CN118398208A

CN118398208A - Atrial fibrillation risk assessment method based on machine learning

Info

Publication number: CN118398208A
Application number: CN202410496449.5A
Authority: CN
Inventors: 应航鹰; 范航平; 王耀; 李煅斌
Original assignee: Sir Run Run Shaw Hospital
Current assignee: Sir Run Run Shaw Hospital
Priority date: 2024-04-24
Filing date: 2024-04-24
Publication date: 2024-07-26

Abstract

The invention relates to an atrial fibrillation risk assessment method based on machine learning. Firstly, collecting and preprocessing multi-source data including Electrocardiogram (ECG), echocardiogram, life style and clinical medical records, wherein the preprocessing step comprises data cleaning, denoising, normalization and missing value processing; then, carrying out feature engineering, including extracting the shape and time features of QRS complex, P wave and T wave from an electrocardiogram, extracting the size and function index of the left atrium from an echocardiogram, and selecting the feature with high predictive value by using a feature selection algorithm; training a single-mode machine learning model for each data type respectively, and adopting a self-adaptive multi-mode fusion algorithm to dynamically adjust the weight of each mode data according to the data characteristics and the quality of different patients so as to optimize the overall prediction performance of the model; finally, algorithm structures and parameters are adjusted through cross validation and algorithm iteration to improve prediction accuracy and generalization capability, wherein the optimization of model parameters and the prevention of overfitting are included.

Description

Atrial fibrillation risk assessment method based on machine learning

Technical Field

The invention relates to a risk assessment method, in particular to a machine learning-based atrial fibrillation risk assessment method.

Background

Current methods for atrial fibrillation (atrial fibrillation) risk assessment include traditional clinical scoring systems, electrocardiographic analysis, and detection of some biomarkers. Although these methods are widely used clinically, there are a number of shortcomings and drawbacks, and limitations of these methods are set forth in detail below. First, traditional clinical scoring systems such as the CHA2DS2-VASc score and the HAS-ble score, which evaluate a patient's risk of developing atrial fibrillation and risk of possibly requiring anticoagulation therapy based primarily on the patient's age, medical history (e.g., hypertension, diabetes, cardiac history, etc.), and other health conditions (e.g., previous stroke or bleeding events). The main disadvantage of these scoring systems is that they are based mainly on statistical data, and cannot adequately take into account individual biomarker differences and lifestyle factors, resulting in limited prediction accuracy and individualization levels. Furthermore, these scoring systems cannot be updated dynamically, not reflecting immediate changes in patient health status, which may lead to overestimation or underestimation of risk in clinical practice.

Second, electrocardiogram (ECG) is one of the common methods of assessing cardiac arrhythmias, particularly in the identification and diagnosis of atrial fibrillation. However, conventional ECG analysis relies on the physician or technician identifying specific waveforms and intervals on the electrocardiogram, which may be subject to subjective judgment with certain errors. While automated electrocardiographic analysis systems exist, these systems generally only recognize standard, typical atrial fibrillation waveforms, and the accuracy of their identification is still limited for atypical waveforms or low-amplitude atrial fibrillation waveforms. Furthermore, an electrocardiogram is a time-dependent test that can only provide electrocardiographic information at the time of the test, and has limited capture capability for intermittent atrial fibrillation. Third, biomarkers currently used in risk assessment of atrial fibrillation, such as myocardial injury markers (cardiac troponin) and cardiac function indicators (type B diuretics), while providing useful information for risk assessment of cardiac patients, are not specific to atrial fibrillation. The level of biomarkers may be affected by a variety of factors, such as renal insufficiency, other heart diseases, etc., and are not specific to atrial fibrillation. Therefore, these markers have limitations on the accuracy of predicting the occurrence of atrial fibrillation, and do not fully accurately reflect the risk of atrial fibrillation.

Furthermore, current risk assessment methods also have shortcomings in terms of data processing and analysis. Many assessment methods fail to take full advantage of modern computing techniques and algorithms, such as big data analysis, machine learning, etc., which have the potential to improve the accuracy of risk assessment by analyzing large amounts of data and identifying complex patterns. This is often ignored by existing methods, resulting in poor efficiency and effectiveness in processing large or complex data. Risk assessment of atrial fibrillation also requires consideration of the overall health condition of the patient and various lifestyle factors such as diet, physical activity, mental health, etc., but the existing assessment methods often only pay attention to specific medical indexes and histories, and neglect comprehensive factors which may influence atrial fibrillation risk. Limitations of this approach may lead to underassessment of the actual risk of the patient, thereby affecting the formulation of preventive and therapeutic strategies.

Disclosure of Invention

The invention aims to provide a machine learning-based atrial fibrillation risk assessment method, so that part of defects and shortages pointed out in the background art are overcome.

The technical scheme adopted by the invention for solving the technical problems comprises the following steps: firstly, collecting and preprocessing multi-source data including Electrocardiogram (ECG), echocardiogram, life style and clinical medical records, wherein the preprocessing step comprises data cleaning, denoising, normalization and missing value processing;

Then, carrying out feature engineering, including extracting the shape and time features of QRS complex, P wave and T wave from an electrocardiogram, extracting the size and function index of the left atrium from an echocardiogram, and selecting the feature with high predictive value by using a feature selection algorithm;

Training a single-mode machine learning model for each data type respectively, and adopting a self-adaptive multi-mode fusion algorithm to dynamically adjust the weight of each mode data according to the data characteristics and the quality of different patients so as to optimize the overall prediction performance of the model;

finally, algorithm structures and parameters are adjusted through cross validation and algorithm iteration to improve prediction accuracy and generalization capability, wherein the optimization of model parameters and the prevention of overfitting are included.

Further, the data cleansing process identifies and corrects outliers and errors in the data by applying a real-time quality assessment model based on a function:

where σ (x _i) is a function of scoring a single data point x _i, α and β are adjustment parameters;

in the denoising step, a denoising formula based on time series analysis is adopted:

designed for electrocardiogram and echocardiographic signals to retain physiological information;

Normalization uses an adaptive approach, normalizes the dataset by computing real-time statistical properties μ (x) and σ (x), and increases the tuning parameter γ to optimize the distribution balance;

The missing value processing uses the supplementary information of the modal data, and the formula is as follows:

M(x₁,x₂)＝x₁·ω+x₂·(1-ω)

Performing filling, wherein ω is a weight adjusted based on the correlation;

Next, a machine learning model is trained using the preprocessed data that includes a single-modality and multi-modality fusion algorithm that passes through an adaptive weight adjustment formula:

The method comprises the steps of realizing that the weight w _i is dynamically adjusted according to the quality of each data source and the contribution to prediction;

Finally, cross-validation and parameter optimization are performed on the model, using a cost function:

the model parameter theta is adjusted, and the prediction accuracy and the generalization capability of the model are improved.

Further, the feature engineering is to perform feature extraction of Electrocardiogram (ECG) and echocardiogram by using a deep learning model, wherein the central electrogram feature extraction uses a model based on a deep Convolutional Neural Network (CNN) to detect and separate QRS complex, P wave and T wave, and the extracted waveform includes fine features of peak shape and trough feature;

The ultrasonic cardiogram feature extraction adopts a combined model of a convolutional neural network and a long-short time memory network (LSTM) to analyze an ultrasonic video sequence, and identifies and quantifies the size and functional index of the left atrium, including the expansion and contraction speed and the volume change rate of the left atrium; next, using a multi-scale feature analysis technique to consider features of the electrocardiograph and echocardiography data at different time windows and frequency resolutions;

And finally, dynamically selecting the feature combination for predicting the atrial fibrillation risk information value by combining the feature importance scores based on a gradient hoisting machine (GBM) or a random forest ensemble learning method.

Further, the ECG signal is preprocessed, including normalization to ensure consistency of the input data, using the formula:

removing high-frequency noise and low-frequency drift in the signal, wherein mu and sigma are filtering parameters;

A deep convolutional neural network comprising a plurality of convolutional layers and a pooling layer is then employed, wherein the convolutional layers use the formula:

C_n(x)＝ρ(ω*x+b)

Feature extraction is carried out, ω is a convolution kernel, which represents convolution operation, b is bias, ρ is a ReLU activation function;

the attention mechanism formula is introduced into the CNN:

to enhance learning of key features of QRS complex, P-wave and T-wave, β being a parameter that adjusts the intensity of attention;

finally, by introducing an output layer and a post-processing algorithm, the formula is utilized:

To refine the detection of waveform features, where a is the effective interval of the waveform and k is an adaptively adjusted parameter for optimizing the estimation of waveform features including time position and amplitude.

Further, the echocardiography feature extraction includes the following steps:

Firstly, preprocessing an echocardiographic video sequence, including frame extraction and scaling, standardization processing, adopting a new formula:

S(x)＝log(1+|x|^γ)

Image enhancement is carried out to improve the visibility of key structures including a left room in the image;

then, a deep Convolutional Neural Network (CNN) is designed and applied to extract the characteristics of each frame of image, and the formula is used:

C_l(x)＝σ(W_l*x+b_l)

Including multiple convolution and pooling layers to capture fine changes in left atrial size and morphology;

The output of the CNN is then processed using a long and short term memory network (LSTM), by an iterative formula:

L_t＝φ(W_f·[h_t-1,x_t]+b_f)

Learning and extracting dynamic functional changes of the left atrium including expansion and contraction rates and volume change rates in the echocardiogram;

finally, combining the outputs of CNN and LSTM, adopting a multi-scale feature analysis technology, and passing through an integral formula:

features of different time windows and frequency resolutions are integrated to improve accuracy of atrial fibrillation risk prediction.

Further, the multi-modal machine learner includes:

firstly, training a single-mode machine learning model for an Electrocardiogram (ECG) and an echocardiography data type respectively; for electrocardiographic data, a deep Convolutional Neural Network (CNN) is used:

M_ECG(x)＝σ(W*x+b)

wherein sigma is a nonlinear activation function, and W and b represent the weight and bias of the model respectively and are used for extracting the characteristics of the QRS complex wave, the P wave and the T wave;

For echocardiographic data, a long-short memory network (LSTM) is used:

M_Echo(x_t)＝Γ(h_t-x,x_t)

Γ denotes an LSTM processing unit for analyzing time-series data of left atrium size and function index;

Then through an adaptive multi-modal fusion algorithm:

The weight alpha _i of each mode data is dynamically adjusted to optimize the performance of the overall prediction model, wherein y _i represents the output of the ith mode model.

Further, the cross-validation and algorithm iterative adjustment algorithm structure and parameter method comprises the following steps:

first a hierarchical cross-validation technique is utilized, wherein the function is used:

Ensuring that each cross-validated subset is consistent with the global dataset over a distribution of cardiac disease categories;

Secondly, using a dynamically adjusted network structure according to specific characteristics of heart diseases, and optimizing a scoring function obtained by an algorithm through a genetic algorithm:

Selecting a network structure to capture atrial fibrillation characteristics in electrocardiography and echocardiography;

then, adjusting model parameters by using Bayesian optimization, and optimizing targets:

where θ represents a model parameter and D represents a cardiac disease dataset, particularly atrial fibrillation related data.

The invention provides various beneficial effects, and key advantages include:

1. The prediction accuracy is improved: complex Electrocardiogram (ECG) and echocardiographic data can be more accurately analyzed and interpreted using advanced machine learning models, such as Convolutional Neural Networks (CNNs) and long-short-term memory networks (LSTM). This approach is capable of identifying subtle patterns and changes, which are critical for early detection and prediction of atrial fibrillation.

2. Comprehensive multi-modal data analysis: by combining different types of medical image and physiological signal data, the invention can comprehensively evaluate the heart health condition of the patient. The multi-modal fusion method improves the richness of data analysis and the comprehensiveness of evaluation, thereby enhancing the capability of predicting atrial fibrillation risks.

3. Dynamic parameter adjustment: by adopting advanced algorithms such as Bayesian optimization to dynamically adjust model parameters, the invention can adaptively optimize the performance of the model, adjust the risk assessment model for specific patients or patient groups, and ensure the accuracy and individuation of the assessment result.

4. Real-time risk assessment: the application of a machine learning model allows for real-time atrial fibrillation risk assessment, supporting the immediate generation of risk scores as patients are examined. This helps the physician to quickly respond in clinical decisions, with early intervention.

Drawings

Fig. 1 is a flowchart of a machine learning-based atrial fibrillation risk assessment method of the present invention.

FIG. 2 is a flow chart of the implementation of the feature engineering of the present invention.

Detailed Description

The following describes the embodiments of the present invention in detail with reference to the drawings.

Machine learning-based atrial fibrillation risk assessment methods involve collecting multi-source data from different sources and performing a series of preprocessing steps to prepare the data for analysis: firstly, collecting various data including Electrocardiogram (ECG), echocardiogram, life style data, clinical medical records and the like, wherein the data has various sources, and each data provides different visual angles and information for evaluating atrial fibrillation risks; next, in the preprocessing stage, firstly, data cleaning is performed, and errors, repeated records or irrelevant records in the data are removed so as to ensure the quality of subsequent analysis; data denoising is another key step, particularly biomedical signals such as electrocardiogram and echocardiogram, denoising ensures that interferences in the signals, such as electrical noise or interferences introduced by operation errors, are removed, and clearer and more accurate biological signals are reserved; the normalization process is a transformation of data from different sources to the same scale, which is critical because it allows data from different measurement devices and technologies to be analyzed together in the same model, while also helping machine learning algorithms to learn and identify patterns in the data more efficiently; finally, it is also important to process missing values, including filling in missing data or deciding how to process unrecoverable information, and correct processing missing values can prevent deviation of data analysis results, ensuring accuracy and reliability of the model.

Feature engineering is a key link that involves extracting features of key biomedical signals from electrocardiography and echocardiography, which are the basis for subsequent machine learning model analysis: firstly, the shape and time characteristics of the QRS complex, the P wave and the T wave are extracted from an electrocardiogram, the waveforms reflect the electrophysiological activity of the heart, the QRS complex represents the depolarization of a ventricle, the P wave and the T wave represent the depolarization of the atrium and the repolarization of the ventricle respectively, and the potential dysfunction of the heart can be revealed by precisely quantifying the shape, the duration, the peak value, the interval and other attributes of the waves; second, features extracted from echocardiography include the size and functional indicators of the left atrium, such as the volume, expansion and contraction rate of the left atrium, which are critical to assessing the load and functional status of the left atrium, as left atrial dysfunction is a key factor in the development of atrial fibrillation; feature selection algorithms, such as information gain based, model based selection, or recursive feature elimination, are then used to identify and select the most valuable features for atrial fibrillation risk prediction.

Subsequent to the feature engineering, the next step is to train a separate modal machine learning model for each data type, such as electrocardiogram, echocardiogram, lifestyle data, and clinical medical records, respectively, which can be specifically learned and optimized for each data type specific characteristics and patterns; then, an adaptive multi-mode fusion algorithm is adopted, the weight of each mode data can be dynamically adjusted according to the data characteristics and the quality of different patients by the algorithm, so that the overall prediction performance of the model is optimized, and the process is particularly important, because the system is allowed to flexibly adjust according to the quality and the integrity of available data in practical application, the data of different sources and types can be effectively combined for use when predicting the atrial fibrillation risk, and the prediction accuracy and reliability are improved; the adaptive multi-mode fusion algorithm enables the whole prediction system to adjust the behavior of the whole prediction system according to specific situations by learning how to best combine the information of different data sources, so that the whole prediction system is efficient and accurate in the face of diversified and complex medical scenes.

The final step is to adjust algorithm structure and parameters through cross-validation and algorithm iteration, which is to improve the prediction accuracy and generalization capability of the model; through cross verification, the model is trained and verified on different data subsets, so that the processing capacity of the model on new data can be evaluated, the model is ensured to perform well on specific training data, and meanwhile, stable performance can be maintained on unseen data; algorithm iterative adjustments include optimizing model structures such as layer numbers, node numbers, activation functions, etc., and adjusting learning parameters such as learning rate, regularization strength, etc., based on the model's performance in a cross-validation process, in order to find an optimal model configuration to maximize its effectiveness; in addition, optimization of model parameters includes employing advanced optimization algorithms such as momentum, adaGrad, RMSprop, etc. to accelerate the learning process and improve the convergence of model training.

Example 1:

data processing in this embodiment is the primary and critical steps, including data cleaning and denoising processes, which ensure accuracy and reliability of subsequent analysis. The implementation and effectiveness of these steps is illustrated by way of example.

A set of Electrocardiogram (ECG) and echocardiographic data is set from a middle-aged male patient at risk of atrial fibrillation, which data is collected by different medical devices during multiple visits. In order to accurately evaluate the atrial fibrillation risk, data cleaning is firstly required by a real-time quality evaluation model.

In this example, the function is defined:

To identify and correct outliers and errors in the data. Where σ (x _i) is a function of scoring a single data point x _i and α and β are adjustment parameters. In practical applications, α can be set between 0.1 and 0.5 to weight the effect of the data points; beta may be set between 1 and 10 to ensure that the model has some tolerance for extremes.

For example, for a series of data points in an electrocardiogram, if a value is found to be significantly above or below the surrounding value (i.e., outlier), the real-time quality assessment model Q (x) will mark those points and modify or delete the data accordingly, thereby cleaning the data set. Next, in the denoising step, a denoising formula based on time series analysis is used:

This formula is particularly useful for processing noise in electrocardiography and echocardiography, such as electrical interference or noise generated during operation. This denoising process can effectively preserve important physiological information in the lower electrocardiogram and echocardiogram, such as heart rate, QRS complex, and waveform characteristics of left atrial size and function, while removing non-physiological noise. By such processing, even in the case of different data quality, it can be ensured that the data analyzed each time is purified and standardized, thereby making the model training based on machine learning more accurate and efficient. The prediction accuracy of the model is improved, and the generalization capability of the model to new and unseen data is enhanced.

The next step in this embodiment is to normalize and missing values of these preprocessed data, to ensure the integrity and consistency of the data preparation stage for the machine learning model to learn effectively.

First, normalization is achieved by an adaptive method. This method relies on calculating in real time the statistical properties of the data set, such as the mean μ (x) and standard deviation σ (x). Specifically, each data point x will passSuch formulas are transformed to ensure standardization of data distribution. Here, an adjustment parameter γ, typically set between 0.01 and 0.1, is also introduced to fine tune the dimensions of each feature, optimizing the distribution balance of the overall dataset. This step is critical because it ensures that different data features have the same influence in the model, avoiding that some larger value features dominate the final learning process.

Next, for the missing value processing, the supplementary information of the modality data is utilized. In a multi-modal data environment, different data sources have complementarity, such as that the electrocardiographic data is incomplete at certain times, and the corresponding echocardiographic data is intact at those times. Therefore, a simple and efficient interpolation formula is employed:

M(x₁,x₂)＝x₁·ω+x₂·(1-ω)

wherein ω is a weight adjusted based on the correlation between two data, and the range of values may be 0.3 to 0.7, dynamically adjusted according to the correlation between actual data. For example, if the correlation between electrocardiographic and echocardiographic data is high, ω can be set close to 0.7, thus relying more on electrocardiographic data for filling of missing values. Through these elaborate steps, the data is effectively standardized and complemented, laying a solid foundation for subsequent feature extraction, model training, and final atrial fibrillation risk assessment.

These preprocessed data are prepared for training machine learning models containing both single-modality and multi-modality fusion algorithms to assess atrial fibrillation risk. The key to this stage is how to adjust the formula by adaptive weights:

and integrating data of different modes to optimize the overall prediction performance.

First, each data type (electrocardiogram, echocardiogram, clinical record, etc.) will train the model alone to capture its specific information and features, which is referred to as unimodal model training. For example, electrocardiographic data is more focused on capturing electrophysiological patterns of the heart, while echocardiographic data provides visual and quantitative information about the structure and function of the heart.

Then, to integrate the outputs of these unimodal models and maximize their complementarity, a multimodal fusion algorithm is applied. In this fusion process, an adaptive weight adjustment formula is used:

Where x _i represents the output of the ith modality model and w _i is the corresponding weight, dynamically adjusted according to the quality of each data source and the contribution to the predicted outcome. The specific value range of the weight can be set to 0.1 to 0.9, and the optimization is carried out according to the actual data quality and the prediction contribution. For example, if the quality of the electrocardiographic data is very high and the diagnostic relevance to atrial fibrillation is strong, the weight w _i thereof may be set high, such as 0.8 or higher; while some clinical records have unstable data quality with lower weights, such as 0.2 or less.

Such dynamic adjustment of weights not only allows the model to exhibit a high degree of flexibility and adaptability to different patients and different situations, but also improves overall prediction accuracy. For example, by analyzing electrocardiographic and echocardiographic data of a specific patient, when an electrocardiographic display atypical heartbeat pattern and echocardiography are found to reveal abnormal left atrial enlargement, the multi-modal fusion model can integrate the information and adjust weights to reflect higher atrial fibrillation risk, thereby providing important information for supporting decision-making for clinic.

In order to ensure that the atrial fibrillation risk assessment model based on machine learning not only has good performance on training data, but also has good generalization capability, the method enters a key stage of model verification and parameter optimization.

The core of this stage is the application of cross-validation and algorithmic adjustment, by means of a cost function:

To optimize the model parameters θ, where m is the total number of samples in the validation dataset, y _j is the actual label for sample j, h _θ(x_j) is the predicted output of the model for sample j. The cost function measures the error between the model predictions and the actual data, with the goal of minimizing this error by adjusting the parameter θ.

The feasibility of this procedure is illustrated in a practical example: the set data set includes electrocardiogram and echocardiographic data of 1000 patients. The dataset was randomly split into 10 parts for 10 fold cross validation. In each round of verification, 9 parts are used for training and 1 part is used for testing. In this way, the behavior of the model on unseen data can be evaluated, ensuring its generalization ability.

During each round of training, a cost function C (θ) is calculated and θ is adjusted by a gradient descent or other optimization algorithm. The initial cost value is set to be 0.25, and the cost value is reduced to be below 0.05 through parameter adjustment of the system, so that the prediction error is obviously reduced, and the prediction accuracy of the model is obviously improved.

Furthermore, the optimization of the model parameters θ is not only a simple tuning value, but also involves changing the structure of the model, such as increasing or decreasing the number of network layers, tuning the activation function, etc. The flexible adjustment strategy, combined with the cross-validation results, enables comprehensive assessment of the impact of different configurations on model performance, thereby finding the optimal model structure and parameter settings. Through these steps, not only is high precision of the model on known data ensured, but also the validity and reliability of the model on new and unknown data are confirmed through a strict verification process.

Example 2:

Feature extraction is a critical step, especially in the processing of complex medical data such as Electrocardiographs (ECG) and echocardiography. The present example details how key features of electrocardiographic data are extracted by technical means around the application of a deep learning model and pre-processing is performed to ensure data quality and consistency.

First, preprocessing of the electrocardiogram signal includes two key steps: normalization and bandpass filtering. The normalization process ensures that all input data has a uniform scale before entering the deep learning model, a step that is critical for subsequent model training and feature extraction. The band-pass filter is used to remove unnecessary noise in the signal, using the formula:

implementation, where μ and σ are filter parameters, typically μ can be set within the main frequency range of the electrocardiogram signal, e.g., 1Hz (for removing baseline wander), while σ is responsible for adjusting the bandwidth of the filter, and the value range can be set to 0.5 to 10Hz, which can effectively remove high frequency noise without losing important information of the signal.

Next, the ECG is feature extracted using a deep Convolutional Neural Network (CNN) based model. The CNN model learns the details and hierarchy in the signal through multi-layered convolution and pooling layers, which is well suited for processing image and time series data. At this stage, the model focuses on detecting and separating QRS complexes, P-waves and T-waves in the electrocardiogram, which are key indicators for diagnosing atrial fibrillation. By designing specific convolution kernels and network structures, CNNs are able to identify peak shapes and valley features of these waveforms, such as peak width, height, spacing, etc.

For example, a set of electrocardiographic data collected from atrial fibrillation patients and non-atrial fibrillation patients is set. By applying the preprocessing and CNN models described above to these data, the characteristics of a particular waveform can be accurately extracted from the ECG signal. These features can then be used to train a classifier that can predict whether the patient is at risk of atrial fibrillation based on the learned waveform features. By the method, the quality of data and the input consistency of a model are improved, valuable medical information can be effectively extracted from complex electrocardiogram signals, and support is provided for early diagnosis and risk assessment of atrial fibrillation.

The present embodiment utilizes preprocessed Electrocardiogram (ECG) data. At this stage, a deep Convolutional Neural Network (CNN) comprising a plurality of convolutional layers and pooled layers is employed, which network architecture is particularly suitable for processing complex data having spatial or temporal correlation, such as electrocardiographic signals.

In the CNN model, the convolution layer uses the formula:

C_n(x)＝ρ(ω*x+b)

To extract features where ω is the convolution kernel, where x represents the convolution operation, b is the bias, and ρ is the ReLU activation function. ReLU (RectifiedLinearUnit) the activation function is a common function for increasing the nonlinearity of the network, which sets all negative values to zero, helps to solve the problem of gradient extinction during training, and accelerates the convergence speed of the network. In particular to the convolution layer, 8 convolution kernels are used, each having a size of 3x3, a step size of 1, and no padding (padding), which is provided to aid in capturing detailed features in the ECG data, such as the shape and size of the QRS complex.

The pooling layer follows the convolution layer and is typically used to reduce feature dimensions and enhance the robustness of the features. Here, using the max-pooling operation, a typical pooling window size is 2x2, which can reduce the amount of computation and avoid overfitting while preserving the most important features.

In order to further improve the learning ability of the model on key features of QRS complex, P wave and T wave in electrocardiogram, an attention mechanism formula is introduced in CNN:

The attention mechanism here enhances the model's attention to these key electrocardiographic features by adjusting the parameter β. For example, setting β in the range of 0.1 to 1 can effectively regulate the degree of focus of attention, and higher β values can make the network more focused on specific, high impact features. Through such a process, the deep learning model can accurately identify abnormal QRS complexes, which are indicative of the risk of atrial fibrillation in a patient. By careful training and optimization of the model, in combination with focusing of the attention mechanism, a highly accurate predictive tool can ultimately be provided in clinical applications.

The next step in this embodiment is to further refine the detection of these features by introducing an output layer and post-processing algorithms to ensure that the information extracted from the medical signals is as accurate as possible for a high quality atrial fibrillation risk assessment.

In deep learning architecture, the output layer is responsible for converting the advanced features extracted from the convolutional and pooling layers into final outputs, e.g., specific parameters for atrial fibrillation diagnosis. The output layer is typically designed as a fully connected layer, and may generate one or more output values from the features extracted by the previous layer, which values represent the final predictions of the model for a particular task.

To further improve the accuracy of feature extraction, a post-processing algorithm was introduced, using the formula:

To refine the detection of waveform features. This integration formula is mathematically implemented by weighted integration of the signal x (t) over an effective waveform interval [ -a, a ], where t ^k is a weight function and k is an adaptively adjusted parameter, which can be adjusted as needed for waveform characteristics. This method is particularly useful for extracting time position and amplitude information of waveforms, the key being to optimize feature extraction by selecting appropriate values of k and a. For example, the value of k may vary from 1 to 3, depending on whether the desired waveform characteristics reflect amplitude or time position more; the size of a should be set to be sufficient to cover the entire QRS complex or P-wave period, e.g., 0.1 seconds to 0.2 seconds.

Electrocardiogram data of a specific patient suspected of atrial fibrillation is set. QRS complex and P-wave features in its electrocardiogram have been initially extracted by the previous models and algorithms. Applying the above post-processing algorithm, it has now been found that the temporal position and amplitude of the QRS complex is similar to that of a typical atrial fibrillation patient. By adjusting k and a, the characteristics of these waveforms, such as the width of the QRS complex and the irregularities of the P wave, are precisely quantified. Such quantified results may be directly input into a final risk assessment model for predicting the risk of atrial fibrillation for the patient.

Example 3:

Feature extraction of echocardiography is a vital link in this embodiment. Not only because echocardiography can directly show structural and functional changes in the heart, particularly the left atrial portion of the heart, but these changes are closely related to atrial fibrillation. The feature extraction process of the echocardiogram comprises a plurality of steps from preprocessing of the video sequence to feature analysis by using a deep learning model, and finally, the comprehensive features are extracted by combining a multi-scale feature analysis technology.

First, the video sequence of the echocardiogram is pre-processed, which includes frame extraction and scaling to ensure that all video frames are processed at a uniform size and resolution. The image is standardized, and a new formula is adopted:

S(x)＝log(1+|x|^γ)

Where γ is a parameter that adjusts the level of image enhancement, typically ranging from 0.3 to 0.8. The function of this formula is to enhance the visibility of critical structures in the image, in particular the boundary and internal structures of the left atrium of the heart, which is critical for subsequent feature analysis.

Next, these pre-processed echocardiographic video sequences are analyzed using a combined model of Convolutional Neural Network (CNN) and long-short-term memory network (LSTM). CNNs are used to extract spatial features from each frame image, such as the size, shape, and other relevant structural features of the left atrium. LSTM is used to analyze these features over time in the video sequence, such as the rate of expansion and contraction of the left atrium and the rate of change of volume. The model combines the spatial feature recognition capability of CNN and the time sequence analysis capability of LSTM, and can comprehensively capture key information about heart functions in echocardiography.

Finally, multi-scale feature analysis techniques are used to further refine feature extraction of electrocardiographic and echocardiographic data. This step involves analyzing the characteristics of the data over different time windows and frequency resolutions so that the functional state of the heart and the risk of atrial fibrillation can be more fully understood. For example, finer or wider patterns of changes may be captured by adjusting the size of the time window, or adjusting the range of the frequency analysis to focus on a particular type of signal change.

Echocardiography of a patient showed slight dilation of the left atrium. Through the method, the definition of the image is enhanced through the S (x) formula, so that the outline of the left room is more obvious. The cnn+lstm model then identifies and tracks the changes in this expansion process throughout the cardiac cycle, quantifying the expansion rate and volume changes. The multi-scale feature analysis further reveals the left atrial expansion versus heart rate variability, providing comprehensive data support for assessing the risk of atrial fibrillation in a patient. Such information is of great value to the clinician and can help make more accurate diagnostic and therapeutic decisions.

The present embodiment is how to design and apply a deep Convolutional Neural Network (CNN) to perform detailed feature extraction on each frame of image. This process is to capture key changes in the size and morphology of the left atrium, which are critical to assessing risk of atrial fibrillation.

The design of deep convolutional neural networks features a multi-layer structure, each layer intended to extract more abstract and complex features from the image. For left atrial feature extraction in echocardiography, the network generally includes the following several core layers:

1. convolution layer (ConvolutionalLayers):

The convolution layer uses equation C _l(x)＝σ(W_l*x+b_l) to extract the image features, where W _l is the weight matrix of the convolution layer/which represents the convolution operation, b _l is the bias term, and σ is the activation function, typically chosen to be ReLU (RectifiedLinearUnit) for its efficiency and effectiveness in training the depth network.

In left-hand feature extraction applications, the primary convolution layer uses a smaller convolution kernel (e.g., 3x3 or 5x 5), which helps capture small features such as edges and contours of the left hand.

2. Pooling layer (PoolingLayers):

The pooling layer is typically applied after the convolution layer to reduce feature dimensions while preserving important feature information. Maximum pooling (MaxPooling) is a common pooling approach that helps the model extract the most salient features and reduces sensitivity to small changes in position in the image. In analyzing the left atrium, the pooling layer may help the model focus on the overall shape and size of the left atrium, while ignoring some unnecessary background information.

Processing each frame of image: the preprocessed echocardiogram is sent to the CNN per frame of image. The convolution layer analyzes the image content layer by layer, going from basic edge and texture information to more complex structures such as the shape and size of the left atrium.

Feature map (FeatureMaps) generation: the feature map of the image is gradually formed each time it is convolved and pooled, each feature map capturing a unique view of the left atrium from different angles and scales. These feature maps can then be used to generate final left-hand feature descriptors that contain key information about the left-hand function and structure.

In a specific case, an echocardiogram of a middle-aged male patient showed slight left atrial expansion. By applying the designed CNN model, the expansion speed and morphological change of the left atrium can be accurately measured and tracked. For example, the maximum expansion rate and specific trend of change in the left atrium over the cardiac cycle are found by modeling, which is critical for assessing the risk of a patient developing severe atrial fibrillation in the future.

The present embodiment performs feature extraction on the echocardiography through a Convolutional Neural Network (CNN), focusing mainly on the size and morphology of the left atrium. Now, to further mine dynamic information behind these static features, such as the expansion and contraction rate and volume change rate of the left atrium, long and short term memory networks (LSTM) were introduced, particularly for deep learning models that process and predict time series data.

The output of CNNs is typically high-level feature representations of each frame of image, which capture the key visual features of the left atrium. These features are arranged in time series order to form sequence data x _t of the LSTM input, where t represents a time step. LSTM processes time series data through its internal gating mechanism (input gate, forget gate, output gate) and can learn long-term dependency effectively. The processing of the echocardiographic data is mainly performed by the following iterative formula:

L_t＝φ(W_f·[h_t-1,x_t]+b_y)

Where L _t is the LSTM cell output at time t, W _f is the weight matrix of the forgetting gate, [ h _t-1,x_t ] is the combination of the previous hidden state and the current input, b _f is the bias term of the forgetting gate, phi is typically a nonlinear activation function such as sigmoid or tanh. This structure enables the LSTM to decide how long to retain past information and when to update that information, which is particularly important for analyzing dynamic changes in the left atrium.

In practical applications, W _f、b_f and other relevant parameters require learning optimization through training data. The weight of the forgetting gate W _f controls the degree of retention of information, which is critical to capturing the continuous changes in left atrial function. The specific values of these parameters are usually adjusted during training, but initially set, the weights range from-0.5 to 0.5, and the bias term b _f ranges from 0 to 1.

Considering a specific clinical scenario, an echocardiographic sequence of one patient shows that the left atrium gradually expands over consecutive cardiac cycles. By CNN, the morphology features of the left atrium in each cardiac cycle have been acquired. Subsequently, LSTM processes these features, learning the rate of left atrial expansion and contraction and its rate of change of volume. For example, LSTM recognizes a gradual trend that predicts an increased risk of atrial fibrillation in terms of left atrial expansion velocity and volume changes. Such dynamic analysis provides valuable diagnostic information to the physician, facilitating early intervention and treatment planning.

By the analysis method combining the CNN and the LSTM, not only the spatial characteristics in the echocardiogram can be extracted, but also the change rule of the characteristics along with time can be revealed, and more comprehensive and accurate support is provided for atrial fibrillation risk assessment.

In order to further improve the accuracy of atrial fibrillation risk prediction, the embodiment applies a multi-scale feature analysis technology. This step aims to integrate features at different time windows and frequency resolutions, and this goal is achieved by a well-designed integration formula.

The heart of multi-scale feature analysis techniques is the ability to extract data features from multiple scales, capturing short-term and long-term trends in changes and their interactions. This is particularly important to understand how heart function changes over time. In this project, the following integral formula is used to integrate the characteristics of the CNN and LSTM outputs:

In this formula, x (t) represents the time series characteristics obtained from CNN and LSTM, Is a gaussian weighting function used to weight the feature, where t represents a time variable and α is a parameter that adjusts the sensitivity of the time scale. The parameter controls the width of the weighting function and thus affects the time frame of the analysis. Typically, the value of α can be set in the range of 0.01 to 0.1, with the specific value depending on the time sensitivity of the analysis and the nature of the data.

In view of the practical use of echocardiographic data in monitoring cardiac function, for example, monitoring left atrial expansion in patients with chronic atrial fibrillation. By means of the aforementioned CNN and LSTM models, morphological and functional changes of the left atrium have been able to be tracked and recorded. Now, using multi-scale feature analysis techniques, these data points can be integrated, particularly in view of the rapid changes that exist during the occurrence of atrial fibrillation.

For example, by adjusting the value of α, changes in left atrial expansion velocity before and after the onset of atrial fibrillation can be analyzed more carefully. If α is smaller, the Gaussian function is wider, meaning that the analysis will be more focused on long-term trends; if α is larger, the Gaussian function is narrower and the analysis will be more focused on short-term variations. This fine-grained control allows capturing small changes that are ignored in routine detection, thereby providing greater predictive accuracy in atrial fibrillation risk assessment.

Example 4:

The adoption of the multi-modal learning method in this embodiment is critical because it allows complementary information to be extracted and utilized from different data sources, thereby enhancing the accuracy and robustness of the predictive model. This process involves two main sources of data, electrocardiogram (ECG) and echocardiography, each of which is a separate training model, and then integrating the outputs of these models for final risk assessment. The following are detailed implementation steps and their feasibility descriptions:

1. mono-modal model training of electrocardiographic data

For electrocardiographic data, deep Convolutional Neural Networks (CNNs) are employed to extract key electrocardiographic features, including QRS complexes, P-waves, and T-waves. These features are critical for diagnosing heart rhythm abnormalities. The formula of the CNN model is as follows:

M_ECG(x)＝σ(W*x+b)

Wherein σ is a nonlinear activation function (e.g., reLU) that is used to increase the nonlinearity of the model and help learn complex features; w and b are the weights and biases of CNN, respectively, and these parameters are learned by training data. For example, the weight W may be initialized to a random small value, and the bias b may be initialized to 0 or a small constant (e.g., 0.01).

2. Application of feature extraction

In practical medical data applications, such as ECG recordings of a patient, irregular P-waves and broad QRS complexes are displayed, which suggests the risk of atrial fibrillation. By applying the CNN model described above, the characteristics of these waveforms can be effectively identified and quantified. The weights W and bias b of the model are adjusted to optimally capture these key waveform features.

3. Single mode model training of echocardiographic data

The processing of echocardiographic data is similar but may use different feature extraction techniques, such as combining CNN with long and short term memory networks (LSTM) etc., specifically to analyze changes in cardiac structure and function.

4. Multimodal feature integration

After training the two unimodal models, the next step is to integrate the outputs of these models. Features extracted from the ECG and echocardiographic data can be fused through a high-level machine learning model (such as a random forest or gradient hoist) to obtain predictions for comprehensive assessment of atrial fibrillation risk.

In clinical tests, electrocardiographic and echocardiographic data of a patient are entered into corresponding models. The electrocardiogram model identifies abnormal P-wave and QRS waveforms, while the ultrasound model shows left atrial enlargement. After integrating this information, the model highly predicts that the patient is at risk of developing atrial fibrillation.

The present embodiment utilizes a long and short term memory network (LSTM) to analyze the size and functional indicators of the left atrium. This step involves a detailed understanding and prediction of the dynamic changes of the left atrium over time, which is critical to assessing a patient's potential risk of atrial fibrillation. The following are detailed implementation steps and their feasibility descriptions:

firstly, echocardiographic data needs to be pre-processed, including frame extraction, denoising, normalization, and appropriate image enhancement, to ensure the quality and consistency of the input data.

LSTM networks are designed to process and analyze time series data, which is particularly important when processing dynamically changing medical image data such as echocardiography. For each frame of processed echocardiographic data, the LSTM model performs a feature analysis by the following formula:

M_Echo(x_t)＝Γ(h_t-1,x_t)

Where Γ represents the LSTM cell, h _t-1 is the hidden state of the previous time step, x _t is the input feature of the current time step. The LSTM unit includes several key components: forget gate, input gate, output gate. These gating mechanisms help the model determine which information should be retained at each time step, which information should be discarded, and how to integrate new input information, which is critical to analyzing dynamic changes (e.g., size changes and functional changes) of the left atrium.

In practical applications, parameters of the LSTM model (such as weights and biases) need to be optimized by training data. For analysis of left atrium size and function, the initialization weights of the LSTM model are set to a small range of random values, e.g., -0.05 to 0.05, so that the model can adapt to different data features at the beginning of learning.

Continuous echocardiographic data for a group of patients is set, showing that the left atrium of the patient gradually expands over the past several months. By applying the LSTM model, accurate measures of the left atrial expansion and contraction rate can be extracted from these data. For example, the model recognizes that the rate of left atrial expansion increases significantly over a particular period, and this change is associated with an increased risk of atrial fibrillation. Through continuous monitoring and analysis of these data, the LSTM model can not only provide immediate readings of left atrial size and function, but also predict future trends in changes. This approach allows the physician to identify potential risks earlier and thus to intervene in a timely manner.

The present embodiment relates to integrating data of different modalities such as Electrocardiogram (ECG) and echocardiogram (Echo). By adopting the self-adaptive multi-mode fusion algorithm, the model is allowed to dynamically adjust weights according to the quality and the correlation of different data types, and the performance of the whole prediction model can be optimized. The following are detailed implementation steps and their feasibility descriptions:

The adaptive multi-modal fusion algorithm is implemented by the following formula:

Where y _i represents the output of the ith modality model (e.g., ECG, echo), and α _i is the weight corresponding to the modality data. These weights are not fixed, but are dynamically adjusted according to the predicted contribution of each modality data in a particular patient or case.

The adjustment of the weights α _i is based on the quality of each modality data and the contribution to the final prediction result. For example, if a modality shows higher predicted relevance or accuracy in a particular case, its weight will be increased. The weights are typically between 0 and 1 and sum to 1, ensuring that the output of all modes is properly considered.

When the algorithm is implemented, the contribution degree of each mode is usually evaluated through methods such as cross validation and the like in the model training stage, and the weight is dynamically adjusted according to the contribution degree. For example, the weights are adjusted according to the performance of each modality on the validation dataset, and an optimal weight configuration is found by an optimization algorithm such as gradient descent.

A set of patient data is set including an electrocardiogram and an echocardiogram. Electrocardiogram models are very adept at capturing electrophysiological changes, while echocardiogram models can provide detailed information about the structure and function of the heart. In the case of a specific patient, the electrocardiogram shows significant P-wave and QRS-complex abnormalities, while the echocardiogram shows significant enlargement of the left atrium.

By applying the fusion algorithm, the information can be integrated to improve the prediction accuracy. If the electrocardiographic data shows a higher correlation in predicting atrial fibrillation, its weight α ₁ is set high, e.g., to 0.7, and the weight α ₂ of the echocardiogram is correspondingly set low, to 0.3. This weight adjustment ensures that the model output reflects maximally the most critical diagnostic information. Through the self-adaptive multi-mode fusion algorithm, the atrial fibrillation risk can be estimated more accurately, more comprehensive and comprehensive data support can be provided for clinical decisions, and the practicability and effectiveness of the model are enhanced.

Example 5:

the embodiment adopts a method of cross-validation and algorithm iterative adjustment of algorithm structure and parameters. This process is particularly focused on the use of hierarchical cross-validation techniques to ensure that each validation subset is consistent with the overall dataset in terms of distribution of cardiac disease categories. The following are detailed implementation steps and their feasibility descriptions:

Hierarchical cross-validation is a special type of cross-validation that ensures that the proportion of categories in each cross-validated fold (fold) is the same as in the overall dataset. This is particularly important for medical data, as disease categories are often unbalanced. In a cardiac disease dataset, some categories (such as atrial fibrillation) are less common than others, and simple random sampling results in some cross-validated subsets lacking sufficient representative samples.

Functions used in the layered cross-validation process:

Is a logistic regression model that estimates the probability that a data point belongs to a class, where x _i represents the characteristics of the data point. This function helps evaluate the distribution of categories in the dataset and assigns samples accordingly to each subset, ensuring that each subset effectively represents the category distribution of the overall dataset.

After the layered cross-validation is completed, the next step is to iteratively adjust the algorithmic structure and parameters of the model to optimize the model performance. This typically involves adjusting the number of network layers, the size of the layers, the learning rate, and other super parameters.

Iterative adjustment of these parameters is based on the performance results of each cross-validation. For example, if the model is found to have low accuracy in predicting a disease of a certain type in a certain verification, it is necessary to adjust the network structure or add more feature learning layers to improve the recognition capability of the model for the disease of the type.

The set of data includes electrocardiogram and echocardiogram data from different patients, some proportion of whom are diagnosed as atrial fibrillation. Through layered cross-validation, it is ensured that each validation subset has representative atrial fibrillation and non-atrial fibrillation samples, so that the performance of the model on each category can be accurately evaluated. For example, during cross-validation, if the model is found to perform poorly in identifying patients with atrial fibrillation, the parameters of the model, including adding feature extraction layers or adjusting activation functions, may be adjusted.

It is critical that the present embodiment optimize the model structure to more accurately capture key features in Electrocardiography (ECG) and echocardiography (Echo). This is achieved using dynamically tuned network architecture and genetic algorithms. The following is a detailed description of the steps, including specific application examples of the actual data introduction and calculation results:

to dynamically select the best network structure, a scoring function is defined:

Where G represents the network structure and y _i is the actual observed value, an Is a model predictive value. This function aims to minimize prediction errors, ensuring that the selected network structure is most effective in predicting atrial fibrillation risk.

Genetic algorithm is a heuristic search algorithm, mimicking the mechanism of natural selection. It optimizes the problem solution by selection, crossover (hybridization) and mutation etc. In application, each parameter of the network structure G (e.g., number of layers, convolution kernel size, activation function type, etc.) is a gene on the chromosome.

The initial population consists of a variety of randomly generated network structures. In each generation, the score S (G) of each network structure is calculated, and the most excellent network structure is selected for hybridization and mutation according to the scores, so that a new generation of network structure is generated.

Through iteration of this process, the algorithm gradually finds the optimal network structure that reduces the prediction error. This process is repeated until a stop condition is met (e.g., a preset number of iterations is reached or the improvement amplitude is below a threshold).

The initial data set is set to include ECG and Echo data for thousands of patients, including detailed label information, such as whether atrial fibrillation is diagnosed. By training and evaluating various network structures with actual data, it can be observed that certain structures perform better in capturing atrial fibrillation characteristics. For example, models with deeper convolutional networks and complex attention mechanisms were found to be more effective in capturing subtle QRS complex changes in ECG and left atrial volume changes in Echo.

The present embodiment uses bayesian optimization to adjust model parameters as a key step, aimed at maximizing the performance of the model, especially for atrial fibrillation features in the cardiac disease dataset. This approach relies on probabilistic models of parameters, with the objective of minimizing the negative logarithm of the likelihood of the parameter, and thus finding the optimal model parameter configuration. The following are detailed steps and application descriptions:

in the bayesian optimization process, the objective is to maximize the objective function, defined herein as:

where θ represents the parameter set of the model and D represents the cardiac disease data set, particularly including data related to atrial fibrillation. The negative log likelihood-log P (theta|D) is used for measuring the fitting degree of the model to the data under the given parameters; the parameter set θ may include network weights, biases, learning rates, and the like.

Bayesian optimization is a global optimization strategy based on a probability model, and is mainly used for optimizing black box functions with high calculation cost. It uses a gaussian process to build a probabilistic model of the objective function and iterates the optimization by selecting parameters that improve the model performance.

In each iteration, the Bayesian optimization algorithm evaluates the current parameter set θ and calculates its performance index, logP (θ|D). The algorithm then updates its probabilistic model about the objective function and determines the next set of optimal parameters.

In practical applications, such as a set of data sets including various electrocardiographs and echocardiography, parameters of the model such as CNN or LSTM can be systematically adjusted through bayesian optimization to find the best network structure and weight configuration, so as to more accurately identify and predict atrial fibrillation.

The optimization process involves adjusting the depth of the CNN layer, the convolution kernel size, the selection of the activation function, etc. Bayesian optimization may take into account the results of all previous iterations and infer the parameter setting that brings the greatest performance improvement.

In the study containing the atrial fibrillation data of the patient, the learning rate and the layer number of the model are adjusted through Bayesian optimization, the initial learning rate is between 0.001 and 0.1, and the layer number is different from 1 to 10. The optimization procedure shows that learning rates of 0.01 and 5 layers of the network structure can give the best predictive performance, and this arrangement will be used in subsequent model training and testing. By the method, model parameters can be optimized, prediction accuracy is improved, and effectiveness and reliability of the model in processing complex heart disease data can be ensured, so that more accurate atrial fibrillation risk assessment is provided for clinic.

The foregoing has shown and described the basic principles, principal features and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made without departing from the spirit and scope of the invention, which is defined in the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims

1. The atrial fibrillation risk assessment method based on machine learning is characterized by comprising the following steps of: firstly, collecting and preprocessing multi-source data including Electrocardiogram (ECG), echocardiogram, life style and clinical medical records, wherein the preprocessing step comprises data cleaning, denoising, normalization and missing value processing;

2. The machine learning based atrial fibrillation risk assessment method as defined in claim 1, wherein the data cleansing process identifies and corrects outliers and errors in the data by applying a real-time quality assessment model based on a function:

M(x₁,x₂)＝x₁·ω+x₂·(1-ω)

Performing filling, wherein ω is a weight adjusted based on the correlation;

3. The machine learning based atrial fibrillation risk assessment method according to claim 1, wherein the feature engineering is feature extraction of Electrocardiogram (ECG) and echocardiogram using a deep learning model, wherein the electrocardiogram feature extraction uses a deep Convolutional Neural Network (CNN) based model to detect and separate QRS complex, P wave and T wave, and the extracted waveform includes fine features of peak shape and trough feature;

4. A machine learning based atrial fibrillation risk assessment method according to claim 3, wherein the ECG signal is pre-processed, including normalization to ensure consistency of the input data, applying a bandpass filter using the formula:

C_n(x)＝ρ(ω*x+b)

the attention mechanism formula is introduced into the CNN:

5. The machine learning based atrial fibrillation risk assessment method as defined in claim 3, wherein the echocardiographic feature extraction comprises the steps of:

S(x)＝log(1+|x|^γ)

C_l(x)＝σ(W_l*x+b_l)

L_t＝φ(W_f·[h_t-1,x_t]+b_f)

6. The machine learning-based atrial fibrillation risk assessment method as defined in claim 1, wherein the multi-modal machine learning party comprises:

M_ECG(x)＝σ(W*x+b)

For echocardiographic data, a long-short memory network (LSTM) is used:

M_Echo(x_t)＝Γ(h_t-1,x_t)

Then through an adaptive multi-modal fusion algorithm:

7. The machine learning based atrial fibrillation risk assessment method as defined in claim 1, wherein the cross-validation and algorithm iterative adjustment algorithm structure and parameters method comprises: