CN-121980358-A - Self-evolution training method and device based on neurophysiologic engine and causal generation, electronic equipment and storage medium

CN121980358ACN 121980358 ACN121980358 ACN 121980358ACN-121980358-A

Abstract

The application relates to the field of intelligent medical treatment, in particular to a self-evolution training method, device, electronic equipment and storage medium based on a neurophysical engine and causal generation, which are used for determining a plurality of real sample magnetocardiogram signals, virtual sample magnetocardiogram signals and corresponding class labels as initial training sets, carrying out model training on a data classification model according to the training sets for a plurality of times in an iterative mode, and determining target class labels with classification effects which do not meet preset requirements after training. And generating pathological parameter information according to the structural causal model and the virtual sample magnetocardiogram signals of the target class labels. And forward solving the obtained virtual sample magnetocardiogram signals through the differential neurophysiologic simulator to update the training set. According to the application, the data classification model is trained for multiple times, the class which is difficult to identify is automatically identified after each training, and the corresponding training sample is automatically added, so that the training is performed again, and the model training effect under the data scarcity scene is ensured.

Inventors

ZHANG YADAN
CUI YANGYANG
YAO LI
XIANG MIN
ZHENG SHIQIANG
CUI HE

Assignees

杭州极弱磁场国家重大科技基础设施研究院

Dates

Publication Date: 20260505
Application Date: 20260403

Claims (10)

1. A self-evolutionary training method based on a neurophysiologic engine and causal generation, the method comprising: Determining a training set comprising a plurality of sample magnetocardiogram signals and corresponding class labels, wherein the sample magnetocardiogram signals comprise real sample magnetocardiogram signals and virtual sample magnetocardiogram signals; taking the training set as an initial training set, and executing the following steps for a plurality of times in an iterative mode: Model training is carried out on the data classification model according to the current training set, and at least one target class label with the classification effect which does not meet the preset requirement after training is determined; Generating corresponding pathological parameter information according to the virtual sample magnetocardiogram signals corresponding to the target class labels according to the structural causal model; forward solving is carried out through a differentiable neural physical simulator based on the pathological parameter information, and corresponding virtual sample magnetocardiogram signals are obtained; and updating the training set according to the virtual sample magnetocardiogram signal and the corresponding target class label.
2. The method of claim 1, wherein the determining a training set comprising a plurality of sample magnetocardiogram signals and corresponding class labels comprises: Collecting real magnetocardiogram signals of a plurality of patients; Performing simulation modeling according to the heart image data of each patient and the corresponding real magnetocardiogram signals to obtain a digital twin generator of each patient; Generating a corresponding magnetocardiogram signal as a virtual magnetocardiogram signal according to each digital twin generator; and taking the real magnetocardiogram signal and the virtual magnetocardiogram signal as sample magnetocardiogram signals, and determining a training set according to the sample magnetocardiogram signals and the corresponding class labels.
3. The method of claim 2, wherein the determining comprises a training set of a plurality of sample magnetocardiogram signals and corresponding class labels, further comprising: And carrying out signal coding on the sample magnetocardiogram signal through a nerve radiation field coder to obtain a high-fidelity sample magnetocardiogram signal.
4. The method according to claim 1, wherein the performing model training on the data classification model according to the current training set and determining at least one target class label whose trained classification effect does not meet a preset requirement comprises: model training is carried out on the data classification model according to the training set; Obtaining a real sample magnetocardiogram signal in the training set and a corresponding class label to obtain a verification set; inputting the sample magnetocardiogram signals in the verification set into a trained data classification model, and outputting corresponding prediction labels and confidence coefficients; and determining a target category label according to the prediction label and the confidence coefficient corresponding to each sample magnetocardiogram signal in the verification set.
5. The method of claim 4, wherein said determining a target class label based on the confidence level and the prediction label corresponding to each of the sample magnetocardiogram signals in the verification set comprises: Searching a high-entropy region according to each sample magnetocardiogram signal in the verification set, the corresponding prediction label and the confidence coefficient; And feeding the high-entropy area obtained through exploration back to the structural causal model, and automatically generating at least one target category label.
6. The method of claim 1, wherein generating corresponding pathological parameter information from virtual sample magnetocardiogram signals corresponding to the target class labels according to a structural causal model comprises: obtaining a virtual sample magnetocardiogram signal corresponding to the target class label in a current training set; Reversely solving the abnormal pathological information based on the virtual sample magnetocardiogram signals corresponding to the target class labels through a differentiable neural physical simulator; and obtaining corresponding pathological parameter information according to the abnormal pathological information and the structural causal model.
7. The method of claim 1, wherein updating the training set based on the virtual sample magnetocardiogram signal and the corresponding target class label comprises: Performing signal coding on the virtual sample magnetocardiogram signal through a nerve radiation field coder to obtain a virtual sample magnetocardiogram signal with high fidelity; And adding the newly added virtual sample magnetocardiogram signals and the corresponding target class labels into the training set.
8. A self-evolving training device based on neurophysiologic engine and causal generation, the device comprising: The training set acquisition module is used for determining a training set comprising a plurality of sample magnetocardiogram signals and corresponding class labels, wherein the sample magnetocardiogram signals comprise real sample magnetocardiogram signals and virtual sample magnetocardiogram signals; The iterative training module is used for taking the training set as an initial training set, and executing the following steps for a plurality of times in an iterative mode: Model training is carried out on the data classification model according to the current training set, and at least one target class label with the classification effect which does not meet the preset requirement after training is determined; Generating corresponding pathological parameter information according to the virtual sample magnetocardiogram signals corresponding to the target class labels according to the structural causal model; forward solving is carried out through a differentiable neural physical simulator based on the pathological parameter information, and corresponding virtual sample magnetocardiogram signals are obtained; and updating the training set according to the virtual sample magnetocardiogram signal and the corresponding target class label.
9. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that the processor implements the steps of the method of any of claims 1 to 7 when the program is executed.
10. A computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the method according to any one of claims 1 to 7.

Description

Self-evolution training method and device based on neurophysiologic engine and causal generation, electronic equipment and storage medium Technical Field The embodiment of the application relates to the field of intelligent medical treatment, and relates to a self-evolution training method, a device, electronic equipment and a storage medium based on a neurophysiologic engine and causal generation. Background Magnetocardiography (MCG) is a non-invasive technique for detecting cardiac electrical activity, which can capture millisecond-level electrical activation propagation process and weak magnetic field changes, and has unique advantages in early arrhythmia recognition, focus positioning and other aspects. In recent years, deep learning has been widely used in MCG signal analysis to achieve automated diagnostics, but its performance is severely dependent on large-scale, high-quality and well-annotated training datasets. However, for rare cardiac pathology patterns (e.g., ventricular tachycardia, latent Brugada phenotype of particular anatomical origin), clinically available samples are extremely limited, resulting in poor generalization and insufficient robustness of the existing AI model. While traditional data enhancement means (such as time shifting and noise adding) lack physiological significance, mainstream generation countermeasure networks (GANs) or diffusion models can generate visually lifelike signals, but often violate biophysical rules, so that false correlation misleads AI learning. Disclosure of Invention In view of this, the self-evolution training method, device, electronic equipment and storage medium based on the neurophysiologic engine and causal generation provided by the embodiment of the application aim to provide a self-evolution training scheme for guaranteeing the accuracy of a classification model in a magnetocardiogram data scarcity scene. The self-evolution training method, the device, the electronic equipment and the storage medium based on the neurophysiologic engine and the causal generation are realized as follows: In one aspect of the embodiments of the present application, a self-evolutionary training method based on a neurophysiologic engine and causal generation is provided, the method comprising: Determining a training set comprising a plurality of sample magnetocardiogram signals and corresponding class labels, wherein the sample magnetocardiogram signals comprise real sample magnetocardiogram signals and virtual sample magnetocardiogram signals; taking the training set as an initial training set, performing the following steps for a plurality of times in an iterative mode: Model training is carried out on the data classification model according to the current training set, and at least one target class label with the classification effect which does not meet the preset requirement after training is determined; generating corresponding pathological parameter information according to the virtual sample magnetocardiogram signals corresponding to the target class labels according to the structural causal model; Forward solving is carried out through a differentiable neural physical simulator based on pathological parameter information, and corresponding virtual sample magnetocardiogram signals are obtained; And updating the training set according to the virtual sample magnetocardiogram signal and the corresponding target class label. In one possible implementation, determining a training set comprising a plurality of sample magnetocardiogram signals and corresponding class labels comprises: Collecting real magnetocardiogram signals of a plurality of patients; Performing simulation modeling according to heart image data of each patient and corresponding real magnetocardiogram signals to obtain a digital twin generator of each patient; generating a corresponding magnetocardiogram signal as a virtual magnetocardiogram signal according to each digital twin generator; and taking the real magnetocardiogram signal and the virtual magnetocardiogram signal as sample magnetocardiogram signals, and determining a training set according to the sample magnetocardiogram signals and the corresponding class labels. In one possible implementation, determining the training set including the plurality of sample magnetocardiogram signals and the corresponding class labels further includes: And (3) carrying out signal coding on the sample magnetocardiogram signal through a nerve radiation field coder to obtain a high-fidelity sample magnetocardiogram signal. In one possible implementation manner, performing model training on the data classification model according to the current training set, and determining at least one target class label with a trained classification effect which does not meet a preset requirement, where the method includes: Model training is carried out on the data classification model according to the training set; acquiring a real sample magnetocardiogram signal in a tra