Method and device for constructing a pneumothorax drainage risk model based on multi-modal physiological and imaging parameters

Abstract

This invention discloses a method and apparatus for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters, belonging to the field of risk model construction technology. The method includes the following steps: multimodal data acquisition, data preprocessing, logistic regression model training, model performance verification and optimization, and model engineering deployment. This invention first acquires a dataset of historical lung biopsy patients containing multimodal parameters and pneumothorax drainage outcome labels. After cleaning, standardization, and validation preprocessing, training data is obtained. Then, a risk prediction model with parameter weights is trained using a logistic regression algorithm. After validation and optimization on an independent dataset, the model is deployed to a prediction system to achieve real-time risk calculation for new patients. This improves the accuracy of pneumothorax drainage risk model construction and solves the problem of low accuracy in existing technologies due to the lack of multi-dimensional parameter fusion and multi-factor analysis methods.

Description

Technical Field

[0001] This invention relates to the field of risk model construction technology, and in particular to a method and apparatus for constructing a risk model for pneumothorax drainage based on multimodal physiological and imaging parameters. Background Technology

[0002] A basic prediction framework was built using traditional logistic regression. Through single-center retrospective medical record data collection, Z-score standardized preprocessing, and maximum likelihood estimation to fit weight coefficients, a risk nomogram model was ultimately generated to achieve probability prediction, aiming to assist clinicians in identifying high-risk patients and developing preventative intervention plans. However, several unavoidable local defects exist in model construction, data fitting, generalization application, and the accuracy of risk prediction. First, there are defects in variable selection and algorithm design. Using only single logistic regression for linear modeling fails to capture the nonlinear relationships and interaction effects between predictive variables. Furthermore, the lack of regularization and feature selection algorithms makes it prone to including redundant variables and causing overfitting. Simultaneously, it ignores implicit high-risk factors such as underlying comorbidities, intraoperative details, and quantitative imaging parameters, resulting in incomplete variable coverage, insufficient model discrimination, and low predictive efficacy. Second, there are defects in data and training. The modeling data comes from multiple sources. In single-center, small-sample retrospective cohorts, there are selection and information biases, poor data labeling consistency, and only basic standardization is used without stratified cross-validation or external cohort validation. The model weight coefficients have poor adaptability, and performance degradation is very likely to occur when promoted across institutions. Thirdly, there are defects in model implementation and practicality. Traditional models are mostly static formulas or nonodal graphs, and do not achieve dynamic calibration of risk thresholds. They are not sensitive enough in predicting special subtypes such as occult pneumothorax and delayed pneumothorax, which can easily lead to false negatives and missed diagnoses, resulting in delayed treatment. They are difficult to meet the needs of rapid and accurate risk stratification in emergency and critical care scenarios. Summary of the Invention

[0003] To address the technical problem of low accuracy in constructing pneumothorax drainage risk models due to the lack of multi-dimensional parameter fusion and multi-factor analysis methods in existing technologies, this invention provides a method and apparatus for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters. The technical solution is as follows: On the one hand, a method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters is provided. This method includes: Step 1, obtaining a historical patient dataset of CT-guided percutaneous lung biopsy. This dataset contains multimodal parameters and outcome labels indicating whether the patient developed drainage-required pneumothorax post-surgery. Multimodal parameters include, but are not limited to, physiological characteristic parameters, lung CT imaging anatomical parameters, and variables related to the biopsy procedure. Step 2, performing full-process preprocessing on the historical patient dataset obtained in Step 1. Preprocessing operations include data cleaning, variable standardization, and input validation. Data cleaning involves removing missing values, outliers, and invalid samples from the dataset. Variable standardization involves normalizing continuous parameters and digitizing categorical parameters. Input validation... To verify whether the parameter values ​​are within a clinically reasonable range, standardized training data is generated. Step three involves inputting the standardized training data obtained in step two into a logistic regression algorithm for model training. This trains the model to learn the independent contribution of each multimodal parameter to the risk of pneumothorax drainage and obtains a risk prediction model containing parameter weights. Step four involves constructing an external validation dataset independent of the standardized training data and performing multidimensional performance evaluation on the risk prediction model trained in step three. If the model performance does not reach the preset evaluation threshold, the process returns to step three to readjust the model training parameters and retrain. If the model performance reaches the preset evaluation threshold, model parameter optimization is complete. Step five involves deploying the risk prediction model validated in step four into the prediction system to achieve real-time calculation of the risk of pneumothorax drainage in new patients.

[0004] On the other hand, a system for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters is provided. This system includes: a multimodal data acquisition module, a data preprocessing module, a logistic regression model training module, a model performance verification and optimization module, and a model engineering deployment module. The multimodal data acquisition module acquires historical patient datasets from CT-guided percutaneous lung biopsy. These datasets contain multimodal parameters and outcome labels indicating whether the patient developed drainage-required pneumothorax post-surgery. Multimodal parameters include, but are not limited to, physiological characteristic parameters, lung CT imaging anatomical parameters, and variables related to the biopsy procedure. The data preprocessing module performs full-process preprocessing on the historical patient datasets acquired in step one. Preprocessing operations include data cleaning, variable standardization, and input validation. Data cleaning removes missing values, outliers, and invalid samples from the dataset. Variable standardization normalizes continuous parameters. The system processes and encodes the categorical parameters digitally, and inputs them for validation to verify whether the parameter values ​​are within a clinically reasonable range, thereby generating standardized training data. The logistic regression model training module inputs the standardized training data obtained in step two into the logistic regression algorithm for model training, learning the independent contribution of each multimodal parameter to the risk of pneumothorax drainage, and obtaining a risk prediction model including parameter weights. The model performance validation and optimization module constructs an external validation dataset independent of the standardized training data, performs multi-dimensional performance evaluation on the risk prediction model trained in step three. If the model performance does not reach the preset evaluation threshold, it returns to step three to readjust the model training parameters and trains again. If the model performance reaches the preset evaluation threshold, the model parameter optimization is complete. The model engineering deployment module deploys the risk prediction model validated in step four into the prediction system to achieve real-time calculation of the risk of pneumothorax drainage in new patients.

[0005] Beneficial effects The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. The method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters provided by this invention integrates multi-dimensional multimodal parameters related to patient physiological characteristics, lung CT imaging anatomy, and puncture operation to construct a pneumothorax drainage risk prediction model. It also clarifies core key predictive variables such as gender, pleural depth, and emphysema. Combined with logistic regression algorithm, it mines the independent contribution of each parameter to the risk and quantifies the weight coefficients, breaking through the limitations of traditional single-factor analysis. This allows the model to comprehensively assess the actual risk of pneumothorax drainage under the combined effect of multiple dimensions of parameters. It ensures the scientificity and accuracy of risk prediction from the perspectives of data foundation and algorithm design, and effectively solves the problem of risk assessment relying on subjective experience and lacking quantitative basis in the prior art.

[0006] 2. This invention establishes a comprehensive dataset preprocessing standard, which eliminates invalid samples through data cleaning, standardizes variables to unify parameters, and validates inputs for clinical rationality. Simultaneously, it constructs an external validation dataset independent of the training data to conduct multi-dimensional performance evaluation. If the threshold is not met, the parameters are retrained and optimized. Furthermore, it employs a maximum likelihood estimation algorithm combined with gradient iterative optimization to accurately solve and determine the model parameters, forming a closed-loop system from data processing to model training, validation, and optimization. This significantly improves the model's fit and generalization ability, avoids prediction bias caused by non-standard data and insufficient model validation, and enables the trained model to possess stable and accurate individualized risk prediction capabilities.

[0007] 3. By engineering and deploying the validated and optimized risk prediction model onto a dedicated prediction system, and with the design of five functional modules—multimodal data acquisition, data preprocessing, model training, performance validation and optimization, and engineering deployment—the model has been successfully transformed from theoretical construction to practical clinical application. It can receive relevant parameters from new patients in real time in clinical settings and quickly calculate the probability of pneumothorax drainage risk. Furthermore, the standardized process of model construction provides a clear implementation path for subsequent model iterations and parameter adjustments, allowing the system to continuously optimize model performance based on new clinical data. This not only meets the needs of real-time and efficient clinical risk assessment but also ensures the long-term applicability and advancement of the prediction system. It provides objective and quantitative decision-making basis for clinicians to formulate preoperative plans and postoperative monitoring strategies, helping to optimize the allocation of medical resources and reduce the medical costs associated with postoperative complications. Attached Figure Description

[0008] 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.

[0009] Figure 1 A flowchart illustrating the method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters, as provided in this application embodiment; Figure 2 A flowchart illustrating the construction of a risk prediction model for a pneumothorax drainage risk model based on multimodal physiological and imaging parameters, as provided in this application embodiment. Figure 3 A schematic diagram of the structure of the device for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters provided in the embodiments of this application. Detailed Implementation

[0010] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0011] like Figure 1 The diagram shown is a flowchart of a method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters, provided in an embodiment of this application. The method includes the following steps: Step 1: Obtain a historical patient dataset of CT-guided percutaneous lung biopsy. This dataset includes multimodal parameters and outcome labels indicating whether the patient developed drainage-required pneumothorax post-procedure. Multimodal parameters include, but are not limited to, physiological characteristic parameters, lung CT imaging anatomical parameters, and variables related to the biopsy procedure. Step 2: Perform full-process preprocessing on the historical patient dataset obtained in Step 1. Preprocessing operations include data cleaning, variable standardization, and input validation. Data cleaning involves removing missing values, outliers, and invalid samples. Variable standardization involves normalizing continuous parameters and digitizing categorical parameters. Input validation verifies whether the parameter values ​​are within clinically reasonable ranges. Step 1: Standardize the training data. Step 2: Input the standardized training data obtained in Step 2 into the logistic regression algorithm for model training to learn the independent contribution of each multimodal parameter to the risk of pneumothorax drainage and obtain a risk prediction model containing parameter weights. Step 3: Construct an external validation dataset independent of the standardized training data and perform multi-dimensional performance evaluation on the risk prediction model trained in Step 3. If the model performance does not reach the preset evaluation threshold, return to Step 3 to readjust the model training parameters and train again. If the model performance reaches the preset evaluation threshold, the model parameter optimization is completed. Step 4: Deploy the risk prediction model that has passed the validation in Step 4 into the prediction system to realize real-time calculation of the risk of pneumothorax drainage in new patients.

[0012] In this embodiment, the data acquisition step is performed first to obtain a historical patient dataset of CT-guided percutaneous lung biopsy. This dataset covers multimodal parameters including physiological characteristic parameters, lung CT imaging anatomical parameters, and puncture operation-related variables. It also includes outcome labels indicating whether patients developed drainage-required pneumothorax post-procedure. This comprehensive, multi-source data acquisition method overcomes the limitations of traditional single-dimensional data collection, laying a complete data foundation for building a comprehensive and accurate risk prediction model and ensuring that the model can comprehensively consider various key factors affecting the risk of pneumothorax drainage. Next, the acquired historical patient dataset undergoes a full-process preprocessing, including data cleaning to remove missing values, outliers, and invalid samples; normalization of continuous parameters; variable standardization through categorical parameter digitization; and parameter validation. Input validation within a clinically reasonable range was performed to generate standardized training data. The standardized preprocessing process effectively resolved issues such as messy raw data, inconsistent formats, and numerical anomalies, improving the quality and consistency of the training data. This avoided model training bias caused by data defects at the data source, allowing subsequent algorithm training to be based on standardized and accurate data sources. The standardized training data was then input into a logistic regression algorithm for model training. The algorithm fitted and learned the independent contribution of each multimodal parameter to the risk of pneumothorax drainage, solving for and obtaining a risk prediction model including parameter weights. The application of the logistic regression algorithm enabled quantitative analysis of the risk contribution of each parameter. This approach breaks away from the traditional reliance on empirical judgments without quantitative evidence, enabling the model to accurately identify the impact of various parameters on the risk of pneumothorax drainage, thus forming an initial risk prediction model with quantitative analysis capabilities. Then, an external validation dataset, independent of the standardized training data, is constructed to perform multi-dimensional performance evaluations on the trained initial risk prediction model. If the model performance does not reach a preset evaluation threshold, the training parameters are readjusted and retrained. Once the threshold is reached, the model parameters are optimized. The construction of the independent validation dataset, along with the multi-dimensional performance evaluation and iterative optimization mechanism, achieves rigorous verification of the model's generalization ability and prediction accuracy, effectively improving the model's fit and accuracy. Stability was ensured, avoiding overfitting or underfitting issues and guaranteeing that the final model possessed clinically applicable predictive performance. Finally, the validated risk prediction model was deployed into the prediction system to achieve real-time calculation of pneumothorax drainage risk in new patients. The engineering deployment of the model completed the transformation from a theoretical model to a clinically practical tool, enabling rapid reception of relevant parameters from new patients and immediate output of risk prediction results in clinical settings. This solved the problem of existing technologies failing to translate research findings into practical applications and real-time evaluation, providing clinicians with objective and efficient quantitative decision-making basis for developing preoperative plans and postoperative monitoring strategies, and achieving accurate and real-time prediction of pneumothorax drainage risk.

[0013] It should be understood that physiological characteristic parameters, lung CT imaging anatomical parameters, and puncture operation-related variables together constitute a comprehensive set of influencing parameters for constructing a risk model of pneumothorax drainage. Physiological characteristic parameters include patient gender and whether emphysema is present. Lung CT imaging anatomical parameters include whether the target lesion is adjacent to the pleura, the pleural depth of the target lesion, whether the target lesion involves the interlobar pleura, and the location of the target lesion in the lung lobe. Puncture operation-related variables include the number of times the puncture needle passes through the pleura and the number of biopsy samples taken.

[0014] It should be noted that, as Figure 2 The diagram shows a flowchart of the risk prediction model construction method for pneumothorax drainage risk model based on multimodal physiological and imaging parameters provided in this application embodiment. The specific process is as follows: starting with the construction of the risk prediction function structure, a linear prediction sub-function and a probability transformation function are first constructed and combined into a complete mathematical structure; then, the function fitting training and parameter solving stage is entered, inputting the observed values ​​of key variables of historical patients, and using the maximum likelihood estimation algorithm to iteratively optimize the undetermined parameters. The maximum likelihood iterative optimization is refined into constructing the overall likelihood function, calculating the gradient, iteratively updating the parameters and verifying convergence, and the gradient descent parameter update is further clarified into setting the initial value and learning rate, updating the parameters according to the gradient and dynamically adjusting the learning rate; finally, the final parameter weights and the baseline constant term obtained after convergence are embedded into the linear prediction sub-function, combined with the probability transformation function, to generate the final risk prediction model with fully determined parameters, which fully covers the entire process from function construction and parameter solving to model generation.

[0015] It should be noted that the specific process for obtaining a risk prediction model that includes parameter weights is as follows: Standardized training data is input into the logistic regression algorithm to train the model, so as to learn the independent contribution of each multimodal parameter to the risk of pneumothorax drainage and obtain a risk prediction model including parameter weights. From the standardized training data, a set of identified key predictive variables were selected for model construction. The key predictive variables included the patient's gender, the relationship between the target lesion and the adjacent pleura, the pleural depth of the target lesion, the number of times the puncture needle passed through the pleura, the number of biopsy samples, whether emphysema was present, whether the interlobar pleura was involved, and the location of the lung lobe where the target lesion was located.

[0016] In this embodiment, during the process of obtaining a risk prediction model containing parameter weights, the following key predictive variables, clinically and research-identified, are precisely selected from the standardized training data: patient gender, the relationship between the target lesion and adjacent pleura, the pleural depth of the target lesion, the number of times the puncture needle passes through the pleura, the number of biopsy samplings, whether emphysema is present, whether interlobar pleura is involved, and the location of the lung lobe where the target lesion is located. These variables serve as the core independent variables for model construction. This operation precisely identifies the core parameters that directly affect the risk of pneumothorax drainage, eliminates interference from irrelevant or secondary variables, significantly improves the targeting and efficiency of model training, and makes the subsequent risk prediction logic of the model more closely aligned with the actual pathogenic factors in clinical practice. Subsequently, the standardized training data is compared with the selected key predictive variables. After matching, the entire dataset is fed into a logistic regression algorithm for model training. Through the fitting operation of the logistic regression algorithm, the independent contribution of each key predictor variable to the risk of pneumothorax drainage is learned through deep learning. At the same time, the parameter weights corresponding to each key predictor variable are solved and determined. Finally, a risk prediction model containing parameter weights is generated. The application of the logistic regression algorithm realizes the quantitative analysis of the risk contribution of each key predictor variable, breaking through the limitation of traditional single-factor analysis that cannot measure the independent effect of multiple parameters. This allows the model to accurately quantify the influence of different factors on the risk of pneumothorax drainage. The resulting risk prediction model with parameter weights also provides core algorithmic and parameter support for the subsequent realization of individualized and accurate probability calculation of pneumothorax drainage risk, ensuring the objectivity and quantification of risk prediction results from the model construction level.

[0017] It should be further explained that obtaining the risk prediction model that includes parameter weights also includes: Based on the logistic regression framework, a risk prediction function structure is constructed with key predictor variables as independent variables and whether or not pneumothorax requiring drainage occurs as the dependent variable. The risk prediction function is fitted and trained using standardized training data to determine the specific weight coefficients corresponding to each key predictor variable in the risk prediction function. Based on the weight coefficients of each key predictor variable, the final risk prediction model is generated.

[0018] In this embodiment, a basic risk prediction function structure is first constructed based on the logistic regression framework, with key predictor variables as independent variables and the occurrence of pneumothorax requiring drainage as a binary dependent variable. The L1 regularization algorithm is introduced to perform feature selection and sparsity constraints on the predictor variables, retaining high-contribution predictor variables while eliminating redundant variables to improve model generalization ability and interpretability. Subsequently, the Z-score standardization method is used to standardize the original training set data to eliminate the interference of differences in variable dimensions and numerical magnitudes on model fitting. Then, the logistic regression objective loss function incorporating the L1 regularization term is iteratively optimized using maximum likelihood estimation combined with coordinate descent. During the iteration process, the regularization weight coefficients and intercept terms corresponding to each key predictor variable are dynamically updated and determined. Simultaneously, cross-validation is used to determine the optimal regularization parameters to balance model fitting accuracy and overfitting risk. Finally, based on the optimized and converged weight coefficients and intercept terms of each key predictor variable, and the selected core feature set, a fused L1 regularization function is generated. A binary risk prediction model for pneumothorax requiring drainage, based on regularization constraints and standardized preprocessing, enables quantitative calculation and risk classification of the probability of occurrence.

[0019] Furthermore, the logistic regression framework constructs a risk prediction function structure with key predictor variables as independent variables and the occurrence of pneumothorax requiring drainage as the dependent variable, including: Construct a linear predictor function whose output is the sum of the products of the values ​​of each key predictor variable and their respective undetermined weight coefficients, and add it to an undetermined baseline constant term; Construct a probability transformation function that takes the output of a linear prediction subfunction as input and maps it to a risk probability value between 0 and 1 through a nonlinear transformation. This probability value represents the predicted likelihood of pneumothorax requiring drainage. By combining the linear prediction sub-function with the probability transformation function, a complete mathematical structure for the risk prediction function is formed.

[0020] In this embodiment, when constructing the risk prediction function structure based on the logistic regression framework, a linear prediction sub-function is first constructed. The values ​​of each key predictor variable are multiplied by their respective undetermined weight coefficients, summed, and then an undetermined baseline constant term is added to obtain the output value of the sub-function. This construction method can linearly integrate the quantitative information of all key predictor variables, accurately reflecting the linear influence of each variable on the risk of pneumothorax drainage. Simultaneously, the setting of the undetermined weight coefficients and the baseline constant term reserves an optimizable parameter space for subsequent fitting training to quantify the independent risk contribution of each variable and calibrate the model's basic prediction baseline, allowing the sub-function to adapt to the actual clinical risk impact patterns. Next, a probability transformation function is constructed, using the output value of the linear prediction sub-function as the sole input. Through nonlinear transformation, it is mapped to a value between 0 and 1, which is the predicted risk probability value of pneumothorax requiring drainage. This nonlinear transformation operation overcomes the limitation that the output value of the linear prediction sub-function has no clear probabilistic meaning, transforming the abstract linear calculation result into a probability value that conforms to clinical risk assessment cognition, allowing the model output results to be intuitively understood and directly used by clinicians. Furthermore, the value between 0 and 1... The range of values ​​strictly matches the mathematical definition of the probability of an event, ensuring the scientific validity and rationality of the prediction results. Finally, the linear prediction sub-function and the probability transformation function are organically combined to form a complete mathematical structure for the risk prediction function. This combination realizes the integrated operational logic of "linearly quantifying the influence of each variable and nonlinearly converting it into a probability value." This allows the complete function structure to retain the good interpretability of the logistic regression algorithm, clearly trace the influence process of each key variable on the final risk probability, and also has the practical value of directly outputting the quantitative risk probability. This lays a rigorous and clinically relevant mathematical foundation for subsequent fitting and training using standardized training data and determining specific parameters, ensuring that the subsequently generated risk prediction model can accurately and scientifically complete the quantitative prediction of pneumothorax drainage risk.

[0021] It needs to be explained that the risk prediction function is fitted and trained based on standardized training data to determine the specific weight coefficients corresponding to each key predictor variable in the risk prediction function, including: The actual observed values ​​of the key predictor variables for each historical patient in the standardized training data are input into the constructed risk prediction model; The maximum likelihood estimation algorithm is adopted to maximize the overall predictive likelihood of the risk prediction function for the actual occurrence of pneumothorax requiring drainage in all historical patients. The algorithm iteratively optimizes and solves all undetermined parameters in the function, including each weight coefficient and the baseline constant term. Obtain and record the final weight coefficient value and the final baseline constant value for each key predictor variable determined after the convergence of the iterative optimization solution process.

[0022] In this embodiment, during the process of fitting and training the risk prediction function based on standardized training data and determining the specific weight coefficients corresponding to each key predictor variable, the actual observed values ​​of the key predictor variables for each historical patient in the standardized training data are first accurately input into the constructed risk prediction model. The standardized observed values ​​eliminate the dimensional differences and abnormal biases of the original data, ensuring that the basic data input into the model is consistent, accurate, and reasonable, thus laying a solid data foundation for the accuracy of subsequent parameter solutions. Simultaneously, the one-to-one input method of patient observation values ​​allows the model to fit based on real clinical data, ensuring that the training results closely match the actual clinical risk impact patterns. Subsequently, the maximum likelihood estimation algorithm is used, with the core objective of maximizing the overall predictive likelihood of the risk prediction function for the actual occurrence of pneumothorax requiring drainage in all historical patients. All undetermined parameters, such as the undetermined weight coefficients and undetermined baseline constants of each key predictor variable in the function, are iteratively optimized and solved. The application of the maximum likelihood estimation algorithm... This approach allows the model to closely match the actual distribution characteristics of historical clinical data during the fitting process, accurately uncovering the intrinsic correlation between each parameter and the risk of pneumothorax drainage. The iterative optimization solution method gradually corrects parameter values, effectively avoiding parameter bias caused by single calculations, and making the solved parameters more realistic. Finally, after the iterative optimization solution process reaches convergence, the final weight coefficient value corresponding to each key predictor variable and the final baseline constant value are accurately obtained and completely recorded. The determination of convergence ensures the stability and effectiveness of parameter solution, avoiding parameter distortion due to insufficient iteration. The accurate recording and solidification of the final parameters not only clarifies the quantitative contribution of each key predictor variable to the risk of pneumothorax drainage, and intuitively reflects the risk impact weight of different variables, but also transforms the risk prediction function from a theoretical structure with undetermined parameters into a practical function with defined parameters. This provides core quantitative parameter support for the subsequent generation of the final risk prediction model and the conduct of accurate risk probability calculations.

[0023] It should be further explained that the specific process for generating the final risk prediction model based on the obtained weight coefficients of each key predictor variable is as follows: The final parameter weights and final baseline constants of each key predictor obtained by iterative optimization are embedded into the undetermined parameter positions of the constructed linear predictor subfunction to form a linear predictor subfunction with completely determined parameter values. By combining the linear prediction sub-function with the constructed probability transformation function, a complete risk prediction function with all parameters determined is obtained. This complete risk prediction function with all parameters determined is the final risk prediction model.

[0024] In this embodiment, when generating the final risk prediction model based on the weight coefficients of each key predictor variable, the final parameter weights and final baseline constants corresponding to each key predictor variable obtained after iterative optimization are first precisely embedded into the undetermined parameter positions in the previously constructed linear predictor function. This completes the parameter assignment and solidification of the linear predictor function, forming a linear predictor function with completely determined parameter values. This operation transforms the original linear predictor function, which only had a theoretical structure, into a practical function that can be directly calculated numerically. The precise embedding of the quantified weights and baseline constants of each key variable ensures that the function can accurately calculate the linear predicted value reflecting the risk of pneumothorax drainage based on the actual parameters of the input patient. Moreover, the quantified parameter weights can accurately reflect the independent contribution of each variable to the risk, laying a precise numerical foundation for the subsequent conversion of risk probability. Then, the linear predictor function with determined parameters is organically combined with the previously constructed probability conversion function to form a complete risk prediction function with fully determined parameters. This complete risk prediction function with fully determined parameters is the final pneumothorax drainage risk prediction model that can be directly applied. This combination method continues the principle of "linear quantified variable influence - The scientific computational logic of "converting nonlinearity into probability values" retains the good interpretability of logistic regression models, clearly tracing the complete calculation process of each key variable from numerical input to risk probability output. It also enables the final model to directly receive individualized patient parameters, quickly calculate and output the probability of pneumothorax drainage risk between 0 and 1. The complete determination of all parameters in the model also ensures the uniqueness, stability and accuracy of the risk prediction results, avoiding prediction bias caused by parameter uncertainty. It provides core model support for realizing individualized and quantitative pneumothorax drainage risk prediction in clinical practice, allowing doctors to formulate targeted preoperative plans and postoperative management strategies based on the accurate probability values ​​output by the model.

[0025] It should be understood that the maximum likelihood estimation algorithm is used to maximize the overall predictive likelihood of the risk prediction function for the actual occurrence of pneumothorax requiring drainage in all historical patients. This involves iteratively optimizing all undetermined parameters in the function, including weighting coefficients and baseline constants, including: Based on the predicted probability of pneumothorax requiring drainage calculated for each historical patient in the standardized training data using the risk prediction function, and combined with the outcome label of whether or not pneumothorax requiring drainage actually occurred, an overall likelihood function is constructed to evaluate the accuracy of the model's predictions. Calculate the gradient of the overall likelihood function with respect to each undetermined weight coefficient and baseline constant term in the risk prediction function; Based on the calculated gradient direction, the values ​​of each undetermined weight coefficient and the baseline constant term are iteratively updated using an optimization algorithm. After each iteration update, the value of the overall likelihood function is recalculated, and it is determined whether its growth has reached the preset convergence criterion. When the convergence criterion is reached, the iteration stops, and the values ​​of the weight coefficients and the baseline constants obtained in the current iteration are used as the final weight coefficient values ​​and the final baseline constant values.

[0026] In this embodiment, when iteratively optimizing the undetermined parameters of the risk prediction function using the maximum likelihood estimation algorithm, the system first constructs an overall likelihood function to evaluate the model's prediction accuracy. This is based on the predicted probability of pneumothorax requiring drainage calculated for each historical patient in the standardized training data using the risk prediction function, combined with the actual outcome label of whether the patient developed pneumothorax requiring drainage post-surgery. The accurate input of the standardized training data ensures the foundation for calculating the prediction probability, while the combination of actual outcome labels allows the overall likelihood function to truly reflect the degree of fit between the model's prediction results and clinical reality. This establishes a clinically relevant evaluation basis for subsequent parameter optimization, ensuring that the goal of parameter optimization always revolves around... The process begins by improving the model's predictive accuracy for clinical situations. Next, the gradient of the overall likelihood function with respect to each undetermined weight coefficient and baseline constant term in the risk prediction function is calculated. Precise gradient calculation clarifies the direction and extent of each undetermined parameter's influence on the overall likelihood function value, providing a scientific and quantitative basis for subsequent parameter iterations. This avoids blind parameter updates and ensures that each parameter adjustment is directed towards improving overall likelihood. Subsequently, based on the calculated gradient direction, an optimization algorithm iteratively updates the values ​​of each undetermined weight coefficient and baseline constant term. The precise guidance of the gradient direction gives the parameter updates a clear objective, allowing for gradual correction. The parameter values ​​are adjusted to improve the overall likelihood function, allowing the risk prediction function to continuously approximate the actual clinical situation of historical patients, thus gradually optimizing the model's ability to predict the risk of pneumothorax drainage. After each parameter iteration update, the overall likelihood function is recalculated, and its growth is rigorously judged to have reached the preset convergence criterion. This recalculation and criterion judgment after each iteration allows for real-time monitoring of the model parameter optimization effect, timely understanding of the overall likelihood improvement status, avoiding waste of computational resources due to excessive iteration, and preventing insufficient parameter optimization due to insufficient iteration. When the growth of the overall likelihood function reaches the preset convergence criterion, iteration is immediately stopped, and the current iteration is updated. The obtained weighting coefficients and baseline constants are determined as the final weighting coefficients and baseline constants. The convergence criteria ensure the stability and effectiveness of the parameter solution, allowing the final parameters to achieve the optimal overall prediction likelihood of the risk prediction function for the actual occurrence of pneumothorax requiring drainage in all historical patients. At this point, the parameters not only match the distribution characteristics of actual clinical data but also enable the risk prediction function to have the best basic prediction performance. This provides core quantitative parameter guarantees for the subsequent generation of an accurate and reliable final risk prediction model, ensuring that the model can accurately explore the intrinsic relationship between key variables and the risk of pneumothorax drainage based on these parameters, and achieve scientific and quantitative prediction of risk.

[0027] It is important to understand that, based on the calculated gradient direction, an optimization algorithm iteratively updates the values ​​of each undetermined weight coefficient and the baseline constant term, including: An initial value is set for each undetermined weight coefficient and baseline constant term, and an initial learning rate parameter is set for the optimization process. Based on the calculated gradient direction, each gradient value is multiplied by the learning rate parameter and then subtracted from the current value of the corresponding undetermined weight coefficient or baseline constant term, thus completing one iteration update. After completing one iteration of the step, the learning rate parameter is adjusted according to the preset strategy; Using the updated values ​​of the weight coefficients and baseline constants, return to calculate the gradient direction and proceed to the next iteration.

[0028] In this embodiment, during the process of iteratively updating the values ​​of each undetermined weight coefficient and baseline constant term using an optimization algorithm based on the calculated gradient direction, a reasonable initial value is first set for each undetermined weight coefficient and baseline constant term. Simultaneously, an appropriate initial learning rate parameter is configured for the entire optimization process. The scientific setting of the initial values ​​lays a stable starting point for parameter iteration, avoiding slow convergence due to excessive initial value deviation. A reasonable initial learning rate configuration controls the step size rhythm of parameter updates, providing a foundation for efficient subsequent iterative optimization. Next, strictly following the previously calculated gradient direction, each gradient value is multiplied by the learning rate parameter, and the product result is subtracted from the current value of the corresponding undetermined weight coefficient or baseline constant term, thus completing a complete parameter iteration update. This calculation method ensures that parameter updates strictly follow the optimization direction guided by the gradient, ensuring that each update progresses towards improving the overall likelihood. Simultaneously, the scaling effect of the learning rate on the gradient value effectively avoids the problems of oscillation and non-convergence caused by excessively large parameter update step sizes or low convergence efficiency caused by excessively small step sizes, ensuring single-step optimization. The effectiveness and rationality of the iteration update are demonstrated. After completing one parameter iteration update, the learning rate parameter is dynamically adjusted according to a preset strategy. The dynamic adaptation of the learning rate can optimize the update step size based on the actual situation of parameter iteration. For example, the step size can be appropriately increased in the early stage of iteration to improve the convergence speed, and the step size can be decreased in the later stage of iteration to make the parameters approach the optimal value. This effectively solves the problem that it is difficult to balance convergence speed and convergence accuracy with a fixed learning rate, and further improves the overall efficiency of iterative optimization. Finally, the new values ​​of each weight coefficient and the baseline constant term after this iteration update are directly used to return to the step of performing gradient direction calculation and enter the next iteration loop. This iterative method allows the parameters to be continuously optimized under the guidance of gradient and the control of dynamic learning rate, gradually correcting numerical deviations and allowing each parameter to continuously approach the optimal value that maximizes the overall likelihood. Through multiple rounds of iterative iteration, the undetermined parameters are finally accurately solved, providing key parameter optimization support for subsequently determining the final parameters of the risk prediction function and generating a high-precision risk prediction model. This ensures that the model parameters can accurately match the clinical data characteristics and improve the prediction accuracy and stability of the model.

[0029] like Figure 3 The diagram shown is a structural schematic of the pneumothorax drainage risk model construction device based on multimodal physiological and imaging parameters provided in this application embodiment. It includes: a multimodal data acquisition module, a data preprocessing module, a logistic regression model training module, a model performance verification and optimization module, and a model engineering deployment module. The multimodal data acquisition module is used to acquire historical patient datasets from CT-guided percutaneous lung biopsy. These historical patient datasets contain multimodal parameters and outcome labels indicating whether the patient developed drainage-required pneumothorax post-surgery. The multimodal parameters include, but are not limited to, physiological characteristic parameters, lung CT image anatomical parameters, and puncture operation-related variable parameters. The data preprocessing module is used to perform full-process preprocessing on the historical patient dataset acquired in step one. Preprocessing operations include data cleaning, variable standardization, and input validation. Data cleaning removes missing values, outliers, and invalid samples from the dataset. Variable standardization converts continuous parameters... The system performs normalization and digitizes the categorical parameters. Input validation verifies whether the parameter values ​​are within a clinically reasonable range to generate standardized training data. The logistic regression model training module inputs the standardized training data obtained in step two into the logistic regression algorithm for model training to learn the independent contribution of each multimodal parameter to the risk of pneumothorax drainage and obtain a risk prediction model containing parameter weights. The model performance validation and optimization module constructs an external validation dataset independent of the standardized training data to perform multi-dimensional performance evaluation on the risk prediction model trained in step three. If the model performance does not reach the preset evaluation threshold, it returns to step three to readjust the model training parameters and trains again. If the model performance reaches the preset evaluation threshold, the model parameter optimization is completed. The model engineering deployment module deploys the risk prediction model validated in step four into the prediction system to achieve real-time calculation of the risk of pneumothorax drainage in new patients.

[0030] In this embodiment, the pneumothorax drainage risk model is constructed and applied through the collaborative operation of a multimodal data acquisition module, a data preprocessing module, a logistic regression model training module, a model performance verification and optimization module, and a model engineering deployment module. First, the multimodal data acquisition module specifically acquires a historical patient dataset of CT-guided percutaneous lung biopsy. This dataset comprehensively includes physiological characteristics and lung CT scans. Multimodal parameters related to imaging anatomy and puncture procedures, along with outcome labels indicating whether postoperative pneumothorax required drainage, were collected from multiple sources across all dimensions. This targeted acquisition of data overcame the limitations of traditional single-source data collection, comprehensively covering various clinical influencing factors of pneumothorax drainage risk. This laid a comprehensive and clinically relevant foundation of raw data for subsequent model construction, ensuring that the model could comprehensively consider various key risk factors. Subsequently, the data preprocessing module performed standardized preprocessing on the collected historical patient dataset, sequentially completing data cleaning to remove missing values, outliers, and invalid samples; variable standardization by normalizing continuous parameters and digitizing categorical parameters; and input validation to verify the clinically reasonable range of parameter values. Finally, standardized training data was generated. This preprocessing effectively solved the problems of messy raw data, inconsistent units, and numerical anomalies, significantly improving data quality and consistency. It avoided model training bias caused by data defects from the data source, allowing subsequent algorithm training to be carried out systematically based on standardized and accurate data sources. Next, the logistic regression model training module received the standardized training data and input it into the logistic regression algorithm for professional model training, using the algorithm to fit the depth... This study mines and learns the independent contribution of various multimodal parameters to the risk of pneumothorax drainage, and simultaneously solves for the weight coefficients corresponding to each parameter. Finally, a risk prediction model containing parameter weights is generated. The targeted application of the logistic regression algorithm enables quantitative analysis of the risk contribution of each parameter, breaking the industry pain point of traditional experience-based judgment without quantitative basis. This allows the model to accurately identify the degree of influence of different parameters on the risk of pneumothorax drainage, forming an initial risk prediction model with quantitative analysis capabilities. Then, the model performance verification and optimization module constructs an external verification dataset independent of the standardized training data based on the historical patient dataset. Multi-dimensional performance evaluation is carried out on the trained initial risk prediction model. If the model performance does not reach the preset evaluation threshold, the logistic regression model training module is fed back with adjustment instructions to readjust the training parameters and retrain. If the threshold is reached, the model parameter optimization is completed. The construction of the independent verification dataset enables strict cross-set verification of the model's generalization ability and prediction accuracy, while the iterative training optimization mechanism can effectively correct model bias, avoid overfitting or underfitting problems, and significantly improve the model's fitting degree and stability, ensuring that the final output model has the predictive performance that meets the requirements of clinical application.Finally, the model engineering deployment module deploys the validated optimal risk prediction model, embedding the model's core algorithms and parameters into the pneumothorax drainage risk prediction system. This achieves seamless integration between the model and the clinical application system, ultimately enabling real-time calculation of pneumothorax drainage risk in new patients. This module completes the crucial transformation from a theoretical model to a clinically practical tool, solving the problems of existing technologies where research results are difficult to implement and real-time clinical risk assessment is impossible. It allows for the rapid reception of individualized parameters from new patients in clinical settings and the immediate output of accurate risk prediction results. This provides clinicians with objective and efficient quantitative decision-making basis for developing preoperative plans and postoperative monitoring strategies. Furthermore, the division of labor and orderly connection of the five modules form a closed-loop system from data collection to model deployment, ensuring the efficiency, scientific rigor, and practicality of model construction and application, and contributing to accurate, real-time prediction and standardized clinical management of pneumothorax drainage risk.

Claims

1. A method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters, characterized in that, Includes the following steps: Step 1: Obtain a historical patient dataset of CT-guided percutaneous lung biopsy. The historical patient dataset contains multimodal parameters and outcome labels indicating whether the corresponding patients developed drainage-required pneumothorax after the procedure. The multimodal parameters include physiological characteristic parameters, lung CT image anatomical parameters, and puncture operation-related variable parameters. Step 2 involves performing full-process preprocessing on the historical patient dataset obtained in Step 1. The preprocessing operations include data cleaning, variable standardization, and input validation. Data cleaning involves removing missing values, outliers, and invalid samples from the dataset. Variable standardization involves normalizing continuous parameters and digitizing categorical parameters. Input validation involves verifying whether the parameter values ​​are within the clinically reasonable range in order to generate standardized training data. Step 3: Input the standardized training data obtained in Step 2 into the logistic regression algorithm to train the model, so as to learn the independent contribution of each multimodal parameter to the risk of pneumothorax drainage and obtain a risk prediction model containing parameter weights. Step 4: Construct an external validation dataset independent of the standardized training data, and evaluate the risk prediction model trained in Step 3 in multiple dimensions. If the model performance does not reach the preset evaluation threshold, return to Step 3 to readjust the model training parameters and train again. If the model performance reaches the preset evaluation threshold, the model parameter optimization is complete. Step 5: Deploy the risk prediction model validated in Step 4 into the prediction system to achieve real-time calculation of the drainage risk of pneumothorax in new patients.

2. The method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters as described in claim 1, characterized in that: The physiological characteristic parameters, lung CT imaging anatomical parameters, and puncture operation-related variable parameters together constitute a set of all-dimensional influencing parameters for constructing a risk model of pneumothorax drainage; The physiological characteristic parameters include the patient's gender and whether or not they have emphysema; The lung CT imaging anatomical parameters include whether the target lesion is adjacent to the pleura, the pleural depth of the target lesion, whether the target lesion involves the interlobar pleura, and the location of the lung lobe where the target lesion is located. The relevant variables and parameters for the puncture procedure include the number of times the puncture needle passes through the pleura and the number of biopsy samples taken.

3. The method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters as described in claim 1, characterized in that: The specific process for obtaining the risk prediction model including parameter weights is as follows: The standardized training data is input into the logistic regression algorithm to train the model, so as to learn the independent contribution of each multimodal parameter to the risk of pneumothorax drainage and obtain a risk prediction model containing parameter weights. From the standardized training data, a set of identified key predictive variables were selected for model construction. These key predictive variables include the patient's gender, the relationship between the target lesion and the adjacent pleura, the pleural depth of the target lesion, the number of times the puncture needle passed through the pleura, the number of biopsy samples taken, whether emphysema is present, whether the interlobar pleura is involved, and the location of the lung lobe where the target lesion is located.

4. The method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters as described in claim 3, characterized in that: The obtained risk prediction model including parameter weights further includes: Based on the logistic regression framework, a risk prediction function structure is constructed with the aforementioned key predictor variables as independent variables and whether or not pneumothorax requiring drainage occurs as the dependent variable. The risk prediction function is fitted and trained using the standardized training data to determine the specific weight coefficients corresponding to each key predictor variable in the risk prediction function. Based on the weight coefficients of each key predictor variable, the final risk prediction model is generated.

5. The method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters as described in claim 3, characterized in that: The logistic regression framework constructs a risk prediction function structure with the key predictor variables as independent variables and the occurrence of pneumothorax requiring drainage as the dependent variable, including: Construct a linear prediction subfunction whose output value is the sum of the products of the values ​​of each key predictor variable and their respective undetermined weight coefficients, and add it to an undetermined baseline constant term; Construct a probability transformation function that takes the output of the linear prediction subfunction as input and maps it to a risk probability value between 0 and 1 through a nonlinear transformation. This probability value represents the predicted likelihood of pneumothorax requiring drainage. The linear prediction sub-function is combined with the probability transformation function to form the complete mathematical structure of the risk prediction function.

6. The method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters as described in claim 5, characterized in that: The step of fitting and training the risk prediction function based on the standardized training data to determine the specific weight coefficients corresponding to each key predictor variable in the risk prediction function includes: The actual observed values ​​of the key predictive variables for each historical patient in the standardized training data are input into the constructed risk prediction model; The maximum likelihood estimation algorithm is adopted to maximize the overall predictive likelihood of the risk prediction function for the actual occurrence of pneumothorax requiring drainage in all historical patients. The algorithm iteratively optimizes and solves all undetermined parameters in the function, including each weight coefficient and the baseline constant term. Obtain and record the final weight coefficient value and the final baseline constant value for each key predictor variable determined after the convergence of the iterative optimization solution process.

7. The method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters as described in claim 6, characterized in that: The specific process for generating the final risk prediction model based on the obtained weight coefficients of each key predictor variable is as follows: The final parameter weights and final baseline constants of each key predictor obtained by iterative optimization are embedded into the undetermined parameter positions of the constructed linear predictor subfunction to form a linear predictor subfunction with completely determined parameter values. By combining the linear prediction sub-function with the constructed probability transformation function, a complete risk prediction function with all parameters determined is obtained. This complete risk prediction function with all parameters determined is the final risk prediction model.

8. The method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters as described in claim 1, characterized in that: The maximum likelihood estimation algorithm is employed to maximize the overall predictive likelihood of the risk prediction function for the actual occurrence of pneumothorax requiring drainage in all historical patients. This involves iterative optimization of all undetermined parameters in the function, including weighting coefficients and baseline constants, including: Based on the risk prediction function, the predicted probability of pneumothorax requiring drainage is calculated for each historical patient in the standardized training data. Combined with the outcome label of whether or not pneumothorax requiring drainage actually occurs, an overall likelihood function is constructed to evaluate the accuracy of the model's prediction. Calculate the gradient of the overall likelihood function with respect to each undetermined weight coefficient and the baseline constant term in the risk prediction function; Based on the calculated gradient direction, the values ​​of each undetermined weight coefficient and the baseline constant term are iteratively updated using an optimization algorithm. After each iteration update, the value of the overall likelihood function is recalculated, and it is determined whether its growth has reached the preset convergence criterion. When the convergence criterion is reached, the iteration stops, and the values ​​of the weight coefficients and the baseline constants obtained in the current iteration are used as the final weight coefficient values ​​and the final baseline constant values.

9. The method for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters as described in claim 1, characterized in that: The step of iteratively updating the values ​​of each undetermined weight coefficient and the baseline constant term using an optimization algorithm based on the calculated gradient direction includes: An initial value is set for each undetermined weight coefficient and the baseline constant term, and an initial learning rate parameter is set for the optimization process. Based on the calculated gradient direction, each gradient value is multiplied by the learning rate parameter and then subtracted from the current value of the corresponding undetermined weight coefficient or baseline constant term, thereby completing one iteration update. After completing one iteration of the update, the learning rate parameter is adjusted according to a preset strategy. Using the updated values ​​of the weight coefficients and baseline constants, return to calculate the gradient direction and proceed to the next iteration.

10. An apparatus for constructing a pneumothorax drainage risk model based on multimodal physiological and imaging parameters as described in any one of claims 1-9, characterized in that, include: The system includes a multimodal data acquisition module, a data preprocessing module, a logistic regression model training module, a model performance verification and optimization module, and a model engineering deployment module. The multimodal data acquisition module is used to acquire historical patient datasets of CT-guided percutaneous lung biopsy. The historical patient datasets include multimodal parameters and outcome labels indicating whether the patients developed drainage-required pneumothorax after the procedure. The multimodal parameters include physiological characteristic parameters, lung CT image anatomical parameters, and puncture operation-related variable parameters. The data preprocessing module is used to perform full-process preprocessing on the historical patient dataset obtained in step one. The preprocessing operations include data cleaning, variable standardization, and input validation. Data cleaning is to remove missing values, outliers, and invalid samples from the dataset. Variable standardization is to normalize continuous parameters and digitize categorical parameters. Input validation is to verify whether the parameter values ​​are within the clinically reasonable range in order to generate standardized training data. The logistic regression model training module is used to input the standardized training data obtained in step two into the logistic regression algorithm for model training, so as to learn the independent contribution of each multimodal parameter to the risk of pneumothorax drainage and obtain a risk prediction model containing parameter weights. The model performance verification and optimization module is used to construct an external verification dataset that is independent of the standardized training data, and to perform multi-dimensional performance evaluation on the risk prediction model trained in step three. If the model performance does not reach the preset evaluation threshold, the module returns to step three to readjust the model training parameters and train again. If the model performance reaches the preset evaluation threshold, the model parameter optimization is completed. The model engineering deployment module is used to deploy the risk prediction model that has passed the verification in step four into the prediction system to realize real-time calculation of the drainage risk of pneumothorax in new patients.

Images