Multi-parameter coronary heart disease adverse event risk prediction method based on machine learning

Abstract

The invention discloses a multi-parameter coronary heart disease adverse event risk prediction method based on machine learning, and relates to the technical field of risk prediction. Compared with a previous risk assessment method for the adverse events of the coronary heart disease, the risk assessment method solves the problems that an existing risk assessment method for the adverse events of the coronary heart disease is single in data source, limited in model accuracy, complex in operation and lack of individuation and real-time performance; by integrating medical image data and personal health historical data of a patient and applying a dynamic graph neural network, the risk of coronary heart disease adverse events of the patient in two years in the future is accurately evaluated, and by deeply explaining risk factors, more accurate diagnosis support and treatment suggestions are provided for doctors, so that the diagnosis and treatment efficiency is improved. Personalized risk assessment and treatment schemes are realized; the operation is simple, the operation process is simplified, and the patient burden is reduced; the system can rapidly process and update data in real time, outputs a real-time risk assessment result, has a feedback mechanism, and provides instant and accurate reference for clinical decisions.

Description

Technical Field

[0001] The present invention relates to the technical field of risk prediction, and in particular to a multi-parameter coronary heart disease adverse event risk prediction method based on machine learning. Background Art

[0002] In the prior art, the risk assessment of adverse events of coronary heart disease mainly relies on traditional clinical assessment methods, such as electrocardiogram (ECG), blood biomarker detection, and doctors' clinical experience. Although these methods can assess patients' coronary heart disease risks to a certain extent, they often have problems such as strong subjectivity, insufficient accuracy, and inability to fully consider the impact of multiple parameters. In addition, traditional risk assessment methods usually require patients to go to the hospital for multiple examinations, which increases the burden on patients and the consumption of medical resources.

[0003] In the existing technology, the coronary heart disease risk assessment system usually includes the following main components:

[0004] Data input module: used to input the patient's clinical data, such as age, gender, family medical history, living habits, etc.

[0005] Data processing module: performs preliminary processing and analysis on input data, such as data cleaning and standardization.

[0006] Risk assessment model: Perform risk assessment on processed data based on statistical methods or simple machine learning algorithms.

[0007] Result output module: outputs the risk assessment results in a visual form for doctors’ reference.

[0008] The risk assessment models in the existing technology usually use traditional statistical methods, such as logistic regression and Cox proportional hazard model. Although these methods can assess the risk of coronary heart disease in patients to a certain extent, they often cannot fully consider the impact of multiple parameters, and the accuracy and generalization ability of the models are limited. In addition, the risk assessment systems in the existing technology often lack in-depth analysis of medical imaging data and cannot fully utilize imaging information to improve the accuracy of risk assessment.

[0009] In order to solve the above problems, the present invention proposes a multi-parameter coronary heart disease adverse event risk prediction method based on machine learning. Summary of the invention

[0010] The purpose of the present invention is to propose a multi-parameter coronary heart disease adverse event risk prediction method based on machine learning to solve the problems raised in the background technology:

[0011] Existing risk assessment methods for adverse coronary heart disease events have problems such as single data source, limited model accuracy, complex operation, lack of personalization and lack of real-time performance.

[0012] In order to achieve the above object, the present invention adopts the following technical solutions:

[0013] A multi-parameter coronary heart disease adverse event risk prediction method based on machine learning, comprising the following steps:

[0014] S1: Collect patients’ medical imaging data and personal health history data, and clean, standardize and format the data;

[0015] S2: Extract key imaging factors and health history factors from the preprocessed data, and use feature selection algorithms to screen out the most influential features for the risk prediction of adverse coronary events;

[0016] S3: used to train the extracted features through machine learning algorithms, build a risk prediction model for adverse events of coronary heart disease, and optimize the model;

[0017] S4: Input new patient data into the trained model, and the model outputs the risk prediction results of the patient experiencing adverse coronary heart disease events in the next two years;

[0018] S5: Update the patient’s latest data in real time and re-predict risk.

[0019] Preferably, the key imaging factors and health history factors extracted in S2 include ALCOHOL_HISTORY, PLAQUE_VOLUME, PLAQUE_LENGTH, MIN_LUMINAL_AREA, MIN_NARROWING_PERCENTAGE, HIG H_RISK_NAPKIN_RING_SIGN, HIGH_RISK_POSITIVE_REMODE L, HIGH_RISK_LARGE_PLAQUE_VOLUME, CT_FFR, DELTA_C T_FFR, FAI, LIPID_VOLUME, LIPID_VOLUME_RATIO, FIBRO LIPID_VOLUME, and FIBROLIPID_VOLUME_RATIO.

[0020] Preferably, in S2, feature selection is performed based on an optimized recursive feature elimination method, as follows:

[0021] Based on the preprocessed data set N = {X, Y}, where X is the feature matrix, X∈R n×p , n is the number of samples, p is the number of initial features; Y is the corresponding target variable vector, Y∈R n ;

[0022] Train a random forest model and calculate the average impurity reduction of each feature in the decision tree as the importance score;

[0023] Train the gradient boosted tree model and calculate the split contribution of each feature in the gradient boosted tree as the importance score;

[0024] Calculate the average importance score of each feature in the random forest model and the gradient boosting tree model to calculate the comprehensive feature importance score, sort the comprehensive importance scores, and obtain the feature importance ranking result π sorted in descending order;

[0025] Set the initial feature selection ratio η0, and the corresponding initial selection feature number m0 = η0 × p;

[0026] Define a performance evaluation index P based on the currently selected feature subset That is, the first m selected in order of importance j A subset of features, initially m j =m0, train a support vector machine model and get the current model performance P j ;

[0027] Increase or decrease the number of features to update m j+1 :

[0028] Add features: Let m j+1 =m j +Δm + , where Δm + For a small positive integer increment, select a new feature subset Train the model and get performance

[0029] Reduce features: Let m j+1 =m j -Δm - , where Δm - Decrement by a small positive integer, selecting a new subset of features Train the model and get performance

[0030] Adaptive Updates:

[0031] if and Then update m j =m j+1 , that is, increase the number of features and continue to the next iteration;

[0032] if and Then update m j =m j+1, that is, reduce the number of features and continue to the next iteration;

[0033] if and Do not update m j , stop the iteration or decide whether to continue based on other stopping conditions;

[0034] The number of features m finally determined according to the above adaptive adjustment process end , select the corresponding feature subset As the final feature selection result.

[0035] Preferably, in S3, the risk prediction model for adverse events of coronary heart disease is constructed as follows:

[0036] Each patient is represented as a node in the graph, and the feature vector of the node includes the feature selection result of the feature selection algorithm; edges are constructed according to the calibration relationship between patients, and the weight of the edge is determined based on the similarity measurement method. The initial graph structure is represented as G = (V, E), where V is the node set and E is the edge set;

[0037] The constructed patient relationship graph structure is initially learned and features extracted through graph neural networks, and the node feature representation is enhanced using the similarity information provided by the walk-based graph kernel function:

[0038] Construct the adjacency matrix A of the graph based on the initial graph structure, where A ab Indicates whether there is an edge connection between node a and node b. If there is an edge, then A ab =1, otherwise A ab =0;

[0039] Calculate the node transition probability matrix R, where R ab Represents the transition probability from node a to node b:

[0040] R=D -1 A

[0041]

[0042] Where D is the degree matrix;

[0043] For each node a, define its random walk feature map in, is the probability distribution of a random walk path starting from node a and going through t steps;

[0044] For two nodes a and b, the graph kernel function based on random walk is defined as k(a,b)=<φ(a),φ(b)>, where <,> represents the inner product operation;

[0045] For all nodes in the graph G, the graph core function values ​​between each two nodes are calculated to obtain the graph core matrix K, where K ab =k(a,b),a,b∈V;

[0046] For each node, its feature vector is updated at each time step t according to the features of its neighboring nodes and the weights of the edges;

[0047]

[0048] Among them, h v,l+1 、h v,l are the feature vectors of node v at the input of the l+1th layer network and the lth layer network respectively; σ is the activation function; w l , b l are the learning parameters of the lth layer network respectively; δ(u) is the set of neighbor nodes of node u; δ(v) is the set of neighbor nodes of node v;

[0049] The weight of the edge is dynamically adjusted based on the similarity of the subsequent examination data of the patient node; the node features and model parameters in the graph neural network are updated based on the gradient descent algorithm;

[0050] Take the structure and node features of the current graph as the state of the reinforcement learning agent, and define the environment state, action, reward, and state transition probability:

[0051] The environment state C represents all possible states of the environment in which the agent is located, that is, c = {c τ}, where c τ represents the state at time τ;

[0052] The action set Q represents the set of all actions of the agent, and the action q∈Q includes selecting different graph neural network parameter adjustment strategies;

[0053] The reward function H(c,q,c') represents the reward that the agent receives when taking action q in state c and transferring to state c'; the reward is set based on the accuracy of the model's prediction of adverse events of coronary heart disease. If the prediction accuracy of the model in the current state improves, the agent receives a positive reward; if the prediction accuracy decreases or the model performance deteriorates, the agent receives a negative reward;

[0054] The state transition probability F(c'|c,q) represents the probability of taking action q in state c to transition to state c';

[0055] Define strategy θ, strategy θ(q|c) represents the probability of taking action q in state c;

[0056] Define the state value function V θ (c) are as follows:

[0057]

[0058] Among them, E θ [·] is the value function; λ∈[0,1] is the discount factor;

[0059] Define the action value function U θ (c) are as follows:

[0060]

[0061] The optimal value function is defined as:

[0062] Optimal state value function:

[0063] Optimal action-value function:

[0064] In this way, we can find the optimal strategy θ * , that is, for all states c∈C, the state value function corresponding to the optimal strategy is satisfied In finding the optimal strategy θ * In the process, the strategy is updated according to the reward feedback.

[0065] Preferably, in S4, the prediction result of the patient's adverse events of coronary heart disease in the next two years is output in a visualized form.

[0066] Preferably, in S4, risk factors are also explained in depth.

[0067] Preferably, in S5, the prediction results are fed back according to the actual situation of the patient through a feedback mechanism, and the prediction model is optimized.

[0068] Compared with the prior art, the present invention provides a multi-parameter coronary heart disease adverse event risk prediction method based on machine learning, which has the following beneficial effects:

[0069] The present invention integrates the patient's medical imaging data and personal health history data and applies a dynamic graph neural network to more accurately assess the patient's risk of adverse coronary heart disease events in the next two years. The dynamic graph neural network, combined with the graph kernel function, can capture the deep structural characteristics of the graph, so that the model can better understand the relationship between patients and the structural characteristics of the entire graph, thereby improving the accuracy of the risk prediction of adverse coronary heart disease events. It is also combined with a reinforcement learning method, and the structure and parameters of the graph neural network, as well as the structure of the graph itself, can be continuously adjusted according to the prediction effect of the model, further improving the accuracy of the risk prediction of adverse coronary heart disease events. It also provides doctors with more accurate diagnostic support and treatment suggestions through in-depth interpretation of risk factors, and realizes personalized risk assessment and treatment plans. It is simple to operate, simplifies the operating process, and reduces the burden on patients. It can also quickly process and update data in real time, output real-time risk assessment results predicted by the model, and has a feedback mechanism to provide immediate and accurate reference for clinical decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is a flow chart of the method mentioned in Example 1 of the present invention;

[0071] Figure 2 This is a schematic diagram of the prediction error results of the model evaluation mentioned in Example 1 of the present invention;

[0072] Figure 3 This is a schematic diagram of the system interface mentioned in Example 2 of the present invention;

[0073] Figure 4 This is a schematic diagram of the data element input area mentioned in Example 2 of the present invention;

[0074] Figure 5 This is a schematic diagram of the model prediction results mentioned in Example 2 of the present invention;

[0075] Figure 6 This is a schematic diagram of the model prediction result analysis mentioned in Example 2 of the present invention;

[0076] Figure 7 This is a schematic diagram of the reference suggestion results mentioned in Example 2 of the present invention. DETAILED DESCRIPTION

[0077] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0078] The present invention integrates the patient's medical imaging data and personal health history data, and applies a dynamic graph neural network to more accurately assess the patient's risk of adverse events of coronary heart disease in the next two years. The dynamic graph neural network, combined with the graph kernel function, can capture the deep structural features of the graph, so that the model can better understand the relationship between patients and the structural features of the entire graph, thereby improving the accuracy of the risk prediction of adverse events of coronary heart disease; it also combines the reinforcement learning method, and can continuously adjust the structure and parameters of the graph neural network and the structure of the graph itself according to the prediction effect of the model, further improving the accuracy of the risk prediction of adverse events of coronary heart disease; it also provides doctors with more accurate diagnostic support and treatment suggestions through in-depth interpretation of risk factors, and realizes personalized risk assessment and treatment plans; it is simple to operate, simplifies the operation process, and reduces the burden on patients; it can also quickly process and update data in real time, output the real-time risk assessment results predicted by the model, and has a feedback mechanism to provide immediate and accurate reference for clinical decision-making. Specifically include the following contents.

[0079] Embodiment 1:

[0080] See also Figure 1-2 The present invention provides a multi-parameter coronary heart disease adverse event risk prediction method based on machine learning, comprising the following steps:

[0081] S1: Collect the patient's medical imaging data and personal health history data, and clean, standardize and format the data; the details are as follows:

[0082] Collect medical imaging data and personal health history data of patients, including basic demographic information (such as age, gender, etc.), medical history (such as history of hypertension, diabetes, hyperlipidemia, etc.), lifestyle factors (such as smoking status, drinking status, exercise habits, etc.), laboratory test results (such as blood lipid levels, blood sugar, liver and kidney function indicators, myocardial enzymes, etc.), electrocardiogram data, cardiac ultrasound results, etc. You can also collect follow-up data of patients to record whether adverse events (such as myocardial infarction, heart failure, arrhythmia, sudden cardiac death, etc.) have occurred and when they occurred. Then clean, standardize and format the collected data. For numerical variables, you can use the mean, median and other methods to fill; for categorical variables, you can use the most common categories to fill. Check the accuracy and consistency of the data and remove obviously wrong or abnormal data points. For example, situations such as age that is not in line with common sense and negative values ​​or gender labeling errors need to be corrected. The data can also be normalized or standardized. Commonly used methods include minimum-maximum normalization (mapping the data to the [0,1] interval) and Z-score standardization (making the data conform to the standard normal distribution). Data preprocessing can ensure the quality and consistency of the data for subsequent analysis and processing.

[0083] S2: Extract key imaging factors and health history factors from the preprocessed data, and use feature selection algorithms to screen out the most influential features for the risk prediction of adverse events of coronary heart disease; the details are as follows:

[0084] The key 15 imaging factors and health history factors were extracted from the preprocessed data, as follows:

[0085] ALCOHOL_HISTORY (alcohol abuse history)

[0086] This is a key clinical variable used to assess whether the patient has a long-term habit of excessive drinking. A history of alcohol abuse is an important risk factor in the prevention and management of cardiovascular disease. 0 means that the patient has no history of alcohol abuse, 1 means that the patient has a history of alcohol abuse, and 9 means that the data is unclear or missing.

[0087] PLAQUE_VOLUME (plaque volume)

[0088] Plaque volume refers to the total volume of plaque on the inner wall of the coronary artery. Plaque is the accumulation of fat, cholesterol, calcium and other substances in the arterial wall. The larger the plaque volume, the higher the possibility of arterial stenosis, thereby increasing the risk of heart disease. It is an important imaging parameter measured by CT, IVUS and other technologies to assess the degree of arterial stenosis and the risk of future cardiovascular events.

[0089] PLAQUE_LENGTH (plaque length)

[0090] Plaque length refers to how long the plaque extends along the inner wall of the artery. The longer the plaque, the greater the obstruction to blood flow and the higher the risk of a cardiac event.

[0091] Plaque length refers to the linear extension length of atherosclerotic plaque on the inner wall of the artery. This indicator helps to understand the distribution characteristics of plaques and evaluate their impact on arterial hemodynamics.

[0092] MIN_LUMINAL_AREA (minimum lumen area)

[0093] The minimum lumen area refers to the minimum cross-sectional area of ​​the arterial lumen under the influence of atherosclerotic plaques. This indicator is the key to assessing the severity of arterial stenosis and directly affects the blood supply to the myocardium.

[0094] MIN_NARROWING_PERCENTAGE (minimum NARROWING percentage)

[0095] The narrowest point percentage refers to the degree of stenosis at the narrowest part of the artery, usually expressed as a percentage. This indicator helps to assess the impact of arterial stenosis on blood flow and is an important basis for determining whether further intervention is needed.

[0096] HIGH_RISK_NAPKIN_RING_SIGN (High risk - napkin ring sign)

[0097] The napkin ring sign is a specific imaging feature that appears as a low-density area surrounding the plaque. This feature is often associated with plaque instability and may indicate that the plaque is prone to rupture and increase the risk of acute cardiovascular events.

[0098] HIGH_RISK_POSITIVE_REMODEL (High Risk - Positive Remodeling)

[0099] Positive remodeling refers to the increase in the outer diameter of the artery during plaque formation. This remodeling may mask the actual volume of the plaque and make the stenosis appear less severe. However, plaques with positive remodeling may be more unstable and increase the risk of cardiac events.

[0100] HIGH_RISK_LARGE_PLAQUE_VOLUME (High risk - large plaque volume)

[0101] Large plaque volume means that the plaque occupies more space and may cause more severe narrowing of the artery, thereby increasing the risk of cardiac events. This indicator is crucial in assessing the clinical significance of plaque.

[0102] CT_FFR (non-invasive fractional flow reserve)

[0103] Noninvasive blood flow reserve fraction (CT FFR) is a hemodynamic parameter obtained by computed tomography (CT) technology, which is used to evaluate the effect of coronary artery stenosis on myocardial blood flow. The measurement of CT FFR helps to determine whether the stenosis is sufficient to cause myocardial ischemia.

[0104] DELTA_CT_FFR (△ non-invasive blood flow reserve fraction)

[0105] △Non-invasive blood flow reserve fraction refers to the changes between CT FFR values ​​measured at different time points or under different conditions. This indicator helps monitor the progression of the disease and the effectiveness of treatment.

[0106] FAI (perivascular fat attenuation index)

[0107] The perivascular fat attenuation index refers to the degree of X-ray attenuation of fat tissue surrounding blood vessels. This index is related to the severity of atherosclerosis and may affect the health of arteries.

[0108] LIPID_VOLUME (lipid component volume)

[0109] Lipid component volume refers to the volume of lipid components in plaques. Lipids are the main component of plaque formation. The larger the lipid component volume, the higher the instability of the plaque may be.

[0110] LIPID_VOLUME_RATIO (lipid volume ratio)

[0111] The lipid volume fraction refers to the proportion of lipid components in the total plaque volume. This ratio helps to assess the composition and possible instability of the plaque.

[0112] FIBROLIPID_VOLUME (Fiber lipid component volume)

[0113] The volume of fibrous lipid components refers to the total volume of fibrous tissue and lipid components in plaques. The volume of this component may affect the mechanical stability and rupture risk of plaques.

[0114] FIBROLIPID_VOLUME_RATIO (fiber lipid component volume ratio)

[0115] The volume ratio of fibrous lipid components refers to the proportion of fibrous lipid components in the total volume of plaques. This ratio helps to assess the structure of plaques and the possible risk of rupture, and is of great significance in guiding clinical treatment decisions.

[0116] By performing feature selection based on the optimized recursive feature elimination method, the most influential features for the risk prediction of adverse events of coronary heart disease are screened out to improve the accuracy and generalization ability of the model. The details are as follows:

[0117] Based on the preprocessed data set N = {X, Y}, where X is the feature matrix, X∈R n×p , n is the number of samples, p is the number of initial features; Y is the corresponding target variable vector, Y∈R n ;

[0118] Train a random forest model and calculate the average impurity reduction of each feature in the decision tree as the importance score; you can also calculate the Gini index or information gain as the importance score.

[0119] Train the gradient boosted tree model and calculate the split contribution of each feature in the gradient boosted tree as the importance score;

[0120] Calculate the average importance score of each feature in the random forest model and the gradient boosting tree model to calculate the comprehensive feature importance score. You can also use more complex methods such as weighted average to calculate the total and feature importance scores based on the performance or priors of different models, sort the comprehensive importance scores, and get the feature importance ranking result π sorted from large to small;

[0121] Set the initial feature selection ratio η0, for example, select the most important 50% features first, and the corresponding initial selection feature number m0 = η0 × p;

[0122] Define a performance evaluation indicator P, which can be the accuracy of the model on the validation set.

[0123] With the currently selected feature subset That is, the first m selected in order of importance j A subset of features, initially m j = m0, train a support vector machine model, or a linear regression model, and obtain the current model performance P j ;

[0124] Increase or decrease the number of features to update m j+1 :

[0125] Add features: Let m j+1 =m j +Δm + , where Δm + Select a new feature subset for a small positive integer increment, such as 1 or 2 Train the model and get performance

[0126] Reduce features: Let m j+1 =m j -Δm - , where Δm - Decrement by a small positive integer, such as 1 or 2, to select a new feature subset Train the model and get performance

[0127] Adaptive Updates:

[0128] if and Then update m j =m j+1 , that is, increasing the number of features is beneficial and continuing to the next iteration;

[0129] if and Then update m j =m j+1 , that is, reducing the number of features is beneficial and continuing to the next iteration;

[0130] if and Do not update m j , indicating that the current number of features is appropriate, stop the iteration or decide whether to continue based on other stopping conditions (such as reaching the maximum number of iterations, etc.);

[0131] The number of features m finally determined according to the above adaptive adjustment process end , select the corresponding feature subset As the final feature selection result.

[0132] S3: Used to train the extracted features through machine learning algorithms, build a risk prediction model for adverse events of coronary heart disease, and optimize the model; the details are as follows:

[0133] Each patient is represented as a node in the graph, and the feature vector of the node includes the feature selection result of the feature selection algorithm; edges are constructed according to the calibration relationship between patients, and the weight of the edge is determined based on the similarity measurement method. The initial graph structure is represented as G = (V, E), where V is the node set and E is the edge set;

[0134] The constructed patient relationship graph structure is initially learned and features extracted through graph neural networks, and the node feature representation is enhanced using the similarity information provided by the walk-based graph kernel function:

[0135] Construct the adjacency matrix A of the graph based on the initial graph structure, where A ab Indicates whether there is an edge connection between node a and node b. If there is an edge, then A ab =1 or A ab The value is the weight of the edge, otherwise A ab =0;

[0136] Calculate the node transition probability matrix R, where R ab Represents the transition probability from node a to node b:

[0137] R=D -1 A

[0138]

[0139] Where D is the degree matrix;

[0140] For each node a, define its random walk feature map in, is the probability distribution of a random walk path starting from node a and going through t steps;

[0141] For two nodes a and b, the graph kernel function based on random walk is defined as k(a,b)=<φ(a),φ(b)>, where <,> represents the inner product operation; the common calculation method of k(a,b) here is Among them, α∈(0,1) is the attenuation factor, which is used to control the contribution weights of random walk paths of different lengths.

[0142] For all nodes in the graph G, the graph core function values ​​between each two nodes are calculated to obtain the graph core matrix K, where K ab =k(a,b),a,b∈V; the final graph core matrix K is a symmetric matrix, and the elements on its diagonal represent the graph core function values ​​of node a itself.

[0143] For each node, its feature vector is updated at each time step t according to the features of its neighboring nodes and the weights of the edges;

[0144]

[0145] Among them, h v,l+1 、h v,l are the feature vectors of node v at the input of the l+1th layer network and the lth layer network respectively; σ is the activation function; w l , b l are the learning parameters of the lth layer network respectively; δ(u) is the set of neighbor nodes of node u; δ(v) is the set of neighbor nodes of node v;

[0146] The weight of the edge is dynamically adjusted based on the similarity of the subsequent examination data of the patient node; the node features and model parameters in the graph neural network are updated based on the gradient descent algorithm;

[0147] Take the structure and node features of the current graph as the state of the reinforcement learning agent, and define the environment state, action, reward, and state transition probability:

[0148] The environment state C represents all possible states of the environment in which the agent is located, that is, c = {c τ}, where c τ represents the state at time τ;

[0149] The action set Q represents the set of all actions of the agent, and the action q∈Q includes selecting different graph neural network parameter adjustment strategies;

[0150] The reward function H(c,q,c') represents the reward that the agent receives when taking action q in state c and transferring to state c'; the reward is set based on the accuracy of the model's prediction of adverse events of coronary heart disease. If the prediction accuracy of the model in the current state improves, the agent receives a positive reward; if the prediction accuracy decreases or the model performance deteriorates, the agent receives a negative reward;

[0151] The state transition probability F(c'|c,q) represents the probability of taking action q in state c to transition to state c';

[0152] Define strategy θ, strategy θ(q|c) represents the probability of taking action q in state c;

[0153] Define the state value function Vθ (c) are as follows:

[0154]

[0155] Among them, E θ [·] is the value function; λ∈[0,1] is the discount factor;

[0156] Define the action value function U θ (c) are as follows:

[0157]

[0158] The optimal value function is defined as:

[0159] Optimal state value function:

[0160] Optimal action-value function:

[0161] In this way, we can find the optimal strategy θ * , that is, for all states c∈C, the state value function corresponding to the optimal strategy is satisfied In finding the optimal strategy θ * In the process, the strategy is updated according to the reward feedback.

[0162] The dataset is divided into a training set and a test set, usually in a certain ratio (such as 70% training set and 30% test set). The selected machine learning model is trained using the training set, and the model parameters are adjusted using the gradient descent algorithm to optimize the model's performance. Cross-validation techniques (such as K-fold cross-validation) can be used to more accurately evaluate the model's performance and select the best parameter combination.

[0163] S4: Input new patient data into the trained model, and the model outputs the risk prediction results of the patient's adverse events of coronary heart disease in the next two years; the details are as follows:

[0164] The prediction results are output in a visual form for doctors’ reference. At the same time, risk factors are explained in depth to provide doctors with more accurate diagnostic support and treatment suggestions.

[0165] The prediction model built with the common graph neural network model is model 1, and the model integrated with reinforcement learning is model 2. The average prediction error statistics of the risk prediction model of adverse events of coronary heart disease in this embodiment are based on different data sets. The specific results can be referred to Figure 2The bar graphs corresponding to each data set correspond to the prediction model of the present application, model one and model two from left to right. It can be seen from the figure that the prediction results of the coronary heart disease adverse event risk prediction model of the present embodiment have a relatively small average prediction error compared with model one and model two, and have a higher prediction accuracy.

[0166] S5: Update the patient's latest data in real time and re-predict the risk. Details are as follows:

[0167] The latest patient data is updated in real time, and risk prediction is re-performed based on this data to ensure the accuracy and timeliness of risk assessment results. Doctors can also provide real-time feedback on risk assessment results based on the actual situation of the patient, and continuously optimize and improve the model.

[0168] Embodiment 2:

[0169] A multi-parameter coronary heart disease adverse event risk prediction method based on machine learning in Application Example 1 is used to build a multi-parameter coronary heart disease adverse event risk prediction system for predicting the risk of coronary heart disease adverse events. By extending and analyzing the patient's medical imaging data and personal health history, it can accurately assess the patient's possibility of experiencing coronary heart disease adverse events within the next two years.

[0170] The corresponding system may include the following modules:

[0171] User Interface (UI) Module

[0172] Function: Provides a simple and intuitive web interface for doctors to input patients' medical imaging and health history information.

[0173] Connection relationship: directly connected to the data input module to receive data entered by the user.

[0174] Data input module

[0175] Function: Receive and store patient data from the user interface module, including medical imaging data and personal health history data.

[0176] Connection relationship: connected with the user interface module and the data preprocessing module, passing the received data to the data preprocessing module.

[0177] Data preprocessing module

[0178] Function: Clean, standardize and format the input data to ensure data quality and consistency.

[0179] Connection relationship: Connected with the data input module and the feature extraction module, passing the preprocessed data to the feature extraction module.

[0180] Feature extraction module

[0181] Function: Extract key imaging and health history factors from preprocessed data.

[0182] Connection relationship: connected with the data preprocessing module and the feature selection module, passing the extracted features to the feature selection module.

[0183] Feature Selection Module

[0184] Function: Through feature selection algorithm, the most influential features for risk prediction of adverse events of coronary heart disease are screened out.

[0185] Connection relationship: Connect with the feature extraction module and the model training module to pass the filtered features to the model training module.

[0186] Model training module

[0187] Function: Use machine learning algorithms to train the screened features and build a risk prediction model for adverse events of coronary heart disease.

[0188] Connection relationship: Connect with the feature selection module and the model optimization module to pass the trained model to the model optimization module.

[0189] Model Optimization Module

[0190] Function: Optimize the model through cross-validation, hyperparameter tuning and other methods to improve the prediction performance and stability of the model.

[0191] Connection relationship: Connect with the model training module and the risk assessment module to pass the optimized model to the risk assessment module.

[0192] Model prediction module

[0193] Function: Input new patient data into the optimized model and output the risk assessment results of patients experiencing adverse coronary heart disease events in the next two years.

[0194] Connection relationship: connected with the model optimization module and the result output module, and the risk assessment results are passed to the result output module.

[0195] Result output module

[0196] Function: Output the risk assessment results in a visual form for doctors' reference.

[0197] Connection relationship: Connect with the risk assessment module and the user interface module to display the output results in the user interface module.

[0198] Real-time update and feedback module

[0199] Function: Update the latest patient data in real time and re-evaluate risk. At the same time, receive feedback from doctors to help the system continuously optimize and improve.

[0200] Connection relationship: Connect with the data input module, risk assessment module and user interface module to realize real-time update and feedback of data.

[0201] Reference Figure 3 When the system is started, the software interface will be clearly divided into three main areas: left, right, and bottom. Each area carries different functions and information to ensure that doctors can quickly get started.

[0202] The area on the left is the data element entry area, which allows doctors to quickly and accurately enter patient imaging elements and personal health elements (including Chinese and English descriptions, input element units, etc.).

[0203] The area on the right is the model prediction result area. Based on the background of the factor model, the probability of the current patient suffering from adverse events of coronary heart disease is calculated and the top 5 main factors affecting the prediction results are given, and displayed in the form of a SHAP legend. This area is the core function of the system. It is necessary to predict the risk probability of the patient suffering from adverse events of coronary heart disease based on the patient's image elements and personal health elements input by the user, through the machine learning model that has been trained in the background, and give the top 5 most influential factors, and call the large language model to compare and explain the patient's predicted risk probability and judgment threshold. First enter the elements and click Prediction, you can refer to Figure 4 , the model prediction gives the results and analysis, please refer to Figure 5-6 , and finally give corresponding reference suggestions, which can be referred to Figure 7 ,.

[0204] The lower area is the simulation sample selection area, which gives three rows of test data by default, so that doctors can quickly evaluate the operation and results of the software. Just click on the corresponding row, and the data in that row will be automatically filled in the information entry area on the left. Then click the prediction button (predict) to see the prediction results on the right. This area gives three rows of test data, so that doctors can quickly evaluate the operation and results of the software. Just click on the corresponding row, and the data in that row will be automatically filled in the information entry area on the left. Then click the prediction button (predict) to see the prediction results on the right.

[0205] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A multi-parameter coronary heart disease adverse event risk prediction method based on machine learning, characterized in that: The steps include: S1: Collect patients’ medical imaging data and personal health history data, and clean, standardize and format the data; S2: Extract key imaging factors and health history factors from the preprocessed data, and use feature selection algorithms to screen out the most influential features for the risk prediction of adverse coronary events; S3: used to train the extracted features through machine learning algorithms, build a risk prediction model for adverse events of coronary heart disease, and optimize the model; S4: Input new patient data into the trained model, and the model outputs the risk prediction results of the patient experiencing adverse coronary heart disease events in the next two years; S5: Update the patient’s latest data in real time and re-predict risk.

2. The method for predicting the risk of adverse events of coronary heart disease based on multi-parameter machine learning according to claim 1, characterized in that: The key imaging factors and health history factors extracted in S2 include ALCOHOL_HISTORY, PLAQUE_VOLUME, PLAQUE_LENGTH, MIN_LUMINAL_AREA, MIN_NARROWING_PERCENTAGE, HIGH_RISK_NAPKIN_RING_SIGN, HIGH_RISK_POSITIVE_REMODEL, HIGH_RISK_LARGE_PLAQUE_VOLUME, CT_FFR, DELTA_CT_FFR, FAI, LIPID_VOLUME, LIPID_VOLUME_RATIO, FIBROLIPID_VOLUME, and FIBROLIPID_VOLUME_RATIO.

3. The method for predicting the risk of adverse events of coronary heart disease based on multi-parameter machine learning according to claim 2, characterized in that: In S2, feature selection is performed based on an optimized recursive feature elimination method, as follows: Based on the preprocessed data set N = {X, Y}, where X is the feature matrix, X∈R n×p , n is the number of samples, p is the number of initial features; Y is the corresponding target variable vector, Y∈R n ; Train a random forest model and calculate the average impurity reduction of each feature in the decision tree as the importance score; Train the gradient boosted tree model and calculate the split contribution of each feature in the gradient boosted tree as the importance score; Calculate the average importance score of each feature in the random forest model and the gradient boosting tree model to calculate the comprehensive feature importance score, sort the comprehensive importance scores, and obtain the feature importance ranking result π sorted from large to small; Set the initial feature selection ratio η0, and the corresponding initial selection feature number m0 = η0 × p; Define a performance evaluation index P based on the currently selected feature subset That is, the first m selected in order of importance j A subset of features, initially m j =m0, train a support vector machine model and get the current model performance P j ; Increase or decrease the number of features to update m j+1 : Add features: Let m j+1 =m j +Δm + , where Δm + For a small positive integer increment, select a new feature subset Train the model and get performance Reduce features: Let m j+1 =m j -Δm - , where Δm - Decrement by a small positive integer, selecting a new subset of features Train the model and get performance Adaptive Updates: if and Then update m j =m j+1 , that is, increase the number of features and continue to the next iteration; if and Then update m j =m j+1 , that is, reduce the number of features and continue to the next iteration; if and Do not update m j , stop the iteration or decide whether to continue based on other stopping conditions; The number of features m finally determined according to the above adaptive adjustment process end , select the corresponding feature subset As the final feature selection result.

4. The method for predicting the risk of adverse events of coronary heart disease based on multi-parameter machine learning according to claim 1, characterized in that: In S3, the risk prediction model for adverse events of coronary heart disease is constructed as follows: Each patient is represented as a node in the graph, and the feature vector of the node includes the feature selection result of the feature selection algorithm; edges are constructed according to the calibration relationship between patients, and the weight of the edge is determined based on the similarity measurement method. The initial graph structure is represented as G = (V, E), where V is the node set and E is the edge set; The constructed patient relationship graph structure is initially learned and features extracted through graph neural networks, and the node feature representation is enhanced using the similarity information provided by the walk-based graph kernel function: Construct the adjacency matrix A of the graph based on the initial graph structure, where A ab Indicates whether there is an edge connection between node a and node b. If there is an edge, then A ab =1, otherwise A ab =0; Calculate the node transition probability matrix R, where R ab Represents the transition probability from node a to node b: R=D -1 A Where D is the degree matrix; For each node a, define its random walk feature map in, is the probability distribution of a random walk path starting from node a and going through t steps; For two nodes a and b, the random walk-based graph kernel function is defined as k(a,b)=<φ(a),φ(b)>, where <,> represents the inner product operation; For all nodes in the graph G, the graph core function values ​​between each node are calculated to obtain the graph core matrix K, where K ab =k(a,b),a,b∈V; For each node, its feature vector is updated at each time step t according to the features of its neighboring nodes and the weights of the edges; Among them, h v,l+1 、h v,l are the feature vectors of node v at the input of the l+1th layer network and the lth layer network respectively; σ is the activation function; w l , b l are the learning parameters of the lth layer network respectively; δ(u) is the set of neighbor nodes of node u; δ(v) is the set of neighbor nodes of node v; The weight of the edge is dynamically adjusted based on the similarity of the subsequent examination data of the patient node; the node features and model parameters in the graph neural network are updated based on the gradient descent algorithm; Take the structure and node features of the current graph as the state of the reinforcement learning agent, and define the environment state, action, reward, and state transition probability: The environment state C represents all possible states of the environment in which the agent is located, that is, c = {c τ }, where c τ represents the state at time τ; The action set Q represents the set of all actions of the agent, and the action q∈Q includes selecting different graph neural network parameter adjustment strategies; The reward function H(c,q,c') represents the reward that the agent receives when taking action q in state c and transferring to state c'; the reward is set based on the accuracy of the model's prediction of adverse events of coronary heart disease. If the prediction accuracy of the model in the current state improves, the agent receives a positive reward; if the prediction accuracy decreases or the model performance deteriorates, the agent receives a negative reward; The state transition probability F(c'|c,q) represents the probability of taking action q in state c to transition to state c'; Define strategy θ, strategy θ(q|c) represents the probability of taking action q in state c; Define the state value function V θ (c) are as follows: Among them, E θ [·] is the value function; λ∈[0,1] is the discount factor; Define the action value function U θ (c) are as follows: The optimal value function is defined as: Optimal state value function: Optimal action-value function: In this way, we can find the optimal strategy θ * , that is, for all states c∈C, the state value function corresponding to the optimal strategy is satisfied In finding the optimal strategy θ * In the process, the strategy is updated according to the reward feedback.

5. The method for predicting the risk of adverse events of coronary heart disease based on multi-parameter machine learning according to claim 1, characterized in that: In S4, the prediction results of adverse coronary heart disease events occurring in the patient in the next two years are output in a visualized form.

6. The method for predicting the risk of adverse events of coronary heart disease based on multi-parameter machine learning according to claim 5, characterized in that: In S4, risk factors are also explained in depth.

7. The method for predicting the risk of adverse events of coronary heart disease based on multi-parameter machine learning according to claim 1, characterized in that: In S5, the prediction results are fed back according to the actual situation of the patient through a feedback mechanism, and the prediction model is optimized.

Images