CN-122025112-A - COPD complication prediction model training method and COPD complication prediction method and device

CN122025112ACN 122025112 ACN122025112 ACN 122025112ACN-122025112-A

Abstract

The invention belongs to the technical field of complication prediction, and provides a training method of a COPD (chronic obstructive pulmonary disease) complication prediction model and a COPD complication prediction method and device aiming at the problem of low accuracy of chronic obstructive pulmonary disease complication prediction, wherein the training method comprises the steps of acquiring multi-mode data of a patient to obtain pre-processed clinical and image characteristics; inputting clinical and image features into a COPD complication prediction model based on energy diffusion and knowledge distillation to obtain a complication classification or risk score, determining a cross entropy loss value based on the complication classification or risk score and a real label of data, determining a model total loss according to the cross entropy loss value and the double knowledge distillation loss value, and adjusting model parameters based on the total loss to obtain a target COPD complication prediction model. According to the invention, through the COPD complication prediction model, the prediction accuracy is improved.

Inventors

YU WEIWEI
WANG QI
XIONG YUHAN

Assignees

华中科技大学同济医学院附属同济医院

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. A method of training a predictive model for COPD complications, the method comprising: Acquiring multi-mode data of a COPD patient and preprocessing the multi-mode data to obtain preprocessed clinical and image characteristics; inputting clinical and image features into a COPD complication prediction model based on energy diffusion and knowledge distillation to obtain a complication classification result or risk score corresponding to the clinical and image features; Determining a cross entropy loss value based on the complication classification result or the risk score and the real label of the data; Determining a total loss value of the model according to the cross entropy loss value and the double knowledge distillation loss value; Based on the total loss value of the model, model parameters of the COPD complication prediction model are adjusted, and the model after the last parameter adjustment is used as a target COPD complication prediction model.
2. The method of claim 1, wherein the energy spread and knowledge distillation based COPD complication prediction model comprises a feature encoding network and a knowledge distillation based energy spread model, wherein: A feature encoding network for mapping the input features into one potential representation of potential space; An energy diffusion model based on knowledge distillation is used to diffuse energy potentially representing in the forward process and predict in the reverse process the type of complications that are caused by COPD and the risk score for that complication.
3. The method according to claim 2, wherein the knowledge-based distillation energy diffusion model is constructed by: defining a forward diffusion process controlled by a set of leachable energy functions that progressively increases noise during the time step T increases from 0 to T and makes the potential representation of time step T z t approach a simple prior; Defining a reverse generation process, controlled by another set or the same set of leachable energy functions as in the forward diffusion process, to progressively denoise from the potential representation z T of time step T back to the potential representation z 0 of the initial time; A noise prediction branch and a complication classification branch are connected at the end of the reverse generation process, so that the complication type or risk score is predicted and output.
4. A method according to claim 3, wherein in the noise prediction branch, a pre-trained MEDICALCLIP model is taken as a teacher model, an energy diffusion model is taken as a student model, and knowledge distillation is performed on the energy diffusion model and the teacher model as the student model, so as to obtain the prediction feature most fitting the potential representation z 0 at the initial moment.
5. The method of claim 4, wherein the inputting the clinical and image features into the COPD complication prediction model based on energy spread and knowledge distillation, results in a complication classification result or risk score corresponding to the clinical and image features, comprises: Mapping the preprocessed clinical and image features into potential representations of initial moments through a feature coding network; Forward energy diffusion is performed on the potential representation of the initial moment through T steps; Introducing KL divergence alignment and gradient matching items of denoising prediction means in each time step in the reverse generation process, determining double knowledge distillation loss, and obtaining a trained noise prediction branch by minimizing the double knowledge distillation loss; denoising the potential representation z t of the time step t output by the trained noise prediction branch to obtain denoising features, inputting the denoising features into the complication classification branch, and outputting the classification result or risk score of the complication.
6. The method of claim 5, wherein the double knowledge distillation loss is: Wherein, the And Is the time step t Dynamically adjusted weight coefficients for balancing the contribution of mean alignment and gradient alignment; Representing the divergence; the denoising prediction probability distribution of the student model at the time step t is as follows; the model is denoising prediction probability distribution of the teacher model based on the advanced feature f large and the time step t; And Representing the denoised prediction gradient of the student model and the teacher model over time step t for the potential representation z t respectively, Representing the denoising predictive mean function of the student model, And representing the denoising prediction mean function of the teacher model.
7. The method according to claim 2, wherein the end of the COPD complication prediction model based on energy spread and knowledge distillation is further provided with a COPD staging diagnostic branch for outputting a risk level of COPD according to the inputted potential representation.
8. A method of predicting COPD complications, comprising: acquiring multi-modal data of a target patient; Inputting the preprocessed multi-mode data into a target COPD complication prediction model for carrying out complication prediction, wherein the target COPD complication prediction model is obtained according to the COPD complication prediction model training method of any one of claims 1-7; judging the current reasoning mode, if the reasoning mode is a diagnosis mode, outputting the complication type and risk score of the target patient, otherwise, judging the complication type of the patient if the reasoning mode is a generation contrast mode.
9. The COPD complication prediction method of claim 8 wherein determining the most likely type of complication in the patient if the inference pattern is a generated contrast pattern comprises: by increasing the noise from a priori or from the initial potential representation z 0 to z t and then reversely sampling for a plurality of times, a sample set under the complication assumption with the category of S is obtained; and determining the type of the complications of the patient by comparing the reconstruction distance of the potential representation of the real data of the patient with each data in a sample set under different types of complications hypothesis and selecting a sample closest to the potential representation of the real data of the patient.
10. A COPD complication prediction apparatus, the apparatus comprising: the acquisition module is used for acquiring multi-mode data of a target patient; The system comprises a detection module, a target COPD complication prediction model, a data processing module and a data processing module, wherein the detection module is used for inputting the preprocessed multi-mode data into the target COPD complication prediction model to conduct complication prediction, and the target COPD complication prediction model is obtained according to the COPD complication prediction model training method of any one of claims 1-7; and the output module is used for outputting the complication category or risk score of the target patient.

Description

COPD complication prediction model training method and COPD complication prediction method and device Technical Field The application relates to the technical field of chronic obstructive pulmonary disease complication prediction, in particular to a COPD complication prediction model training method and a COPD complication prediction method and device. Background COPD is an irreversible chronic airway inflammatory disease, comprising two manifestations of chronic bronchitis and emphysema. The key feature is that airway obstruction results in a restricted flow of breathing gas and the condition is progressively aggravated over time. Chronic Obstructive Pulmonary Disease (COPD) has become a worldwide high-grade respiratory disease, particularly in the elderly population, the high disability rate and high mortality rate of COPD making it a significant challenge for clinical management. The complications of COPD are of a wide variety and the course of the disease is complex, which presents great difficulties for clinical diagnosis and patient management. Common complications such as pulmonary heart disease, respiratory failure, pneumonia and the like not only seriously affect the life quality of patients, but also increase the burden of medical systems. Traditional COPD complication diagnostic methods typically rely on a single type of features, such as blood oxygen saturation, lung function testing, etc., while ignoring the potentially highly coupled relationships between multimodal data (including biochemical indicators, imaging features, time series information, etc.). Furthermore, clinical manifestations of COPD patients often present non-gaussian distributions and there is a significant imbalance in the data, which makes it difficult for predictive models based on simple assumptions to effectively capture complex disease changes, resulting in low current COPD complications predictive rates. In order to cope with the challenge, the invention can more flexibly simulate the noise injection and denoising process by describing forward and reverse diffusion distribution through the diffusion model guided by the energy function in the multi-mode data-driven COPD complication identification and risk assessment scene, and accurately classify the stage, severity and the like of the complications in a potential space by combining with multi-task output. The energy function can better adapt to non-Gaussian distribution and allow the existence of a multimodal structure, so that the multi-modal data fusion method has strong expression capacity in the aspect of multi-modal data fusion, and can provide higher sensitivity and precision in critical tasks such as early detection, rare complication diagnosis and the like. Meanwhile, by means of a pre-trained large model knowledge distillation technology, the prediction accuracy of the model is further improved, and different requirements of clinic on instant diagnosis and subsequent deep analysis are met. Disclosure of Invention In view of the above problems, the embodiment of the application provides a training method of a COPD complication prediction model, a COPD complication prediction method and a COPD complication prediction device, which aim at solving the problem of low accuracy of chronic obstructive pulmonary disease complication prediction, and the accuracy of prediction is improved by establishing a COPD complication prediction model based on energy diffusion and knowledge distillation. In a first aspect, embodiments of the present application provide a method for training a COPD complication prediction model, the method comprising: Acquiring multi-mode data of a COPD patient and preprocessing the multi-mode data to obtain preprocessed clinical and image characteristics; inputting clinical and image features into a COPD complication prediction model based on energy diffusion and knowledge distillation to obtain a complication classification result or risk score corresponding to the clinical and image features; Determining a cross entropy loss value based on the complication classification result or the risk score and the real label of the data; Determining a total loss value of the model according to the cross entropy loss value and the double knowledge distillation loss value; Based on the total loss value of the model, model parameters of the COPD complication prediction model are adjusted, and the model after the last parameter adjustment is used as a target COPD complication prediction model. Optionally, the COPD complication prediction model based on energy diffusion and knowledge distillation comprises a feature encoding network and an energy diffusion model based on knowledge distillation, wherein: A feature encoding network for mapping the input features into one potential representation of potential space; An energy diffusion model based on knowledge distillation is used to diffuse energy potentially representing in the forward process and predict in the