Knowledge-guided and intervention-response driven vascular inflammation prediction method

Abstract

The present application relates to a kind of vascular inflammation prediction method based on knowledge guide and intervention-response drive, belong to artificial intelligence and medical cross technical field.It includes the following steps: acquisition multivariate physiological measurement sequence, completes missing value preprocessing using treatment-aware time series linear interpolation;Physiological and treatment sequence are explicitly decoupled into short-term acute path and long-term cumulative path, respectively through exponential decay model, low-pass filter modeling and are fused by gate mechanism, obtain treatment dynamic representation;Heterogeneous dynamic graph is constructed by fusing medical priori and data learning, executes graph convolution update node representation after treatment modulation adjacency matrix;Finally, the representation of treatment dynamics and graph enhancement is fused, and the TBR prediction value is output by attention weighting and time point specific regression.The present application can effectively separate acute and chronic treatment effects, inject clinical knowledge constraints, and is suitable for tumor survivor cardiovascular risk stratification, vascular toxicity mechanism analysis and personalized diagnosis and treatment management.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and medicine, specifically relating to a method for predicting vascular inflammation based on knowledge guidance and intervention-response driven approaches. Background Technology

[0002] Cardiovascular disease (CVD) is a leading cause of death worldwide, primarily driven by atherosclerosis. Vascular inflammation precedes plaque formation and can be quantified using the target-to-background ratio (TBR) of 18F-FDG PET / CT as a validating biomarker. In cancer patients, while chemotherapy, radiotherapy, and immunotherapy significantly improve survival, they may induce treatment-related vascular inflammation with a complex and heterogeneous temporal trajectory. With the increasing number of cancer survivors, accurately predicting post-treatment TBR levels is crucial for cardiovascular risk stratification and understanding the mechanisms of treatment-related vascular toxicity.

[0003] Traditional time-series prediction methods are mostly based on statistical regression or simple recurrent networks, assessing vascular status through linear transformations of physiological indicator sequences. These methods have significant limitations when dealing with irregular and sparse longitudinal data, easily leading to effects confounding, insufficient generalization, or physiological inconsistencies. Although deep learning models such as LSTM, GRU, and Transformer have made progress in time-series prediction, conventional models often treat the intervention process as a black box, lacking causal separation and exhibiting insufficient cross-patient generalization ability. Existing research based on graph neural network (GNN) frameworks, while attempting to approximate physiological dependence to improve robustness, still falls short in decoupling acute and chronic pathways, multimodal fusion, and efficient injection of clinical priors. Therefore, a novel TBR prediction method is urgently needed that combines the multi-scale effect separation capabilities of deep time-series networks with knowledge-guided dynamic graph physiological causal modeling. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a knowledge-guided and intervention-response driven method for predicting vascular inflammation, comprising the following steps: S1. Collect multivariate physiological measurement sequences and perform preprocessing to obtain preprocessed multivariate physiological measurement sequences; S2. The preprocessed multivariate physiological measurement sequences and treatment sequences are explicitly decomposed into short-term acute pathway (SAP) and long-term cumulative pathway (LCP). For short-term acute pathway (SAP), the acute effect term is calculated using a physiologically inspired exponential decay model. Use a gated loop unit (GRU) to update the hidden state. For long-term cumulative pathway (LCP), treatment sequences are categorized into drug-based and non-drug-based classes, and their cumulative contributions are calculated separately to obtain the total cumulative contribution. Update the hidden state using a low-pass filter. By fusing the two pathways through a gating mechanism, a decoupled dynamic representation of treatment is obtained. ; S3. Constructing a heterogeneous dynamic graph It includes physiological nodes and treatment nodes, and the adjacency matrix integrates the prior and learned parts of medical knowledge; the modulation adjacency matrix is ​​obtained by applying treatment modulation; graph convolution is performed on the physiological subgraph to update the node representation; S4. The decoupled treatment dynamic representation and the node representation updated by graph convolution are fused to obtain the graph-enhanced representation of each physiological node; the graph-enhanced representation of each physiological node is weighted through an attention mechanism to obtain the patient-level representation; time-specific regression is used to map the patient-level representation to the target TBR value to obtain the TBR prediction value. .

[0005] Further, in step S1, the multivariate physiological measurement sequence includes vital signs, blood biochemical indicators, inflammatory indicators, and metabolic indicators; the vital signs include systolic blood pressure, diastolic blood pressure, and heart rate; the blood biochemical indicators include total cholesterol, low-density lipoprotein, high-density lipoprotein, and triglycerides; the inflammatory indicators include C-reactive protein and white blood cell count; and the metabolic indicators include fasting blood glucose and glycated hemoglobin.

[0006] Furthermore, in step S1, missing values ​​are supplemented by introducing a temporal linear interpolation method that incorporates drug and treatment-specific adjustment terms, as expressed in the following formula: , in, , These represent the nearest observation time points before and after the time point corresponding to the missing value, respectively; , These represent the variables in a multivariate physiological measurement sequence. At the observation time point , The corresponding observations; Representing variables The reconstructed value; Indicates time The collection of medications that the patient is currently using; Indicates a specific drug; This represents the set of non-pharmacological treatments the patient is receiving at time t; Indicates specific treatment; Indicates drug-dependent variables Clinical impact weight; Indicates the effect of treatment on variables Clinical impact weight; To ensure physiological rationality, all reconstructed values ​​are constrained within the therapeutic perception reference range, as expressed by the following formula: , in, Representing variables The constraint value; Represents the clipping function; , Representing variables respectively The lower and upper limits of clinical acceptance.

[0007] Furthermore, in step S2, the SAP modeling process for short-term acute pathways is as follows: set up Indicates the patient At time step The intensity of the most recent treatment event is obtained by normalizing the treatment dose, as shown in the following formula: , in, Indicates the actual dosage. This indicates the standard dose recommended by clinical guidelines; Acute effect item The formula is expressed as: , , in, Indicates the current follow-up time point. Indicates the time of the most recent treatment. Indicates time difference, Indicates the rate of decay of acute effects. Represents a nonlinear mapping function; The hidden state of the short-term path is updated using a gated recurrent unit, as expressed by the following formula: , in, Indicates the time of short-term acute pathways The hidden state vector is used to capture the immediate dynamic changes triggered by the recent treatment event, with initial values... Let it be the zero vector; Indicates the patient At follow-up time points Treatment input information; Indicates from Physiological measurement subsequences extracted from it.

[0008] Furthermore, in step S2, the long-term cumulative path (LCP) modeling process is as follows: Treatment events are categorized into two types: pharmacological treatment and non-pharmacological treatment; for the first The formula for the cumulative contribution of each drug treatment event is as follows: , in, Indicates the first The cumulative contribution of each drug treatment; Indicates the first Time of onset of drug treatment; Indicates the standardized dose threshold; This represents the learnable cumulative decay rate, used to characterize the process by which the cumulative effect of drug treatment gradually decays over time. Indicates the first The time of the drug treatment event The cumulative contribution; This is a dose saturation function; For non-pharmacological treatment events, the formula for cumulative contribution is as follows: , in, Indicates the first The cumulative contribution of non-pharmacological treatments; The baseline strength parameter indicating the impact of non-pharmacological treatments on physiological indicators; Indicates the time decay rate of the effect of non-pharmacological treatment; The total cumulative effect is calculated based on the cumulative contribution of drug therapy and the cumulative contribution of non-drug therapy, as expressed by the following formula: , in, Indicates time The total cumulative effect of all historical treatment events at any given moment is used to characterize the combined impact of drug therapy and non-drug therapy on the patient's physiological state at the current point in time. Represents a set of drug treatment events; This represents a set of non-pharmacological treatment events; The long-term path hidden state is updated using a low-pass filtering mechanism, as expressed by the following formula: , in, Indicates the long-term cumulative path over time The hidden state vector; Represents the smoothing factor. This represents a multilayer perceptron.

[0009] Furthermore, in step S2, the short-term path hidden state and the long-term path hidden state are fused through a gating mechanism to obtain the decoupled treatment dynamic representation, as shown in the following formula: , , in, This represents the gating vector, used to dynamically determine the fusion weights of short-term acute information and long-term cumulative information at the current time point; Represents the Sigmoid function; Represents the gate weight matrix; This represents the dynamic representation of treatment after decoupling; This indicates element-wise multiplication.

[0010] Furthermore, in step S3, at each follow-up time Constructing heterogeneous dynamic graphs : , in, Represents a set of physiological nodes; Represents the set of treatment nodes; Indicates time Directed interaction relationships between nodes; adjacency matrix It is broken down into a prior knowledge component and a learnable component, as expressed by the following formula: , in, This represents a fixed prior matrix pre-constructed based on clinical medical knowledge; Represents a learnable matrix; The effects of treatment on physiological indicators are dynamically modulated by adjusting the edge weights between physiological nodes in the adjacency matrix in the following way: According to the current time Decoupling of treatment dynamics For each treatment node Calculate modulation intensity ; Based on modulation strength The original edge weights between physiological nodes are multiplicatively modulated, and the specific formula is as follows: , in, This represents the adjacency matrix after treatment modulation; Indicates time From nodes in a time-heterogeneous dynamic graph To the node The right of the border; Represents physiological nodes; Indicates treatment nodes; For each treatment Learnable scalar coefficients; This represents the modulation function generated based on the decoupled treatment dynamics representation; Indicates an indicator function; Graph convolution is performed on the physiological subgraph using modulated adjacency to obtain the node representation updated by graph convolution. .

[0011] Furthermore, the decoupled treatment dynamic representation and the node representation updated by graph convolution are concatenated and linearly mapped to obtain the graph-enhanced representation of each physiological node; the graph-enhanced representation of each physiological node is weighted through an attention mechanism to obtain the patient-level representation. Patient levels were represented using time-specific regression. Mapping to the target TBR value, we obtain the patient's... time TBR prediction value The formula is expressed as follows: , in, (.) indicates a time-specific regression function; Represents the learnable weight matrix for a specific regression at a given time point; This represents the bias vector for a specific regression at a given time point.

[0012] The advantages of this invention are: This invention integrates a knowledge-guided preprocessing mechanism with a dual-path temporal causal decoupling strategy. First, it accurately reconstructs irregular longitudinal physiological trajectories through clinical knowledge-modulated missing value imputation. Then, it separates treatment dynamics using short-term acute pathways (SAP) and long-term cumulative pathways (LCP), and combines a dynamic graph convolution module to constrain the heterogeneous propagation of physiological dependence. This overcomes the shortcomings of traditional models in handling time-effect entanglement in intervention-response asymmetric systems, insufficient clinical interpretability, and poor robustness to sparse data. Furthermore, it optimizes representation learning through gated fusion and a time-aware multimodal adaptive mechanism, dynamically adjusting the weights of acute and chronic effects, effectively avoiding the limitations of traditional models. The method is prone to problems such as physiological inconsistency, insufficient generalization, or ambiguous boundaries. Using real lung cancer cohort data for training and evaluation, experimental verification has shown that the present invention outperforms comparative models such as LSTM, Transformer, and graph neural networks in key indicators such as MAE, MSE, and R². It can not only efficiently predict the degree of vascular inflammation induced by chemotherapy, radiotherapy, and other treatments, but also provide interpretive output consistent with clinical response patterns. It has strong cross-patient adaptability and is applicable to multiple fields with high requirements for time-series physiological prediction, such as cardiovascular risk stratification of cancer survivors, toxicity mechanism analysis, and personalized management. It has both practicality and wide application value. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0014] Figure 1 This is a flowchart of the steps of the method of the present invention. Detailed Implementation

[0015] 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.

[0016] Example 1 In this embodiment, as Figure 1 As shown, this invention provides a knowledge-guided and intervention-response driven method for predicting vascular inflammation, with specific steps including: S1. Collect multivariate physiological measurement sequences and perform preprocessing to obtain preprocessed multivariate physiological measurement sequences; Specifically, the multivariate physiological measurement sequence includes vital signs, blood biochemical indicators, inflammatory indicators, and metabolic indicators; the vital signs include systolic blood pressure, diastolic blood pressure, and heart rate; the blood biochemical indicators include total cholesterol, low-density lipoprotein, high-density lipoprotein, and triglycerides; the inflammatory indicators include C-reactive protein and white blood cell count; and the metabolic indicators include fasting blood glucose and glycated hemoglobin.

[0017] Specifically, missing values ​​are supplemented by a time-series linear interpolation method that introduces drug and treatment-specific adjustments, as expressed in the following formula: , in, , These represent the nearest observation time points before and after the time point corresponding to the missing value, respectively; , These represent the variables in a multivariate physiological measurement sequence. At the observation time point , The corresponding observations; Representing variables The reconstructed value; Indicates time A collection of medications that the patient is currently using; Indicates a specific drug; This represents the set of non-pharmacological treatments the patient is receiving at time t; Indicates specific treatment; Indicates drug-dependent variables Clinical impact weight; Indicates the effect of treatment on variables Clinical impact weight; Weight and The settings are based on the typical therapeutic effects well-reported in the clinical literature; they are derived from the average clinical effects of drugs / treatments on corresponding physiological variables in large randomized controlled trials and meta-analyses, and can be easily obtained and verified through standard medical knowledge. For example, the adjustment weights for statins (at commonly used doses) on low-density lipoprotein cholesterol (LDL) are set as follows: , These weights are explicitly set as fixed parameters in this invention, with the aim of embedding reliable clinical prior knowledge into the model and ensuring the reconstructed values. The physiological rationality and explainability.

[0018] To ensure physiological rationality, all reconstructed values ​​are constrained within the therapeutic perception reference range, as expressed by the following formula: , in, Representing variables The constraint value; Represents the clipping function; , Representing variables respectively The clinically acceptable lower and upper limits can be obtained according to guidelines. For example, the systolic blood pressure range is 90–140 mmHg.

[0019] S2. The preprocessed multivariate physiological measurement sequences and treatment sequences are explicitly decomposed into short-term acute pathway (SAP) and long-term cumulative pathway (LCP). For short-term acute pathway (SAP), the acute effect term is calculated using a physiologically inspired exponential decay model. Update the hidden state using a gated loop unit (GRU). For long-term cumulative pathway (LCP), treatment sequences are categorized into drug-based and non-drug-based classes, and their cumulative contributions are calculated separately to obtain the total cumulative contribution. Update the hidden state using a low-pass filter. By fusing the two pathways through a gating mechanism, a decoupled dynamic representation of treatment is obtained. ; Specifically, the SAP modeling process for short-term acute pathways is as follows: set up Indicates the patient At time step The intensity of the most recent treatment event is obtained by normalizing the treatment dose, as shown in the following formula: , in, Indicates the actual dosage. This indicates the standard dose recommended by clinical guidelines; if multiple drugs or treatments are used in combination (such as the common combination of chemotherapy and targeted / immunotherapy in lung cancer), then... Calculated using the summation and normalization method: , in, The set of drugs / treatments included in the most recent treatment event. Indicates the first The actual dosage administered in this drug treatment event. Indicates the first The standard dose recommended by clinical guidelines for this type of drug treatment event.

[0020] Acute effect item The formula is expressed as: , , in, Indicates the current follow-up time point. Indicates the time of the most recent treatment. Indicates time difference, Indicates the acute effect attenuation rate. Represents a nonlinear mapping function; Nonlinear mapping function This is used to simulate the physiological phenomenon that the effect gradually approaches saturation as the therapeutic dose increases, and its form is as follows: , in This is a saturation adjustment parameter used to control the saturation rate of the dose effect. Its initial value is determined based on typical dose-response saturation characteristics in pharmacodynamic literature. In one embodiment, in modeling acute effects related to chemotherapy or radiotherapy in lung cancer, the standard reference dose is typically much lower than the dose level that produces significant saturation; therefore, this invention will... The initial value was set to 0.1 (corresponding to a half-saturation dose of approximately 10 times the reference dose). This setting ensures that the effect increases approximately linearly within the commonly used clinical dose range (1–3 times the standard dose), while gradually saturating at higher doses, consistent with clinical observations of most cytotoxic drugs or radiotherapy. This value was then optimized as a learnable parameter during model training.

[0021] The hidden state of the short-run path is updated using a gated cyclic unit, as shown in the following formula: , in, Indicates the time of short-term acute pathways The hidden state vector is used to capture the immediate dynamic changes triggered by the recent treatment event, with initial values... Let it be the zero vector; Indicates the patient At follow-up time points Treatment input information, including the type and dosage of treatment received during this period, such as the type of chemotherapy drug, dosage, or radiotherapy dosage; Indicates from Physiological measurement subsequences extracted from it, including blood pressure, heart rate, blood lipids or inflammatory markers; In one embodiment, a lung cancer patient receives cisplatin chemotherapy 75 mg / m² and thoracic radiotherapy 20 Gy before a 12-month follow-up, then the treatment input... This can be represented as a treatment vector containing "cisplatin dose 75 mg / m²" and "radiotherapy dose 20 Gy"; simultaneously, at this follow-up time point, the patient's systolic blood pressure was 130 mmHg, heart rate was 80 bpm, and C-reactive protein (CRP) was 3.2 mg / L. These physiological indicators constitute a physiological measurement sequence. .

[0022] Specifically, the long-term cumulative path (LCP) modeling process is as follows: Treatment events are categorized into two types: pharmacological treatment and non-pharmacological treatment; for the first The formula for the cumulative contribution of each drug treatment event is as follows: , in, Indicates the first The cumulative contribution of each drug treatment; Indicates the first Time of onset of drug treatment; This represents the standardized dose threshold derived from the guidelines; This represents the learnable cumulative decay rate initialized within the guidelines' scope, used to characterize the gradual decay of the cumulative effect of drug treatment over time. Indicates the first The time of the drug treatment event The cumulative contribution; This is the dose saturation function, and its specific expression corresponds to the nonlinear mapping function in the Short-Term Acute Pathway (SAP). Totally consistent.

[0023] In one embodiment, the decay half-life is first determined based on the typical duration of cumulative drug toxicity (such as myelosuppression) as defined in clinical guidelines (such as NCCN). Then, using the formula: Initial values ​​were calculated. These initial values ​​ensured that the model's starting point conformed to clinical physiological patterns and were optimized as learnable parameters during model training. Taking the most common platinum-based chemotherapy regimen for lung cancer (cisplatin / carboplatin + pemetrexed or paclitaxel) as an example, clinical guidelines indicate that the cumulative toxicity of this type of chemotherapy (such as grade II–III peripheral neuropathy) usually appears after 4–6 cycles of treatment and gradually subsides within 2–4 months after treatment completion. Therefore, this invention uses the half-life... Set to 5 months, and then calculate. The initial value is: .

[0024] For non-pharmacological treatment events, the formula for cumulative contribution is as follows: , in, Indicates the first The cumulative contribution of non-pharmacological treatments; The baseline intensity parameter representing the impact of non-pharmacological treatment on physiological indicators is determined based on statistical results of changes in inflammatory markers (such as C-reactive protein) from medical knowledge. In one embodiment, for thoracic radiotherapy for lung cancer, a typical radiotherapy dose (45–60 Gy) typically leads to an approximately 1-fold increase in CRP levels; therefore, this invention will... The initial value is set to 1.0 (normalization strength), which can be optimized as a learnable parameter during model training. Indicates the time decay rate of the effect of non-pharmacological treatment; The initial value is set based on the duration range of inflammation caused by radiotherapy or other non-pharmacological treatments in medical knowledge. In one embodiment, the specific determination method is as follows: first, based on the typical half-life of the inflammatory response in clinical literature. Then, using the formula: Initial values ​​were calculated. Taking chest radiotherapy, a common treatment for lung cancer patients, as an example, inflammatory responses such as radiation pneumonitis typically peak 4–12 weeks after radiotherapy, and then gradually weaken over 2–4 months. Based on this clinical timeline, this invention defines the half-life... Set to 3 months, and then calculate. The initial value is: , This design allows the cumulative effect of non-pharmacological treatment to decay to approximately 50% after 3 months and to approximately 25% after 6 months, consistent with time curves observed in radiation oncology literature. The duration of this inflammatory effect can be easily obtained and verified by consulting ESTRO, RTOG, or NCCN guidelines for the management of radiation therapy toxicities.

[0025] The total cumulative effect is calculated based on the cumulative contribution of drug therapy and the cumulative contribution of non-drug therapy, as expressed by the following formula: , in, Indicates time The total cumulative effect of all historical treatment events at any given moment is used to characterize the combined impact of drug therapy and non-drug therapy on the patient's physiological state at the current point in time. Represents a set of drug treatment events; This represents a set of non-pharmacological treatment events; The long-term path hidden state is updated using a low-pass filtering mechanism, as expressed by the following formula: , in, Indicates the long-term cumulative path over time The hidden state vector, initial value Let it be the zero vector; In one embodiment, a smoothing factor is used, based on medical knowledge that vascular inflammation or physiological changes caused by radiotherapy or long-term chemotherapy typically stabilize gradually over approximately 3–6 months. Initialize to 0.4; This represents a multilayer perceptron.

[0026] Specifically, the short-term path hidden state and the long-term path hidden state are fused through a gating mechanism to obtain a decoupled treatment dynamic representation, as shown in the following formula: , , in, This represents the gating vector, used to dynamically determine the fusion weights of short-term acute information and long-term cumulative information at the current time point; Represents the Sigmoid function; Represents the gate weight matrix; This represents the dynamic representation of treatment after decoupling; This indicates element-wise multiplication.

[0027] S3. Constructing a heterogeneous dynamic graph It includes physiological nodes and treatment nodes, and the adjacency matrix integrates the prior and learned parts of medical knowledge; the modulation adjacency matrix is ​​obtained by applying treatment modulation; graph convolution is performed on the physiological subgraph to update the node representation; Specifically, at each follow-up time To characterize the effects of treatment on physiological indicators and the relationships between physiological variables, this invention constructs a heterogeneous dynamic graph. : , in, It represents a set of physiological nodes, including physiological measurement indicators such as blood pressure, heart rate, blood lipids, and CRP; This refers to a set of treatment nodes, including interventions such as chemotherapy, radiotherapy, and targeted therapy. Indicates time The directed interaction relationships between nodes are used to encode directed asymmetric interactions: physiological nodes can influence each other, while therapeutic nodes only act as information sources to intervene in physiological nodes, forming a unidirectional causal flow; adjacency matrix It is broken down into a prior knowledge component and a learnable component, as expressed by the following formula: , in, This represents a fixed prior matrix pre-constructed based on clinical medical knowledge, which mainly encodes reliable domain knowledge (such as known physiological correlations and typical drug-indicator causal relationships). This represents a learnable matrix, a supplementary matrix learned from patient follow-up data through model training, used to capture individual patient differences, unknown associations, or subtle adjustments reflected in the data; Knowledge Prior Matrix The primary encoding domain expert knowledge and typical relationships in clinical guidelines are pre-constructed based on clinical medical knowledge. In one embodiment, the edge weight from systolic blood pressure to diastolic blood pressure is set to 0.9; the edge weight from cisplatin chemotherapy to white blood cell count is set to 0.8. Learnable matrix To capture individual patient differences, unknown associations, or data-driven subtle adjustments, the model learns dynamically from patient follow-up data through model training. In one embodiment, for a hypertensive patient, the edge weight between CRP and blood pressure is further learned to increase from a priori 0.5 to 0.75; when chemotherapy and immunotherapy are used in combination, the synergistic effect on CRP is learned from a priori 0 to 0.4.

[0028] To ensure unidirectional causal flow, treatment nodes act only as sources, and physiological nodes are not allowed to pass information to treatment nodes. The effect of treatment on physiological indicators is dynamically modulated by adjusting the edge weights between physiological nodes in the adjacency matrix in the following way: According to the current time Decoupling of treatment dynamics For each treatment node Calculate modulation intensity The formula is expressed as follows: , in, Indicates treatment node The modulation intensity function is used to calculate the modulation coefficient of the treatment on the edge weights between physiological nodes based on the treatment dynamic information at the current moment; Indicates treatment node The corresponding learnable weight vector is used to linearly project the decoupled treatment dynamic representation into the modulation intensity of the treatment; This represents the hyperbolic tangent activation function, whose output range is fixed at [...]. The interval [1,1] allows for the smooth compression of the calculated original modulation value into this interval.

[0029] Then, for physiological nodes The original edge weights between them are multiplicatively modulated, and the specific formula is as follows: , in, This represents the adjacency matrix after treatment modulation; Indicates time From nodes in a time-heterogeneous dynamic graph To the node The right of the border; Represents physiological nodes; Indicates treatment nodes; For each treatment A learnable scalar coefficient is used to adjust the strength of its influence on physiological edges, and its initial value is set based on prior clinical knowledge (in one embodiment, the effect of statins on lipid-related edges). (Initial value: 0.8–1.2) This represents the modulation function generated based on the decoupled treatment dynamics representation. ,in, express Corresponding treatment The amount, Represents the hyperbolic tangent function; This indicates the indicator function, if a treatment node exists. To physiological nodes If an edge is connected to a given edge, the value is 1; otherwise, it is 0.

[0030] The specific working mechanism of modulation is as follows: when the patient is in time When a patient receives a treatment (such as chemotherapy or radiation therapy), the model calculates the modulation intensity based on the decoupled dynamic representation of that treatment. If the treatment targets a certain physiological node If there is an impact, then it is determined through scalar coefficients. For all from The edge weights between the starting physiological nodes are multiplicatively enhanced or inhibited. When the treatment intensity is high and When the correlation is greater than 0, the strength of the association between physiological nodes is amplified accordingly, thus better characterizing the propagation effect of treatment-induced physiological changes such as vascular inflammation between nodes; conversely, when... When the value is less than 0, some associations are suppressed. This mechanism, while maintaining the unidirectional information flow of treatment nodes, enables dynamic and personalized modulation of physiological causal relationships by treatment intervention.

[0031] Graph convolution is performed on the physiological subgraph using modulated adjacency to obtain the node representation updated by graph convolution. The formula for graph convolution update is expressed as follows: , in, Representing the Physiological nodes in layer graph convolution The hidden representation, Representing physiological nodes The set of neighboring nodes, Representativeness normalization factor to ensure stable graph convolution values. Representing the Layer graph convolution weights, Represents a nonlinear activation function S4. The decoupled treatment dynamic representation and the node representation updated by graph convolution are fused to obtain the graph-enhanced representation of each physiological node; the graph-enhanced representation of each physiological node is weighted through an attention mechanism to obtain the patient-level representation; time-specific regression is used to map the patient-level representation to the target TBR value to obtain the TBR prediction value. .

[0032] Specifically, the decoupled treatment dynamic representation and the node representation updated by graph convolution are concatenated and linearly mapped to obtain the graph-enhanced representation of each physiological node, as shown in the following formula: , in, Indicates time Physiological nodes Graph augmentation representation; Indicates the patient In time Physiological nodes Decoupling representation; This represents the node representation after graph convolution. , Let represent the learnable weight matrix and bias vector of the linear mapping, respectively; the graph augmentation representation of each physiological node is weighted through an attention mechanism to obtain the patient-level representation. The formula is expressed as follows: , in, Represents a node Attention weights are used to represent the importance of this physiological indicator in the patient's overall condition; Represents the attention vector; Indicates time Physiological nodes Enhanced graph representation; time-specific regression is used to represent patient levels. Mapping to the target TBR value, we obtain the patient's... time TBR prediction value The formula is expressed as follows: , in, (.) represents a time-specific regression function used to map patient-level representations to TBR predictions at the current follow-up time point; Represents the learnable weight matrix for a specific regression at a given time point; This represents the bias vector for a specific regression at a given time point.

[0033] Example 2 In this embodiment, a 62-year-old male non-small cell lung cancer (NSCLC, stage IV) patient is used as an example to fully illustrate all the steps of the method of the present invention and the meaning and calculation process of each key value.

[0034] At the 12-month follow-up, the patient received a combination of cisplatin chemotherapy 75 mg / m² and chest radiotherapy 20 Gy, while also taking oral atorvastatin 20 mg / day.

[0035] The LDL value was missing at this time point, but other physiological indicators were observable.

[0036] Step a) Knowledge-guided follow-up data preprocessing The patient's LDL value was missing. Temporal interpolation was initially performed: the LDL level observed at the previous follow-up (t=11) was 3.4 mmol / L, and the LDL level observed at the next follow-up (t=13) was 2.9 mmol / L. Linear interpolation was then performed using the following formula: , Then, the therapeutic effect is reconstructed: , in: (The clinical impact weight of statins on LDL, meaning: at the usual dose, they can reduce LDL by an average of 0.25 mmol / L); (Clinical weighting of cisplatin's effect on LDL, meaning: slight increase) (The effect of chest radiotherapy on LDL is negligible); Substituting into the calculation, we get: , Physiological rationality constraints: , Final physiological measurement sequence (Partial indicators): SBP = 130 mmHg, DBP = 85 mmHg, heart rate = 78 bpm, LDL = 2.95 mmol / L, CRP = 4.8 mg / L, blood glucose = 5.6 mmol / L.

[0037] Step b) Dynamic perception of therapeutic effects b-1) Short-term acute pathway (SAP) Intensity of most recent treatment: , Where: 75: Actual cisplatin dose received by the patient (unit: mg / m²); 100: Standard cisplatin dose recommended by clinical guidelines (unit: mg / m²); 20: Actual chest radiotherapy dose received by the patient (unit: Gy); 60: Reference standard dose for chest radiotherapy in clinical guidelines (unit: Gy). The normalized intensity of this combination therapy was 1.083 (greater than 1 indicates that the intensity of this treatment was slightly higher than the standard single-drug level).

[0038] The acute effect item is calculated using the following formula: , in, (γ=0.1, meaning: saturation adjustment parameter, corresponding to a half-saturation dose of 10 times the standard dose), λa=0.05 (meaning: short-term acute effect decay rate, corresponding to a half-life of about 14 days), Δt=1 (meaning: the time interval between the current follow-up time point and the most recent treatment, unit: month).

[0039] Substituting into the calculation, we get: , The intensity of the short-term acute physiological effect produced by this treatment at the current time point is 0.808 (approximately 81% residual inflammatory response).

[0040] b-2) Long-Term Cumulative Pathway (LCP) The cumulative contribution of drug therapy is calculated using the following formula: , Substitution (Half-life: 3 months) ,have to: ; Cumulative contribution of non-pharmacological treatments: , in (Meaning: Baseline intensity parameters for radiotherapy, normalized reference values).

[0041] b-3) Dual-path integration The gating vector is calculated using the following formula: , The short-term path accounted for 0.65% of the weight in this fusion (which was dominated by acute effects). The final decoupled treatment dynamics are represented as follows: .

[0042] Step c) Knowledge-guided dynamic graph convolution Taking the edge pointing from CRP to white blood cell count as an example, the original edge weight After treatment and adjustment: , Step d) Final prediction After fusion, the predicted arterial TBR value for this time point was 1.82, output by the regression head.

[0043] Example 3 In this embodiment, the performance of different time-series prediction models is compared using various regression metrics. These metrics include MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and R² (Coefficient of Determination). MAE measures the average absolute difference between the predicted value and the actual TBR; a smaller value indicates higher prediction accuracy. RMSE measures the square root of the prediction error; a smaller value indicates better model stability. R² measures the model's explanatory power for data variation; a value closer to 1 indicates a better fit.

[0044] Data collection was coordinated by the hospital's information center, with seven experts collaborating on data annotation and verification to ensure accuracy and reliability. The dataset includes 434 lung cancer patients (aged 25-79 years), with a total of 3465 follow-up time points, grouped by treatment duration at 6, 12, 18, and 24 months. Each patient underwent at least one PET / CT scan before and after treatment. Data includes image-derived TBR values, physiological parameters (e.g., HDL-C, triglycerides, creatinine, glucose), pathology reports, demographic data, and detailed treatment records.

[0045] As shown in Tables 1, 2, and 3, the longitudinal follow-up data of lung cancer patients consist of tables of multivariate physiological measurements collected at irregular time intervals. Each follow-up record also includes treatment-related information.

[0046] Table 1 Baseline data before treatment Table 2 Data during treatment Table 3 Post-treatment data This embodiment selects nine methods for comparison: Irregular multivariate temporal methods (HyperIMTS, TRANS): handle sparse longitudinal data through temporal modeling, but do not explicitly separate acute and chronic effects; Transformer-based temporal models (DyTed): capture sequence dependencies, but the uniform representation of dependencies leads to effect entanglement; Dynamic and heterogeneous graph neural networks (DynaGraph, TGNN4I, DHGAS): improve heterogeneous and temporal representation learning, but assume homogeneous propagation and treat treatment as a common feature; Knowledge-enhanced graph methods (KEDGN, KGDNNet, GraphCare): incorporate knowledge-enhanced propagation, but lack clinically grounded propagation constraints and acute-chronic decoupling.

[0047] Table 4 Comparison of predicted comprehensive performance of cross-sectional TBR Table 5 Comparison of MAE at different follow-up stages As shown in Tables 4 and 5, based on these experimental results, the present invention can draw the following conclusions: Irregular multivariate time series methods generally perform poorly in handling sparse data, exhibiting high MAE and RMSE but low R², indicating limited predictive effectiveness in intervention-response asymmetric systems, especially during long-term follow-up. Transformer-based time series models perform well, particularly in capturing sequence dependencies and reducing some errors, but still suffer from effect entanglement issues, with an overall R² below 0.7. Dynamic and heterogeneous graph neural networks perform moderately, exhibiting large prediction errors due to a lack of knowledge constraints, resulting in poor MAE and RMSE. Knowledge-enhanced graph methods perform well, but fail to decouple acute and chronic paths, leading to physiological inconsistencies in sparse data, and MAE and RMSE still have room for improvement. The KIR-Net method of this invention performs best, with an overall MAE of 0.1186, RMSE of 0.1678, and R² of 0.781. Compared to the best knowledge-enhanced graph baseline KEDGN, it reduces MAE by 11.4%, RMSE by 7.9%, and improves R² by 6.1%. Throughout the single follow-up phase, KIR-Net consistently outperformed all baselines, for example, reducing MAE by 18.9% (from 0.1335 to 0.1081 in KEDGN) and 29.0% (from 0.1523 to 0.1081 in HyperIMTS) at 6 months and by 28.3% (from 0.1604 to 0.1150 in KEDGN) at 24 months, demonstrating its early sensitivity and long-term robustness.

[0048] In summary, the TBR prediction model based on knowledge-guided intervention-response networks has significant advantages in predicting treatment-induced vascular inflammation, and is particularly suitable for sparse and irregular longitudinal clinical data and scenarios with asymmetric acute and chronic effects, such as cardiovascular risk stratification of cancer survivors and toxicity assessment, which have high requirements for time-series physiological prediction.

[0049] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge-guided and intervention-response driven method for predicting vascular inflammation, characterized in that, Includes the following steps: S1. Collect and preprocess multivariate physiological measurement sequences to obtain preprocessed multivariate physiological measurement sequences; the multivariate physiological measurement sequences include vital signs, blood biochemical indicators, inflammatory indicators, and metabolic indicators; the vital signs include systolic blood pressure, diastolic blood pressure, and heart rate; the blood biochemical indicators include total cholesterol, low-density lipoprotein, high-density lipoprotein, and triglycerides; the inflammatory indicators include C-reactive protein and white blood cell count; the metabolic indicators include fasting blood glucose and glycated hemoglobin; S2. The preprocessed multivariate physiological measurement sequences and treatment sequences are explicitly decomposed into short-term acute pathway (SAP) and long-term cumulative pathway (LCP). For short-term acute pathway (SAP), the acute effect term is calculated using a physiologically inspired exponential decay model. Update the hidden state using a gated loop unit (GRU). For long-term cumulative pathway (LCP), treatment sequences are categorized into drug-based and non-drug-based classes, and their cumulative contributions are calculated separately to obtain the total cumulative contribution. Update the hidden state using a low-pass filter. By fusing the short-term and long-term path hidden states through a gating mechanism, a decoupled dynamic representation of treatment is obtained. ; S3, at each follow-up time Constructing heterogeneous dynamic graphs : ,in, Represents a set of physiological nodes; Represents the set of treatment nodes; Indicates time Directed interaction relationships between nodes; adjacency matrix It is broken down into a prior knowledge component and a learnable component, as expressed by the following formula: ， in, This represents a fixed prior matrix pre-constructed based on clinical medical knowledge; Represents a learnable matrix; The effects of treatment on physiological indicators are dynamically modulated by adjusting the edge weights between physiological nodes in the adjacency matrix in the following way: According to the current time Decoupling of treatment dynamics For each treatment node Calculate modulation intensity ; Based on modulation strength The original edge weights between physiological nodes are multiplicatively modulated, and the specific formula is as follows: ， in, This represents the adjacency matrix after treatment modulation; Indicates time From nodes in a time-heterogeneous dynamic graph To the node The right of the border; Represents physiological nodes; Indicates treatment nodes; For each treatment Learnable scalar coefficients; This represents the modulation function generated based on the decoupled treatment dynamics representation; Indicates an indicator function; Then, graph convolution is performed on the physiological subgraph with modulated adjacency to obtain the node representation updated by graph convolution. ; S4. The decoupled treatment dynamic representation and the node representation updated by graph convolution are fused to obtain the graph-enhanced representation of each physiological node; the graph-enhanced representation of each physiological node is weighted through an attention mechanism to obtain the patient-level representation; time-specific regression is used to map the patient-level representation to the target TBR value to obtain the TBR prediction value. .

2. The knowledge-guided and intervention-response driven vascular inflammation prediction method according to claim 1, characterized in that, In step S1, missing values ​​are supplemented by a time-series linear interpolation method that introduces drug and treatment-specific adjustment terms, as shown in the following formula: ， in, , These represent the nearest observation time points before and after the time point corresponding to the missing value, respectively; , These represent the variables in a multivariate physiological measurement sequence. At the observation time point , The corresponding observations; Representing variables The reconstructed value; Indicates time A collection of medications that the patient is currently using; Indicates a specific drug; This represents the set of non-pharmacological treatments the patient is receiving at time t; Indicates specific treatment; Indicates drug-dependent variables Clinical impact weight; Indicates the effect of treatment on variables Clinical impact weight; All reconstructed values ​​are constrained within the therapeutic perception reference range, as expressed by the following formula: ， in, Representing variables The constraint value; Represents the clipping function; , Representing variables respectively The lower and upper limits of clinical acceptance.

3. The knowledge-guided and intervention-response driven vascular inflammation prediction method according to claim 2, characterized in that, In step S2, the SAP modeling process for short-term acute pathways is as follows: set up Indicates the patient At time step The intensity of the most recent treatment event is obtained by normalizing the treatment dose, as shown in the following formula: ， in, Indicates the actual dosage. This indicates the standard dose recommended by clinical guidelines; Acute effect item The formula is expressed as: ， ， in, Indicates the current follow-up time point. Indicates the time of the most recent treatment. Indicates time difference, Indicates the acute effect attenuation rate. Represents a nonlinear mapping function; The hidden state of the short-term path is updated using a gated recurrent unit, as expressed by the following formula: ， in, Indicates the time of short-term acute pathways The hidden state vector, initial value Let it be the zero vector; Indicates the patient At follow-up time points Treatment input information; Indicates from Physiological measurement subsequences extracted from it.

4. The knowledge-guided and intervention-response driven method for predicting vascular inflammation according to claim 1, characterized in that, In step S2, the long-term cumulative path (LCP) modeling process is as follows: Treatment events are categorized into two types: pharmacological treatment and non-pharmacological treatment; for the first... The formula for the cumulative contribution of each drug treatment event is as follows: ， in, Indicates the first The cumulative contribution of each drug treatment; Indicates the first Time of onset of drug treatment; Indicates the standardized dose threshold; This represents the learnable cumulative decay rate, used to characterize the process by which the cumulative effect of drug treatment gradually decays over time. Indicates the first The time of the drug treatment event The cumulative contribution; This is a dose saturation function; For non-pharmacological treatment events, the formula for cumulative contribution is as follows: ， in, Indicates the first The cumulative contribution of non-pharmacological treatments; The baseline strength parameter indicating the impact of non-pharmacological treatments on physiological indicators; Indicates the time decay rate of the effect of non-pharmacological treatment; The total cumulative effect is calculated based on the cumulative contribution of drug therapy and the cumulative contribution of non-drug therapy, as expressed by the following formula: ， in, Indicates time The total cumulative effect of all historical treatment events at any given moment; Represents a set of drug treatment events; This represents a set of non-pharmacological treatment events; The long-term path hidden state is updated using a low-pass filtering mechanism, as expressed by the following formula: ， in, Indicates the long-term cumulative path over time The hidden state vector; Indicates the smoothing factor; This represents a multilayer perceptron.

5. The knowledge-guided and intervention-response driven method for predicting vascular inflammation according to claim 1, characterized in that, In step S2, the short-term path hidden state and the long-term path hidden state are fused through a gating mechanism to obtain the decoupled treatment dynamic representation, as shown in the following formula: ， ， in, Represents the gate vector; Represents the Sigmoid function; Represents the gate weight matrix; This represents the dynamic representation of treatment after decoupling; This indicates element-wise multiplication.

6. The knowledge-guided and intervention-response driven method for predicting vascular inflammation according to claim 1, characterized in that, The decoupled treatment dynamics representation and the node representation updated by graph convolution are concatenated and linearly mapped to obtain the graph-enhanced representation of each physiological node. The graph-enhanced representation of each physiological node is then weighted through an attention mechanism to obtain the patient-level representation. ; Patient levels were represented using time-specific regression. Mapping to the target TBR value, we obtain the patient's... time TBR prediction value The formula is expressed as follows: ， in, (.) indicates a time-specific regression function; Represents the learnable weight matrix for a specific regression at a given time point; This represents the bias vector for a specific regression at a given time point.

Images