A method, device, medium and program product for constructing a refractory mycoplasma pneumoniae pneumonia prediction model

Abstract

This invention belongs to the field of medical auxiliary prediction model construction technology. Current prediction models for refractory Mycoplasma pneumoniae pneumonia are all multimodal and rely on time-consuming experimental results, making them unsuitable for the clinical need to quickly determine the probability of having refractory Mycoplasma pneumoniae pneumonia. This invention proposes a method, device, medium, and program product for constructing a prediction model for refractory Mycoplasma pneumoniae pneumonia. By developing a Transformer-based deep learning framework, it achieves automated and efficient risk identification of RMPP using baseline non-contrast chest CT without additional examinations. Furthermore, a first prediction model is constructed based on the risk probability values ​​output by the Transformer-based deep learning framework, and a second prediction model is constructed based on the risk probability values ​​and six clinical data points. The predicted probabilities of each model show good consistency with the observed results.

Description

Technical Field

[0001] This invention belongs to the field of medical auxiliary prediction model construction technology, specifically relating to a method, equipment, medium and program product for constructing a prediction model for refractory mycoplasma pneumonia. Background Technology

[0002] Mycoplasma pneumoniae pneumonia (MPP) is a leading cause of community-acquired pneumonia in children. Despite the standard of care being macrolides, the increasing prevalence of macrolide-resistant strains, particularly in East Asia, presents a significant clinical challenge. Consequently, a considerable proportion of children develop refractory MPP (RMPP), characterized by persistent fever and progressive radiographic findings despite treatment. RMPP frequently requires second-line antibiotics or immunomodulators and is associated with an increased risk of severe long-term sequelae, including pulmonary necrosis and obliterative bronchiolitis. Therefore, accurate identification of pediatric RMPP is crucial for guiding timely clinical intervention.

[0003] However, accurate and timely identification of RMPP remains challenging. While chest X-ray is the standard first-line imaging modality for two-way MPP, computed tomography (CT) is often required for patients with unresolved clinical symptoms or suspected severe parenchymal lung involvement. Following CT, clinicians must rapidly determine the probability of RMPP to inform critical decisions regarding corticosteroid initiation or antibiotic escalation. In this acute clinical setting, although various RMPP prediction models have been developed, most current models are multimodal and rely on time-consuming laboratory results, manual data entry, subjective radiological interpretation, or complex image preprocessing. These limitations hinder clinical application and fail to meet the demands of real-time decision-making in the high-pressure environment of pediatric settings.

[0004] Therefore, there is an urgent clinical need for a simplified and timely risk stratification model to address acute cases. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method, device, medium, and program product for constructing a predictive model for refractory mycoplasma pneumoniae pneumonia. This invention utilizes a Transformer-based deep learning framework (trans-DLF) to achieve automated and efficient risk identification of RMPP (Recurrent Risk Prediction Processing) using baseline non-contrast-enhanced chest 3D CT volumetric data. For patients who have already undergone chest 3D CT examination due to clinical concern, this model enables rapid bedside risk stratification, reduces the burden of emergency decision-making, and ensures timely intervention to prevent serious complications.

[0006] The first aspect of this invention provides a CT-based method for predicting refractory mycoplasma pneumoniae pneumonia, comprising: The patient's original three-dimensional CT volume data of the chest were standardized to obtain lung parenchyma volume data; Based on the lung parenchyma volume data, slice groups are divided and feature extraction and spatial aggregation are performed to generate slice feature sequences; The sliced ​​feature sequence is input into the Transformer encoder to perform global dependency modeling and obtain global context features; The global context features are input into the classifier to generate risk probability values ​​for refractory mycoplasma pneumonia.

[0007] Furthermore, the standardization processing of the patient's original three-dimensional CT volume data of the chest to obtain lung parenchyma volume data includes: The lung segmentation model was used to segment the patient's original three-dimensional CT volume data of the chest and remove irrelevant anatomical regions to obtain the mask of the left and right lung parenchyma regions. The mask of the left and right lung parenchyma regions is fused and enhanced with the corresponding original three-dimensional CT volume data of the chest to obtain the fused volume data. The fused volume data is then cropped and spatially normalized to obtain standardized lung parenchyma volume data with uniform dimensions.

[0008] Furthermore, the lung segmentation model includes any one of U-Net, V-Net, UNet++, and MK-Unet.

[0009] Furthermore, the step of dividing the lung parenchyma volume data into slice groups, extracting features, and spatially aggregating them to generate slice feature sequences includes: Along the axis of the lung parenchyma volume data, every N consecutive slices are divided into a slice group; where N is a natural number greater than 1. A two-dimensional convolutional neural network is used to extract features from each slice group to obtain the lung feature map corresponding to each slice group; Spatial aggregation is performed on each of the lung feature maps to generate a feature vector of the corresponding local morphology and texture features of the lung; The feature vectors of all slice groups are integrated to obtain the slice feature sequence.

[0010] Furthermore, the Transformer encoder includes an M-layer coding unit; Each coding unit includes a multi-head self-attention module for capturing non-local slice dependencies in the slice feature sequence, a position coding module coupled to the multi-head attention module for introducing the original spatial position information of the slices in the slice feature sequence along the axis to the multi-head self-attention module, and a residual connection feedforward network module coupled to the position coding module for refining the features in the slice feature sequence.

[0011] Furthermore, the multi-head self-attention module captures non-local slice dependencies in the slice feature sequence by calculating the relationship between any M-th feature vector and the (M+p)-th feature vector in the slice feature sequence. Here, M is a natural number greater than or equal to 1, and p is a natural number greater than or equal to 1. Furthermore, the position encoding module employs relative position encoding to maintain the spatial order of the slices along the z-axis.

[0012] Furthermore, the two-dimensional convolutional neural network is any one of ResNet-18, MobileNetV2, MobileNetV3, EfficientNet, and ResNeXt.

[0013] A second aspect of this invention provides a method for constructing a predictive model for refractory mycoplasma pneumonia, comprising: The raw three-dimensional CT volume data of the chest of the training set samples are processed according to any of the methods described above to obtain the risk probability value of each training set sample. Obtain the real disease outcomes for each training set sample; Among them, the actual disease outcome includes whether the patient has refractory mycoplasma pneumoniae pneumonia; The regression analysis model is trained using the risk probability value and the actual disease outcome. When the trained model reaches the training stopping condition, a predictive model for refractory mycoplasma pneumonia is obtained, namely the first predictive model.

[0014] Furthermore, the regression analysis model is a logistic regression model or a Cox proportional hazards regression model.

[0015] A third aspect of this invention provides a method for constructing a predictive model for refractory mycoplasma pneumonia, comprising: The raw three-dimensional CT volume data of the chest of the training set samples are processed according to any of the methods described above to obtain the risk probability value of each training set sample. Obtain the real disease outcomes and clinical data for each training set sample; Among them, the actual disease outcome includes whether the patient has refractory mycoplasma pneumoniae pneumonia; clinical data include the duration of fever before consultation, oxygen support requirement, duration of cough before consultation, type of severe / critical pneumonia, body temperature and hemoglobin; The regression analysis model is trained using the aforementioned risk probability values, actual disease outcomes, and clinical data. When the trained model reaches the training stopping condition, a predictive model for refractory mycoplasma pneumoniae pneumonia, i.e., the second predictive model, is obtained.

[0016] Furthermore, the regression analysis model is a logistic regression model or a Cox proportional hazards regression model.

[0017] A fourth aspect of the present invention provides a method for predicting refractory mycoplasma pneumonia, implemented using the first prediction model described above, comprising: Obtain raw three-dimensional CT volume data of the patient's chest, and process the raw three-dimensional CT volume data of the chest using any of the methods disclosed in the first aspect of this application to obtain a risk probability value; The risk probability value is input into the first prediction model to obtain the first prediction probability. Based on the first prediction probability, an auxiliary prediction result of the probability of the patient developing refractory mycoplasma pneumonia is output.

[0018] The fifth aspect of this invention provides a method for predicting refractory mycoplasma pneumoniae pneumonia, comprising: Obtain raw three-dimensional CT volume data and clinical data of the patient to be tested, and process the raw three-dimensional CT volume data of the chest using the method disclosed in any of the first aspects of this application to obtain a risk probability value; The risk probability value and clinical data are input into the second prediction model to obtain the second prediction probability. Based on the second prediction probability, the auxiliary prediction result of the probability of the patient developing refractory mycoplasma pneumonia is output.

[0019] A sixth aspect of the present invention provides a computer device, the device comprising: a memory and a processor; the memory being used to store a computer program; the processor executing the computer program to implement the steps of the method described in any of the preceding claims.

[0020] A seventh aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0021] The eighth aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0022] Compared with the prior art, the advantages of this invention are as follows: 1. This invention develops and validates a rapid, automated framework, Trans-DLF, for accurate risk stratification of RMPP in children. The Trans-DLF framework eliminates the need for time-consuming laboratory data when predicting risk for RMPP in children, ensuring reliable risk stratification at the time of image acquisition. This allows clinical decisions (e.g., initiating corticosteroid use) to be made immediately after CT scans. Furthermore, the Trans-DLF framework proposed in this invention improves clinical workflows under high-pressure environments and provides real-time support as a reliable, stand-alone tool when rapid intervention is required.

[0023] 2. This application, based on the Trans-DLF framework, can effectively assess the risk of RMPP (Recovery of Respiratory Damage Prolapse) using routine CT scan data, eliminating the need for additional examinations and simplifying the assessment process. This invention constructs a first prediction model based on the risk probability values ​​output by the Trans-DLF framework, achieving risk prediction through this model. Furthermore, this invention constructs a second prediction model based on the risk probability values ​​output by the Trans-DLF framework and six clinical data points, achieving risk prediction through this second model. The predicted probabilities of the above models show good consistency with the observed results. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the method flow provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the method flow provided in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the method flow provided in Embodiment 3 of the present invention; Figure 4 This is a schematic diagram of the method flow provided in Embodiment 4 of the present invention; Figure 5 This is a schematic diagram of the method flow provided in Embodiment 5 of the present invention; Figure 6 This is a schematic diagram of the computer device provided in Embodiment 6 of the present invention; Figure 7 This is a schematic diagram of the architecture of an exemplary computing device provided in Embodiment 6 of the present invention; Figure 8 This is a schematic diagram of the storage medium provided in Embodiment 7 of the present invention; Figure 9 This is a flowchart of the research process of the present invention, in which CT is computed tomography, CBC is complete blood cell count, LDH is lactate dehydrogenase, RMPP is refractory mycoplasma pneumonia, MLP is multilayer perceptron, RelPE is relative position encoding, Trans-DLF is a deep learning framework based on Transformer, 3D-CNN is a three-dimensional convolutional neural network, CRP is C-reactive protein, NR is neutrophil ratio, PFD is the duration of fever before medical visit, and DCA is decision curve analysis. Figure 10 This is an overview of the patient inclusion and exclusion criteria and dataset construction of the present invention; Figure 11 This invention provides a nomogram for predicting multimodal refractory mycoplasma pneumoniae pneumonia. Figure 12 The performance of different prediction models on risk stratification of RMPP, among which, Figure 12 AD is the receiver operating characteristic (ROC) curve. Figure 12 EH is a waterfall plot showing the predicted RMPP probability of trans-DLF in the corresponding cohort. Cases were classified using an optimal cutoff value (horizontal baseline) of 0.5174. Blue bars below the baseline represent correctly classified non-RMPP cases, orange bars above the baseline represent correctly classified RMPP cases, and bars crossing the baseline represent misclassified cases. In the figure, 3D-CNN is a 3D convolutional neural network; Clinical is a clinical model; and Nomogram is a nomogram predicting multimodal refractory Mycoplasma pneumoniae pneumonia. Figure 13 The clinical benefits and calibration performance of trans-DLF, 3D convolutional neural networks (3D-CNN), clinical models, and multimodal refractory mycoplasma pneumoniae prediction nomograms are discussed; among them, Figure 13 AE represents decision curve analysis. Figure 13 F represents the calibration plot of trans-DLF across all test sets; in the plot, 3D-CNN represents a 3D convolutional neural network; Clinical represents the clinical model; Nomogram represents the nomogram of predictions for multimodal refractory Mycoplasma pneumoniae pneumonia; Training represents the training set; Validation represents the validation set; ITC represents the internal test set; ETC1 represents external test cohort 1; ETC2 represents external test cohort 2; Ideal represents the ideal calibration (45-degree diagonal line). Figure 14 The results are subgroup analysis results of trans-DLF provided by this invention in a joint external test queue, where the red dashed line represents a reference value of 0.8 on the y-axis.

[0026] Figure 15These are the original axial CT images of RMPP patients and non-RMPP patients, along with corresponding gradient-weighted class activation mapping (Grad-CAM) visualizations; among them, Figure 15 A and B are the original axial CT images and corresponding Grad-CAM visualizations of RMPP patients, respectively. Figure 15 C and D are the original axial CT images and corresponding Grad-CAM visualizations of non-RMPP patients, respectively; In the diagram: 2000, Computer device; 2010, Processor; 2020, Memory; 3000, Computing device; 3010, Bus; 3020, CPU; 3030, Read-only memory; 3040, Random access memory; 3050, Communication port; 3060, Input / output component; 3070, Hard disk; 3080, User interface; 4010, Computer-readable instructions; 4020, Computer-readable storage medium. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0028] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Example 1 like Figure 1 As shown, this embodiment of the invention provides a CT-based method for predicting refractory mycoplasma pneumoniae pneumonia, specifically including: S101: Standardize the patient's original three-dimensional CT volume data of the chest to obtain lung parenchyma volume data; In this embodiment of the invention, the standardization processing of the patient's original three-dimensional CT volume data of the chest to obtain lung parenchyma volume data includes: S1011: Using a lung segmentation model, the patient's original three-dimensional CT volume data of the chest is segmented and irrelevant anatomical regions (such as the chest wall, mediastinum, etc.) are removed to obtain the mask of the left and right lung parenchyma regions. S1012: The left and right lung parenchyma region masks are fused and enhanced with the corresponding original three-dimensional CT volume data of the chest (e.g., by adding voxels) to obtain the fused volume data. S1013: The fused volume data is cropped and spatially normalized to obtain standardized lung parenchyma volume data with uniform size.

[0031] Specifically, the lung segmentation model includes any one of U-Net, V-Net, UNet++, and MK-Unet.

[0032] S102: Based on the lung parenchyma volume data, divide the slice groups, extract features, and spatially aggregate them to generate slice feature sequences; In this embodiment of the invention, the step of dividing the lung parenchyma volume data into slice groups, extracting features, and spatially aggregating them to generate slice feature sequences includes: S1021: Along the axis of the lung parenchyma volume data, divide every N consecutive slices into a slice group; where N is a natural number greater than 1; S1022: Use a two-dimensional convolutional neural network to extract features from each slice group to obtain the lung feature map corresponding to each slice group; Specifically, the two-dimensional convolutional neural network is any one of ResNet-18, MobileNetV2, MobileNetV3, EfficientNet, and ResNeXt.

[0033] S1023: Spatial aggregation is performed on each of the lung feature maps to generate a feature vector of the corresponding local morphology and texture features of the lung; S1024: Integrate the feature vectors of all slice groups to obtain the slice feature sequence.

[0034] S103: Input the sliced ​​feature sequence into the Transformer encoder to perform global dependency modeling and obtain global context features; In this embodiment of the invention, the Transformer encoder includes an M-layer coding unit; Each coding unit includes a multi-head self-attention module for capturing non-local slice dependencies in the slice feature sequence, a position coding module coupled to the multi-head attention module for introducing the original spatial position information of the slices in the slice feature sequence along the axis to the multi-head self-attention module, and a residual connection feedforward network module coupled to the position coding module for refining the features in the slice feature sequence.

[0035] Specifically, the multi-head self-attention module captures non-local slice dependencies in the slice feature sequence by calculating the relationship between any M-th feature vector and the (M+p)-th feature vector in the slice feature sequence. Here, M is a natural number greater than or equal to 1, and p is a natural number greater than or equal to 1; in specific implementations, M can be 1, and p can be 1, meaning the multi-head self-attention module is used to calculate the relationship between any two feature vectors in the slice feature sequence, achieving dynamic learning of non-local slice dependency relationships. Specifically, the position encoding module uses relative position encoding (RelPE) to maintain the spatial order of the slices along the z-axis.

[0036] In some embodiments, M is a natural number greater than 3, preferably 6.

[0037] In some more specific embodiments, the basic data processing flow of the multi-head self-attention module, positional encoding module, and residual connection feedforward network module within the M-layer encoding unit is as follows: the slice feature sequence of S102 is directly input into the first layer encoding unit of the Transformer encoder, and each layer encoding unit processes the data in the following order, with the specific steps for data processing as follows: The input is a slice feature sequence, which can be assumed to have dimensions [L,D], where L is the number of slice groups and D is the feature vector dimension of each slice. Before being input into the multi-head self-attention module, the position encoding module will first mark the spatial order information of the slice feature sequence. Since the Transformer architecture itself does not contain sequence order information, this application utilizes a positional encoding module to label the original spatial order information of the slices in the axial direction, thereby determining the original position of the slices by introducing spatial information. In this application, Relative Positional Encoding (RelPE) is used to calculate the relative distance between slices, such as the offset of the i-th slice from the j-th slice on the z-axis, and this information is incorporated into the attention calculation to obtain a slice feature sequence with added positional encoding information. It is worth noting that during the interaction process, the positional encoding information is usually not directly added to the input slice feature sequence, but rather introduced as a bias term when the multi-head self-attention module calculates the attention weights to dynamically adjust the weights of the dependencies between slices. When the positional encoding information is integrated into the attention mechanism, it enables subsequent self-attention to perceive the spatial context of the slices.

[0038] The slice feature sequence with added positional encoding information is input into a multi-head self-attention module. The module splits the input features into multiple "heads," each independently calculating self-attention. First, a query Q, key K, and value V vector are generated through a linear transformation. Then, an attention score is calculated based on the relative positional bias calculated using the query Q, key K, value V vectors and the relative positional encoding. These scores are weighted and summed to obtain the output of each head. The outputs of all heads are concatenated and linearly transformed to obtain the self-attention output, resulting in a feature sequence with dimensions [L,D] containing global slice dependencies. This feature sequence is then subjected to residual connection and normalization to stabilize the feature distribution. In this application, the relative positional encoding is tightly coupled with the multi-head self-attention module. It is not simply added to the input but serves as a bias term in the attention calculation, dynamically adjusting the weights of interactions between slices to better maintain the axial spatial order of the relative positional relationships of the slices.

[0039] The residual connection feedforward network module performs nonlinear transformation and refinement on the feature series of each slice to enhance the model's expressive power. This module comprises at least two fully connected layers, with an activation function in between. Specifically, the normalized feature sequence is input into the residual connection feedforward network module for feature refinement, resulting in a refined feature sequence with unchanged dimensions. This refined feature sequence is then further residual-connected and layer-normalized with its input to obtain the output of the first-layer encoding unit.

[0040] The above steps are iteratively processed in the M-layer coding unit. The input of the next layer is the output of the previous layer, until the output after processing by the M-layer coding unit is the global context feature.

[0041] S104: Input global contextual features into the classifier to generate risk probability values ​​for refractory mycoplasma pneumonia.

[0042] In this embodiment, the classifier is a multilayer perceptron.

[0043] Example 2 like Figure 2 As shown in the figure, this embodiment of the invention provides a method for constructing a predictive model for refractory mycoplasma pneumonia, specifically including: S201. Process the original three-dimensional CT volume data of the chest of the training set samples according to the method of Example 1 to obtain the risk probability value of each training set sample. S202. Obtain the true disease outcomes for each training set sample; Among them, the actual disease outcome includes whether the patient has refractory mycoplasma pneumoniae pneumonia; S203. The regression analysis model is trained using the risk probability value and the actual disease outcome. When the trained model reaches the training stopping condition, the refractory mycoplasma pneumoniae pneumonia prediction model is obtained, namely the first prediction model.

[0044] Specifically, the regression analysis model is a logistic regression model or a Cox proportional hazards regression model.

[0045] Specifically, training stopping conditions include, but are not limited to: based on the number of iterations (e.g., setting a fixed training threshold in advance, and stopping training once that threshold is reached), based on the loss function (stopping when the loss function value no longer decreases or the decrease is less than a threshold), and based on performance metrics (such as accuracy, recall, AUC, etc.) no longer improving or the improvement is less than a threshold.

[0046] Example 3 like Figure 3 As shown in the figure, this embodiment of the invention provides a method for constructing a predictive model for refractory mycoplasma pneumonia, specifically including: S301. Process the original three-dimensional CT volume data of the chest of the training set samples according to the method of Example 1 to obtain the risk probability value of each training set sample. S302. Obtain the real disease outcomes and clinical data for each training set sample; Among them, the actual disease outcome includes whether the patient has refractory mycoplasma pneumoniae pneumonia; clinical data include the duration of fever before consultation, oxygen support requirement, duration of cough before consultation, type of severe / critical pneumonia, body temperature and hemoglobin; S303. The regression analysis model is trained using the risk probability value, the actual disease outcome, and clinical data. When the trained model reaches the training stopping condition, a predictive model for refractory mycoplasma pneumoniae pneumonia, i.e., the second predictive model, is obtained.

[0047] Specifically, the regression analysis model can be any one of logistic regression, Cox proportional hazards regression, or machine learning algorithms (such as LightGBM, Random Forest, etc.). When using logistic regression, the model is fitted using methods such as maximum likelihood estimation to obtain the regression coefficients of each variable, then linear predicted values ​​are calculated to obtain the prediction model, which is then converted into probability values ​​for constructing a nomogram. When using Cox proportional hazards regression, the algorithm estimates the impact of each variable on disease risk, outputs regression coefficients to obtain the prediction model, and constructs a nomogram based on the regression coefficients to show the risk scores corresponding to different combinations of variables. In complex data scenarios, such as when there are many parameters, common machine learning algorithms can be used to train the parameters and true outcome labels to obtain the prediction model. Interpretability methods (such as SHAP values) are also used to extract important variables and their impacts, and finally, a nomogram is constructed based on the results.

[0048] Specifically, training stopping conditions include, but are not limited to: based on the number of iterations (e.g., setting a fixed training threshold in advance, and stopping training once that threshold is reached), based on the loss function (stopping when the loss function value no longer decreases or the decrease is less than a threshold), and based on performance metrics (such as accuracy, recall, AUC, etc.) no longer improving or the improvement is less than a threshold.

[0049] Example 4 like Figure 4 As shown, this embodiment of the invention provides a method for predicting refractory mycoplasma pneumoniae pneumonia, implemented using the first prediction model of Embodiment 2, including: S401. Obtain the original three-dimensional CT volume data of the patient's chest, and process the original three-dimensional CT volume data of the chest using the method provided in Example 1 to obtain the risk probability value. S402. Input the risk probability value into the first prediction model to obtain the first prediction probability, and output the auxiliary prediction result of the probability of the patient to be tested developing refractory mycoplasma pneumonia based on the first prediction probability.

[0050] Example 5 like Figure 5 As shown, this embodiment of the invention provides a method for predicting refractory mycoplasma pneumoniae pneumonia, implemented using the second prediction model of Embodiment 3, including: S501. Obtain the original three-dimensional CT volume data and clinical data of the patient to be tested, and process the original three-dimensional CT volume data of the chest using the method provided in Example 1 to obtain the risk probability value. S502. Input the risk probability value and clinical data into the second prediction model to obtain the second prediction probability. Based on the second prediction probability, output the auxiliary prediction result of the probability of the patient to be tested developing refractory mycoplasma pneumonia.

[0051] In some more specific embodiments, when the predicted probability is greater than a preset threshold, an auxiliary prediction result indicating a high risk of RMPP is output; when the predicted probability is less than the preset threshold, an auxiliary prediction result indicating a low risk of RMPP is output. The auxiliary prediction results output based on the predicted probability include, but are not limited to, paper or electronic reports. These results are obtained by the intelligent machine based on the subject's relevant data and are intended only as a reference for medical personnel, not as the subject's final diagnostic result.

[0052] In some embodiments, the preset threshold is obtained through training on training set samples. It can be a specific threshold or an interval range. The specific form is not specifically limited in this embodiment.

[0053] Example 6 like Figure 6 As shown, an embodiment of the present invention provides a computer device, the device 2000 may include: one or more processors 2010 and one or more memories 2020; wherein, the memories store computer-readable code, which, when run by the one or more processors 2010, can execute the methods described above.

[0054] The processor in this embodiment can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, operations, and logic block diagrams disclosed in this embodiment. The general-purpose processor can be a microprocessor or any conventional processor, and can be based on an x86 or ARM architecture.

[0055] In general, the various exemplary embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, firmware, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device. When aspects of embodiments of this disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it will be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as non-limiting examples in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0056] For example, the method or apparatus according to embodiments of this disclosure can also be used by means of Figure 7 The architecture of the computing device 3000 shown is used for implementation. For example... Figure 7As shown, the computing device 3000 may include a bus 3010, one or more CPUs 3020, a read-only memory (ROM) 3030, a random access memory (RAM) 3040, a communication port 3050 connected to a network, an input / output component 3060, a hard disk 3070, etc. The storage devices in the computing device 3000, such as the ROM 3030 or the hard disk 3070, may store various data or files used for processing and / or communication of the methods provided in this disclosure, as well as program instructions executed by the CPU. The computing device 3000 may also include a user interface 3080. Of course, Figure 7 The architecture shown is merely exemplary and can be omitted as needed when implementing different devices. Figure 7 One or more components in the computing device shown.

[0057] Example 7 like Figure 8 As shown, this embodiment of the invention provides a computer-readable storage medium 4020 on which computer-readable instructions 4010 are stored. When the computer-readable instructions 4010 are executed by a processor, the methods according to embodiments of this disclosure described with reference to the above drawings can be performed.

[0058] The computer-readable storage medium in the embodiments of this disclosure may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may be random access memory (RAM) used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous interconnected dynamic random access memory (SLDRAM), and direct memory bus random access memory (DR RAM). It should be noted that the memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0059] Example 8 This invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the methods described above.

[0060] Example 9 This invention provides a CT-based predictive system for refractory mycoplasma pneumoniae pneumonia, used to implement the steps of the method described in Embodiment 1, including: The first processing module standardizes the patient's original three-dimensional CT volume data of the chest to obtain lung parenchyma volume data. The slice feature sequence generation module is used to divide slice groups based on the lung parenchyma volume data, extract features, and spatially aggregate them to generate slice feature sequences. The encoder modeling module is used to input the slice feature sequence into the Transformer encoder to perform global dependency modeling and obtain global context features. The Transformer encoder includes M layers of encoding units. Each layer of encoding unit includes: a multi-head self-attention module for capturing non-local dependencies between slices, a position encoding module coupled to the multi-head self-attention module for introducing the original spatial position information of the slice in the axial direction to the multi-head self-attention module, and a residual connection feedforward network module coupled to the position encoding module for feature refinement. The classification and output module is used to input the global context features into the classifier to generate a risk probability value for refractory mycoplasma pneumonia.

[0061] Example 10 This invention provides a system for constructing a predictive model for refractory mycoplasma pneumonia, used to implement the steps described in Embodiment 2, including: The second processing module is used to process the original three-dimensional CT volume data of the chest of the training set samples according to the method of Example 1 to obtain the risk probability value of each training set sample. The first acquisition module is used to acquire the real disease outcome of each training set sample, including whether the sample has refractory mycoplasma pneumoniae pneumonia. The first prediction model training module is used to train the regression analysis model using the risk probability value and the actual disease outcome. When the trained model reaches the training stopping condition, the prediction model for refractory mycoplasma pneumoniae pneumonia, namely the first prediction model, is obtained.

[0062] Example 11 This invention provides a system for constructing a predictive model for refractory mycoplasma pneumonia, used to implement the steps of the method described in Embodiment 3, including: The third processing module is used to process the original three-dimensional CT volume data of the chest of the training set samples according to the method of Example 1 to obtain the risk probability value of each training set sample. The second acquisition module is used to acquire the real disease outcomes and clinical data for each training set sample; The second prediction model training module is used to train the regression analysis model using the risk probability value, the actual disease outcome, and clinical data. When the trained model reaches the training stopping condition, the prediction model for refractory mycoplasma pneumoniae pneumonia, i.e., the second prediction model, is obtained.

[0063] Example 12 This invention provides a predictive system for refractory mycoplasma pneumoniae pneumonia, used to implement the steps of the method described in Example 4, including: The third acquisition module is used to acquire the original three-dimensional CT volume data of the patient's chest and process the original three-dimensional CT volume data of the chest using the method provided in Example 1 to obtain the risk probability value. The first prediction module is used to input the risk probability value into the first prediction model to obtain the first prediction probability, and output the auxiliary prediction result of the probability of the patient to be tested developing refractory mycoplasma pneumonia based on the first prediction probability.

[0064] Example 13 This invention provides a predictive system for refractory mycoplasma pneumonia, used to implement the steps of the method described in Example 5, including: The fourth acquisition module is used to acquire the original three-dimensional CT volume data and clinical data of the patient to be tested, and to process the original three-dimensional CT volume data of the chest using the method provided in Example 1 to obtain the risk probability value. The second prediction module is used to input the risk probability value and clinical data into the second prediction model to obtain the second prediction probability, and output the auxiliary prediction result of the probability of the patient developing refractory mycoplasma pneumonia based on the second prediction probability.

[0065] To more clearly describe the research process of using a CT-based deep learning model for automatic risk stratification of refractory mycoplasma pneumoniae pneumonia in children, this invention uses a study of non-enhanced chest CT data from 1224 children with mycoplasma pneumoniae pneumonia as an example for illustrative description. However, this description is only one use case and is not intended to limit the scope of protection of this data processing method.

[0066] 1. Materials and Methods This study was conducted in accordance with the Declaration of Helsinki and approved by the institutional review committees of all participating hospitals. Because this study employed a retrospective design, informed consent was not required from individuals.

[0067] 1.1. Research Design and Participants This retrospective, multicenter study consecutively recruited pediatric patients with multiple pediatric pediatric malpractice (MPP) from three hospitals representing different levels of healthcare. The development dataset used was built from data from the National Pediatric Referral Center (Center 1) and included 785 inpatients enrolled between August 2019 and December 2023. These patients were randomly assigned in an approximately 3:1:1 ratio to the training set (n=506), validation set (n=140), and internal test set (n=139). To assess generalizability, two independent cohorts were used for external validation: external cohort 1 consisted of 331 patients recruited from a regional tertiary hospital (Center 2) between February 2018 and April 2024, and external cohort 2 consisted of 108 inpatients and outpatients recruited from a large comprehensive general hospital (Center 3) between June 2023 and March 2025. The overall flowchart and study flowchart are shown below. Figure 9 and Figure 10 As shown.

[0068] Inclusion criteria were: (1) age > 28 days and ≤ 18 years; (2) clinically confirmed MPP; (3) availability of at least one non-contrast chest CT scan during the initial disease assessment; and (4) a record of symptom resolution and radiographic absorption.

[0069] Exclusion criteria were: (1) no thin-slice (≤1.5 mm) CT; (2) poor quality CT or unavailable medical digital imaging and communication (DICOM) data; (3) chest CT performed after diagnosis of RMPP; (4) missing complete blood count (CBC) and key inflammatory biomarkers; (5) pre-existing immunodeficiency or immunosuppressive therapy before admission; and (6) requiring mechanical ventilation after admission.

[0070] 1.2. Collection of Clinical and Radiological Data Thin-section (≤1.5 mm), non-contrast chest CT scans and corresponding CT reports were retrieved from the institutional imaging archives and communication systems of each hospital. For all patients, the first available baseline CT scan obtained prior to the RMPP diagnosis was retrieved. The decision to perform a CT scan was made by the treating pediatrician when guidelines were met. These guidelines included: concern about severe lung involvement (such as extensive consolidation or pleural effusion), refractory clinical course 48–72 hours after targeted therapy, need to rule out alternative diagnoses, and inconsistencies between clinical presentation and chest X-ray or assessment of recurrent pneumonia.

[0071] Demographic, clinical, and laboratory data, including complete blood counts and inflammatory markers such as C-reactive protein (CRP), lactate dehydrogenase (LDH), and D-dimer levels, are collected from the electronic hospital information system. Clinical and laboratory data are collected upon admission of inpatients and at the first visit of outpatients.

[0072] RMPP is defined as persistent fever, worsening clinical symptoms, radiographic progression, or extrapulmonary complications occurring at least 7 days after macrolide therapy. Disease severity is classified as mild or severe / critical, and is defined according to the 2023 Chinese Pediatric MPP Guidelines. RMPP status and disease severity are categorized by experienced pediatric pulmonologists.

[0073] 1.3. Framework Construction Based on Deep Learning 1) Framework Architecture This invention develops a Transformer-based deep learning framework for predicting refractory mycoplasma pneumoniae pneumonia (RMPP) from pediatric chest CT scans. The framework aims to capture global inter-slice dependencies and spatial contextual relationships across the entire lung volume.

[0074] 2) Lung segmentation and pretreatment First, a lung segmentation model was used to process all CT volumes to isolate the lung parenchyma and eliminate irrelevant anatomical regions such as the chest wall and mediastinum, reducing background interference and accurately segmenting the left and right lung parenchyma regions. Voxel addition was then used to add each lung mask to the corresponding CT image to enhance parenchymal contrast and disease localization. Subsequently, the volumes within the masks were cropped, retaining only the regions containing lung tissue. All cropped volumes were resampled to a uniform resolution of 192×256×256 voxels to ensure spatial consistency across all cases.

[0075] 3) Slice grouping and feature encoding To effectively simulate inter-slice continuity while maintaining manageable computational cost, every three consecutive slices were grouped into a unit. Each 3-slice input (1×3×256×256) was fed into a ResNet-18 backbone pre-trained on ImageNet and fine-tuned on a research cohort for 2D feature extraction. This process generated a series of 64 feature embeddings, each representing local morphological and textural features of the lung. The pre-trained ResNet-18 network was pre-trained on ImageNet and fine-tuned with medical image data.

[0076] 4) Feature modeling based on Transformer The core of this invention's architecture is a slice-level Transformer encoder designed to model long-range context dependencies. For each of the 64 encoded slice groups, max pooling is used to globally aggregate spatial features into [1×256] vectors. This embedding sequence is then processed by a six-layer Transformer encoder. The encoder utilizes a Multi-Head Self-Attention (MHSA) module to capture non-local inter-slice dependencies, a Relative Position Encoding (RelPE) module to maintain spatial order along the z-axis, and a residual connection feedforward network with remaining connections for feature refinement.

[0077] This attention-based mechanism enables the model to dynamically learn the interactions between different lung regions and understand how distributed inflammation patterns contribute to the development of RMPP. Finally, the Transformer's output is fed into a multilayer perceptron (MLP) classifier to generate a risk probability value for the current CT scan sample belonging to RMPP.

[0078] 1.4. Construction of clinical model, Trans-DLF model and multimodal refractory mycoplasma pneumoniae pneumonia prediction nomogram To ensure data quality, systematically missing variables were excluded. Systematically missing variables refer to variables with a missing rate exceeding 80% in the outpatient subgroup of external test cohort 2 due to differences in data collection settings. The remaining missing values ​​in each cohort were independently imputed using the predicted mean matching method to prevent data leakage. This invention constructs a Trans-DLF model (i.e., the first prediction model proposed in Example 2) using the risk probability values ​​output in section 1.3, with the specific construction method referring to Example 2. Simultaneously, potential indicators are first screened using univariate logistic regression, followed by inverse multivariate selection to obtain independent indicators. The regression analysis model is then trained solely using these independent indicators and their corresponding real disease outcomes to construct a clinical model. Finally, the aforementioned independent indicators are combined with the risk probability values ​​output in section 1.3 to construct a multimodal predictive nomogram for refractory mycoplasma pneumoniae pneumonia (i.e., the second prediction model provided in Example 3), with the specific construction method referring to Example 3.

[0079] 1.5. Statistical Analysis This invention uses a 3D convolutional neural network (3D-CNN) as a baseline to validate the effectiveness of Transformer in modeling long-range dependencies. In the 3D-CNN model, all slice features are stacked along the channel dimension and fed into 3D convolutional layers to jointly learn spatial channel representations. Unlike the Transformer, which explicitly models inter-slice relationships through self-attention, 3D-CNN captures only local volumetric dependencies. This invention compares the performance of the three models constructed in Section 1.4 with the 3D-CNN model to highlight the superiority of the Transformer in representing global spatial dependencies related to disease progression.

[0080] Specifically, the performance of different models was evaluated by calculating the area under the receiver operating characteristic (AUC) curve. The optimal threshold was determined by maximizing the Youden index in the validation cohort. Secondary measures included accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Two thousand repeated determinations were performed using a stratified bootstrapping method, and their 95% confidence intervals (CIs) were reported. Significant differences in AUC between models were assessed using the DeLong test. Model calibration, representing the consistency between predicted probabilities and observed outcomes, was evaluated using calibration curve analysis using the “rms” package in R. Ideal calibration was considered to be closely aligned with the 45° line. Clinical utility was estimated using decision curve analysis (DCA) to quantify benefits at different threshold probabilities.

[0081] To assess the robustness of each model, this invention stratifies patients based on age, disease severity, pre-visit fever duration (PFD), treatment environment, and initial levels of C-reactive protein and neutrophil ratio (NR), and performs subgroup analysis. In the validation cohort, stratification thresholds for laboratory biomarkers are determined using receiver operating characteristic (ROC) curve analysis and the Youden index in the validation cohort. The age threshold is based on the age distribution of the cohort and known MPP epidemiology. To improve model interpretability, this application uses gradient-weighted class activation mapping (Grad-CAM) to visualize key regions in CT scans that have the greatest impact on risk stratification.

[0082] All statistical analyses were performed using R software (version 4.3.1). Continuous variables were compared using the Student t-test or the Mann-Whitney U test, while categorical variables were compared using the chi-square test or Fisher test as appropriate. The Bonferroni method was used to adjust for p-values ​​in multiple comparisons. A two-sided p-value <0.05 was considered statistically significant.

[0083] 2. Results 2.1. Characteristics of the research population This multicenter study included 1224 pediatric patients, comprising 538 patients with recurrent pulmonary embolism (RMPP) and 686 non-RMPP patients. The median age of the entire cohort was 6.83 years (interquartile range [IQR] 5.0–8.6), with a balanced sex distribution (609 males, 49.8%; 615 females, 50.2%). Baseline CT scans were performed at a median of 7 days after symptom onset (IQR 5–9) and 4 days after initiation of standard macrolide therapy (IQR 3–6). Multivariate logistic regression in the training cohort identified six independent indices of RMPP (p<0.05), as shown in Table 1.

[0084] Table 1 Six Independent Indicators of RMPP ; Table 1 shows that the six independent indicators are pre-visit fever duration (PFD), oxygen support requirement (ROS), pre-visit cough duration (PCD), severe / critical pneumonia type, body temperature (T), and hemoglobin (HGB). A multimodal predictive nomogram for mycoplasma pneumoniae pneumonia was then developed using these six independent indicators and trans-DLF risk probability values. Figure 11 As shown.

[0085] pass Figure 11 As can be seen, the nomogram has 10 rows, with rows 2-8 representing the included predictive indicators. The scores of the 7 predictive indicators are added to the total score in row 9 and correspond to the RMPP predictive probability (i.e., the second predictive probability) in row 10. The nomogram shows the percentage risk of RMPP.

[0086] The following is an example of how to use the multimodal refractory mycoplasma pneumoniae pneumonia prediction nomogram: For example, to predict the risk of a patient having RMPP based on the multimodal mycoplasma pneumoniae pneumonia prediction nomogram, the patient's original three-dimensional CT volume data of the chest was first obtained, and the risk probability value was obtained as 0.3 using the method in Example 1. The patient's pre-visit fever duration (PFD, in days) was 6 days, oxygen support requirement (ROS) was 1, pre-visit cough duration (PCD, in days) was 4 days, severe / critical pneumonia type was 0 / severe pneumonia, body temperature (T) was 38.5℃, and hemoglobin (HGB) was 135. The individual scores for each prediction indicator are as follows: risk probability value 30, PFD score 12, ROS score 5, PCD score 12, Severe / Critical pneumonia score 0, PFD score 12, ROS score 5, PCD score 12, and Severe / Critical pneumonia score 12. The patient's Pneumonia score was 0, T score was 10, and HGB score was 10; the total score for all predictive indicators was 79; the predicted probability of RMPP for this patient was less than 0.05, indicating that the patient had a low risk of RMPP.

[0087] 2.2. Model Performance like Figure 12 As shown, the proposed Trans-DLF model exhibits strong discriminative ability against RMPP in all cohorts. In the training cohort, the Trans-DLF model has an AUC of 0.97 (95% CI 0.96–0.98), accuracy of 0.92, sensitivity of 0.96, and specificity of 0.87. In the validation cohort, the model maintains high performance with an AUC of 0.91 (95% CI 0.86–0.96), and in the internal test cohort, the AUC is 0.90 (95% CI 0.85–0.95). Importantly, the Trans-DLF model demonstrates good generalization ability, achieving AUCs of 0.89 (95% CI 0.84–0.94) and 0.89 (95% CI 0.82–0.95) in external test cohorts 1 and 2, respectively. The AUC of the 3D-CNN baseline is moderately lower than that of the inverse DLF model.

[0088] The AUC of the 3D-CNN baseline model was slightly lower than that of the Trans-DLF model, but the difference was not statistically significant (AUC range: 0.82–0.95). In contrast, the clinical model showed significantly poorer discriminative ability; according to the DeLong test, the AUC for all cohorts ranged from 0.65 to 0.76 (all p < 0.001 compared to Trans-DLF), indicating the limitations of relying solely on routine clinical variables. Furthermore, the AUC of Trans-DLF was comparable to that of the multimodal refractory Mycoplasma pneumoniae prediction nomogram, with no statistically significant difference observed (both p > 0.05).

[0089] DCA further confirms the superior clinical utility of Trans-DLF. Figure 13 AE). Through Figure 13 AD showed that Trans-DLF provided greater net gains than 3D-CNN and clinical models over a wide range of threshold probabilities, while maintaining net gains comparable to multimodal refractory Mycoplasma pneumoniae prediction nomograms. Furthermore, calibration curve analysis ( Figure 13 F) shows good agreement between the predicted probability and the observed results.

[0090] 2.3. Subgroup Analysis To further evaluate the robustness of Trans-DLF, this invention performed subgroup analyses on the pooled external test cohorts, analyzing key clinical and demographic variables ( Figure 14 ).pass Figure 14 The AUC for Trans-DLF was 0.87 (95% CI 0.77–0.98) in patients under 5 years old, 0.90 (95% CI 0.86–0.94) in patients aged 5–12 years, and 0.81 (95% CI 0.66–0.96) in patients over 12 years. Furthermore, the model showed similar stability after stratification by clinical indicators, including PFD, CRP levels, and neutrophil ratio (NR) levels; the cutoff values ​​for PFD, CRP, and NR were 6.5 days, 18.44 mg / L, and 67.34%, respectively. Notably, the model maintained robust predictive ability in the outpatient subgroup (n=74), with an AUC of 0.87 (95% CI 0.78–0.96).

[0091] 2.4. Model Interpretability By using Grad-CAM to visualize pixel-level importance in color-coded heatmaps, the decision-making process of the Trans-DLF model is revealed, and radiological features crucial for risk assessment are highlighted. Figure 15 ).pass Figure 15 The results show that in RMPP cases, the model focuses on extensive consolidation, a recognized radiographic hallmark of refractory disease. Conversely, in non-RMPP cases characterized by smaller consolidation and ground-glass opacities (GGO), the model selectively focuses on areas of consolidation while assigning significantly lower weights to GGO areas.

[0092] The exemplary embodiments of this disclosure described in detail above are merely illustrative and not restrictive. Those skilled in the art will understand that various modifications and combinations can be made to these embodiments or their features without departing from the principles and spirit of this disclosure, and such modifications should fall within the scope of this disclosure.

Claims

1. A CT-based method for predicting refractory mycoplasma pneumoniae pneumonia, characterized in that, include: The patient's original three-dimensional CT volume data of the chest were standardized to obtain lung parenchyma volume data; Based on the lung parenchyma volume data, slice groups are divided and feature extraction and spatial aggregation are performed to generate slice feature sequences; The sliced ​​feature sequence is input into the Transformer encoder to perform global dependency modeling and obtain global context features; The global context features are input into the classifier to generate risk probability values ​​for refractory mycoplasma pneumonia.

2. The CT-based method for predicting refractory mycoplasma pneumoniae pneumonia according to claim 1, characterized in that, The process of dividing the lung parenchyma volume data into slice groups, extracting features, and spatially aggregating them to generate slice feature sequences includes: Along the axis of the lung parenchyma volume data, every N consecutive slices are divided into a slice group; A two-dimensional convolutional neural network is used to extract features from each slice group to obtain the lung feature map corresponding to each slice group; Spatial aggregation is performed on each of the lung feature maps to generate a feature vector of the corresponding local morphology and texture features of the lung; The feature vectors of all slice groups are integrated to obtain the slice feature sequence.

3. The CT-based method for predicting refractory mycoplasma pneumoniae pneumonia according to claim 1, characterized in that, The Transformer encoder includes an M-layer encoding unit; Each coding unit includes a multi-head self-attention module for capturing non-local slice dependencies in the slice feature sequence, a position coding module coupled to the multi-head attention module for introducing the original spatial position information of the slices in the slice feature sequence along the axis to the multi-head self-attention module, and a residual connection feedforward network module coupled to the position coding module for refining the features in the slice feature sequence.

4. The CT-based method for predicting refractory mycoplasma pneumoniae pneumonia according to claim 1, characterized in that, The multi-head self-attention module captures non-local slice dependencies in the slice feature sequence by calculating the relationship between any Mth feature vector and the (M+p)th feature vector in the slice feature sequence.

5. A method for constructing a predictive model for refractory mycoplasma pneumoniae pneumonia, characterized in that, include: The method according to any one of claims 1-4 processes the original three-dimensional CT volume data of the chest of the training set samples to obtain the risk probability value of each training set sample; Obtain the real disease outcomes for each training set sample; Among them, the actual disease outcome includes whether the patient has refractory mycoplasma pneumoniae pneumonia; The regression analysis model is trained using the aforementioned risk probability values ​​and actual disease outcomes to obtain a predictive model for refractory mycoplasma pneumonia, namely the first predictive model.

6. A method for constructing a predictive model for refractory mycoplasma pneumoniae pneumonia, characterized in that, include: The method according to any one of claims 1-4 processes the original three-dimensional CT volume data of the chest of the training set samples to obtain the risk probability value of each training set sample; Obtain the real disease outcomes and clinical data for each training set sample; Among them, the actual disease outcome includes whether the patient has refractory mycoplasma pneumoniae pneumonia; clinical data includes the duration of fever before consultation, oxygen support requirement, duration of cough before consultation, type of severe / critical pneumonia, body temperature and hemoglobin. The regression analysis model was trained using the aforementioned risk probability values, actual disease outcomes, and clinical data to obtain a predictive model for refractory mycoplasma pneumoniae pneumonia, namely the second predictive model.

7. A method for predicting refractory mycoplasma pneumoniae pneumonia, characterized in that, include: Obtain the raw three-dimensional CT volume data of the chest of the patient to be tested, and process the raw three-dimensional CT volume data of the chest using the method described in any one of claims 1-4 to obtain a risk probability value; The risk probability value is input into the first prediction model of claim 5 to obtain the first prediction probability, and the auxiliary prediction result of the probability of the patient to be tested developing refractory mycoplasma pneumonia is output based on the first prediction probability.

8. A method for predicting refractory mycoplasma pneumoniae pneumonia, characterized in that, include: Obtain raw three-dimensional CT volume data and clinical data of the patient to be tested, and process the raw three-dimensional CT volume data of the chest using the method described in any one of claims 1-4 to obtain a risk probability value; The risk probability value and clinical data are input into the second prediction model described in claim 6 to obtain the second prediction probability. Based on the second prediction probability, an auxiliary prediction result of the probability of the patient to be tested developing refractory mycoplasma pneumonia is output.

9. A computer device, characterized in that, The device includes: a memory and a processor; the memory is used to store a computer program; the processor executes the computer program to implement the steps of the method according to any one of claims 1-8.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-8.

Images