Cervical cancer precancerous lesion diagnosis method based on machine learning

Abstract

The invention relates to a machine learning-based cervical precancerous lesion diagnosis method, which comprises the following steps of: based on historical clinical data, analyzing by using a statistical method to obtain risk factor variables significantly related to cervical precancerous lesions; constructing a plurality of diagnosis models based on different machine learning algorithms by taking the screened risk factor variables as input variables and the cervical cancer precancerous lesion probabilities as output variables; on the basis of historical clinical data, optimal diagnosis models with the best prediction effect for different levels of precancerous lesions of the cervical cancer are screened out; and collecting clinical data of a patient, and respectively inputting the clinical data into each preferable diagnosis model to predict the probability of the precancerous lesions of the cervical cancer at different levels so as to obtain a precancerous lesion grading prediction result. According to the invention, the level and probability of precancerous lesions of the neck can be effectively predicted, unnecessary biopsy is reduced, and missed diagnosis of high-level lesions is avoided.

Description

Technical Field

[0001] This invention relates to the field of smart medical technology, and in particular to a machine learning-based method for diagnosing precancerous lesions of the cervix. Background Technology

[0002] Cervical cancer is the fourth most common cancer among women worldwide, after breast cancer, colorectal cancer, and lung cancer. A 2024 domestic study showed that the incidence and mortality rates of cervical cancer have increased significantly in people over 35 years of age in recent years, with the largest increase in the 45-49 age group. In low- and middle-income countries, due to limited preventative measures, the disease is often diagnosed only after it has progressed to an advanced stage, and access to treatment for this advanced stage can be very limited, leading to higher mortality rates in resource-scarce regions. Therefore, the prevention and treatment of cervical cancer is a global public health issue and one of the major public health challenges that urgently needs to be addressed in my country.

[0003] The primary cause of cervical cancer and precancerous lesions is persistent infection with HPV (Human Papillomavirus), especially HRHPV (High-risk HPV). The progression from HRHPV infection to precancerous cervical changes (CIN1-CIN2-CIN3) and finally to cervical cancer is a lengthy process, taking years or even decades. Currently, cervical cytology and HPV testing are key strategies for cervical cancer screening globally. In regions with sufficient resources, combined testing has become one of the preferred methods for cervical cancer screening. However, combined testing suffers from insufficient specificity and a high false-positive rate in practical applications. Furthermore, existing methods lack risk stratification tools, which can easily lead to overdiagnosis and overtreatment. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a machine learning-based diagnostic method for cervical precancerous lesions, which can effectively predict the grade and probability of cervical precancerous lesions, reduce unnecessary biopsies and avoid missing high-grade lesions.

[0005] The technical solution adopted by this invention to solve its technical problem is: to provide a machine learning-based method for diagnosing precancerous lesions of the cervix, comprising:

[0006] Based on historical clinical data, statistical methods were used to analyze and identify risk factor variables that were significantly associated with precancerous lesions of the cervix.

[0007] Using the selected risk factor variables as input variables and the probability of cervical precancerous lesions as output variables, multiple diagnostic models based on different machine learning algorithms were constructed.

[0008] Based on historical clinical data, the best diagnostic models for predicting different levels of cervical precancerous lesions were selected.

[0009] Patient clinical data were collected and input into various optimized diagnostic models to predict the probability of different grades of cervical precancerous lesions, thus obtaining the grading prediction results of cervical precancerous lesions.

[0010] Furthermore, before the step of constructing multiple diagnostic models based on different machine learning algorithms using the selected risk factor variables as input variables and the probability of cervical precancerous lesions as output variables, the method further includes a step of encoding the classification results in the risk factor variables into corresponding numerical values ​​as the values ​​of the corresponding classification variables.

[0011] Furthermore, the different grades of cervical precancerous lesions include CIN2+ and CIN3+.

[0012] Furthermore, the preferred diagnostic models for CIN2+ and CIN3+ are both constructed based on the logistic regression algorithm.

[0013] Furthermore, it also includes:

[0014] Obtain sensitivity, NPV, and PPV data for the optimal CIN2+ diagnostic model;

[0015] The maximum log-odds ratio that makes the model’s sensitivity and NPV greater than their respective set thresholds is set as the low-risk cutoff value, and the minimum log-odds ratio that makes the model’s PPV greater than its set threshold is set as the high-risk cutoff value.

[0016] A nomogram was plotted, which included scores for all risk factor variables and the lesion probability predicted by the preferred diagnostic model for CIN2, and low-risk and high-risk cutoff values ​​were introduced to guide clinical decision-making.

[0017] Furthermore, after obtaining the cervical precancerous lesion grading prediction result, the method further includes:

[0018] Obtain the regression coefficients of each risk factor variable in the CIN2+ optimal diagnostic model;

[0019] Calculate the product of the regression coefficient and the value for each risk factor variable to obtain the corresponding score;

[0020] The scores of all risk factor variables are summed to obtain the patient's total CIN2+ risk score.

[0021] Furthermore, the risk factor variables include age, TCT classification, HPV type, colposcopy impression, and cervical transformation zone type.

[0022] Furthermore, based on historical clinical data, the preferred diagnostic models with the best predictive effect for different levels of cervical precancerous lesions are selected, including:

[0023] Historical clinical data was divided into training datasets and validation datasets;

[0024] All diagnostic models were trained using the training dataset to obtain multiple candidate diagnostic models for predicting different levels of cervical precancerous lesions.

[0025] All candidate diagnostic models were validated using a validation dataset, and the best-performing diagnostic model for different levels of cervical precancerous lesions was selected.

[0026] Furthermore, the statistical analysis used to identify risk factor variables significantly associated with precancerous cervical lesions includes:

[0027] For continuous variables, their distributions are presented in terms of median and interquartile ranges, and nonparametric tests are used to compare differences in the distributions.

[0028] For categorical variables, percentages are used to represent their distribution, and chi-square tests are used to compare differences in the distributions.

[0029] Variables with P < 0.05 were selected as risk factors significantly associated with cervical precancerous lesions.

[0030] Furthermore, the historical clinical data includes patient basic information questionnaires, clinical history, HPV test results, cervical cytology results, colposcopy diagnosis and cervical transformation zone type, as well as pathological results and pathological grades.

[0031] Beneficial effects

[0032] By employing the above-mentioned technical solutions, this invention has the following advantages and positive effects compared with existing technologies: This invention utilizes statistical methods to determine risk factor variables significantly associated with precancerous cervical lesions, and selects diagnostic models with the best predictive effect for these risk factor variables. This overcomes the limitations of traditional combined testing that relies solely on single or dual indicators, significantly improving the ability to differentiate high-grade lesions (CIN2+, CIN3+), effectively distinguishing true lesions from false positive results, and reducing over-biopsy and treatment due to insufficient specificity. Furthermore, by setting dual cutoff values ​​for low and high risk, and combining visualization tools such as nomograms and integral tables, this invention transforms complex risk probabilities into quantifiable clinical decision indicators. The low-risk cutoff value accurately identifies low-risk individuals who do not require immediate biopsy, avoiding a "one-size-fits-all" approach to testing. The high-risk cutoff value identifies high-risk groups, ensuring timely intervention for high-grade lesions. This achieves a stratified strategy for personalized diagnosis, helping clinicians quickly assess patient risk levels, shorten decision-making time, and improve diagnostic efficiency. Attached Figure Description

[0033] Figure 1 This is a flowchart of an embodiment of the present invention;

[0034] Figure 2 This is a line diagram of an embodiment of the present invention. Detailed Implementation

[0035] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0036] The embodiments of the present invention relate to a machine learning-based method for diagnosing precancerous lesions of the cervix, such as... Figure 1 As shown, it includes:

[0037] Based on historical clinical data, statistical methods were used to analyze and identify risk factor variables that were significantly associated with precancerous lesions of the cervix.

[0038] Using the selected risk factor variables as input variables and the probability of cervical precancerous lesions as output variables, multiple diagnostic models based on different machine learning algorithms were constructed.

[0039] Based on historical clinical data, the best diagnostic models for predicting different levels of cervical precancerous lesions were selected.

[0040] Patient clinical data were collected and input into various optimized diagnostic models to predict the probability of different grades of cervical precancerous lesions, thus obtaining the grading prediction results of cervical precancerous lesions.

[0041] Cervical intraepithelial neoplasia (CIN) is a general term for precancerous lesions of the cervix, reflecting abnormal proliferation of squamous epithelial cells in the cervix. It is classified into three grades: CIN1, CIN2, and CIN3, based on the proportion of abnormal cells in the epithelial layer. Therefore, CIN2+ and CIN3+ can be selected for probability prediction and diagnosis of precancerous lesions of the cervix. Specifically: CIN2+ refers to CIN2 and higher grade lesions (including CIN2 and CIN3), belonging to high-grade squamous intraepithelial lesions (HSIL); CIN3+ refers to CIN3 and carcinoma in situ.

[0042] In clinical treatment strategies, cervical biopsy and conservative treatment are generally recommended for CIN2+, while immediate treatment is strongly recommended for CIN3+ to prevent progression to cervical cancer. Therefore, diagnostic models can be developed separately for CIN2+ and CIN3+. The CIN2+ model aims to screen a broader range of high-grade lesions (including CIN2 and CIN3), focusing on early detection of potential risks and avoiding missed diagnoses of any high-grade lesions. The CIN3+ model focuses on more severe CIN3 and above lesions, aiming to accurately identify high-risk groups and guide clinicians to adopt more aggressive treatment strategies.

[0043] Historical clinical data can include patient basic information, clinical history, HPV test results, cervical cytology results, colposcopy diagnosis and cervical transformation zone type, and pathology results. To ensure data integrity, data cleaning can be performed, and missing and outlier values ​​can be added and corrected. Additionally, the patient's pathological grade can be determined based on the pathology report.

[0044] The collected information was compiled into an Excel spreadsheet, and each patient's information was coded. The following methods can be used:

[0045] a) Patient age is a continuous variable and no special treatment is required;

[0046] b) Cervical cytology results of patients: NILM was coded as 0, ASC-US as 1, LSIL as 2, ASC-H as 3, while HSIL, carcinoma in situ and cervical squamous cell carcinoma were all classified as HSIL+ (High-Grade Squamous Intraepithelial Lesion and above) and coded as 4, as a multi-category variable;

[0047] c) Patient's HPV test results: To ensure the stability of the model, HRHPV types other than HPV16 were classified as "other HPV". For the HPV status variable, the state of not being infected with HPV was 0, the state of being infected with HPV16 only was 1, the state of being infected with both HPV16 and "other HPV" was 2, and the state of being infected with "other HPV" only was 3, as a multi-category variable;

[0048] d) Patient's colposcopy impression: "No abnormal findings" is coded as 0, "LSIL findings" is coded as 1, "HSIL findings" is coded as 2, and "cervical cancer findings" is coded as 3, as a multi-category variable;

[0049] e) Cervical transformation zone type of the patient: Cervical transformation zone type I is coded as 1, cervical transformation zone type II is coded as 2, and cervical transformation zone type III is coded as 3, as a multi-category variable.

[0050] When screening risk factor variables, the following statistical analysis methods can be used:

[0051] To compare the variables across datasets, continuous variables were presented with median and quartile ranges, and nonparametric tests (Mann-Whitney U test) were used to compare differences in distribution. For categorical variables, including TCT results, HPV type, colposcopy impression, cervical transformation zone type, CIN1+, CIN2+, and CIN3+, percentages were used to represent distributions, and chi-square tests were used to compare differences in distribution. A p-value less than 0.05 was considered statistically significant. All analyses were performed using IBM SPSS Statistical 25 software, and p-values ​​< 0.05 were used as the significance criterion.

[0052] Ultimately, the variables included in the construction of the diagnostic model were age, TCT, HPV type, colposcopy impression, and cervical transformation zone type.

[0053] When modeling, the following machine learning algorithms can be selected:

[0054] a) Support Vector Machine (SVM) modeling was employed, exploring linear, radial basis function (RBF), and polynomial kernel functions. The performance of each kernel function was optimized by adjusting the penalty parameter. Ultimately, by comparing models built with different kernel functions, the linear kernel function was found to perform best in the diagnostic models for all grades of cervical lesions; therefore, it was adopted in the analysis. Specifically, the penalty parameter was 10 for the CIN2+ risk model and 40 for the CIN3+ risk model.

[0055] b) Logistic regression, modeled conventionally, without parameter adjustment.

[0056] c) Neural network: Adjust the number of hidden layers to achieve optimal performance. The CIN2+ risk model has 8 hidden layers, and the CIN3+ risk model has 9 hidden layers.

[0057] d) Random forest, with the number of generated trees set to 1000. The error rate plot was used to check and confirm that the number of generated trees was sufficient to make the model tend to stabilize.

[0058] e) Decision tree, modeled conventionally, without parameter adjustments.

[0059] The above algorithms can be filtered using the following methods:

[0060] Historical clinical data was divided into training datasets and validation datasets;

[0061] All diagnostic models were trained using the training dataset to obtain multiple candidate diagnostic models for predicting different levels of cervical precancerous lesions.

[0062] All candidate diagnostic models were validated using a validation dataset, and the best-performing diagnostic model for different levels of cervical precancerous lesions was selected.

[0063] Specifically, selection can be based on AUC. In this implementation, a logistic regression algorithm is ultimately chosen to construct the optimal diagnostic model for CIN2+ and CIN3+. Logistic regression is a statistical learning method used for binary classification problems. Its goal is to predict the probability of cervical precancerous lesion grade CIN2+ or CIN3+ by inputting variables including age, TCT, HPV type, colposcopy impression, and cervical transformation zone type. The model is fitted using the following formula: Calculate the probability of risk Where log-odds is the log-odds ratio, The probability of a person being infected (value range 0-1). For the intercept term, For the first The regression coefficients of each variable represent the degree of influence of that variable on risk. The input variables are used. The coefficients of each term in the formula can be fitted using the training dataset, and then the corresponding risk probabilities can be calculated based on the collected patient clinical data.

[0064] .

[0065] In some preferred embodiments, to better classify and diagnose precancerous lesions of the cervix, a CIN2+ preferred diagnostic model can be used to set low-risk and high-risk cutoff values. The low-risk cutoff value is used to screen out low-risk individuals for whom a biopsy is not required and a missed diagnosis is highly likely; therefore, the model's sensitivity and negative predictive value have high priority. The high-risk cutoff value is used to identify high-risk individuals for whom a biopsy is highly necessary; therefore, the model's positive predictive value has high priority. Specifically, the maximum log-odds ratio at which the model's sensitivity and NPV exceed their respective set thresholds can be set as the low-risk cutoff value, and the minimum log-odds ratio at which the model's PPV exceeds its set threshold can be set as the high-risk cutoff value.

[0066] In addition, nomograms can be generated, which include scores for all risk factor variables and the lesion probability predicted by the optimal diagnostic model for CIN2. Low-risk and high-risk cutoff values ​​are introduced to guide clinical decision-making. Besides nomograms, an integral method can be used to quickly assess a patient's CIN2+ risk. By calculating the integral scores corresponding to the patient's five key variables and adding these scores together, the patient's total risk score can be quickly obtained, and the patient's CIN2+ risk can be assessed.

[0067] Specifically, it can be calculated using the following methods:

[0068] Obtain the regression coefficients of each risk factor variable in the CIN2+ optimal diagnostic model;

[0069] Calculate the regression coefficient for each risk factor variable. Product of values To obtain the corresponding score;

[0070] The scores of all risk factor variables are summed to obtain the patient's total CIN2+ risk score.

[0071] This method is simple and convenient, requiring no complex calculation tools, and can be quickly applied in clinical practice. By comparing the calculated total score with the scores of the two cutoff values, clinicians can quickly determine the patient's risk level.

[0072] A preferred embodiment 1 of this implementation focuses on patients who underwent colposcopy at a hospital between December 2016 and December 2024.

[0073] The included data consisted of: basic patient information, clinical history, HPV test results, cervical cytology results, colposcopy diagnosis, cervical transformation zone type, and cervical biopsy pathology, totaling 17,201 cases. Further inclusion criteria were used: 1. Based on the 2019 ASCCP guidelines for cervical screening, patients meeting either of the following criteria were included in the study: ① HPV 16 / 18 infection regardless of cytology abnormalities; ② Abnormal cytology (≥ASC-US) regardless of HPV infection; 2. Patients underwent colposcopy and pathological biopsy at the hospital where the study was conducted; 3. Complete and valid HPV test results, cervical cytology results, colposcopy diagnosis, cervical transformation zone type, and biopsy pathology were available. The exclusion criteria were as follows: 1. History of cervical resection or hysterectomy; 2. Colposcopy performed during pregnancy; 3. Cervical cytology results of AGC, AIS, or suspected adenocarcinoma, or any other evidence of suspected adenocarcinoma; 4. History of other malignant tumors. A total of 6681 cases met the criteria and were included in subsequent studies.

[0074] 3471 data points were randomly divided into a training set (N=2429) and an internal validation set (N=1042) in a 7:3 ratio, and 3210 data points were used as the external validation set.

[0075] The collected information was compiled into an Excel spreadsheet, and the patient's information was coded as follows:

[0076] a) Patient age was a continuous variable and no special treatment was given;

[0077] b) Cervical cytology results of patients: NILM was coded as 0, ASC-US as 1, LSIL as 2, ASC-H as 3, while HSIL, carcinoma in situ and cervical squamous cell carcinoma were all classified as HSIL+ (High-Grade Squamous Intraepithelial Lesion and above) and coded as 4, as a multi-category variable;

[0078] c) Patient's HPV test results: To ensure the stability of the model, the study classified HRHPV types other than HPV16 as "other HPV". For the HPV status variable, the state of not being infected with HPV was 0, the state of being infected with HPV16 only was 1, the state of being infected with both HPV16 and "other HPV" was 2, and the state of being infected with "other HPV" only was 3, as a multi-category variable;

[0079] d) Patient's colposcopy impression: "No abnormal findings" is coded as 0, "LSIL findings" is coded as 1, "HSIL findings" is coded as 2, and "cervical cancer findings" is coded as 3, as a multi-category variable;

[0080] e) Cervical transformation zone type of the patient: Cervical transformation zone type I is coded as 1, cervical transformation zone type II is coded as 2, and cervical transformation zone type III is coded as 3, as a multi-category variable.

[0081] Based on the sample size calculation method for the prediction model, this study estimates that the required sample sizes for the CIN2+ and CIN3+ diagnostic models are 449 and 598, respectively. Meanwhile, the training set sample size is 2429.

[0082] The study compared various variables across datasets. Continuous variables were presented with median and quartile ranges, and nonparametric tests (Mann-Whitney U test) were used to compare differences in distribution. For categorical variables, including TCT results, HPV type, colposcopy impression, cervical transformation zone type, CIN1+, CIN2+, and CIN3+, percentages were used to represent distributions, and chi-square tests were used to compare differences in distribution. A p-value less than 0.05 was considered statistically significant.

[0083] All the above analyses were performed using IBM SPSS Statistic 25 software, and statistical results were considered to be significant with P < 0.05. Table 1 below shows the variables suitable for inclusion in the diagnostic model after screening for statistically significant risk factors for CIN2+ and CIN3+ lesions using LASSO regression, based on univariate and multivariate analyses.

[0084]

[0085] The model was constructed and the AUC value was calculated using RStudio (R4.3.0). Variables included in the diagnostic model construction were age, TCT, HPV type, colposcopy impression, and cervical transformation zone type.

[0086] The following machine learning methods were used to model CIN2+ and CIN3+ respectively:

[0087] a) During the support vector machine (SVM) modeling process, linear kernel functions, radial basis function (RBF) kernel functions, and polynomial kernel functions were tested, and the model performance of each kernel function was optimized by adjusting the penalty parameter. Ultimately, by comparing the models built with different kernel functions, the linear kernel function was found to perform best in the diagnostic models for all grades of cervical lesions; therefore, the linear kernel function was used in the analysis. Specifically, the penalty parameter was 10 for the CIN2+ risk model and 40 for the CIN3+ risk model.

[0088] b) Logistic regression was modeled conventionally without parameter adjustment.

[0089] c) Neural Networks: The number of hidden layers in the neural network needs to be adjusted to achieve optimal performance. The number of hidden layers in the CIN2+ risk model is 8, and the number of hidden layers in the CIN3+ risk model is 9.

[0090] d) The number of generating trees in the random forest was set to 1000. The error rate plot was used to check and confirm that the number of generating trees was sufficient to make the model tend to stabilize.

[0091] e) The decision tree was modeled conventionally without parameter adjustment.

[0092] This embodiment constructed and validated a diagnostic model for CIN2+ lesions. On the training set, Random Forest had the highest AUC (0.898), performing best; Logistic Regression and Neural Network both had an AUC of 0.847, showing comparable performance; Support Vector Machine (SVM) had an AUC of 0.830; and Decision Tree had the lowest AUC (0.793). On the internal validation set, Logistic Regression and Neural Network both had an AUC of 0.865, tied for first place; Random Forest had an AUC of 0.834; SVM had an AUC of 0.837; and Decision Tree had an AUC of 0.797. On the external validation set, Logistic Regression had the best AUC (0.840); Neural Network had an AUC of 0.838; SVM had an AUC of 0.821; Random Forest had an AUC of 0.822; and Decision Tree had the lowest AUC (0.780) (Table 2-1). Therefore, the optimal model for CIN2+ diagnosis is the Logistic Regression model.

[0093]

[0094] The same method was used to construct and validate a diagnostic model for CIN3+ lesions. On the training set, Random Forest showed the highest AUC (0.947), performing best; Neural Network had an AUC of 0.901, followed by Logistic Regression (0.893); Support Vector Machine (0.878); and Decision Tree had the lowest AUC (0.778). On the internal validation set, Logistic Regression performed best with an AUC of 0.898; Neural Network had an AUC of 0.883; Support Vector Machine (0.882); Random Forest had an AUC of 0.862; and Decision Tree had an AUC of 0.797. On the external validation set, Logistic Regression performed best with an AUC of 0.877; Random Forest also had an AUC of 0.877, tied with Logistic Regression; Neural Network had an AUC of 0.872; Support Vector Machine (0.843); and Decision Tree had the lowest AUC (0.780) (Table 2-2). Therefore, logistic regression is the optimal model algorithm for CIN3+ diagnosis.

[0095]

[0096] Based on the results of the model above, two cutoff values, low risk and high risk, were determined to construct a new diagnostic method. Both cutoff values ​​are derived from the risk score (in logistic regression) given by the diagnostic model for each case in the training set. It's worth it.

[0097] First, to accurately screen low-risk patients, based on a model sensitivity greater than 90% and NPV greater than 95%, and to minimize the possibility of missed diagnoses, the model's linear predictive value (i.e., The analysis revealed a low-risk cutoff value of -2.59. Based on the risk probability of CIN2+ output by the model, a PPV greater than 80% in the training set was used as the criterion for determining the cutoff value (high risk), and the final cutoff value (high risk) was determined to be 0.63.

[0098] Tables 3-1, 3-2, and 3-3 list the clinical implications of using a cutoff value (low risk) for cervical biopsy necessity screening, and compare this diagnostic method with biopsy methods recommended by the ASCCP guidelines. If biopsies are postponed for low-risk patients diagnosed by the diagnostic model, while biopsies are performed on other patients, the sensitivity for detecting CIN2+ lesions in the training set is 90.10%, specificity is 51.77%, NPV is 96.21%, PPV is 27.79%, and accuracy is 58.32%; in the internal validation set, the sensitivity is 90.75%, specificity is 52.36%, NPV is 96.60%, PPV is 27.50%, and accuracy is 58.73%; and in the external validation set, the sensitivity is 86.43%, specificity is 55.81%, NPV is 95.70%, PPV is 26.56%, and accuracy is 60.59%.

[0099]

[0100] This diagnostic method was compared with the method recommended by the ASCCP guidelines (biopsy for any of the following criteria: TCT HSIL+, HPV16 or 18 positive, and colposcopy showing high-grade lesions). In the training set, the logistic regression model outperformed the ASCCP recommended method in sensitivity, specificity, NPV, PPV, and accuracy. Similar results were observed in the internal validation set. In the external validation set, the logistic regression algorithm showed slightly lower sensitivity than the ASCCP recommended method (87.59% vs. 86.43%), but higher sensitivity on other metrics, and significantly higher specificity and accuracy. In conclusion, the new diagnostic method has potential in reducing the number of biopsies and increasing the detection rate of CIN2+ lesions.

[0101] To make the diagnostic model more intuitive and easier for clinical application, a nomogram was drawn, such as... Figure 1 As shown in the nomogram, each variable included in the model (such as TCT results, HPV type, colposcopy impression, etc.) has a corresponding score axis. Based on the patient's actual situation, the corresponding score can be found on each variable's axis, and these scores are summed to obtain the total score. Below the total score in the nomogram, the corresponding values ​​and the accurate risk probability of CIN2+ are arranged in parallel, facilitating quick reading and understanding of the results. Furthermore, two key cutoff values ​​are clearly marked on the scale: -2.59 corresponds to the low-risk cutoff value, and 0.63 corresponds to the high-risk cutoff value. These two cutoff values ​​are used to identify low-risk and high-risk patient groups, respectively, providing clear reference for clinical decision-making.

[0102] In addition to nomograms, the study also provides an integral method for quickly assessing a patient's CIN2+ risk, as shown in Table 3-4. By calculating the integral scores corresponding to the patient's five key variables and summing these scores, a patient's total risk score can be quickly obtained, thus assessing the patient's CIN2+ risk. This method is simple and convenient, requiring no complex calculation tools, and can be quickly applied in clinical practice. By comparing the total score with the scores at two cutoff values ​​(score for the low-risk cutoff: 39.766, score for the high-risk cutoff: 46.100), clinicians can quickly determine the patient's risk level.

[0103]

Claims

1. A machine learning-based diagnostic method for cervical precancerous lesions, characterized in that, include: Based on historical clinical data, statistical methods were used to analyze and obtain risk factor variables that were significantly associated with cervical precancerous lesions. Using the selected risk factor variables as input variables and the probability of cervical precancerous lesions as output variables, multiple diagnostic models based on different machine learning algorithms were constructed. Based on historical clinical data, the best diagnostic models for predicting different levels of cervical precancerous lesions were selected. Patient clinical data were collected and input into various optimized diagnostic models to predict the probability of different grades of cervical precancerous lesions, thus obtaining the grading prediction results of cervical precancerous lesions.

2. The method according to claim 1, characterized in that, Before the step of constructing multiple diagnostic models based on different machine learning algorithms using the selected risk factor variables as input variables and the probability of cervical precancerous lesions as output variables, the method further includes the step of encoding the classification results in the risk factor variables into corresponding numerical values ​​as the values ​​of the corresponding classification variables.

3. The method according to claim 2, characterized in that, The different grades of cervical precancerous lesions include CIN2+ and CIN3+.

4. The method according to claim 3, characterized in that, The preferred diagnostic models for CIN2+ and CIN3+ are all built based on the logistic regression algorithm.

5. The method according to claim 4, characterized in that, Also includes: Obtain sensitivity, NPV, and PPV data for the optimal CIN2+ diagnostic model; The maximum log-odds ratio that makes the model’s sensitivity and NPV greater than their respective set thresholds is set as the low-risk cutoff value, and the minimum log-odds ratio that makes the model’s PPV greater than its set threshold is set as the high-risk cutoff value. A nomogram was plotted, which included scores for all risk factor variables and the lesion probability predicted by the preferred diagnostic model for CIN2, and low-risk and high-risk cutoff values ​​were introduced to guide clinical decision-making.

6. The method according to claim 4, characterized in that, Following the step of obtaining the cervical precancerous lesion grading prediction result, the method further includes: Obtain the regression coefficients of each risk factor variable in the CIN2+ optimal diagnostic model; Calculate the product of the regression coefficient and the value for each risk factor variable to obtain the corresponding score; The scores of all risk factor variables are summed to obtain the patient's total CIN2+ risk score.

7. The method according to claim 1, characterized in that, The risk factor variables include age, TCT classification, HPV type, colposcopy impression, and cervical transformation zone type.

8. The method according to claim 1, characterized in that, Based on historical clinical data, the preferred diagnostic models with the best predictive effect for different levels of cervical precancerous lesions were selected, including: Historical clinical data was divided into training datasets and validation datasets; All diagnostic models were trained using the training dataset to obtain multiple candidate diagnostic models for predicting different levels of cervical precancerous lesions. All candidate diagnostic models were validated using a validation dataset, and the best-performing diagnostic model for different levels of cervical precancerous lesions was selected.

9. The method according to claim 1, characterized in that, The statistical analysis used to identify risk factors significantly associated with precancerous cervical lesions included: For continuous variables, their distributions are presented in terms of median and interquartile ranges, and nonparametric tests are used to compare differences in the distributions. For categorical variables, percentages are used to represent their distribution, and chi-square tests are used to compare differences in the distributions. Variables with P < 0.05 were selected as risk factors significantly associated with cervical precancerous lesions.

10. The method according to claim 1, characterized in that, The historical clinical data includes patient basic information questionnaires, clinical history, HPV test results, cervical cytology results, colposcopy diagnosis and cervical transformation zone type, as well as pathological results and pathological grades.

Images