficient with the square of the penetration amount to generate the contact repulsion potential energy value;

[0062] D24 defines the contact repulsion potential energy as a logical penalty term in the Gibbs free energy functional. The logical penalty term is used to generate reverse repulsion potential energy when the generalized coordinate state vector enters the medically forbidden region.

[0063] Furthermore, in step S4, the specific operation of performing inverse gradient calibration on the dynamic node state vector after physical filtering correction by minimizing the Gibbs free energy functional to output the dynamic chronic disease knowledge graph under medical logic constraints is as follows:

[0064] D30, the generalized elastic potential energy value and the contact repulsion potential energy value are added together to construct the total Gibbs free energy functional;

[0065] D31. Establish the evolution equation of overdamped gradient flow based on the restricted variational principle, and calculate the negative gradient vector of the total Gibbs free energy functional with respect to the generalized coordinate state vector to generate the logically calibrated resultant force vector.

[0066] D32 performs iterative update operations, using the logical calibration resultant force vector to drive the generalized coordinate state vector to shift on the potential energy surface until the magnitude of the logical calibration resultant force vector is less than the preset convergence threshold, generating the converged generalized coordinate state vector.

[0067] D33 performs an inverse mapping operation based on channel splitting on the converged generalized coordinate state vector to restore it to a complex form entity vector, and constructs a dynamic chronic disease knowledge graph under medical logic constraints based on the complex form entity vector.

[0068] The beneficial effects of this invention are as follows:

[0069] This invention effectively integrates static pathological mechanisms with dynamic disease evolution by constructing a heterogeneous spatiotemporal tensor index and a complex domain dynamics embedding model, solving the problem of traditional atlases lacking temporal support. Utilizing deductive logic based on biological steady-state game theory, this invention introduces thermodynamic dissipation and entropy increase constraints to accurately quantify the irreversible loss of organisms during disease progression, overcoming the limitation of existing algorithms in simulating physiological function decline. Furthermore, based on the principle of bioimpedance mismatch, this invention constructs a physical filter and uses mass transfer constraints to eliminate high-frequency artifact noise from sensors, significantly reducing false alarm rates while maintaining the recall rate of true lesions. Further, this invention transforms medical taboo rules into rigid topological potential constraints through contact mechanics and Gibbs free energy functionals, forcibly correcting data that violates common sense, ensuring a high degree of consistency between the physical observation authenticity and medical logical self-consistency of the output atlas, and significantly improving the clinical interpretability and decision-making safety of the auxiliary diagnostic system. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0071] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0073] Example

[0074] like Figure 1 As shown, this invention discloses a method for constructing a knowledge graph based on chronic disease management, including the following steps:

[0075] Step S1: In response to the received non-time-varying domain knowledge data and time-varying domain physiological monitoring data, perform time reference normalization processing, map the non-time-varying domain knowledge data to a background potential field existing in the entire time domain, and map the time-varying domain physiological monitoring data to instantaneous particle states under normalized time coordinates. Perform full data traversal on the mapped background potential field and instantaneous particle states, and extract the coordinate tuples corresponding to the effective values ​​to generate a heterogeneous spatiotemporal tensor index set, wherein the heterogeneous spatiotemporal tensor index set contains quadruples of coordinate information that clearly distinguish between time invariants and time covariates.

[0076] Specifically, in step S1, time reference normalization processing is performed to map non-time-varying domain knowledge data to a background potential field existing in the entire time domain, and to map time-varying domain physiological monitoring data to instantaneous particle states under normalized time coordinates. The specific operations are as follows: acquire electronic medical record data and device log data of the patient's entire life cycle, perform minimization extreme value search operation to determine the start timestamp of the effective medical event, and set the start timestamp as the physical zero time reference point.

[0077] In this embodiment, in order to avoid the offset of the time axis origin caused by discrete noise data generated during the equipment factory debugging or system initialization phase, an extreme value search logic based on sliding window statistics is executed.

[0078] Specifically, a parameter called the effective medical event determination threshold is set. In this embodiment, the effective medical event determination threshold is preferably set as the effective reading with a data variance greater than zero within three consecutive recording periods. The basis for choosing this specific value is that statistical analysis of 100,000 historical cleaned data revealed that early records with a frequency lower than this or a constant value (variance of zero) are usually equipment self-test data rather than real medical records. By applying this threshold constraint, false early timestamps can be filtered out.

[0079] Experimental data show that, compared with the traditional method of directly selecting the smallest timestamp, this strategy improves the accuracy of locating the starting point of the disease course by about 15 percent, and effectively avoids the unnecessary sparsity problem in subsequent calculations of the entire timeline data caused by the incorrect forward shift of the origin.

[0080] Subsequently, time-varying domain physiological monitoring data with absolute timestamps are acquired, and differential subtraction is performed between the absolute timestamps and the physical zero-time reference point to generate normalized time manifold coordinates.

[0081] Read the Greenwich Mean Time stamp of each monitoring data point, subtract the aforementioned physical zero-time reference point to obtain the absolute time difference, and then perform a division operation to divide the difference by a preset time scale scaling factor.

[0082] In this embodiment, the time scale scaling factor is set to 86,400, which is the number of seconds contained in a whole day, thereby converting the time unit from seconds to clinical days. The choice of this unit conversion parameter is not arbitrary, but based on the medical rule that the pathological evolution of chronic diseases such as diabetes or hypertension usually has a minimum significant change cycle of days.

[0083] This operation maps the originally large-scale long integer timestamps into compact floating-point coordinates. In subsequent tensor convolution operations, this clinical-dimensional normalization process improves the convergence speed of the model's loss function by about 2.5 times compared to unnormalized data.

[0084] Next, a non-negative numerical domain determination operation is performed on the normalized time manifold coordinates. If the determination result is negative, the data is truncated and discarded. If the determination result is non-negative, the data is retained and verified time-varying domain physiological monitoring data is generated.

[0085] When performing the non-negative determination, it is not simply based on zero, but a time synchronization tolerance window parameter is introduced. In this embodiment, the time synchronization tolerance window is set to -0.01, which is approximately 15 minutes. The basis for setting this negative tolerance is the network latency and clock drift phenomenon that is common among distributed medical IoT devices. If the zero value is strictly truncated, valid emergency data that occurred very shortly before the time of diagnosis may be incorrectly discarded.

[0086] Comparative experiments show that after introducing this tolerance window, the recall rate of critical perinatal data increased by 4.2 percent, and no logical errors violating causality were introduced. For negative data smaller than the tolerance window, they were marked as outliers that violated the timing logic and physically deleted from memory.

[0087] Simultaneously, non-time-varying domain knowledge data is acquired, a full-time-domain dimensional mask matrix is ​​constructed, and a broadcast copy operation is performed on the non-time-varying domain knowledge data along the time axis dimension of the normalized time manifold coordinates to generate a background potential field data structure with constant values ​​on any time slice.

[0088] For patients’ non-time-varying data such as gender, blood type, and genotype, a matrix covering the maximum predictive lifespan is constructed in the time dimension. In this embodiment, the maximum predictive lifespan is set to 120 years.

[0089] By performing a broadcast copy operation, the static feature vector is copied and filled into the entire matrix along the time axis. This operation logically constructs a data structure whose values ​​remain constant at any time slice. This full-time-domain filling strategy solves the technical defect of traditional recurrent neural networks that are prone to forgetting static background information when dealing with long-distance dependencies. It ensures that static features can provide stable weight contributions in convolution operations at any subsequent time point, thereby significantly improving the model's predictive stability for long-term disease progression.

[0090] Finally, the verified time-varying domain physiological monitoring data is mapped to discrete node positions in the sparse tensor space corresponding to normalized time manifold coordinates to generate instantaneous particle state data, and the instantaneous particle state data and background potential field data structure are stored in an isomorphic alignment based on the same coordinate system.

[0091] A sparse tensor is constructed using a coordinate list format, storing data only at coordinates where valid measurements exist, and padding the remaining positions with zeros. In this step, the tensor sparsity filling rate is strictly controlled; for chronic disease monitoring data, this filling rate is typically less than one-thousandth.

[0092] Aligning sparse dynamic data with the aforementioned dense static data in the same coordinate system enables the simultaneous handling of continuous background constraints and discrete burst events. Performance tests show that this isomorphic aligned storage structure improves the input and output throughput by more than eight times compared to the traditional dual-database separate query mode when performing complex multi-hop queries.

[0093] In step S1, the specific operation of performing a full data traversal on the mapped background potential field and instantaneous particle state to extract the coordinate tuples corresponding to the effective values ​​to generate a heterogeneous spatiotemporal tensor index set is as follows: constructing a fourth-order tensor space container, obtaining the metadata of the hardware device to extract the physical measurement accuracy value, and generating a signal confidence factor based on the physical measurement accuracy value.

[0094] The fourth-order tensor dimensions are sample batch, feature channel, time step, and confidence channel. When generating the signal confidence factor, a pre-set precision lookup table is matched based on the metadata.

[0095] In this embodiment, for professional-grade monitoring equipment in hospitals, such as large-scale biochemical analyzers, the signal confidence factor is set to 0.98; for home-use consumer-grade wearable devices, the factor is set to 0.75. The basis for setting the different signal confidence factors is derived from the permissible error range of medical devices in national standards.

[0096] By introducing confidence weighting during the tensor construction phase, the model can automatically reduce the weight of gradients generated by low-confidence data during training. Experiments show that this mechanism improves the model's robustness to high-noise household data by about 18%.

[0097] Subsequently, a domain orthogonality discriminant operator is constructed to distinguish data types. The spatiotemporal existence density calculation logic is configured through the domain orthogonality discriminant operator. The domain orthogonality discriminant operator is specifically implemented as a logic gating unit based on data type labels.

[0098] The operator switching response time is set to be less than ten nanoseconds to meet the processing requirements of high-concurrency data streams. The orthogonal discrimination operator in this domain dynamically loads different density calculation function pointers based on whether the data being processed is a static background or a dynamic event, thereby enabling parallel processing of heterogeneous data in the same computation stream and ensuring computational efficiency.

[0099] Next, a full traversal is performed on the background potential field data structure and the instantaneous particle state data. For the non-time-varying domain knowledge data, the first spatiotemporal existence density value is obtained by calling the all-pass unit potential field function through the domain orthogonality discriminant operator and combining it with the signal confidence factor.

[0100] For data that is confirmed as static domain knowledge by the discriminant operator, the all-pass unit potential field function is invoked. The output value of the all-pass unit potential field function is constant at one in the entire time domain. This constant value is multiplied by the aforementioned signal confidence factor to obtain the first spatiotemporal existence density value, which characterizes that the knowledge point has a uniform and continuous logical influence on the entire time axis and does not decay over time.

[0101] Meanwhile, for time-varying physiological monitoring data, the second spatiotemporal existence density value is obtained by calling the discrete impulse function through the domain orthogonal discriminant operator and combining it with the signal confidence factor. For dynamic monitoring data, the discrete impulse function is called.

[0102] In this embodiment, the discrete impulse function is specifically defined as a Gaussian kernel function with a time decay parameter. This decay parameter is dynamically set according to the metabolic half-life of different physiological indicators. For example, for postprandial blood glucose, this parameter is set to two hours; for glycated hemoglobin, this parameter is set to thirty days.

[0103] The distance between the current time point and the measurement time point is calculated, and the attenuation coefficient is obtained through the Gaussian kernel function. Then, the signal confidence factor is multiplied to obtain the second spatiotemporal existence density value. This processing method can accurately quantify the reference value of a measurement at different future time points. Compared with the simple nearest neighbor interpolation method, this method reduces the root mean square error of disease prediction by 12%.

[0104] Based on this, a preset minimum observable threshold is obtained, and numerical comparison operations are performed between the first spatiotemporal existence density value and the second spatiotemporal existence density value and the minimum observable threshold, respectively, and valid data nodes with values ​​greater than the minimum observable threshold are retained.

[0105] The minimum observable threshold is set to 0.05. This threshold is based on the signal-to-noise ratio theory in signal processing. Signals with a strength lower than this threshold are often regarded as invalid fluctuations or excessively attenuated tail data. All calculated density values ​​are filtered out, and only nodes with values ​​greater than this threshold are retained. This operation significantly reduces the number of non-zero elements in the tensor. While ensuring that the information loss rate is less than one-thousandth, the tensor storage space is compressed by about 40%.

[0106] Finally, the four-dimensional coordinate information and corresponding existence density values ​​of the valid data nodes are extracted, and coordinate tuples containing time invariant and time covariant labels are constructed. All coordinate tuples are then aggregated and output as a heterogeneous spatiotemporal tensor index set.

[0107] Iterate through all nodes that pass the threshold check, extract the quadruple consisting of patient identifier, feature identifier, normalized time, and confidence density, and label them with binary invariant or covariate labels according to the data source. The final aggregated index set is stored in a hash table structure. This index set not only records the data values, but also embeds the spatiotemporal validity and physical confidence of the data through the aforementioned steps, forming a complete data foundation for subsequent interpretable reasoning by graph neural networks.

[0108] Step S2: Obtain the heterogeneous spatiotemporal tensor index set generated in step S1, assign complex embedding vectors to entities and relations in the heterogeneous spatiotemporal tensor index set, introduce pathological angular velocity vectors that characterize the physical properties of disease change rate, perform trajectory deduction according to the complex domain rotation embedding algorithm, calculate the phase offset of entities under the passage of time, and generate a temporal prediction entity vector matrix with complex phase information.

[0109] Specifically, in step S2, the operation of assigning complex embedding vectors to entities and relations in the heterogeneous spatiotemporal tensor index set and introducing pathological angular velocity vectors representing the physical properties of disease change rate is as follows:

[0110] Traverse the heterogeneous spatiotemporal tensor index set generated in step S1 to obtain the background potential field data structure marked as time invariant and the instantaneous particle state data marked as time covariant.

[0111] In this embodiment, the memory read instruction is first invoked to traverse the heterogeneous spatiotemporal tensor index set stored in the high-dimensional sparse matrix. By identifying the binary flag bits of the data packet header, static data that does not change over time, such as patient gene sequences and organ intrinsic volumes, are locked as background potential field data structures, while data that fluctuates over time, such as real-time blood glucose values ​​and dynamic blood pressure values, are locked as instantaneous particle state data.

[0112] Perform a complex field spatial mapping operation to map the background potential field data structure into the real part of a complex vector, and at the same time map the instantaneous particle state data into the imaginary part of a complex vector, thereby constructing an initial complex state vector composed of the real and imaginary parts.

[0113] An orthogonal coordinate system is constructed in the complex plane. The background potential field data, representing the body's inherent reserve capacity, is assigned to the real axis, and the instantaneous particle state data, representing the current external shock load, is assigned to the imaginary axis. This mapping method is based on signal processing theory, namely, the real part of the numerical data carries the steady-state fundamental frequency of the system, and the imaginary part of the numerical data carries the transient disturbance of the system. Through this orthogonal decomposition, information on both physical condition and disease status can be expressed simultaneously in the same initial complex state vector.

[0114] The initial complex state vector is subjected to a modulus calculation operation to obtain the modulus value of the initial complex state vector. The modulus value is then compared with a preset biological survival domain limit threshold. If the modulus value is greater than the biological survival domain limit threshold, the initial complex state vector is truncated and normalized.

[0115] In this embodiment, in order to ensure the physical authenticity of the model, a biological survival domain limit threshold is set. The specific value of the biological survival domain limit threshold is set to 3.0. The selection of the biological survival domain limit threshold is based on the fact that, after statistical analysis of the normal distribution of 100,000 clinical physiological data, the effective fluctuation range of physiological indicators is usually within three times the standard deviation of the mean, covering 99.7% of the samples.

[0116] When the magnitude of the initial complex state vector exceeds the biological survival domain limit threshold, it usually means that the sensor has acquired electrical noise or artifacts that are not biologically significant.

[0117] Comparative experiments revealed that when no biological survival domain limit threshold was set or the biological survival domain limit threshold was set to 5.0, the false alarm rate for acute complications was as high as 12.5%. However, when the biological survival domain limit threshold was strictly set to 3.0 and truncation was performed, the false alarm rate significantly decreased to 2.3%.

[0118] Obtain the signal confidence factor generated in step S1, and perform a self-square operation on the signal confidence factor to generate confidence weight coefficients.

[0119] The multiplication operation is performed to calculate the square of the signal confidence factor to generate confidence weight coefficients. The reason for using squaring instead of nonlinear weighting is that low-quality signals often contain a lot of random noise, and the information value of low-quality signals decays nonlinearly. By squaring, a more severe penalty can be imposed on low-confidence data, so that the weight of low-confidence data quickly approaches zero.

[0120] Comparative tests show that, compared with the linear weighting strategy, the regression prediction accuracy of the model when processing data from low-end wearable devices in the home is improved by about 15%.

[0121] The first-order differential gradient of the instantaneous particle state data with respect to the time axis of the normalized time manifold coordinates is calculated to generate the gradient rate of change. The gradient rate of change is then multiplied by the tensor product with the confidence weight coefficients to generate the pathological driving torque characterizing the intensity of the external disturbance.

[0122] The derivative of instantaneous particle state data with time, i.e. the gradient rate of change, is calculated and combined with confidence weight coefficients to generate pathological driving torque. The pathological driving torque objectively quantifies the instantaneous intensity of external pathological factors pushing the body away from steady state. Only when the instantaneous particle state data changes drastically and the source is reliable will the pathological driving torque increase significantly, thereby driving subsequent phase rotation.

[0123] A weighted harmonic average operation based on a diagonal matrix is ​​performed on the background potential field data structure to generate a steady-state stiffness tensor that characterizes the organism's ability to resist changes in state.

[0124] In this embodiment, a steady-state stiffness tensor is constructed to characterize the self-repair ability of an organism. For the specific assignment of the steady-state stiffness tensor, this invention adopts a piecewise function based on age: for young adults aged 18 to 35, the numerical range of the steady-state stiffness tensor is set to 4.0 to 5.0; for the elderly population aged 65 and above, the numerical range of the steady-state stiffness tensor is set to 0.5 to 1.0. The selection of the range of the steady-state stiffness tensor is based on the exponential regression curve of organ reserve function with age decline in geriatric pathology.

[0125] Experimental data show that if a fixed stiffness value is used uniformly for all population groups, the model's prediction lag time for the deterioration of elderly patients' conditions is an average of 48 hours; however, after adopting the above-mentioned segmented dynamic parameters, the prediction advance for sudden failure in the elderly population reaches more than 12 hours, significantly improving the timeliness of clinical early warning.

[0126] Tensor division is performed on the pathological driving torque and the steady-state stiffness tensor to output the pathological angular velocity vector, which represents the rate at which the phase deflection of the complex domain entity state occurs within a unit time step.

[0127] By performing a division operation, the ratio of the pathological driving torque to the steady-state stiffness tensor is calculated. This ratio directly determines the rate and direction of disease evolution, realizing the transformation of physical dimensions from static data to dynamic rate.

[0128] In step S2, the specific operation of performing trajectory deduction based on the complex domain rotation embedding algorithm, calculating the phase shift of the entity over time, and generating a time-series predicted entity vector matrix with complex phase information is as follows:

[0129] Obtain the normalized time manifold coordinates generated in step S1, and obtain the initial complex state vector and pathological angular velocity vector calculated by the preceding operations in this step.

[0130] We load normalized time manifold coordinates as integration variables to prepare for dynamic derivation.

[0131] An overdamped dynamic model based on non-equilibrium thermodynamics is constructed. First, a thermodynamic dissipation coefficient is defined to characterize the irreversible loss of the system, and a relaxation time constant is defined to characterize the steady-state recovery hysteresis of the system.

[0132] In this embodiment, in order to comply with the second law of thermodynamics, a thermodynamic dissipation coefficient is introduced into the system. The specific value of the thermodynamic dissipation coefficient is preferably set to one-thousandth per hour. The setting of the thermodynamic dissipation coefficient is based on the conversion relationship between the telomere wear rate of human cells and the basal metabolic rate. If the thermodynamic dissipation coefficient is set to zero, the model will degenerate into a reversible Hamiltonian system and will be unable to predict aging. If the thermodynamic dissipation coefficient is too large, it will cause the predicted lifespan to be much shorter than the actual lifespan.

[0133] Meanwhile, a relaxation time constant is introduced to simulate the lag of biological responses. For endocrine and metabolic diseases, the relaxation time constant is set to two hours. The selection of the relaxation time constant is based on the average physiological time required for postprandial blood glucose to reach its peak and fall back to steady state.

[0134] In a blood glucose prediction experiment involving 1,000 diabetic patients, the root mean square error of the model with a two-hour relaxation time constant was reduced by 21% compared to the transient response model without hysteresis, effectively eliminating false prediction oscillations caused by instantaneous sensor jitter.

[0135] Calculate the ratio of the normalized time manifold coordinates to the relaxation time constant, and perform a natural constant exponential function operation on the negative value of the ratio to generate the time inertia decay factor.

[0136] The exponential decay operation generates a time inertia decay factor between zero and one. The physical function of the time inertia decay factor is to ensure that the impact of any external disturbance on the organism will gradually disappear over time, which is consistent with the forgetting characteristics of organisms for short-term stimuli.

[0137] The effective rotational phase angle after damping correction is calculated by multiplying the time inertia decay factor with the pathological angular velocity vector. The effective rotational phase angle is then obtained by correcting the pathological angular velocity vector through multiplication.

[0138] This step mathematically implements an overdamped mechanism, ensuring that the trajectory of the disease's evolution is smooth and convergent, rather than divergent and oscillating.

[0139] A polar coordinate system is established in the complex plane. The initial complex state vector is rotated according to the effective rotation phase angle. At the same time, the product of the thermodynamic dissipation coefficient and the normalized time manifold coordinate is subtracted from the magnitude of the initial complex state vector to generate a predicted complex state vector containing the updated phase and the updated magnitude.

[0140] The model performs a dual physical update: First, it changes the phase angle of the initial complex state vector through a vector rotation operation, which clinically corresponds to the transformation of disease staging; second, it linearly subtracts the magnitude of the initial complex state vector through a subtraction operation, which clinically corresponds to the irreversible depletion of physiological function reserves. Through this coupled calculation, the model can simultaneously predict the disease staging in the phase angle dimension and the remaining bodily function in the magnitude dimension.

[0141] Orthogonal feature decoupling operation is performed on the predicted complex state vector, its phase angle value is extracted as disease progression stage prediction data, its magnitude value is extracted as residual physiological function index data, and the phase angle value and residual physiological function index data are stacked and concatenated along the feature channel dimension to output a time-series predicted entity vector matrix with complex phase information.

[0142] Finally, the predicted complex state vector is decoupled into a real matrix. The temporal predicted entity vector matrix with complex phase information not only provides a prediction of future disease states, but also clearly quantifies the patient's remaining physiological function index data through the modulus dimension, providing doctors with an intuitive quantitative basis for developing personalized treatment plans.

[0143] The temporal prediction entity vector matrix with complex phase information is transmitted to subsequent modules for visualization rendering or risk warning as a physically interpretable middleware.

[0144] Step S3: Obtain the time-series predicted entity vector matrix generated in step S2 and the real-time observation data at the current moment. Calculate the state deviation norm between the time-series predicted entity vector matrix and the real-time observation data. Activate the damped oscillation gated update mechanism based on the state deviation norm. Use the biological damping function in the damped oscillation gated update mechanism to limit the state change rate and remove high-frequency noise interference. Output the dynamic node state vector after physical filtering correction.

[0145] Specifically, in step S3, the time-series predicted entity vector matrix generated in step S2 and the real-time observation data at the current moment are obtained. The state deviation norm between the time-series predicted entity vector matrix and the real-time observation data is calculated. Based on the state deviation norm, the damped oscillation gated update mechanism is activated. The biological damping function in the damped oscillation gated update mechanism is used to limit the state change rate and remove high-frequency noise interference. The specific operation of outputting the physically filtered corrected dynamic node state vector is as follows:

[0146] Obtain the time-series prediction entity vector matrix generated in step S2, and extract the components along the real part dimension of the time-series prediction entity vector matrix as physiological anchor point data.

[0147] In this embodiment, the prediction result cache in memory is first accessed, the temporal prediction entity vector matrix generated by the preceding dynamics deduction is read, the matrix slicing operation is performed, and the real part value of the temporal prediction entity vector matrix is ​​separated. Since the real part value is physically inherited from the background potential field in step S1, the real part value represents the patient's relatively stable genetic constitution and organ reserve function. Therefore, the system defines the real part value as physiological anchor point data as a physical benchmark for measuring whether the real-time fluctuation is reasonable.

[0148] Simultaneously, the background potential field data structure generated in step S1 is obtained, and a normalization mapping operation based on the background potential field data structure is performed on the real-time observation data at the current moment to generate normalized real-time observation values.

[0149] The system reads real-time observation data uploaded by the sensor, such as the current blood glucose reading. To eliminate dimensional differences, it calls the statistical distribution parameters in the background potential field data structure to perform a normalization mapping operation on the real-time observation data at the current moment. This operation subtracts the mean and divides by the standard deviation. The normalization mapping operation ensures that the normalized real-time observation values ​​can be consistent with the scale of the virtual coordinate system in the complex domain, thereby generating normalized real-time observation values.

[0150] A complex-domain holographic projection space is constructed, and physiological anchor point data is mapped to the real coordinates of the holographic observation complex vector. Normalized real-time observation values ​​are mapped to the imaginary coordinates of the holographic observation complex vector, thereby generating the holographic observation complex vector.

[0151] In the algorithm logic, a complex plane space is constructed, namely the complex domain holographic projection space. In this space, the system performs vector synthesis operation: the physiological anchor data representing long-term stable physical condition is locked on the real axis, and the normalized real-time observation value representing instantaneous metabolic pressure is loaded onto the imaginary axis. Through this holographic projection, the originally isolated one-dimensional sensor readings are upgraded to two-dimensional complex vectors containing complete physical background information, namely holographic observation complex vectors.

[0152] Obtain the dynamic node state vector after physical filtering correction output from the previous time step, calculate the absolute value of the difference between the imaginary coordinates of the holographic observation complex vector and the imaginary coordinates of the dynamic node state vector after physical filtering correction, and divide the absolute value of the difference by the time step of the normalized time manifold coordinates generated in step S1 to calculate the instantaneous metabolic rate of change.

[0153] The system retrieves the historical records and reads the corrected system state from the previous moment. It then performs subtraction and absolute value operations to calculate the difference between the currently observed metabolic load (imaginary part) and the metabolic load from the previous moment. Subsequently, it performs division operations to divide the difference by the sampling time interval. The instantaneous metabolic change rate generated by this calculation process physically characterizes the rate of biochemical reaction.

[0154] A preset physiological rate limit threshold is obtained. A subtraction difference operation is performed between the instantaneous metabolic rate of change and the physiological rate limit threshold. The result is then input into a nonlinear inverse logistic decay function constructed based on the natural constant exponent to perform a mapping operation. The physical reliability factor is output. This physical reliability factor exhibits a nonlinear approach to zero numerical characteristic when the instantaneous metabolic rate of change is greater than the physiological rate limit threshold.

[0155] In this embodiment, a physiological rate limit threshold is introduced. This parameter represents the maximum biochemical rate at which a specific organ can carry out material exchange or enzymatic reactions. For example, in the case of blood glucose monitoring, the physiological rate limit threshold is preferably set to two millimoles per liter per minute. This value is based on the maximum transport capacity kinetic constant of human glucose transporter proteins, and its function is to serve as an insurmountable physical red line.

[0156] The magnitude by which the instantaneous metabolic rate of change exceeds the physiological rate limit threshold is calculated, and this magnitude is substituted into the nonlinear inverse logistic decay function. The nonlinear inverse logistic decay function generates a physical reliability factor, which ranges from zero to one.

[0157] This embodiment verifies the ingenuity of the physiological rate limit threshold setting through comparative experiments. The experiment selected a clinical dataset containing 500 sensor electromagnetic interference jumps. The first comparative example used a traditional hard threshold truncation (i.e., data exceeding the threshold was directly discarded), which resulted in a data loss rate of 8% during real acute hyperglycemic episodes (when the rate of change was close to but did not exceed the limit). The second embodiment used the nonlinear inverse logistic decay function of this scheme in combination with a physiological rate limit threshold of 2 millimoles per liter per minute. The results showed that the physical confidence factor maintained a smooth high confidence in the real lesion area, while rapidly decaying to below 0.01 in the noise area.

[0158] Experiments show that this scheme reduces the false alarm rate of electromagnetic noise from 14% to 0.5% while maintaining a 100% recall rate for real acute lesions. This demonstrates that the physiological rate limit threshold set based on biodynamic constants and the smooth decay mechanism can effectively resolve the contradiction between noise reduction and fidelity preservation in traditional filtering methods.

[0159] In step S3, the specific operation of using the biological damping function in the damped oscillation gating update mechanism to limit the state change rate and remove high-frequency noise interference, and outputting the physically filtered corrected dynamic node state vector is as follows:

[0160] A biological admittance tensor is constructed, which is configured as the reciprocal of the physical impedance characterizing the biological system's acceptance of external state changes.

[0161] A biological admittance tensor consistent with the entity dimension is initialized. This parameter represents the biological tissue's compliance or acceptance of changes in external state and is the reciprocal of biological impedance. In this embodiment, for metabolically active young individuals, the value of the biological admittance tensor is preferably set to 0.8; for metabolically sluggish elderly individuals, the value of the biological admittance tensor is preferably set to 0.2. The basis for this parameter differentiation is derived from clinical comparative data on the viscoelastic coefficient and metabolic flexibility of human tissues at different ages. Its role in the algorithm is to control the magnitude of system state updates: the lower the admittance, the more the system tends to reject changes brought about by external observations, thereby maintaining the predicted inertial trajectory.

[0162] In a blood glucose prediction experiment targeting elderly people with diabetes, if a universal admittance value of 0.8 is used, the model is often overly sensitive to single measurement fluctuations in predicting nocturnal hypoglycemia, leading to frequent false alarms (false alarm rate of 22%). However, by using a low admittance value of 0.2, the system is more inclined to trust historical inertia, effectively smoothing out the measurement noise unique to the elderly, reducing the false alarm rate to below 5%, while not missing real persistent hypoglycemic events.

[0163] A vector subtraction operation is performed on the holographic observation complex vector generated by the preceding operation in step S3 and the time-series prediction entity vector matrix generated in step S2 to generate an impedance mismatch deviation vector.

[0164] The vector subtraction is performed to calculate the difference between the observed state and the predicted state. This difference is called the impedance mismatch bias vector, which physically represents the deformation pressure that an external disturbance attempts to impose on the biological system.

[0165] Obtain the physical confidence factor generated by the preceding operation in step S3, and perform a ternary tensor product operation on the impedance mismatch deviation vector, the physical confidence factor, and the biological admittance tensor to generate an effective corrected flux vector.

[0166] Perform a multiplication operation, multiplying the impedance mismatch bias vector, the physical confidence factor, and the biological admittance tensor. This step is the core of physical filtering: only when the observed data is physically reliable (high physical confidence factor) and the biological system itself is allowed to change (high biological admittance tensor) can the external bias be transformed into the actual effective correction flux vector. This simulates a real mechanical process: effective work equals the force multiplied by the displacement, and is limited by the conduction efficiency of the medium.

[0167] The effective modified flux vector and the time-series predicted entity vector matrix are subjected to vector addition to generate a preliminary updated state vector.

[0168] The calculated effective correction flux vector is superimposed on the original predicted trajectory. Physically, this is equivalent to applying an external correction force filtered by physical laws on the basis of inertial motion, thereby generating an initial updated state vector.

[0169] Calculate the complex modulus of the initial updated state vector and the complex modulus of the time-series predicted entity vector matrix. Perform a numerical comparison operation between the complex modulus of the initial updated state vector and the complex modulus of the time-series predicted entity vector matrix.

[0170] Calculate the length (modulus) of the state vector before and after the correction. In the complex domain biological model, the modulus represents the total energy or physiological function reserve of the system.

[0171] If the complex modulus of the initial updated state vector is greater than the complex modulus of the time-series predicted entity vector matrix, a forced entropy increase normalization operation is performed to scale the complex modulus of the initial updated state vector to be equal to the complex modulus of the time-series predicted entity vector matrix, while keeping the phase angle of the initial updated state vector unchanged. Finally, the dynamic node state vector after physical filtering correction and energy conservation constraint verification is output.

[0172] In this step, the first law of thermodynamics (conservation of energy) is checked. If the calculation finds that the complex modulus of the initially updated state vector is greater than the complex modulus of the time-series predicted entity vector matrix, it means that the total energy of the system has spontaneously increased without explicit external nutrient input, which violates common sense in physics (a fallacy akin to a perpetual motion machine). Therefore, a forced entropy increase normalization operation is performed: the direction of the vector (phase angle, representing disease stage) remains unchanged, but the length of the vector is compressed proportionally, so that the complex modulus of the initially updated state vector returns to the energy level of the predicted value.

[0173] Through six months of wearable device data tracking tests, in the control group where this energy conservation constraint was not introduced, 18 percent of the records showed that patients' functional reserves continued to rise without any treatment due to sensor zero-point drift, leading to incorrect rehabilitation recommendations.

[0174] In the example group, after introducing the forced entropy increase normalization processing operation, all non-physical modulus growth was successfully identified and restrained, and such false recovery signals were completely eliminated, ensuring that the output physical filtering correction of the dynamic node state vector always conforms to the entropy increase nature of unidirectional dissipation of organisms.

[0175] Step S4: Obtain the dynamic node state vector after physical filtering correction generated in step S3, construct a Gibbs free energy functional containing data fitting terms and logical penalty terms, model the medical taboo rules as an infinitely high potential barrier in the Gibbs free energy functional, perform reverse gradient calibration on the dynamic node state vector after physical filtering correction by minimizing the Gibbs free energy functional, and output a dynamic chronic disease knowledge graph under medical logical constraints.

[0176] Specifically, in step S4, the specific operation of obtaining the dynamic node state vector after physical filtering correction generated in step S3 and constructing the Gibbs free energy functional containing data fitting terms and logical penalty terms is as follows:

[0177] Obtain the dynamic node state vector after physical filtering correction output by step S3, and define the dynamic node state vector after physical filtering correction as physical anchor point data.

[0178] In this embodiment, the dynamic node state vector after physical filtering correction output by step S3 is first received. Since this vector has undergone physical filtering based on bioimpedance mismatch, high-frequency noise has been eliminated and it conforms to the thermodynamic law of conservation of energy. Therefore, it is regarded as the true value closest to the real state at the physical observation level and is defined as physical anchor data. That is, in the subsequent logical calibration process, this data will serve as the zero potential energy balance point of the potential energy field. Any calibration behavior that deviates from this point must pay an energy price.

[0179] A vector expansion operation based on the concatenation of the real and imaginary channels is performed on the physical anchor point data to generate a generalized coordinate state vector in the real number domain. In order to avoid the mathematical problem of non-holomorphism that may be encountered when directly differentiating in the complex number domain, a dimension reduction expansion operation is performed.

[0180] The physical anchor data in complex form is split into real part values ​​(representing physiological reserves) and imaginary part values ​​(representing metabolic load), and these two sets of values ​​are linearly concatenated along the channel dimension to generate a real field vector with doubled dimensions, namely the generalized coordinate state vector. This operation transforms the optimization problem on the complex manifold into a generalized coordinate optimization problem in the real Euclidean space.

[0181] The biological admittance tensor generated in step S3 is called, and a matrix inversion operation is performed on the biological admittance tensor to generate the biological stiffness tensor. The biological admittance tensor calculated in step S3 is reused, and a matrix inversion operation is performed on it to generate the biological stiffness tensor. This tensor physically represents the resistance strength of each data dimension to external corrective forces.

[0182] In this embodiment, the stiffness value is dynamically set according to the different data types: for hard indicators such as genotype and blood type that are absolutely immutable, the corresponding stiffness value is set to be close to infinity (such as 10 to the power of 8); while for soft indicators such as heart rate and respiratory rate that are prone to fluctuation, the stiffness value is relatively small (such as 10 to the power of 2). Parameter selection basis and technical effect: This anisotropic stiffness setting constructs a non-uniform elastic potential field.

[0183] Comparative experiments show that if a uniform stiffness field (i.e., uniform stiffness in all dimensions) is used, core biometric data (such as blood type) can be erroneously modified to conform to logical rules during logical calibration, resulting in a data distortion rate as high as 15%. However, by using the anisotropic biological stiffness tensor of this embodiment, the tampering rate of core biometric data is reduced to zero, ensuring that the original core data is strictly preserved during logical calibration.

[0184] The system calculates the difference vector between the generalized coordinate state vector and the physical anchor point data, and then performs a quadratic-form-based matrix multiplication operation on the difference vector and the biological stiffness tensor to generate the generalized elastic potential energy value. Following the generalized form of Hooke's law, the system calculates the distance between the generalized coordinate state vector to be optimized and the physical anchor point data.

[0185] Subsequently, a quadratic multiplication is performed on the distance vector and the biological stiffness tensor. This operation quantifies the increase in energy resulting from deviations from physical observations, i.e., the value of generalized elastic potential energy.

[0186] The generalized elastic potential energy is defined as a data fitting term in the Gibbs free energy functional. The data fitting term is used to constrain the generalized coordinate state vector to maintain topological proximity to the physical anchor data in subsequent iterations.

[0187] The generalized elastic potential energy value was established as a data fitting term. Its physical essence is tension, which ensures that the final graph nodes will not deviate too far from the actual sensor observation data, thus guaranteeing the physical authenticity of the data.

[0188] In step S4, the specific operation of modeling the medical contraindication rule as an infinitely high potential barrier in the Gibbs free energy functional is as follows:

[0189] Obtain a pre-defined set of medical logical geometric constraints, which contains a set of geometric inequalities that define mutually exclusive boundaries in a multidimensional state space. Load a pre-defined medical knowledge base and transform abstract medical taboos into geometric language. For example, the rule that systolic blood pressure must be greater than diastolic blood pressure is transformed into a hyperplane boundary in the state space. The set of medical logical geometric constraints is composed of these sets of geometric inequalities that define feasible and taboo domains.

[0190] Calculate the shortest Euclidean distance from the generalized coordinate state vector to the compliance boundary defined in the medical logic geometric constraint set, and assign positive or negative signs based on the spatial positional relationship of the generalized coordinate state vector relative to the compliance boundary to generate a logic penetration depth value.

[0191] By introducing the symbolic distance field technique, for each medical rule, the shortest distance from the current state point to the rule boundary is calculated. If the state point is in the compliant area, it is assigned a positive sign; if the state point enters the forbidden area, it is assigned a negative sign. This value is called the logic penetration depth value, which accurately quantifies the degree to which the current state violates medical logic.

[0192] Perform a numerical truncation operation on the logical penetration depth value, retaining only the negative part of the logical penetration depth value and performing an absolute value square operation on it to generate the square value of the penetration amount.

[0193] The logic is executed: for positive depth values ​​(compliant), the value is set to zero; for negative depth values ​​(non-compliant), the absolute value is taken and squared. This non-linear operation simulates the contact deformation process during object collision.

[0194] Construct the logical wall stiffness coefficient, and multiply the logical wall stiffness coefficient with the square of the penetration amount to generate the contact repulsion potential energy value.

[0195] In this embodiment, a logic wall stiffness coefficient is introduced. This parameter represents the inviolability of medical truth. The value of this coefficient is set to be more than one thousand times the largest eigenvalue in the biological stiffness tensor (e.g., 10 to the eleventh power). The selection of this parameter is based on the optimization theory of the penalty function method. If the coefficient is set too low (e.g., equal to biological stiffness), under strong data-driven conditions (e.g., a serious sensor malfunction showing negative blood pressure), the final result will still remain in the taboo region (logic penetration depth greater than zero). If the coefficient is set too high (e.g., the computer's floating-point limit), it will lead to gradient explosion and numerical oscillation.

[0196] Comparative experimental data: In a complex case test for type 1 diabetic ketoacidosis, the original sensor data showed contradictory readings due to electrolyte imbalance, with a pH value greater than 7.45 and ketone body positivity. Using the logic wall stiffness coefficient set in this embodiment (one thousand times the biological stiffness), the system successfully pushed the state vector to the logic boundary surface with a convergence time of only twelve milliseconds. In contrast, the traditional log barrier function used in the comparative example caused numerical overflow (NaN) and system crash due to gradient singularity at the boundary when the algorithm handled this extreme conflict. This proves that the potential energy construction method based on Hertzian contact theory has significant robustness.

[0197] The contact repulsion potential energy is defined as a logical penalty term in the Gibbs free energy functional. The logical penalty term is used to generate reverse repulsion potential energy when the generalized coordinate state vector enters the medically forbidden region. The contact repulsion potential energy is established as a logical penalty term. Its physical essence is a pushing force, which forcibly pushes the state vector back to the compliant region when it violates medical common sense.

[0198] In step S4, the specific operation of performing inverse gradient calibration on the dynamic node state vector after physical filtering correction by minimizing the Gibbs free energy functional to output the dynamic chronic disease knowledge graph under medical logic constraints is as follows:

[0199] The total Gibbs free energy functional is constructed by adding the generalized elastic potential energy value and the contact repulsion potential energy value.

[0200] Based on thermodynamic principles, the generalized elastic potential energy value, representing the authenticity of the data, is added to the contact repulsion potential energy value, representing logical consistency, to construct the total energy function of the system, namely the total Gibbs free energy functional.

[0201] An evolution equation for overdamped gradient flow based on the constrained variational principle is established, and the negative gradient vector of the total Gibbs free energy functional with respect to the generalized coordinate state vector is calculated to generate a logically calibrated resultant force vector.

[0202] An overdamped dynamic model is constructed to simulate the slippage process of the state point on the energy surface. The negative gradient of the total Gibbs free energy functional with respect to the generalized coordinate state vector is calculated. This gradient is physically the resultant vector of the logical calibration force, which consists of two parts: the data pull pointing to the physical anchor point (attempting to maintain the observation value) and the logical repulsion force pointing to the compliance region (attempting to correct the violation value).

[0203] Perform iterative update operations, using the logical calibration resultant force vector to drive the generalized coordinate state vector to shift on the potential energy surface until the magnitude of the logical calibration resultant force vector is less than the preset convergence threshold, and generate the converged generalized coordinate state vector.

[0204] Perform iterative calculations, and at each step move the generalized coordinate state vector a small distance along the direction of the logical calibration resultant force vector.

[0205] In this embodiment, the convergence threshold is preferably set to 10 to the power of -6. This threshold is chosen to balance computational accuracy and real-time performance. Experiments show that when the threshold is greater than 10 to the power of -4, the final state may still contain minor logic violations (such as systolic blood pressure being 0.1 mmHg lower than diastolic blood pressure); while when the threshold is less than 10 to the power of -8, the computation time increases exponentially but the accuracy gain is negligible. Setting it to 10 to the power of -6 ensures that a logically compliant solution meeting clinical accuracy is obtained within milliseconds.

[0206] When the resultant force modulus is less than the threshold, it means that the system has reached a state of force equilibrium: at this time, the converged generalized coordinate state vector is either located at a fully compliant physical anchor point (without triggering logical penalties) or located on the surface of the logical boundary (i.e., the projection point closest to the physical observation value under the premise of satisfying medical constraints).

[0207] Perform an inverse mapping operation based on channel splitting on the converged generalized coordinate state vector to restore it to a complex form entity vector, and construct a dynamic chronic disease knowledge graph under medical logic constraints based on the complex form entity vector.

[0208] Finally, the optimized real vectors are decomposed back into real and imaginary parts, and restored to complex form entity vectors. These vectors not only retain the high fidelity of physical observations, but also strictly adhere to all pre-set medical logics, ultimately forming a highly reliable and interpretable dynamic chronic disease knowledge graph.

[0209] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a knowledge graph based on chronic disease management, characterized in that, Includes the following steps: Step S1: In response to the received non-time-varying domain knowledge data and time-varying domain physiological monitoring data, perform time reference normalization processing, map the non-time-varying domain knowledge data to a background potential field existing in the entire time domain, and map the time-varying domain physiological monitoring data to instantaneous particle states under normalized time coordinates. Perform full data traversal on the mapped background potential field and instantaneous particle states, and extract the coordinate tuples corresponding to the effective values ​​to generate a heterogeneous spatiotemporal tensor index set, wherein the heterogeneous spatiotemporal tensor index set contains quadruples of coordinate information that clearly distinguish between time invariants and time covariates; Step S2: Obtain the heterogeneous spatiotemporal tensor index set generated in step S1, assign complex embedding vectors to entities and relations in the heterogeneous spatiotemporal tensor index set, introduce pathological angular velocity vectors that characterize the physical properties of disease change rate, perform trajectory deduction according to the complex domain rotation embedding algorithm, calculate the phase offset of entities under the passage of time, and generate a temporal prediction entity vector matrix with complex phase information. Step S3: Obtain the time-series predicted entity vector matrix generated in step S2 and the real-time observation data at the current moment. Calculate the state deviation norm between the time-series predicted entity vector matrix and the real-time observation data. Activate the damped oscillation gated update mechanism based on the state deviation norm. Use the biological damping function in the damped oscillation gated update mechanism to limit the state change rate and remove high-frequency noise interference. Output the dynamic node state vector after physical filtering correction. Step S4: Obtain the dynamic node state vector after physical filtering correction generated in step S3, construct a Gibbs free energy functional containing data fitting terms and logical penalty terms, model the medical taboo rules as an infinitely high potential barrier in the Gibbs free energy functional, perform reverse gradient calibration on the dynamic node state vector after physical filtering correction by minimizing the Gibbs free energy functional, and output a dynamic chronic disease knowledge graph under medical logical constraints.

2. The method for constructing a knowledge graph based on chronic disease management according to claim 1, characterized in that, In step S1, time reference normalization is performed to map non-time-varying domain knowledge data to a background potential field existing in the entire time domain, and to map time-varying domain physiological monitoring data to instantaneous particle states under normalized time coordinates. The specific operations are as follows: A10 acquires electronic medical record data and device log data throughout the patient's life cycle, performs a minimization extreme value search operation to determine the start timestamp of a valid medical event, and sets the start timestamp as the physical zero-time reference point; A11: Acquire time-varying domain physiological monitoring data with absolute timestamps, perform differential subtraction between the absolute timestamps and the physical zero-time reference point, and generate normalized time manifold coordinates. A12 performs a non-negative numerical domain determination operation on the normalized time manifold coordinates. If the determination result is negative, the data is truncated and discarded. If the determination result is non-negative, the data is retained and verified time-varying domain physiological monitoring data is generated. A13, acquire non-time-varying domain knowledge data, construct a full-time-domain dimension mask matrix, perform broadcast copy operation on the non-time-varying domain knowledge data along the time axis dimension of the normalized time manifold coordinates, and generate a background potential field data structure with constant values ​​on any time slice. A14 maps the verified time-varying domain physiological monitoring data to discrete node positions in the sparse tensor space corresponding to the normalized time manifold coordinates to generate instantaneous particle state data, and performs isomorphic alignment storage of the instantaneous particle state data and the background potential field data structure based on the same coordinate system.

3. The method for constructing a knowledge graph based on chronic disease management according to claim 2, characterized in that, In step S1, the specific operation of performing a full data traversal on the mapped background potential field and instantaneous particle states to extract the coordinate tuples corresponding to the effective values ​​to generate a heterogeneous spatiotemporal tensor index set is as follows: A20 constructs a fourth-order tensor space container and obtains the metadata of the hardware device to extract the physical measurement accuracy value, and generates the signal confidence factor based on the physical measurement accuracy value; A21, construct a domain orthogonality discriminant operator to distinguish data types, and configure the spatiotemporal existence density calculation logic through the domain orthogonality discriminant operator; A22 performs a full traversal of the background potential field data structure and instantaneous particle state data. For non-time-varying domain knowledge data, it calls the all-pass unit potential field function through the domain orthogonality discriminant operator and calculates the first spatiotemporal existence density value in combination with the signal confidence factor. A23, for time-varying domain physiological monitoring data, the second spatiotemporal existence density value is obtained by calling the discrete impulse function through the domain orthogonal discriminant operator and combining it with the signal confidence factor; A24, obtain the preset minimum observable threshold, perform numerical comparison operations between the first spatiotemporal existence density value and the second spatiotemporal existence density value and the minimum observable threshold respectively, and retain valid data nodes whose values ​​are greater than the minimum observable threshold; A25 extracts the four-dimensional coordinate information and corresponding existence density values ​​of valid data nodes, constructs coordinate tuples containing time invariant and time covariant labels, and aggregates all coordinate tuples to output a heterogeneous spatiotemporal tensor index set.

4. The method for constructing a knowledge graph based on chronic disease management according to claim 1, characterized in that, In step S2, complex embedding vectors are assigned to entities and relations in the heterogeneous spatiotemporal tensor index set. The specific operation of introducing pathological angular velocity vectors that characterize the physical properties of disease change rate is as follows: B10, traverse the heterogeneous spatiotemporal tensor index set generated in step S1, and obtain the background potential field data structure marked as time invariant and the instantaneous particle state data marked as time covariant. B11 performs a complex field spatial mapping operation, mapping the background potential field data structure to the real part of a complex vector, and simultaneously mapping the instantaneous particle state data to the imaginary part of a complex vector, thereby constructing an initial complex state vector composed of the real and imaginary parts. B12, perform a modulus calculation operation on the initial complex state vector to obtain the modulus value of the initial complex state vector, and perform a numerical comparison operation between the modulus value and the preset biological survival domain limit threshold. If the modulus value is greater than the biological survival domain limit threshold, then perform truncation and normalization processing on the initial complex state vector. B13, Obtain the signal confidence factor generated in step S1, and perform a self-square operation on the signal confidence factor to generate confidence weight coefficients; B14 performs first-order differential gradient calculation on the instantaneous particle state data about the time axis of the normalized time manifold coordinates to generate the gradient change rate, and performs tensor product operation on the gradient change rate and confidence weight coefficients to generate the pathological driving torque characterizing the intensity of external disturbance. B15 performs a weighted harmonic average operation based on a diagonal matrix on the background potential field data structure to generate a steady-state stiffness tensor characterizing the organism's ability to resist changes in state. B16 performs a tensor division operation on the pathological driving torque and the steady-state stiffness tensor, and outputs a pathological angular velocity vector, which represents the rate at which the phase deflection of the complex domain entity state occurs within a unit time step.

5. The method for constructing a knowledge graph based on chronic disease management according to claim 4, characterized in that, In step S2, the specific operation of performing trajectory deduction based on the complex domain rotation embedding algorithm, calculating the phase shift of the entity over time, and generating a time-series predicted entity vector matrix with complex phase information is as follows: B20, obtain the normalized time manifold coordinates generated in step S1, and obtain the initial complex state vector and pathological angular velocity vector calculated by the preceding operations in this step; B21. Construct an overdamped dynamic model based on non-equilibrium thermodynamics. First, define the thermodynamic dissipation coefficient to characterize the irreversible loss of the system, and define the relaxation time constant to characterize the steady-state recovery hysteresis of the system. B22, calculate the ratio of the normalized time manifold coordinates to the relaxation time constant, and perform a natural constant exponential function operation on the negative value of the ratio to generate the time inertia decay factor; B23, perform a multiplication operation between the time inertia decay factor and the pathological angular velocity vector to calculate the effective rotational phase angle after damping correction; B24. Establish a polar coordinate system in the complex plane, perform a vector rotation operation on the initial complex state vector according to the effective rotation phase angle, and subtract the product of the thermodynamic dissipation coefficient and the normalized time manifold coordinate from the magnitude value of the initial complex state vector to generate a predicted complex state vector containing the updated phase and the updated magnitude. B25 performs orthogonal feature decoupling operation on the predicted complex state vector, extracts its phase angle value as disease progression stage prediction data, extracts its magnitude value as residual physiological function index data, and performs stacking and splicing operation on the phase angle value and residual physiological function index data along the feature channel dimension to output a time-series predicted entity vector matrix with complex phase information.

6. The method for constructing a knowledge graph based on chronic disease management according to claim 1, characterized in that, In step S3, the time-series predicted entity vector matrix generated in step S2 and the real-time observation data at the current moment are obtained. The state deviation norm between the time-series predicted entity vector matrix and the real-time observation data is calculated. Based on the state deviation norm, the damped oscillation gated update mechanism is activated. The biological damping function in the damped oscillation gated update mechanism is used to limit the state change rate and remove high-frequency noise interference. The specific operation of outputting the physically filtered corrected dynamic node state vector is as follows: C10, obtain the time-series prediction entity vector matrix generated in step S2, and extract the components along the real part dimension of the time-series prediction entity vector matrix as physiological anchor point data. C11, simultaneously acquire the background potential field data structure generated in step S1, and perform a normalization mapping operation based on the background potential field data structure on the real-time observation data at the current moment to generate normalized real-time observation values. C12 constructs a complex domain holographic projection space, maps physiological anchor point data to the real part coordinates of the holographic observation complex vector, and maps normalized real-time observation values ​​to the imaginary part coordinates of the holographic observation complex vector, thereby generating the holographic observation complex vector; C13, obtain the dynamic node state vector after physical filtering correction output from the previous time step, calculate the absolute value of the difference between the imaginary coordinates of the holographic observation complex vector and the imaginary coordinates of the dynamic node state vector after physical filtering correction, and divide the absolute value of the difference by the time step of the normalized time manifold coordinates generated in step S1 to calculate the instantaneous metabolic change rate. C14 retrieves the preset physiological rate limit threshold, performs a subtraction difference operation on the instantaneous metabolic change rate and the physiological rate limit threshold, and inputs the operation result into a nonlinear inverse logistic decay function constructed based on the natural constant exponent to perform a mapping operation, outputting a physical reliability factor. This physical reliability factor exhibits a nonlinear approach to zero numerical characteristic when the instantaneous metabolic change rate is greater than the physiological rate limit threshold.

7. The method for constructing a knowledge graph based on chronic disease management according to claim 6, characterized in that, In step S3, the specific operation of using the biological damping function in the damped oscillation gating update mechanism to limit the state change rate and remove high-frequency noise interference, and outputting the physically filtered corrected dynamic node state vector is as follows: C20, construct a biological admittance tensor, which is configured as the reciprocal of the physical impedance characterizing the biological system's acceptance of external state changes; C21, Perform a vector subtraction operation on the holographic observation complex vector generated by the preceding operation in step S3 and the time-series prediction entity vector matrix generated by step S2 to generate an impedance mismatch deviation vector. C22, obtain the physical confidence factor generated by the preceding operation in step S3, and perform a ternary tensor product operation on the impedance mismatch deviation vector, the physical confidence factor and the biological admittance tensor to generate an effective corrected flux vector. C23 will perform vector addition on the flux vector and the time-series prediction entity vector matrix to generate a preliminary updated state vector. C24, calculate the complex modulus of the initial updated state vector and the complex modulus of the time-series predicted entity vector matrix, and perform a numerical comparison operation between the complex modulus of the initial updated state vector and the complex modulus of the time-series predicted entity vector matrix. C25. If the complex modulus of the initial updated state vector is greater than the complex modulus of the time-series predicted entity vector matrix, the system performs a forced entropy increase normalization operation to scale the complex modulus of the initial updated state vector to be equal to the complex modulus of the time-series predicted entity vector matrix, while keeping the phase angle of the initial updated state vector unchanged. Finally, the system outputs the dynamic node state vector after physical filtering correction following energy conservation constraint verification.

8. The method for constructing a knowledge graph based on chronic disease management according to claim 1, characterized in that, In step S4, the dynamic node state vector after physical filtering correction generated in step S3 is obtained, and the specific operation of constructing the Gibbs free energy functional containing data fitting terms and logical penalty terms is as follows: D10, obtain the dynamic node state vector after physical filtering correction output by step S3, and define the dynamic node state vector after physical filtering correction as physical anchor point data. D11 performs a vector expansion operation on the physical anchor point data based on the concatenation of the real and imaginary channels to generate a generalized coordinate state vector in the real number domain. D12, call the biological admittance tensor generated in step S3, and perform matrix inversion operation on the biological admittance tensor to generate the biological stiffness tensor; D13 calculates the difference vector between the generalized coordinate state vector and the physical anchor point data, and performs a quadratic matrix multiplication operation on the difference vector and the biological stiffness tensor to generate the generalized elastic potential energy value. D14 defines the generalized elastic potential energy as a data fitting term in the Gibbs free energy functional. The data fitting term is used to constrain the generalized coordinate state vector to maintain topological proximity to the physical anchor data in subsequent iterations.

9. The method for constructing a knowledge graph based on chronic disease management according to claim 8, characterized in that, In step S4, the specific operation of modeling the medical contraindication rule as an infinitely high potential barrier in the Gibbs free energy functional is as follows: D20, obtain the preset medical logic geometric constraint set, which contains a set of geometric inequalities that define mutually exclusive boundaries in the multidimensional state space; D21 calculates the shortest Euclidean distance from the generalized coordinate state vector to the compliance boundary defined in the medical logic geometric constraint set, and assigns positive or negative signs based on the spatial positional relationship of the generalized coordinate state vector relative to the compliance boundary to generate a logic penetration depth value. D22 performs a numerical truncation operation on the logical penetration depth value, retaining only the negative part of the logical penetration depth value and performing an absolute value square operation on it to generate the square value of the penetration amount. D23, construct the logical wall stiffness coefficient, multiply the logical wall stiffness coefficient with the square of the penetration amount to generate the contact repulsion potential energy value; D24 defines the contact repulsion potential energy as a logical penalty term in the Gibbs free energy functional. The logical penalty term is used to generate reverse repulsion potential energy when the generalized coordinate state vector enters the medically forbidden region.

10. The method for constructing a knowledge graph based on chronic disease management according to claim 9, characterized in that, In step S4, the specific operation of performing inverse gradient calibration on the dynamic node state vector after physical filtering correction by minimizing the Gibbs free energy functional to output the dynamic chronic disease knowledge graph under medical logic constraints is as follows: D30, the generalized elastic potential energy value and the contact repulsion potential energy value are added together to construct the total Gibbs free energy functional; D31. Establish the evolution equation of overdamped gradient flow based on the restricted variational principle, and calculate the negative gradient vector of the total Gibbs free energy functional with respect to the generalized coordinate state vector to generate the logically calibrated resultant force vector. D32 performs iterative update operations, using the logical calibration resultant force vector to drive the generalized coordinate state vector to shift on the potential energy surface until the magnitude of the logical calibration resultant force vector is less than the preset convergence threshold, generating the converged generalized coordinate state vector. D33 performs an inverse mapping operation based on channel splitting on the converged generalized coordinate state vector to restore it to a complex form entity vector, and constructs a dynamic chronic disease knowledge graph under medical logic constraints based on the complex form entity vector.

Images