Lung-cancer screening model construction apparatus, lung-cancer screening apparatus, and device and medium

Abstract

The present application relates to a lung-cancer screening model construction apparatus, a lung-cancer screening apparatus, and a device and a medium. The lung-cancer screening model construction apparatus comprises: a sample acquisition and data collection module, which is used for acquiring benign pulmonary nodule patient samples, and collecting baseline parameters of the benign pulmonary nodule patient samples, wherein the baseline parameters comprise radial growths of pulmonary nodules and serum thymidine kinase 1; a model training module, which is used for performing model training by using the baseline parameters as independent variables and using, as dependent variables, risks of lung cancer newly diagnosed within three years after baseline visits; and a model evaluation and verification module, which is used for performing prediction efficiency evaluation and internal verification on the trained model, so as to obtain a risk prediction model for newly diagnosed lung cancer of benign pulmonary nodule patients within three years. The present application can be used for predicting the risk of progression from benign pulmonary nodule to lung cancer in a large-sample Chinese population.

Description

Lung cancer screening model construction device, lung cancer screening device, equipment and medium Technical Field

[0001] The present application relates to the field of intelligent medical detection technology, and in particular to a device for constructing a lung cancer screening model, a lung cancer screening device, equipment, and a medium. Background Art

[0002] Lung cancer is a leading cause of global health burden, with an estimated 1.8 million deaths worldwide in 2020 and an estimated 700,000 deaths in my country in the same year. Therefore, early detection and treatment are crucial to reduce lung cancer mortality.

[0003] Low-dose computed tomography (LDCT) is a commonly used method for detecting pulmonary nodules. However, using LDCT to assess the risk of cancer in pulmonary nodules remains challenging. The false-positive rate is as high as 23.3% during two-year follow-up, which can lead to unnecessary radiation exposure, invasive biopsies, financial costs, and patient anxiety. Methods for assessing the risk of pulmonary nodules after LDCT examination need to be improved to optimize follow-up strategies.

[0004] Currently, my country's lung cancer screening guidelines recommend that further examinations be determined solely based on the radial and density of lung nodules measured by LDCT. European and British guidelines recommend the use of prediction models, which can significantly reduce the false positive rate. However, due to differences in the distribution of predictive factors, these foreign models are not applicable to Chinese people. In addition, the published Chinese population models have limited applicability due to their small sample size (<1000). Therefore, there is an urgent need for a prediction model construction technology based on a large sample of the Chinese population to predict the risk of lung cancer development in patients with lung nodules who do not have lung cancer.

[0005] Summary of the Invention

[0006] In order to predict the risk of lung nodules in patients without lung cancer developing lung cancer, the present application provides a device for constructing a lung cancer screening model, a lung cancer screening device, equipment and a medium.

[0007] In a first aspect, the present application provides a device for constructing a lung cancer screening model using the following technical solutions:

[0008] A device for constructing a lung cancer screening model, comprising:

[0009] A sample acquisition and data collection module is used to obtain samples from patients with lung nodules who do not have lung cancer and collect baseline parameters of the samples from patients with lung nodules; the baseline parameters include radial lung nodules and serum thymidine kinase 1;

[0010] a model training module, configured to perform model training using the baseline parameters as independent variables and the risk of new lung cancer within three years after the baseline visit as a dependent variable;

[0011] The model evaluation and verification module is used to evaluate the predictive effectiveness and internally verify the trained model, thereby obtaining a risk prediction model for new lung cancer within 3 years in patients without cancerous lung nodules.

[0012] By adopting the above-mentioned technical means, especially using the baseline parameters (including lung nodule radial and serum thymidine kinase 1) as independent variables, the model is trained with the risk of new lung cancer within 3 years after the baseline visit as the dependent variable, so that it can be used for a large sample of Chinese people to achieve the risk prediction of lung cancer in patients with lung nodules without lung cancer. In other words, this application can be used to predict the probability of new lung cancer in the next 3 years instead of diagnosing existing lung cancer, because patients with lung nodules who do not currently have lung cancer also have the risk of developing lung cancer in the future as the disease progresses, and there is currently no technology that can predict this. This application is the first practical tool that can predict the medium- and short-term risk of lung cancer. In addition, in this application, the model constructed using lung nodule radial and serum thymidine kinase 1 can further improve the sensitivity and specificity of lung cancer risk prediction; in addition, the short-term lung cancer risk prediction model in this application can detect high-risk factors in patients at an early stage and provide relevant intervention measures, thereby promoting early diagnosis and treatment, and will also bring the following benefits: Early detection and diagnosis: By using the short-term lung cancer risk prediction model, doctors can detect patients who may be at high risk earlier and conduct relevant examinations and diagnoses in a timely manner. This will help to detect lung cancer as early as possible, provide earlier treatment opportunities, and improve patient survival rates; Personalized treatment plans: The prediction model can provide each patient with personalized treatment recommendations and plans based on the patient's individual characteristics and risk factors. This will not only improve the treatment effect, but also reduce unnecessary treatment processes or drug side effects; Efficient resource utilization: By using the risk prediction model of this application, medical resources can be more targeted and allocated to those patient groups, reducing the waste of resources, including beds, manpower, equipment, etc., and improving the efficiency of medical services in the face of a large number of patients; Cost reduction: Early diagnosis and treatment are usually more effective than late treatment, and can reduce the cost required for subsequent treatment. Moreover, through the use of risk prediction models, the health risks of lung cancer patients can be better managed and the economic burden on the medical system can be reduced.

[0013] Preferably, in the model training module, model training specifically includes: fitting a logistic regression model and using LASSO regularization to constrain the parameters of the model.

[0014] More preferably, the use of LASSO regularization to constrain model parameters includes: controlling the sparsity of the model by adjusting the regularization parameters, thereby screening out the independent variable features that have the strongest impact on the dependent variable.

[0015] In this application, LASSO adds an L1 regularization term to automatically shrink the coefficients of some unimportant features to zero, thereby achieving automatic variable selection. In contrast, stepwise regression requires manual setting of thresholds for entering and exiting the model, which requires multiple iterations. LASSO, on the other hand, can screen out features that have a significant impact on the target variable in one go, making it simpler and more efficient.

[0016] Preferably, the risk prediction model for new lung cancer in patients without cancerous lung nodules within 3 years is:

[0017] Wherein, the lp=-7.066+lp(serum thymidine kinase 1)+lp(pulmonary nodule radial)

[0018] (Linear prediction value) lp(pulmonary nodule radial direction) = 0.3392*pulmonary nodule radial direction;

[0019] (Linear Prediction Value) lp(serum thymidine kinase 1) = 0.3355*serum thymidine kinase 1.

[0020] By adopting the above technical solution and using the risk prediction model for new lung cancer within 3 years in patients without cancerous lung nodules obtained in this application, the risk probability of new lung cancer within 3 years in patients without cancerous lung nodules can be accurately calculated.

[0021] Preferably, in the sample acquisition and data collection module, the samples exclude individuals with baseline cancer, individuals with a history of tumors, and individuals with a history of thoracic surgery at baseline;

[0022] and / or

[0023] In the sample acquisition and data collection module, the samples are selected from individuals who are 30 years old or above and have first-time lung nodules detected, have received at least one follow-up, have no lung cancer or lung cancer history, and have no history of thoracic surgery at baseline.

[0024] By adopting the above-mentioned technical means, the samples selected for modeling in this application exclude individuals with cancer at baseline, individuals with a history of tumors, and individuals with a history of thoracic surgery at baseline, so that this application can be applied to patients with lung nodules without a history of tumors rather than patients who already have tumors. The group of patients with lung nodules is large, with an incidence rate of 35.5%, and a certain risk of lung cancer (0.54%). Therefore, this application can help prevent a large number of new cases of lung cancer. In this application, selecting samples in which lung nodules are detected for the first time can avoid interference with lung cancer risk prediction caused by interventions that may be received after the nodule is detected, such as anti-inflammatory treatment; in addition, selecting samples that are 30 years old or above and samples that have received at least one follow-up can ensure that the samples have follow-up information and can confirm whether there is a new diagnosis of lung cancer during the follow-up to avoid misclassification, thereby improving the accuracy of the model constructed by the construction device.

[0025] Preferably, in the model evaluation and verification module, when evaluating the predictive effectiveness of the trained model, the predictive effectiveness of the model is evaluated by evaluating the discrimination and calibration of the model;

[0026] Specifically, the discrimination is evaluated by the area under the receiver operating characteristic (ROC) curve (AUC) and its 95% CI; this makes it more robust to class-imbalanced datasets, so that the results are not affected by class imbalance and can better reflect the performance of the model (whereas usually in the case of class imbalance, indicators such as accuracy may be misleading); in addition, the results are not affected by threshold selection, and a more comprehensive model evaluation can be provided for different application scenarios and needs; finally, the discrimination is evaluated by the area under the receiver operating characteristic (ROC) curve (AUC) and its 95% CI, so that it can be directly applied to multi-class classification problems without the need for additional modification or conversion.

[0027] and / or

[0028] The calibration is assessed using the ratio of the model's predicted probability to the actual probability (E:O), the optimal calibration intercept (CITL), and the calibration slope, providing a multi-dimensional assessment and a more comprehensive understanding of the model's calibration performance. Specifically, the ratio metric provides a visual representation of the model's calibration, the calibration slope measures how the model's calibration performance changes with confidence level, and the CITL can be used to evaluate the model's calibration performance on large datasets. Furthermore, the calibration, assessed using the ratio of the model's predicted probability to the actual probability (E:O), the optimal calibration intercept (CITL), and the calibration slope, provides a more accurate assessment of the model's calibration performance. Because each metric has its own unique advantages and limitations, the combined use of these metrics can overcome their shortcomings and provide a more comprehensive and accurate assessment. Furthermore, these metrics are interpretable and can intuitively demonstrate the model's calibration performance. For example, the ratio metric indicates the degree of agreement between the model's predictions and the actual probability, the calibration slope shows how the model's calibration performance changes with confidence level, and the reliability curves and reliability histograms provided by the CITL provide a visual representation of the model's calibration performance. Finally, using these metrics together can help detect potential problems in the model, such as overconfidence or overcautiousness. Model deficiencies can be identified by comparing the ratio of predicted to actual probabilities, analyzing changes in the calibration slope, and evaluating the calibration performance of the model on large datasets.

[0029] Preferably, in the model evaluation and verification module, a ten-fold cross validation is used to perform internal verification of the model; specifically, the following steps are performed:

[0030] Divide the data set consisting of sample data into ten subsets of equal size;

[0031] For each subset, take it as the test set in turn, and combine the remaining nine subsets as the training set;

[0032] Fit a logistic regression model on the training set;

[0033] The fitted model is used to predict the test set, and the performance indicators of the model on the test set are calculated, including the area under the receiver operating characteristic curve, the ratio of the model predicted probability to the actual probability, the optimal calibration intercept and the calibration slope.

[0034] Repeat steps 1 and 2 for each subset, using them as test sets in turn. Use the fitted model to predict the test set until each subset is used as a test set, and record the performance indicators each time.

[0035] For each performance index, the mean and standard deviation of ten tests were calculated to evaluate the average performance and stability of the model.

[0036] By using the above technical means to validate the model, each subset was validated once, thus fully utilizing the entire dataset and reducing the bias introduced by improper data partitioning. In addition, because each sample was used as part of the final evaluation results, the model's performance indicators were more stable and reliable, which could better reflect the model's generalization ability. In addition, by using the above method to continuously train and validate on different training sets and validation sets, it was possible to promptly identify situations where the model performed well on the training set but poorly on the validation set, thus avoiding the model from overfitting the data. Finally, by using the above method to internally validate the model, the optimal parameter configuration (i.e., the baseline parameters of lung nodule radial and serum thymidine kinase 1) could be found from 11 variables to improve model performance.

[0037] In a second aspect, the present application provides a lung cancer screening device that adopts the following technical solutions:

[0038] A lung cancer screening device comprises a lung cancer screening model constructed by the lung cancer screening model construction device described in any of the above items.

[0039] In a third aspect, the present application provides an electronic device that adopts the following technical solution:

[0040] An electronic device includes a memory and a processor, wherein the memory stores a lung cancer screening model constructed by a construction device for implementing a lung cancer screening model as described in any of the above items and can be loaded and executed by the processor.

[0041] In a fourth aspect, the present application provides a computer-readable storage medium that employs the following technical solutions:

[0042] A computer-readable storage medium stores a lung cancer screening model constructed by a construction device for implementing a lung cancer screening model as described in any of the above items and capable of being loaded and executed by a processor.

[0043] In summary, this application includes at least one of the following beneficial technical effects:

[0044] In this application, in particular, the baseline parameters (including lung nodule radial and serum thymidine kinase 1) are used as independent variables, and the risk of new lung cancer within 3 years after the baseline visit is used as the dependent variable for model training, so that it can be used for a large sample of Chinese people to predict the risk of lung cancer in patients with lung nodules without lung cancer. In other words, this application can be used to predict the probability of new lung cancer in the next 3 years rather than diagnose existing lung cancer, because patients with lung nodules who do not currently have lung cancer also have the risk of developing lung cancer in the future as the disease progresses, and there is currently no technology that can predict this. This application is the first practical tool that can predict the medium- and short-term risk of lung cancer. In addition, in this application, the model constructed using lung nodule radial and serum thymidine kinase 1 can further improve the sensitivity and specificity of lung cancer risk prediction;

[0045] The sample selected for modeling in this application excluded individuals with baseline cancer, individuals with a history of tumors, and individuals with a history of thoracic surgery at baseline, allowing this application to be applied to patients with lung nodules without a history of tumors rather than those with existing tumors. The patient population with lung nodules is large, with an incidence rate of 35.5% and a certain risk of lung cancer (0.54%). Therefore, this application can help prevent a large number of new lung cancer cases. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] FIG1 is a block diagram of a device for constructing a lung cancer screening model according to an embodiment of the present application.

[0047] Figure 2 is a flow chart for screening research samples in the experimental example.

[0048] Figure 3 is a schematic diagram of model construction and verification in the experimental example, where (a) is the LASSO result; (b) is the 3-year risk prediction model and the receiver operating characteristic (ROC) curve of each predictive factor; (c) is the calibration curve of the model in internal validation; and (d) is the ROC curve of the model in internal validation.

[0049] Figure 4 Threshold and decision curve diagram; where ac is the Youden index corresponding to the 3-year lung cancer risk prediction model, lung nodule radial and serum thymidine kinase 1 thresholds, respectively; (d) is the model decision curve. DETAILED DESCRIPTION

[0050] The present application is further described in detail below with reference to Figures 1 to 4 .

[0051] The present application discloses a device for constructing a lung cancer screening model. Referring to FIG1 , a device for constructing a lung cancer screening model includes:

[0052] A sample acquisition and data collection module is used to obtain samples from patients with lung nodules who do not have lung cancer and to collect baseline parameters of the samples from patients with lung nodules;

[0053] The sample excluded individuals with cancer at baseline, individuals with a history of tumors, and individuals with a history of thoracic surgery at baseline;

[0054] and / or

[0055] In the sample acquisition and data collection module, the samples are selected from individuals who are 30 years old or older and have a first-time lung nodule detection, have no lung cancer or lung cancer history, have received at least one follow-up visit (if the individual has no record of lung cancer during the follow-up period, the follow-up period shall be at least 3 years), and have no history of thoracic surgery at baseline;

[0056] The baseline parameters include the radial direction of the lung nodules and serum thymidine kinase 1. Specifically, when training the model, the baseline parameters include: age, gender, smoking status, three serum biomarkers - serum thymidine kinase 1 (STK1p), carcinoembryonic antigen (CEA) and alpha-fetoprotein (AFP), four clinical characteristics - body mass index (BMI), hypertension, hyperglycemia and dyslipidemia, and the radial direction of the lung nodules; wherein the three serum biomarkers refer to the concentration of serum thymidine kinase 1 (STK1p) in the blood, the concentration of carcinoembryonic antigen (CEA) in the blood, and the concentration of alpha-fetoprotein (AFP) in the blood;

[0057] The baseline parameters can be obtained by:

[0058] Age, sex, and smoking status were self-reported by patients and recorded by their physicians. For patients whose smoking status was not documented, the physician's written notes were further reviewed. Patients whose smoking status could not be confirmed were not included in the study.

[0059] Three serum biomarkers, including STK1p, CEA, and AFP, were collected at baseline. STK1p was detected using an enhanced chemiluminescence assay (CIS-2) using a TK1-IgY polyclonal antibody. Serum CEA and AFP were measured using chemiluminescence (ALINITY ci-series).

[0060] Four clinical characteristics were included, including body mass index (BMI), hypertension, hyperglycemia, and dyslipidemia. BMI (kg / m2) was calculated based on self-reported height and weight. Hypertension was defined as high systolic blood pressure (≥130 mmHg) or high diastolic blood pressure (≥80 mmHg) according to a medical record. Hyperglycemia was defined as high fasting blood glucose (≥6.1 mmol / L) or high 2-hour postprandial blood glucose (≥7.8 mmol / L) according to a medical record. Dyslipidemia was defined as abnormal levels of total cholesterol (≥6.2 mmol / L), triglycerides (≥11.3 mmol / L), low-density lipoprotein cholesterol (≥4.14 mmol / L), or high-density lipoprotein cholesterol (<1.04 mmol / L) according to a medical record.

[0061] LDCT scans are routinely performed and analyzed by hospital radiologists according to existing guidelines. The radial diameter of pulmonary nodules can be obtained from the LDCT report. The radial diameter (mm) of a nodule is defined as the maximum diameter of the solid area for part-solid and solid nodules or the maximum diameter for ground-glass nodules. For patients with multiple nodules, the maximum diameter of the largest nodule is used.

[0062] a model training module, configured to perform model training using the baseline parameters as independent variables and the risk of new lung cancer within three years after the baseline visit as a dependent variable;

[0063] Lung cancer was defined based on ICD-9 and ICD-10 codes recorded in the hospital information system. Individuals without a code were considered to have no lung cancer. For patients with new-onset lung cancer, the study endpoint was the date of lung cancer diagnosis. For patients without new-onset lung cancer, the study endpoint was the date of the last visit after the baseline visit.

[0064] Specifically, the model training includes: fitting a logistic regression model and using LASSO regularization to constrain the parameters of the model;

[0065] The fitting logistic regression model includes: establishing a logistic regression model using a LASSO regression algorithm, wherein the goal of LASSO is to minimize the following loss function: L(β)=-loglikelihood(y,Xβ)+α*∑|β|; wherein loglikelihood(y,Xβ) represents the log-likelihood loss of logistic regression; α is a hyperparameter for controlling the regularization strength, and the optimal α value can be selected by cross-validation or other methods; ∑|β| represents the L1 regularization term of the parameter, and the model parameter β can be obtained by fitting the model using the selected optimal hyperparameter α value.

[0066] The LASSO regression algorithm is used in this application to establish a logistic regression model, which has the following advantages:

[0067] 1. Feature selection: LASSO can achieve automatic feature variable selection by performing L1 regularization on the parameters, so that the coefficients of some features tend to zero; this can reduce unimportant features in the model and improve the interpretability and generalization ability of the model; compared with existing methods such as stepwise regression, which require manual setting of thresholds for entering and exiting the model and require multiple iterations, LASSO in this application can screen out features that have a significant impact on the target variable at one time, which is simpler and more efficient.

[0068] 2. Model Regularization: LASSO can control model complexity and prevent overfitting by adjusting the regularization parameter. Additionally, L1 regularization can make the model more sparse, reducing unnecessary features and thus improving the model's generalization ability. However, existing stepwise regression algorithms do not have a clear sparsity feature.

[0069] 3. Handling Collinearity: LASSO can effectively handle collinearity between features. That is, when there is a high correlation between input features, LASSO can select one feature as the representative and reduce the coefficients of other related features to zero. This can reduce the impact of multicollinearity on the model and improve modeling efficiency. However, stepwise regression cannot directly handle multicollinearity and may select one relevant feature while ignoring other relevant features.

[0070] 4. Interpretability: LASSO’s constraints on model parameters can make the model more interpretable. Since L1 regularization reduces the coefficients of some features to zero, it is easy to see which features have a greater or lesser impact on the target variable.

[0071] 5. Stability: LASSO can still perform model fitting stably with fewer samples and a large number of features, and has better performance.

[0072] 6. LASSO is very flexible and can be adjusted according to the characteristics of the specific problem and dataset to obtain the best model performance and sparsity. However, stepwise regression does not have clear parameters to adjust.

[0073] In summary, compared with the stepwise regression method, the use of the LASSO regression algorithm to establish a logistic regression model in this application has the advantages of automatic selection of feature variables, model regularization, sparsity, handling of collinearity, interpretability and stability, which can help improve the performance and explanatory power of the logistic regression model.

[0074] The method of using LASSO regularization to constrain the model parameters includes: controlling the sparsity of the model by adjusting the regularization parameters, thereby screening the independent variable features that have the strongest impact on the dependent variable. In this application, the independent variable features that have the strongest impact on the dependent variable are pulmonary nodule radial and serum thymidine kinase 1;

[0075] Specifically, by using techniques such as cross-validation to select the optimal regularization parameter, and then by increasing the regularization parameter, the sparsity of the model can be increased, that is, the number of non-zero coefficients in the model can be reduced; conversely, reducing the regularization parameter can reduce the sparsity of the model, that is, the number of non-zero coefficients in the model can be increased, thereby screening out the independent variable characteristics that have the strongest impact on the dependent variable.

[0076] By using the above method, the coefficients of some unimportant features can be shrunk to zero, thus achieving variable selection. That is, LASSO can automatically screen out features that have a significant impact on the target variable, reduce redundant features in the model, and improve the interpretability of the model.

[0077] The model evaluation and validation module is used to evaluate the predictive efficacy and internally validate the trained model, thereby obtaining a model for predicting the risk of new lung cancer within 3 years in patients without cancerous lung nodules;

[0078] In this example, the risk prediction model for new lung cancer in patients without cancerous lung nodules within 3 years is obtained as follows:

[0079] Wherein, the lp=-7.066+lp(serum thymidine kinase 1)+lp(pulmonary nodule radial)

[0080] (Linear prediction value) lp(pulmonary nodule radial direction) = 0.3392*pulmonary nodule radial direction;

[0081] (Linear Prediction Value) lp(serum thymidine kinase 1) = 0.3355*serum thymidine kinase 1.

[0082] In this embodiment, when evaluating the predictive effectiveness of the trained model, its predictive effectiveness is evaluated by evaluating the discrimination and calibration of the model;

[0083] Specifically, the discrimination was evaluated by the area under the receiver operating characteristic (ROC) curve (AUC) and its 95% CI;

[0084] and / or

[0085] The calibration was assessed by the ratio of model predicted probability to actual probability (E:O), the optimal calibration intercept (CITL), and the calibration slope.

[0086] Calibration plots can be used to visually compare model-predicted probabilities with actual probabilities.

[0087] In this embodiment, a ten-fold cross validation is used to perform internal validation on the model, specifically including:

[0088] S1, divide the data set consisting of sample data into ten subsets of equal size;

[0089] S2, for each subset, take it as the test set in turn, and combine the remaining nine subsets as the training set;

[0090] S3, fit the logistic regression model on the training set;

[0091] S4, use the fitted model to predict the test set and calculate the performance indicators of the model on the test set, including the area under the receiver operating characteristic curve, the ratio of the model predicted probability to the actual probability, the optimal calibration intercept and the calibration slope.

[0092] S5 repeats steps S2-S4 until each subset is used as a test set, and records the performance indicators each time;

[0093] S6, for each performance index, the mean and standard deviation of ten tests are calculated to evaluate the average performance and stability of the model.

[0094] Those skilled in the art will clearly understand that for the sake of convenience and brevity in description, only the division of the above-mentioned functional modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0095] This embodiment also discloses a lung cancer screening device, comprising a lung cancer screening model constructed by the lung cancer screening model construction device described in any one of the above items.

[0096] This embodiment also discloses an electronic device, comprising a memory and a processor, wherein the memory stores a lung cancer screening model constructed by the lung cancer screening model construction device as described in any one of the above items and can be loaded and executed by the processor.

[0097] Among them, the electronic device can be an electronic device such as a desktop computer, a laptop computer or a cloud server, and the electronic device includes but is not limited to a processor and a memory. For example, the electronic device can also include input and output devices, network access devices and buses, etc.

[0098] The processor in the present application may include one or more processing cores. The processor calls the data stored in the memory by running or executing the instructions, programs, code sets or instruction sets stored in the memory, performs the various functions of the present application and processes data. The processor can be at least one of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller and a microprocessor. It is understandable that for different devices, the electronic device for realizing the above-mentioned processor function can also be other, and the embodiments of the present application are not specifically limited.

[0099] Among them, the memory can be an internal storage unit of the electronic device, such as the hard disk or memory of the electronic device, or it can be an external storage device of the electronic device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD) or flash memory card (FC) equipped on the electronic device. In addition, the memory can also be a combination of an internal storage unit and an external storage device of the electronic device. The memory is used to store computer programs and other programs and data required by the electronic device. The memory can also be used to temporarily store data that has been output or is to be output. This application does not impose any restrictions on this.

[0100] This embodiment also discloses a computer-readable storage medium that stores a lung cancer screening model constructed by the lung cancer screening model construction device as described in any one of the above items and that can be loaded and executed by a processor.

[0101] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0102] Experimental example:

[0103] 1. Sample Acquisition and Data Collection

[0104] The health management center provides routine health examinations, including chest LDCT scans and serological tests. In this experimental case, 13,609 patients with lung nodules from January 1, 2018 to December 31, 2021 were retrospectively screened from the hospital's electronic medical records (EMR); of them, 6,841 people (50.27%) met the inclusion criteria (as shown in Figure 2) - patients aged 30 or above with the first detection of lung nodules. The baseline visit was defined as the visit at which the lung nodule was first detected. The baseline visit had complete candidate predictive factor information and had received at least 1 follow-up visit; if the patient had no record of lung cancer during the follow-up period, the follow-up should be at least 3 years. A total of 6768 (49.73%) patients were excluded: 93 (0.68%) had a baseline cancer (International Classification of Diseases (ICD)-9 or ICD-10 code) or a history of thoracic surgery, 1,117 (8.21%) had incomplete baseline information, and 5,558 (40.84%) had no follow-up records. The median age of the 6841 patients enrolled (51 years (IQR = 45-59) vs 53 years (45-61)) and the proportion of women (40.32% vs 39.74%) were similar to those of the 13,609 patients enrolled before screening. The median follow-up duration was 3.00 years (IQR: 3.00-3.40). The age ranged from 31 to 91 years. During follow-up, 52 patients (0.76%) developed new lung cancer. Patients with and without new lung cancer had similar age, sex, smoking, and clinical characteristics at baseline (Table 1). Patients with new lung cancer had higher nodule radial (p < 0.001) and STK1p (p < 0.001) levels at baseline than those without new lung cancer (Table 1), indicating that the selection of lung nodule radial and serum thymidine kinase 1 levels are effective factors for predicting lung cancer.

[0105] Table 1 Baseline characteristics of the study subjects

[0106] Note: Data are medians (interquartile ranges) or n (%). For smoking indicators, data outside parentheses are sample sizes, and data inside parentheses are percentages. For pulmonary nodule radial distribution, data outside parentheses are medians, and data inside parentheses are interquartile ranges.

[0107] The P value represents the significance of the difference between the groups with and without new lung cancer, with p < 0.05 indicating a significant difference. For continuous variables such as age, the P value was obtained using the Wilcoxon rank sum test; for categorical variables such as sex, the P value was obtained using the chi-square test.

[0108] 2. Prediction Model Construction

[0109] The LASSO algorithm selected lung nodule radial size and STK1p from 11 candidate indicators as predictive factors for the logistic model (as shown in Figure 3a and Table 2).

[0110] Table 2

[0111] Univariate and multivariate logistic regression analyses (Table 3) showed that both predictors were associated with a 3-year risk of lung cancer. The model had an area under the curve (AUC) of 0.92 (95% CI: 0.87, 0.97), significantly higher than the nodule radial region (0.89 (0.84, 0.95)) and STK1p (0.74 (0.68, 0.80), respectively) (P < 0.001) (Figure 3b). The model also demonstrated excellent performance in internal validation with its E:O ratio (0.998), CITL (0.002), calibration slope (0.95) (Figure 3c), and AUC (0.92) (Figure 3d).

[0112] Table 3 Predictive factors screened by LASSO procedure to construct logistic regression model

[0113] Table 3 shows that the odds ratio of radial lung nodules and serum thymidine kinase 1 is greater than 1 and the p value is <0.05, indicating that these two factors are significantly positively correlated with lung cancer.

[0114] In logistic regression, a constant is a variable with a coefficient of 0. It is an independent variable with no dependent variable, and its coefficient is set to 0 in the model. Constants are often used in linear regression models to control the effects of other independent variables on the dependent variable.

[0115] The coefficient in logistic regression is a parameter used in the model to measure the degree to which the independent variable affects the dependent variable. For binary logistic regression, the coefficient represents the effect of the independent variable on the log-odds ratio of the dependent variable. Specifically, when the value of the independent variable increases by one unit, the log-odds ratio will increase or decrease accordingly, depending on whether the coefficient is positive or negative.

[0116] Specifically, in logistic regression, the coefficients are obtained by maximum likelihood estimation. The specific steps are as follows:

[0117] Suppose there is a binary classification problem where the target variable takes two values, usually denoted as 0 and 1;

[0118] The linear combination is converted into a probability value using the logistic function (also known as the sigmoid function). The formula for the logistic function is as follows: P(Y=1)=1 / (1+exp(-z));

[0119] Where P(Y=1) represents the probability that the target variable is equal to 1, exp represents the natural exponential function, and z represents the linear combination;

[0120] According to the principle of maximum likelihood estimation, the coefficients that make the observed data have the highest probability of occurrence are selected. The goal of maximum likelihood estimation is to maximize the likelihood function of all samples, that is, to maximize the product of likelihood functions: L(β)=P(y=1|x)^y*P(y=0|x)^(1-y),

[0121] Where β represents the coefficient of the model, x represents the feature vector, and y represents the actual observed target variable.

[0122] Take the logarithm of the likelihood function to facilitate calculation and optimization: log(L(β))=Σ(y*log(P(y=1|x))+(1-y)*log(P(y=0|x))).

[0123] Using an optimization algorithm (e.g., gradient descent, Newton-Raphson, etc.), find the coefficients that maximize the log-likelihood function.

[0124] Based on the Youden index (as shown in Figure 4), the optimal threshold for the model's predicted probability is 0.06%. The sensitivity and specificity at this threshold are 88.5% (false negative rate 11.5%) and 89.0% (false positive rate 11.0%), respectively, which are higher than the optimal thresholds for using nodule radial or STK1p alone (as shown in Table 4). The decision curve of the model is shown in Figure 4d. Between the thresholds of 0% and 50%, the net benefit of model-based intervention is always higher than no intervention or intervention of all samples (that is, the effect of intervention on the two factors selected in this application - "pulmonary nodule radial and serum thymidine kinase 1" is higher than no intervention, or intervention of all factors; among them, intervention of all samples means assuming that follow-up diagnosis and treatment are provided for all samples, such as PET-CT, etc.). The 3-year probability of new lung cancer in the high-risk group (with a predicted probability higher than 0.06%) was higher than that in the low-risk group (OR=44.75 [95%CI: 19.06, 105.05]; P<0.001), indicating that the corresponding model-predicted high-risk probability in the population with new lung cancer was 44.75 times higher than that of the low-risk group.

[0125] Table 4 Optimal thresholds

[0126] Note: AUC = area under the receiver operating characteristic (ROC) curve; 95% CI = 95% confidence interval.

[0127] In specific implementation, a web application can be developed on the terminal based on the model, and then the probability assessment of new lung cancer within 3 years can be provided for individual lung nodule patients based on the input predictive factors.

[0128] The technology of this application can be applied to scenarios such as primary medical institutions, physical examination centers, health management centers, and public health institutions for early screening of lung cancer.

[0129] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the devices and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A device for constructing a lung cancer screening model, characterized in that: include: A sample acquisition and data collection module is used to obtain samples from patients with lung nodules who do not have lung cancer, and to collect baseline parameters of the samples from patients with lung nodules; the baseline parameters include radial lung nodules and serum thymidine kinase 1; A model training module, for performing model training using the baseline parameters as independent variables and the risk of new lung cancer within 3 years after the baseline visit as a dependent variable; The model evaluation and verification module is used to evaluate the predictive effectiveness and internally verify the trained model, thereby obtaining a risk prediction model for new lung cancer within 3 years in patients without cancerous lung nodules.

2. The device for constructing a lung cancer screening model according to claim 1, characterized in that: In the model training module, model training specifically includes: fitting a logistic regression model, and using LASSO regularization to constrain the parameters of the model.

3. The device for constructing a lung cancer screening model according to claim 2, characterized in that: The method of using LASSO regularization to constrain model parameters includes: controlling the sparsity of the model by adjusting the regularization parameters, thereby screening out the independent variable features that have the strongest impact on the dependent variable.

4. The device for constructing a lung cancer screening model according to any one of claims 1 to 3, characterized in that: The risk prediction model for new lung cancer in patients without cancerous lung nodules within 3 years is: Wherein, the lp=-7.066+lp(serum thymidine kinase 1)+lp(pulmonary nodule radial) lp(pulmonary nodule radial direction)=0.3392*pulmonary nodule radial direction; lp(serum thymidine kinase 1)=0.3355*serum thymidine kinase 1.

5. The device for constructing a lung cancer screening model according to claim 1, characterized in that: In the sample acquisition and data collection module, the sample excludes individuals with cancer at baseline, individuals with a history of tumors, and individuals with a history of thoracic surgery at baseline; and / or In the sample acquisition and data collection module, the samples are selected from individuals who are 30 years old or above and have lung nodules detected for the first time, have received at least one follow-up, have no lung cancer or history of lung cancer, and have no history of thoracic surgery at baseline.

6. The device for constructing a lung cancer screening model according to claim 1, characterized in that: In the model evaluation and verification module, when evaluating the predictive performance of the trained model, the predictive performance of the model is evaluated by evaluating the discrimination and calibration of the model; Specifically, the discrimination was evaluated by the area under the receiver operating characteristic curve and its 95% CI; and / or The calibration was assessed by the ratio of the model predicted probability to the actual probability, the optimal calibration intercept and the calibration slope.

7. The device for constructing a lung cancer screening model according to claim 1, characterized in that: In the model evaluation and verification module, a ten-fold cross validation is used to perform internal verification of the model; specifically, the following steps are performed: Divide the data set consisting of sample data into ten subsets of equal size; For each subset, take it as the test set in turn, and combine the remaining nine subsets as the training set; Fit a logistic regression model on the training set; The fitted model was used to predict the test set, and the performance indicators of the model on the test set were calculated, including the area under the receiver operating characteristic curve, the ratio of the model predicted probability to the actual probability, the optimal calibration intercept, and the calibration slope. Repeat steps 1 to 3 for each subset, using it as a test set in turn. Use the fitted model to predict the test set until each subset is used as a test set, and record the performance indicators each time. For each performance index, the mean and standard deviation of ten tests were calculated to evaluate the average performance and stability of the model.

8. A lung cancer screening device, characterized in that: A lung cancer screening model constructed by the device for constructing a lung cancer screening model as described in any one of claims 1-7.

9. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory stores a lung cancer screening model that can be loaded and executed by the processor to implement the lung cancer screening model constructed by the construction device as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The device stores a lung cancer screening model that can be loaded and executed by a processor and is constructed by a construction device for implementing a lung cancer screening model as described in any one of claims 1 to 7.

Images