Self-evolution training method and device based on neuro-physical engine and causal generation, electronic equipment and storage medium

Abstract

The application relates to the field of intelligent medical treatment, in particular to a self-evolution training method and device based on a neural physical engine and causal generation, an electronic device and a storage medium, a plurality of real sample magnetocardiogram signals and virtual sample magnetocardiogram signals and corresponding category labels are determined as an initial training set, a data classification model is trained in an iterative manner according to the training set multiple times, and a target category label whose classification effect after training does not meet preset requirements is determined. According to a structural causal model, pathological parameter information of a virtual sample magnetocardiogram signal of the target category label is generated. A virtual sample magnetocardiogram signal obtained by forward solving of the pathological parameter information by means of a differentiable neural physical simulator is used to update the training set. The data classification model is trained for multiple rounds, and after each training, a category that is difficult to identify is automatically identified, and corresponding training samples are automatically added, so as to be trained again, thereby guaranteeing the model training effect in a data scarce scene.

Description

Technical Field

[0001] This application relates to the field of intelligent healthcare, and includes, but is not limited to, a self-evolving training method, apparatus, electronic device, and storage medium based on a neurophysical engine and causal generation. Background Technology

[0002] Magnetocardiography (MCG) is a non-invasive technique for detecting cardiac electrical activity, capable of capturing millisecond-level electrical excitation propagation processes and subtle magnetic field changes, offering unique advantages in early arrhythmia identification and lesion localization. In recent years, deep learning has been widely applied to MCG signal analysis for automated diagnosis, but its performance heavily relies on large-scale, high-quality, and well-labeled training datasets. However, for rare cardiac pathological patterns (such as ventricular tachycardia of specific anatomical origin and the occult Brugada phenotype), clinically available samples are extremely limited, resulting in poor generalization ability and insufficient robustness of existing AI models. Traditional data augmentation techniques (such as time shifting and noise addition) lack physiological significance; while mainstream generative adversarial networks (GANs) or diffusion models, although capable of generating visually realistic signals, often violate biophysical laws, causing "spurious correlations" that mislead AI learning. Summary of the Invention

[0003] In view of this, the self-evolving training method, apparatus, electronic device and storage medium based on neurophysical engine and causal generation provided in the embodiments of this application aim to provide a self-evolving training scheme that ensures the accuracy of classification models in scenarios where magnetocardiogram data is scarce.

[0004] The self-evolving training method, apparatus, electronic device, and storage medium based on neurophysical engines and causal generation provided in this application are implemented as follows: One aspect of this application provides a self-evolutionary training method based on a neurophysical engine and causal generation, the method comprising: A training set is determined, which includes multiple sample magnetocardiogram (MCC) signals and corresponding category labels. The sample MCC signals include real sample MCC signals and virtual sample MCC signals. Using the training set as the initial training set, the following steps are executed iteratively multiple times: The data classification model is trained based on the current training set, and at least one target category label whose classification effect after training does not meet the preset requirements is identified. Based on the structural causal model, corresponding pathological parameter information is generated from the magnetocardiogram signals of virtual samples corresponding to the target category labels; By using a differentiable neurophysical simulator to perform a forward solution based on pathological parameter information, the corresponding virtual sample magnetocardiogram signal is obtained. The training set is updated based on the virtual sample magnetocardiogram signals and the corresponding target category labels.

[0005] In one possible implementation, a training set is determined, comprising multiple sample magnetocardiogram signals and corresponding category labels, including: Real magnetocardiogram signals were collected from multiple patients; Based on the cardiac imaging data and corresponding real magnetocardiogram signals of each patient, a simulation model is created to obtain a digital twin generator for each patient; Each digital twin generator generates a corresponding magnetocardiogram (MCG) signal as a virtual MCG signal; Real and virtual magnetocardiogram (MCG) signals are used as sample MCG signals, and the training set is determined based on the sample MCG signals and their corresponding category labels.

[0006] In one possible implementation, determining a training set that includes multiple sample magnetocardiogram signals and corresponding category labels further includes: The sample magnetocardiogram signal is encoded by a neural radiation field encoder to obtain a high-fidelity sample magnetocardiogram signal.

[0007] In one possible implementation, a data classification model is trained based on the current training set, and at least one target category label whose post-training classification performance does not meet preset requirements is identified, including: Train the data classification model based on the training set; The validation set is obtained by acquiring real sample magnetocardiogram signals and corresponding category labels from the training set. Input the magnetocardiogram signals of the samples in the validation set into the trained data classification model, and output the corresponding predicted labels and confidence scores; The target category label is determined based on the predicted label and confidence level corresponding to the magnetocardiogram signal of each sample in the validation set.

[0008] In one possible implementation, the target category label is determined based on the predicted label and confidence level corresponding to the magnetocardiogram signal of each sample in the validation set, including: High-entropy regions are explored based on the magnetocardiogram signals of each sample in the validation set and the corresponding predicted labels and confidence levels. The high-entropy regions discovered during exploration are fed back into the structural causal model to automatically generate at least one target category label.

[0009] In one possible implementation, pathological parameter information is generated based on the magnetic resonance imaging (MRI) signal of the virtual sample corresponding to the target category label, according to a structural causal model. This includes: Obtain the magnetocardiogram (MCG) signals of virtual samples corresponding to the target category labels in the current training set; Abnormal pathological information is obtained by inversely solving the magnetocardiogram signal of the virtual sample corresponding to the target category label using a differentiable neurophysical simulator; Based on abnormal pathological information and structural causal models, corresponding pathological parameter information is obtained.

[0010] In one possible implementation, the training set is updated based on the virtual sample magnetocardiogram signal and the corresponding target category label, including: The virtual sample magnetocardiogram signal is encoded by a neural radiation field encoder to obtain a high-fidelity virtual sample magnetocardiogram signal. The newly added virtual sample magnetocardiogram signals and their corresponding target category labels are added to the training set.

[0011] Another aspect of the embodiments of this application provides a self-evolving training device based on a neurophysical engine and causal generation, the device comprising: The training set acquisition module is used to determine a training set that includes multiple sample magnetocardiogram (MCC) signals and corresponding category labels. The sample MCC signals include real sample MCC signals and virtual sample MCC signals. The iterative training module is used to take the training set as the initial training set and execute the following steps iteratively multiple times: The data classification model is trained based on the current training set, and at least one target category label whose classification effect after training does not meet the preset requirements is identified. Based on the structural causal model, corresponding pathological parameter information is generated from the magnetocardiogram signals of virtual samples corresponding to the target category labels; By using a differentiable neurophysical simulator to perform a forward solution based on pathological parameter information, the corresponding virtual sample magnetocardiogram signal is obtained. The training set is updated based on the virtual sample magnetocardiogram signals and the corresponding target category labels.

[0012] In one possible implementation, the training set acquisition module is further used for: Real magnetocardiogram signals were collected from multiple patients; Based on the cardiac imaging data and corresponding real magnetocardiogram signals of each patient, a simulation model is created to obtain a digital twin generator for each patient; Each digital twin generator generates a corresponding magnetocardiogram (MCG) signal as a virtual MCG signal; Real and virtual magnetocardiogram (MCG) signals are used as sample MCG signals, and the training set is determined based on the sample MCG signals and their corresponding category labels.

[0013] In one possible implementation, the training set acquisition module is further used for: The sample magnetocardiogram signal is encoded by a neural radiation field encoder to obtain a high-fidelity sample magnetocardiogram signal.

[0014] In one possible implementation, the iterative training module is further used for: Train the data classification model based on the training set; The validation set is obtained by acquiring real sample magnetocardiogram signals and corresponding category labels from the training set. Input the magnetocardiogram signals of the samples in the validation set into the trained data classification model, and output the corresponding predicted labels and confidence scores; The target category label is determined based on the predicted label and confidence level corresponding to the magnetocardiogram signal of each sample in the validation set.

[0015] In one possible implementation, the iterative training module is further used for: High-entropy regions are explored based on the magnetocardiogram signals of each sample in the validation set and the corresponding predicted labels and confidence levels. The high-entropy regions discovered during exploration are fed back into the structural causal model to automatically generate at least one target category label.

[0016] In one possible implementation, the iterative training module is further used for: Obtain the magnetocardiogram (MCG) signals of virtual samples corresponding to the target category labels in the current training set; Abnormal pathological information is obtained by inversely solving the magnetocardiogram signal of the virtual sample corresponding to the target category label using a differentiable neurophysical simulator; Based on abnormal pathological information and structural causal models, corresponding pathological parameter information is obtained.

[0017] In one possible implementation, the iterative training module is further used for: The virtual sample magnetocardiogram signal is encoded by a neural radiation field encoder to obtain a high-fidelity virtual sample magnetocardiogram signal. The newly added virtual sample magnetocardiogram signals and their corresponding target category labels are added to the training set.

[0018] The electronic device provided in this application includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the method described in this application.

[0019] The computer-readable storage medium provided in this application embodiment stores a computer program thereon, which, when executed by a processor, implements the method provided in this application embodiment.

[0020] In this embodiment, multiple real sample magnetocardiogram (MCC) signals and virtual sample MCC signals, along with corresponding category labels, are determined as an initial training set. The data classification model is trained iteratively multiple times based on this training set, and target category labels whose classification performance does not meet preset requirements are identified. Pathological parameter information is generated from the virtual sample MCC signals of the target category labels using a structured causal model. The virtual sample MCC signals obtained by forward solving the pathological parameter information using a differentiable neurophysical simulator are then used to update the training set. This application performs multiple rounds of training on the data classification model and automatically identifies difficult-to-identify categories after each training iteration, automatically increasing the number of training samples to improve the data volume of difficult samples for retraining, thus ensuring effective model training in data-scarce scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart is shown below illustrating a self-evolutionary training method based on a neurophysical engine and causal generation according to an embodiment of this application; Figure 2 This diagram illustrates an embodiment of obtaining pathological parameter information according to this application. Figure 3 A schematic diagram is shown of a self-evolutionary training process based on a neurophysical engine and causal generation according to an embodiment of this application; Figure 4 This illustration shows an application scenario of a self-evolutionary training method based on a neurophysical engine and causal generation according to an embodiment of this application. Figure 5 This diagram illustrates a self-evolutionary training device based on a neurophysical engine and causal generation according to an embodiment of this application. Figure 6 A schematic diagram of an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the specific technical solutions of this application will be further described in detail below with reference to the accompanying drawings of the embodiments of this application. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0025] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0026] It should be noted that the terms "first, second, third" used in the embodiments of this application are used to distinguish similar or different objects and do not represent a specific order of objects. It can be understood that "first, second, third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0027] The self-evolutionary training method based on neurophysical engines and causal generation in this application embodiment can be executed by any type of electronic device, including but not limited to mobile phones, wearable devices (such as smartwatches, smart bracelets, smart glasses, etc.), tablet computers, laptops, in-vehicle terminals, PCs (Personal Computers), etc. The functions implemented by this method can be achieved by the processor in the electronic device calling program code. Of course, the program code can be stored in a computer storage medium. Therefore, the electronic device includes at least a processor and a storage medium.

[0028] The self-evolutionary training method based on neurophysical engines and causal generation in this application can be used to train any model for classifying magnetocardiogram (MCG) signals. For example, this application can be applied to training a MCG signal classification model for clinical diagnosis and decision support in hospitals. Alternatively, it can be applied to training a MCG signal classification model for pharmaceutical research and teaching in schools or research institutions.

[0029] However, the aforementioned application scenarios all require magnetocardiographic signals from specific pathological states as training data. But for rare cardiac pathological patterns (such as ventricular tachycardia of specific anatomical origin or the occult Brugada phenotype), clinically available samples are extremely limited, resulting in poor generalization ability and insufficient robustness of the trained classification models. Traditional data augmentation techniques (such as time shifting and noise addition) lack physiological significance; while mainstream generative adversarial networks (GANs) or diffusion models, although capable of generating visually realistic signals, often violate biophysical laws, causing "spurious correlations" that mislead AI learning.

[0030] Research on related technologies has attempted to combine electrophysiological simulations to generate synthetic MCG data, but certain technical shortcomings still exist, such as: disconnect between simulation and AI: the PDE solver is not differentiable and cannot support end-to-end forward propagation; lack of causal constraints in the generation process: random perturbation parameters can easily generate data that does not conform to medical logic; static data generation: it is impossible to dynamically adjust the output content according to the weaknesses of the AI ​​model; data silo problem: it is difficult for hospitals to share sensitive information, which restricts multi-center collaborative modeling; coarse-grained spatial representation: only discrete channel signals are used, ignoring the continuous field characteristics of MCG.

[0031] Therefore, the technical problem solved by the embodiments of this application is how to improve the training effect of the data classification model for magnetocardiogram signals in scenarios with scarce data.

[0032] The following describes in detail the self-evolutionary training scheme based on neurophysical engine and causal generation of embodiments of this application with reference to the accompanying drawings.

[0033] Figure 1 A flowchart illustrating a self-evolutionary training method based on a neurophysical engine and causal generation according to an embodiment of this application is shown. Figure 1 As shown, the self-evolutionary training method based on neurophysical engine and causal generation in this application embodiment may include the following steps S10-S50.

[0034] For ease of description, the self-evolutionary training method based on neurophysical engines and causal generation in the embodiments of this application is described using an electronic device as the execution subject. It should be understood that the execution subject in the embodiments of this application can also be a processor or chip in an electronic device, and the embodiments of this application do not impose any limitations.

[0035] Step S10: Determine a training set that includes multiple sample magnetocardiogram signals and corresponding category labels.

[0036] In one possible implementation, the electronic device can acquire multiple training data sets to train the data classification model, resulting in a corresponding training set. The training set includes multiple sets of training data, each set consisting of sample magnetocardiogram (MCG) signals and corresponding class labels. The training data can include both real and virtual data. That is, the sample MCG signals include both real and virtual sample MCG signals, and the training data can include either real sample MCG signals and corresponding class labels, or virtual sample MCG signals and corresponding class labels.

[0037] In some embodiments, the real magnetocardiogram (MCG) signals in this application can be obtained by directly acquiring the patient's MCG signals using a MCG signal acquisition device, and further manually labeled by doctors or researchers according to the patient's condition to obtain corresponding category labels. Virtual MCG signals can be generated by constructing a digital twin generator for the patient, and the corresponding category label can be determined based on the patient's pathological information. For example, an electronic device can acquire real MCG signals from multiple patients. Then, simulation modeling is performed based on the cardiac imaging data and corresponding real MCG signals of each patient to obtain a digital twin generator for each patient. A corresponding MCG signal is generated based on each digital twin generator as a virtual MCG signal. The real and virtual MCG signals are used as sample MCG signals, and a training set is determined based on the sample MCG signals and corresponding category labels.

[0038] Optionally, after acquiring real magnetocardiogram signals from multiple patients, the electronic device can also obtain cardiac imaging data such as MRI (Magnetic Resonance Imaging) or CT (Computed Tomography) scans for each patient. Furthermore, it can acquire the patient's corresponding basic clinical information, such as age, gender, and BMI (Body Mass Index). The electronic device can use a deep segmentation network to extract the three-dimensional geometric structure of the heart based on the aforementioned cardiac imaging data, and can also segment regions such as the atria, ventricles, and conduction bundles. After completing the geometric segmentation, the electronic device can obtain a digital twin generator for each patient by establishing an individualized bidomain reaction-diffusion equation model. The bidomain reaction-diffusion equation model can be represented as:

[0039] in, It is the intracellular conductivity tensor, used to represent the ease and directionality of conducting current within cardiomyocytes. It is the extracellular conductivity tensor, used to represent the ease and directionality of conducting current outside myocardial cells. The membrane capacitance per unit area is used to represent the ability of a cell membrane to store electrical charge. Transmembrane potential, used to represent intracellular potential. extracellular potential The difference, A vector representing the gated state variables of various ion channels. Total ionic current, representing the sum of currents flowing through all ion channels, exchangers, and pumps on the cell membrane. The stimulating current is used to characterize the externally applied pacing current. This is the gradient / divergence operator.

[0040] The electronic device can use a Bayesian inversion algorithm to calibrate parameters such as local conduction velocity, action potential duration (APD), and anisotropy ratio to initialize the digital twin generator for each patient. Furthermore, upon acquiring new, realistic magnetocardiogram signals from the patient, the electronic device can automatically correct the parameters of the two-domain reaction-diffusion equation model to maintain synchronization between the digital twin generator and the patient's actual cardiac state.

[0041] In some embodiments, when constructing a digital twin generator for each patient, the electronic device acknowledges the challenges of personalized parameter inversion and proposes a hierarchical or group-based prior-guided approach. For example, in embodiments of this application, the electronic device can first establish a parameter probability distribution based on large population datasets, and then use limited MCG data for each patient to perform Bayesian inference on this prior distribution to obtain personalized parameter estimates, reducing reliance on data integrity.

[0042] Optionally, after acquiring the digital twin generators for each patient, the electronic device can generate simulated magnetocardiogram (MCG) signals based on each digital twin generator to obtain the corresponding MCG signals as virtual MCG signals. Further, the electronic device can use the category label corresponding to the real MCG signal that generated the patient's digital twin generator as the category label corresponding to the virtual MCG signal generated by that digital twin generator.

[0043] Furthermore, after acquiring the virtual magnetocardiogram (MCG) signal generated by the digital twin generator, the electronic device can also filter the generated virtual MCG signal. Optionally, the filtering strategy for the virtual MCG signal in this embodiment may include: 1) physical consistency check (simulation residual); 2) model adversarial evaluation (data that is only confused by the target model but easily identified by a robust teacher model may be of poor quality); 3) diversity check (avoiding the generation of a large number of similar samples).

[0044] In this embodiment, a double-blind experiment can be designed to evaluate the generated patient digital twin generator, allowing experts to determine the authenticity of the pathological features of the generated MCG. Specifically, for each real patient, a real MCG is obtained, and a corresponding "twin MCG" is generated by the digital twin generator. When the virtual and real magnetocardiogram (MCG) signals generated by the digital twin generator are submitted to experts without distinction, the experts do not know whether each number corresponds to real or generated data (this can be achieved through independent third-party coding). Experts then evaluate each data point based on its authenticity, the accuracy of pathological features, physiological rationality, and diagnostic consistency.

[0045] In other embodiments, after acquiring both real and virtual sample magnetocardiogram (MCC) signals, the electronic device can encode the acquired MCC signals to obtain high-fidelity MCC signals adapted to different device models through reconstruction / completion. For example, a neural radiation field encoder can be used to encode the sample MCC signals and reconstruct / complete them to obtain high-fidelity MCC signals. This neural radiation field encoder is used to upgrade the current discrete sample MCC channel signals into a continuous three-dimensional magnetic field spatiotemporal function. .

[0046] Optionally, for each sample magnetocardiogram signal, the electronic device can input a neural radiance field encoder, which uses a neural radiance field architecture to extract the spatial coordinates from the sample magnetocardiogram signal. and time Converted into a local magnetic field vector High-fidelity sample magnetocardiogram (MCC) signals are then output. This network structure can employ MLP+positional encoding, and the loss function during training includes... For example, a primary-level hospital may only have a 16-lead magnetocardiogram (MCG) device, which can reconstruct an equivalent 64-lead signal using NeRF-MCG for inference in advanced neural network algorithms.

[0047] Optionally, the NeRF architecture in this embodiment of the application has been adaptively improved for electromagnetic field characteristics. This improvement includes reconstructing the output representation, that is, changing the standard NeRF output color and density to a magnetic field vector, forming a vector-valued radiation field. Simultaneously, to capture dynamic processes, the network additionally receives temporal input, and the output becomes a spatiotemporal function. Based on the data fitting loss, this application explicitly adds Maxwell's equations constraint terms, which may include Gauss's magnetic law, Ampere's law, and boundary decay terms. Furthermore, physical guidance can be added to the network architecture. This physical guidance includes introducing physically aware features into the intermediate layers of the MLP, such as gradient information of the input spatial coordinates, or adding a linear combination layer of known basis functions to better represent the field's changing patterns. It may also include adopting a hierarchical training strategy, that is, first training to satisfy the main physical constraints, and then gradually adding more complex source term constraints. To further improve physical consistency and computational efficiency, this embodiment of the application can also design a dedicated field encoding network with built-in Maxwell constraints. Its core idea is to directly embed physical laws into the network's forward propagation process, rather than relying solely on loss function constraints.

[0048] Based on the improved NeRF-MCG architecture described above, this application's embodiments, through customized output representation and multi-physics constraint loss functions, can better adapt to the mathematical characteristics of electromagnetic field signals. Furthermore, the proposed dedicated field coding network can embed physical laws more deeply into the network architecture, achieving a higher degree of physical consistency, more efficient training and inference, and stronger generalization capabilities, thus providing a superior technical path for accurate modeling of bioelectromagnetic fields. Based on these technical features, this application's embodiments can achieve zero-sample interpolation of arbitrary sample magnetocardiogram signals to new sensor locations through a neural radiation field encoder, adapting to different device models.

[0049] In different application scenarios, data classification models can handle different classification tasks, and therefore the category labels in the training set in this application embodiment can be different. For example, when the data classification model is a diagnostic classification model based on magnetocardiogram (MCG) signals for assisted diagnostic classification, the category labels can be diagnostic classification labels, such as normal heart rhythm, Brugada syndrome, catecholamine-sensitive polymorphic ventricular tachycardia, arrhythmogenic right ventricular cardiomyopathy, idiopathic ventricular tachycardia, long QT syndrome, and myocardial ischemia / infarction, etc. When the data classification model is a risk stratification model based on MCG signals for risk warning, the category labels can be risk classification labels, such as low / medium / high risk of sudden cardiac death, probability of arrhythmic events: <5% / 5-20% / >20%, electrical storm tendency: yes / no, etc.

[0050] After determining the above training set, the embodiments of this application can use the training set as the initial training set and execute the following steps S20-S50 multiple times in an iterative manner until the convergence condition is met and the iteration process ends.

[0051] Optionally, embodiments of this application may also design a dynamic curriculum to trigger the termination of the self-evolutionary iteration. This dynamic accumulation initially generates common and typical pathological samples, and as the model's capabilities improve, it gradually generates more complex and marginal "high-entropy" samples. Simultaneously, the electronic device can pre-set early stopping criteria, such as determining that the convergence condition is met and ending the iteration process when the validation set performance plateaus or the marginal benefit of generated data for performance improvement is too low.

[0052] Step S20: Train the data classification model based on the current training set, and determine at least one target category label whose classification effect after training does not meet the preset requirements.

[0053] In one possible implementation, the electronic device first trains the data classification model using the training set for each iteration. This training process involves inputting the magnetocardiogram (MCG) signals of each sample in the training set into the data classification model to obtain corresponding predicted labels. Then, the model loss is calculated based on the predicted labels and the corresponding category labels in the training set, and the model parameters are optimized. After optimizing the data classification model parameters for this round of training, at least one target category label whose post-training classification performance does not meet preset requirements is further identified.

[0054] Optionally, the electronic device can first train the data classification model based on the training set. Then, it acquires the magnetocardiogram (MCG) signals of real samples from the training set and their corresponding class labels to obtain a validation set. The MCG signals of samples from the validation set are then input into the trained data classification model, which outputs the corresponding predicted labels and confidence scores. The target class label is determined based on the predicted labels and confidence scores of each sample's MCG signal in the validation set. This confidence score characterizes the reliability of the predicted label.

[0055] In some embodiments, the electronic device can statistically analyze the deviation between various predicted labels and their corresponding confidence levels and the true labels, perform uncertainty scoring on each category label, and determine that a category label is a difficult sample whose classification performance does not meet the preset requirements if the uncertainty score is greater than a preset threshold. This category label is then used as the target category label. The identification method for this target category label is simple, the identification efficiency is high, and the computational load is small.

[0056] In other embodiments, the electronic device can also automatically generate target category labels using a trained structural causal model. Specifically, the electronic device can explore high-entropy regions based on the magnetocardiogram signals of each sample in the validation set and the corresponding predicted labels and confidence levels. The explored high-entropy regions are then fed back to the structural causal model to automatically generate at least one target category label. This process allows the structural causal model to generate intervention instructions / constraint sampling ranges, thereby generating pathological parameter samples related to the target category label.

[0057] In this embodiment, the process of exploring high-entropy regions can be implemented using a gradient-based active learning strategy. For example, for category labels with poor classification performance, the electronic device can calculate the gradient of the classification loss with respect to the pathological parameter θ corresponding to that category label using a differentiable simulator. θ L). In the parameter space, sampling can be performed along the gradient direction or along the eigenvector direction of the Hessian matrix. The parameter perturbations in these directions are most likely to cause a decrease in the model's classification confidence, thereby accurately locating the "high-entropy region".

[0058] Optionally, this structural causal model (SCM) is used to describe a key causal pathway in cardiac electrical activity: ST-segment elevation. V1-V3 ←f( Conduction, body temperature <36°C), persistent (Stability of the folded loop, tendency for wavefront breaking). Given a defined structural causal model, electronic devices can design a causal diffusion model based on this model. During the denoising process of the causal diffusion model, a causal intervention layer (Do-Calculus Layer) is introduced to enforce compliance with... The causal dependence. This causal diffusion model supports observational models: simulating natural variation, and intervention models: Three generation modes—including a counterfactual model and a counterfactual model ("What would his MCG be like if this patient didn't have fibrosis?")—are used in scenarios such as virtual clinical trials, treatment outcome prediction, and educational demonstrations. A structured generative network composed of a structured causal model and a causal diffusion model is used to positively generate pathological state parameters with causal plausibility.

[0059] In this embodiment, the structural causal model acts as a high-level parameterized generation controller, while the causal diffusion model serves as the underlying network performing the generation task. The causal intervention layer of the causal diffusion model plays different roles in different application scenarios. For example, in the causal diffusion model's denoising network, this causal intervention layer is used to freeze or scale latent variables representing specific pathological factors.

[0060] After acquiring the magnetocardiogram signals and corresponding predicted labels and confidence levels of each sample in the verification set, the electronic device automatically identifies the category labels with low confidence / high error rate and, through a differentiable neurophysical engine, reverse-maps them back to the high-dimensional pathological parameter space of the cardiac digital twin (such as lesion location, conduction velocity, ion channel parameters, etc.). Within this continuous and interpretable parameter space, regions of parameter combinations that can cause model confusion and misjudgment are identified, termed "high-entropy regions." For example, when a data classification model cannot distinguish between "focal ventricular tachycardia of the right ventricular outflow tract" and "early Brugada syndrome," reverse reasoning reveals that the model is most prone to confusion when the Ito current enhancement and lesion automaticity overlap within a specific interval. This parameter interval is the "high-entropy region." In this embodiment, after identifying the high-entropy region, the result is fed back to a structural causal model, which directly proposes challenging counterfactual questions to create difficult samples as target category labels.

[0061] Step S30: Generate corresponding pathological parameter information based on the magnetic cardiogram signal of the virtual sample corresponding to the target category label according to the structural causal model.

[0062] In one possible implementation, the electronic device can generate corresponding pathological parameter information based on the magnetocardiogram (MCG) signal of a virtual sample corresponding to the target category label after determining at least one target category label. Optionally, the generation of this pathological parameter information can be achieved jointly by a differentiable neurophysical simulator and a structural causal model. For example, the electronic device can first acquire at least one virtual sample MCG signal corresponding to the target category label in the current training set, and then use a differentiable neurophysical simulator to perform inverse solving based on the virtual sample MCG signal corresponding to the target category label to obtain abnormal pathological information. Then, based on the abnormal pathological information and the structural causal model, the corresponding pathological parameter information is obtained.

[0063] In some embodiments, the differentiable neurophysical simulator is used to replace the traditional finite element method (PDE) solver with a graph neural operator (GNO) or a Fourier neural operator (FNO). During training, the neural operator learns from the initial stimulus point, lesion parameters, etc. Spatiotemporal distribution of transmembrane potential The mapping relationship. All operations of this differentiable neurophysical simulator are differentiable, supporting gradient feedback from the output magnetocardiogram signal to the input parameters. It can perform both backward and forward reasoning solutions. Backward reasoning is used for precise inverse problem solving, such as identifying which lesion configuration is most likely to cause abnormal magnetocardiogram signals in a sample. Forward reasoning is used for gradient-guided parameter search, i.e., quickly matching unknown cases.

[0064] In other words, the reverse reasoning process involves using gradient descent to optimize a set of abnormal pathological information (such as the location of the reentry loop and the degree of conduction block) most likely leading to an abnormality, given an abnormal sample magnetocardiogram signal. This directly provides doctors with interpretable mechanistic hypotheses for diagnosis, such as: "AI judges this to be Brugada syndrome, and the underlying electrophysiological mechanism is likely an increase in right ventricular outflow tract Ito current and a slowing of conduction." The forward reasoning process uses a differentiable neurophysiological simulator to "translate" or "render" these parameters into realistic transmembrane potentials and electromagnetic field distributions that conform to physical laws, i.e., Vm(x, t). Without it, the parameters generated by the generative network are just a string of meaningless numbers and cannot be transformed into a reliable sample magnetocardiogram signal.

[0065] Optionally, embodiments of this application can also establish a multi-level, multi-modal cross-validation system to check whether the abnormal pathological information obtained through the above-mentioned inverse solution is consistent with the clinical diagnosis or interventional electrophysiological examination results. The output of the system's inverse solution is a set of pathological parameters θ (such as 3D coordinates of the lesion, conduction velocity, degree of ion channel dysfunction, etc.). The core of the validation is to transform these parameters into a form that can be directly compared with clinical examination results. It can verify the accuracy of pathological location localization, such as the isthmus of reentry and focal origin, through spatial and functional localization, and verify whether the pathogenic mechanism inferred by the system conforms to known clinical diagnoses and pathophysiology through mechanism and parameter verification.

[0066] During the training of the differentiable neurophysical simulator, this embodiment of the application can use a large amount of {θ_i, Vm_i(x,t)} paired data generated by a traditional high-precision solver (as the "gold standard") to train the neural operator. After training, this neural operator can approximately solve the Bidomain equations at a speed several orders of magnitude faster than traditional solvers while maintaining differentiability. The neural operator (GNO / FNO) obtained by this training method can learn the "physical laws of cardiac electrophysiology".

[0067] Figure 2 This diagram illustrates an embodiment of obtaining pathological parameter information according to this application. Figure 2 As shown in this embodiment, the electronic device can generate corresponding pathological parameter information based on the virtual sample magnetocardiogram (MCG) signal corresponding to the target category label after determining at least one target category label. The generation of this pathological parameter information can be achieved by acquiring the virtual sample MCG signal corresponding to the target category label in the current training set, and then using a differentiable neurophysical simulator to perform inverse solving based on the virtual sample MCG signal to obtain abnormal pathological information. A corresponding generation instruction is determined based on the abnormal pathological information; for example, in the case of Brugada syndrome, the generation instruction would be: generate a Brugada syndrome case. Furthermore, a causal diffusion model is designed based on the causal graph constraints provided by the structured causal model, and a causal intervention layer is introduced during the denoising process to enforce the causal dependency of do(X)→Y. Finally, causally reasonable pathological parameter information is obtained, which may include at least one pathological parameter such as current density and conduction velocity.

[0068] Step S40: The corresponding virtual sample magnetocardiogram signal is obtained by performing a forward solution based on the pathological parameter information using a differentiable neurophysical simulator.

[0069] In one possible implementation, after generating at least one pathological parameter information for at least one target category label based on the above method, the electronic device can perform forward solving based on the pathological parameter information using a differentiable neurophysical simulator to obtain the corresponding virtual sample magnetocardiogram signal. This forward reasoning process is used by the differentiable neurophysical simulator to "translate" or "render" these parameters into a realistic transmembrane potential and electromagnetic field distribution that conforms to physical laws, i.e., Vm(x, t). Without it, the parameters generated by the generative network are merely a string of meaningless numbers and cannot be transformed into a reliable sample magnetocardiogram signal.

[0070] Step S50: Update the training set based on the virtual sample magnetocardiogram signal and the corresponding target category label.

[0071] In one possible implementation, after determining multiple virtual sample magnetocardiogram (MCG) signals in the current iteration process, the electronic device also encodes the virtual sample MCG signals using a neural radiation field encoder to obtain high-fidelity virtual sample MCG signals. Then, the newly added virtual sample MCG signals and their corresponding target category labels are added to the training set to update the training set for the current iteration process.

[0072] Optionally, after updating the training set in the current iteration, the electronic device can control the entry into the next iteration to optimize the model parameters again. Specifically, in this embodiment, the electronic device can determine whether the constraints are met after training the data classification model in each iteration. If the constraints are met, the iteration process stops, and steps S30-S50 do not need to be executed further.

[0073] Figure 3 This diagram illustrates a self-evolutionary training process based on a neurophysical engine and causal generation, according to an embodiment of this application. Figure 3 As shown, embodiments of this application can construct an "AI diagnostic model" based on self-evolutionary training using a neurophysical engine and causal generation. A self-playing system of "digital twin generator".

[0074] The AI ​​diagnostic model is a data classification model. The training of the model starts from the initialization stage. First, a data classification model is established based on limited real sample magnetocardiogram signals and basic pathological knowledge to obtain a preliminary AI diagnostic model. At the same time, the digital twin system generates synthetic data covering common pathological patterns as virtual sample magnetocardiogram signals.

[0075] After entering the formal training cycle, each iteration includes four key stages: First, the AI ​​diagnostic model is trained on a mixed dataset (real sample magnetocardiogram signals and virtual sample magnetocardiogram signals); second, its performance is evaluated on an independent validation set (real sample magnetocardiogram signals), and through uncertainty analysis and error case analysis, the cognitive blind spots of the AI ​​diagnostic model are accurately located—for example, the model may be found to have insufficient recognition rate for temperature-sensitive Brugada syndrome or complex arrhythmias; next, the system initiates a reverse reasoning mechanism, using a differentiable neurophysical simulator to map these weaknesses to a high-dimensional pathological parameter space, identifying "high-entropy regions" (i.e., parameter combinations with the lowest AI confidence); then, the causal generation network generates targeted challenge samples based on the precise coordinates of these high-entropy regions, under the constraints of a structured causal model. These samples are both consistent with medical logic and precisely at the AI ​​decision boundary; finally, these carefully designed "marginal cases" are included in the next round of training dataset, while the course learning mechanism dynamically adjusts the training difficulty, consistency regularization ensures that the model does not overly rely on generated data features, and the gradient inversion layer eliminates data domain bias.

[0076] Based on the above iterative training process, the entire system is like a tireless medical mentor, constantly creating "tailor-made" exam questions for AI—from single pathologies to complex mechanisms, from ideal conditions to high-noise environments. AI continues to evolve in dealing with these increasingly challenging virtual cases, and ultimately achieves or even surpasses the effects of traditional big data training in data-scarce scenarios such as rare disease diagnosis.

[0077] Figure 4 This diagram illustrates an application scenario of a self-evolutionary training method based on a neurophysical engine and causal generation, according to an embodiment of this application. Figure 4 As shown in the embodiments of this application, a Federated Generative Learning (FGL) framework can also be established to assist different medical institutions in deploying local self-evolutionary training platforms as lightweight generative agents. A central server periodically collects data uploaded by each evolutionary training platform. This data includes latent space prior distribution statistics (mean, covariance), rather than the raw data or complete weights. The central server can aggregate and generate a globally universal generator, supporting knowledge transfer across populations and devices. The global aggregation result is then returned to the self-evolutionary training platform for model training, resulting in the corresponding data classification model. The data uploaded by each platform can also include personalized request information; for example, a hospital might request the generation of a rare disease dataset that matches the characteristics of its local population. Differential privacy noise and homomorphic encryption communication protocols can be incorporated into the information transmission between platforms to ensure data security.

[0078] Optionally, based on this federated generative learning collaborative framework, embodiments of this application can achieve lightweight deployment of the algorithm model through a cloud-edge collaborative scheme. That is, embodiments of this application can place the most resource-intensive causal generation and global model aggregation in the cloud, while each hospital node only deploys a lightweight local classification model, error diagnosis module, and basic simulation query interface, and obtains customized augmented data from the cloud through a request-response method.

[0079] Based on this federated generative learning collaborative framework, global rare disease data co-construction and sharing are achieved, breaking down data silos. This architecture cleverly replaces "data-level collaboration" with "generative collaboration," achieving a four-in-one goal of "privacy protection, data creation, knowledge sharing, and continuous evolution." It fundamentally solves privacy and compliance barriers, enabling knowledge flow between different platforms without altering the data itself. Simultaneously, the synthetic data generated through global aggregation covers a wider range of pathological phenotypes and population characteristics, with quality far exceeding that of a single center. Furthermore, through the global knowledge dissemination, each center can generate synthetic data more suited to the characteristics of its local patient population, achieving a "fusion of general knowledge and local characteristics."

[0080] After training the data classification model, this embodiment can also test the self-evolutionary trained data classification model on a completely independent, scarce real-world rare disease clinical dataset. This testing process can be compared with data classification models trained using traditional augmentation methods or methods without self-evolutionary training. The accuracy of each model's classification results is compared and evaluated to determine the effectiveness of the data classification model obtained through the self-evolutionary training method.

[0081] Based on the aforementioned technical features, this application embodiment fundamentally solves the bottleneck of scarce rare cardiac pathology data by constructing a magnetocardiogram (MCG) digital twin system that integrates differentiable neurophysical simulation, causal generation, and federated learning. This allows for the training of reliable data classification models even in data-scarce scenarios. This solution significantly alleviates the problem of scarce rare cardiac pathology data and improves the sufficiency and generalization ability of the data classification model training. The generated virtual training data possesses clear causal interpretability and high reliability, making it suitable for medical education, mechanism research, and virtual drug trials. Furthermore, targeted training using high-entropy region identification enhances the model's robustness to difficult identifications and optimizes model performance. This further promotes the leap of magnetocardiography technology towards "digital twin + intelligent decision-making" and constructs a sustainably evolving medical AI ecosystem, contributing to the development of precision medicine.

[0082] It should be understood that although the steps in the above flowcharts are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the above flowcharts may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0083] Based on the foregoing embodiments, this application provides a self-evolving training device based on a neurophysical engine and causal generation. The device includes various modules and units included in each module, which can be implemented by a processor; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field-programmable gate array (FPGA), etc.

[0084] Figure 5 This diagram illustrates a self-evolutionary training device based on a neurophysical engine and causal generation, according to an embodiment of this application. Figure 5 As shown, the self-evolving training device based on a neurophysical engine and causal generation in this application includes: The training set acquisition module 50 is used to determine a training set that includes multiple sample magnetocardiogram signals and corresponding category labels. The sample magnetocardiogram signals include real sample magnetocardiogram signals and virtual sample magnetocardiogram signals. Iterative training module 51 is used to take the training set as the initial training set and execute the following steps iteratively multiple times: Train the data classification model based on the current training set, and identify at least one target category label whose classification performance does not meet the preset requirements after training. Based on the structural causal model, corresponding pathological parameter information is generated from the magnetocardiogram signals of virtual samples corresponding to the target category labels; By using a differentiable neurophysical simulator to perform a forward solution based on pathological parameter information, the corresponding virtual sample magnetocardiogram signal is obtained. The training set is updated based on the virtual sample magnetocardiogram signals and the corresponding target category labels.

[0085] In one possible implementation, the training set acquisition module 50 is further used for: Collect real magnetocardiogram signals from multiple patients; Based on the cardiac imaging data and corresponding real magnetocardiogram signals of each patient, a simulation model is created to obtain a digital twin generator for each patient; Each digital twin generator generates a corresponding magnetocardiogram (MCG) signal as a virtual MCG signal; Real and virtual magnetocardiogram (MCG) signals are used as sample MCG signals, and the training set is determined based on the sample MCG signals and their corresponding category labels.

[0086] In one possible implementation, the training set acquisition module 50 is further used for: The sample magnetocardiogram signal is encoded by a neural radiation field encoder to obtain a high-fidelity sample magnetocardiogram signal.

[0087] In one possible implementation, the iterative training module 51 is further used for: Train the data classification model based on the training set; The validation set is obtained by acquiring real sample magnetocardiogram signals and corresponding category labels from the training set. Input the magnetocardiogram signals of the samples in the validation set into the trained data classification model, and output the corresponding predicted labels and confidence scores; The target category label is determined based on the predicted label and confidence level corresponding to the magnetocardiogram signal of each sample in the validation set.

[0088] In one possible implementation, the iterative training module 51 is further used for: High-entropy regions are explored based on the magnetocardiogram signals of each sample in the validation set and the corresponding predicted labels and confidence levels. The high-entropy regions discovered during exploration are fed back into the structural causal model to automatically generate at least one target category label.

[0089] In one possible implementation, the iterative training module 51 is further used for: Obtain the magnetocardiogram (MCG) signals of virtual samples corresponding to the target category labels in the current training set; Abnormal pathological information is obtained by inversely solving the magnetocardiogram signal of the virtual sample corresponding to the target category label using a differentiable neurophysical simulator; Based on abnormal pathological information and structural causal models, corresponding pathological parameter information is obtained.

[0090] In one possible implementation, the iterative training module 51 is further used for: The virtual sample magnetocardiogram signal is encoded by a neural radiation field encoder to obtain a high-fidelity virtual sample magnetocardiogram signal. The newly added virtual sample magnetocardiogram signals and their corresponding target category labels are added to the training set.

[0091] The descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0092] It should be noted that, in the embodiments of this application... Figure 5 The module division shown in the self-evolutionary training device based on neurophysical engines and causal generation is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this application can be integrated into a single processing unit, exist as separate physical entities, or be integrated into a single unit. The integrated units can be implemented in hardware, as software functional units, or a combination of both.

[0093] It should be noted that, in the embodiments of this application, if the above-described methods are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0094] Figure 6 A schematic diagram of an electronic device according to an embodiment of this application is shown. For example... Figure 6 As shown in the figure, this application provides an electronic device, which can be a server, and its internal structure diagram can be as follows. Figure 6 As shown, the electronic device includes a processor 620, a memory, and a transceiver 640 connected via a system bus 610. The processor 620 provides computing and control capabilities. The memory includes a non-volatile storage medium 631 and internal memory 632. The non-volatile storage medium 631 stores an operating system, computer programs, and a database. The internal memory 632 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium 631. The database stores data. The transceiver 640 communicates with external terminals via a network connection. The computer program, when executed by the processor 620, implements the methods described above.

[0095] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor 620, implements the steps of the method provided in the above embodiments.

[0096] This application provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the steps in the method provided in the above-described method embodiments.

[0097] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0098] In one possible implementation, the apparatus provided in this application can be implemented as a computer program, which can be configured as follows: Figure 6 The device operates on the electronic device shown. The memory of the electronic device can store the various program modules that make up the above-described apparatus. The computer program composed of the various program modules causes the processor 620 to execute the steps of the methods in the various embodiments of this application described in this specification.

[0099] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium, storage medium, and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0100] It should be understood that the phrases "one embodiment," "an embodiment," or "some embodiments" mentioned throughout the specification mean that a specific feature, structure, or characteristic related to an embodiment is included in at least one embodiment of this application. Therefore, phrases such as "in one possible implementation," "in one embodiment," or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. The descriptions of the various embodiments above tend to emphasize the differences between the various embodiments; their similarities or commonalities can be referred to mutually, and for the sake of brevity, they will not be repeated here.

[0101] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three kinds of relationships. For example, object A and / or object B can represent three situations: object A exists alone, object A and object B exist simultaneously, and object B exists alone.

[0102] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0103] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or modules can be electrical, mechanical, or other forms.

[0104] The modules described above as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0105] In addition, each functional module in the various embodiments of this application can be integrated into one processing unit, or each module can be a separate unit, or two or more modules can be integrated into one unit; the integrated modules can be implemented in hardware or in the form of hardware plus software functional units.

[0106] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0107] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0108] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0109] The features disclosed in the several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0110] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0111] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A self-evolutionary training method based on neurophysical engines and causal generation, characterized in that, The method includes: A training set is determined, comprising multiple sample magnetocardiogram (MCC) signals and corresponding category labels, wherein the sample MCC signals include real sample MCC signals and virtual sample MCC signals; Using the training set as the initial training set, the following steps are executed iteratively multiple times: The data classification model is trained based on the current training set, and at least one target category label whose classification effect after training does not meet the preset requirements is identified. Based on the structural causal model, corresponding pathological parameter information is generated from the magnetocardiogram signal of the virtual sample corresponding to the target category label; The corresponding virtual sample magnetocardiogram signal is obtained by forward solving based on the pathological parameter information using a differentiable neurophysical simulator. The differentiable neurophysical simulator is a neural operator. The training set is updated based on the virtual sample magnetocardiogram signals and the corresponding target category labels.

2. The method according to claim 1, characterized in that, The determination of the training set includes multiple sample magnetocardiogram signals and corresponding category labels, including: Real magnetocardiogram signals were collected from multiple patients; Based on the cardiac imaging data and corresponding real magnetocardiogram signals of each patient, a simulation model is performed to obtain a digital twin generator for each patient; Each of the digital twin generators generates a corresponding magnetocardiogram signal as a virtual magnetocardiogram signal; The real magnetocardiogram (MCG) signal and the virtual MCG signal are used as sample MCG signals, and the training set is determined based on the sample MCG signals and the corresponding category labels.

3. The method according to claim 2, characterized in that, The determination of the training set, which includes multiple sample magnetocardiogram signals and corresponding category labels, also includes: The sample magnetocardiogram signal is encoded by a neural radiation field encoder to obtain a high-fidelity sample magnetocardiogram signal.

4. The method according to claim 1, characterized in that, The step of training the data classification model based on the current training set and determining at least one target category label whose classification performance after training does not meet the preset requirements includes: The data classification model is trained based on the training set. The validation set is obtained by acquiring the magnetocardiogram signals of real samples in the training set and their corresponding category labels. The sample magnetocardiogram signals from the validation set are input into the trained data classification model, which outputs the corresponding predicted labels and confidence scores. The target category label is determined based on the predicted label and confidence level corresponding to the magnetocardiogram signal of each sample in the validation set.

5. The method according to claim 4, characterized in that, The step of determining the target category label based on the predicted label and confidence level corresponding to the magnetocardiogram signal of each sample in the validation set includes: High-entropy regions are explored based on the magnetocardiogram signals of each sample in the validation set and the corresponding predicted labels and confidence levels. The high-entropy regions discovered during the exploration are fed back into the structural causal model to automatically generate at least one target category label.

6. The method according to claim 1, characterized in that, The step of generating corresponding pathological parameter information based on the virtual sample magnetocardiogram signal corresponding to the target category label according to the structural causal model includes: Obtain the magnetocardiogram signal of the virtual sample corresponding to the target category label in the current training set; Abnormal pathological information is obtained by inversely solving the magnetocardiogram signal of the virtual sample corresponding to the target category label using a differentiable neurophysical simulator. Based on the abnormal pathological information and the structural causal model, the corresponding pathological parameter information is obtained.

7. The method according to claim 1, characterized in that, The step of updating the training set based on the virtual sample magnetocardiogram signal and the corresponding target category label includes: The virtual sample magnetocardiogram signal is encoded by a neural radiation field encoder to obtain a high-fidelity virtual sample magnetocardiogram signal. The newly added virtual sample magnetocardiogram signals and their corresponding target category labels are added to the training set.

8. A self-evolving training device based on a neurophysical engine and causal generation, characterized in that, The device includes: The training set acquisition module is used to determine a training set including multiple sample magnetocardiogram signals and corresponding category labels, wherein the sample magnetocardiogram signals include real sample magnetocardiogram signals and virtual sample magnetocardiogram signals; The iterative training module is used to take the training set as the initial training set and execute the following steps iteratively multiple times: The data classification model is trained based on the current training set, and at least one target category label whose classification effect after training does not meet the preset requirements is identified. Based on the structural causal model, corresponding pathological parameter information is generated from the magnetocardiogram signal of the virtual sample corresponding to the target category label; The corresponding virtual sample magnetocardiogram signal is obtained by forward solving based on the pathological parameter information using a differentiable neurophysical simulator. The differentiable neurophysical simulator is a neural operator. The training set is updated based on the virtual sample magnetocardiogram signals and the corresponding target category labels.

9. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images