A two-stage dynamic risk prediction system and method for cervical diseases based on multi-dimensional data

Abstract

The present application relates to the technical field of medical data processing, and particularly relates to a cervical disease two-stage dynamic risk prediction method based on multidimensional data, comprising the following steps: step S1, data acquisition and structuring: a plurality of dimensions of patient data are collected and structured by questionnaire, and a cervical disease database is constructed; step S2, first-stage model construction: for a general screening population, a first prediction model is constructed by using the questionnaire survey data collected in step S1; step S3, second-stage model construction: for a population with HPV positive or other suspicious indications obtained by the first prediction model, high-dimensional clinical data collected in step S1 are integrated, and a second prediction model is constructed by using an ensemble learning strategy; step S4, model series early warning output; and step S5, decision support generation. The present application integrates low-cost and easily-obtained multidimensional data, constructs a scientific risk stratification model, and realizes two-stage and continuous dynamic early warning.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, specifically to a two-stage dynamic risk prediction system and method for cervical diseases based on multi-dimensional data. Background Technology

[0002] As a malignant tumor that seriously threatens women's health, cervical cancer is a long-term and continuous process, which provides a crucial window for screening and intervention. Currently, the widely used cervical cancer screening strategy in clinical practice mainly relies on the combined screening of human papillomavirus (HPV) testing and liquid-based thin-layer cytology (TCT) testing. However, this strategy has certain limitations, primarily including the high cost of HPV testing, making it difficult to implement in resource-limited areas and large populations. Furthermore, HPV testing struggles to distinguish high-risk groups, exhibiting insufficient accuracy. While HPV infection is widespread, only a small percentage progress to precancerous lesions or cancer, making it difficult to accurately differentiate between transient infections and persistent infections that are about to progress—leading to unnecessary colposcopy referrals, psychological burdens, and wasted medical resources.

[0003] Furthermore, traditional screening results (positive / negative) are static, failing to fully utilize multi-dimensional information such as the patient's pregnancy history and behavioral habits for individualized and dynamic risk stratification assessment. They also lack continuity, resulting in a lack of a continuous decision support chain based on quantitative risk between screening and subsequent risk management and intervention recommendations. To address these issues, there is an urgent need for a low-cost, high-precision, multi-source information integration capability for a two-stage dynamic risk prediction system and method for cervical diseases. This would enable dynamic risk stratification-assisted assessment and facilitate earlier intervention and resource optimization in cervical disease prevention and control. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology by providing a two-stage dynamic risk prediction system and method for cervical diseases based on multi-dimensional data. By integrating multi-dimensional clinical data and using machine learning models, the system and method realize a two-stage, dynamic, and graded early warning from HPV infection risk to high-grade lesion (HSIL) risk, and provide personalized services and management in the form of a mobile application (APP) based on a regional network.

[0005] The technical problem to be solved by this invention is achieved through the following technical solution: This invention discloses a two-stage dynamic risk prediction method for cervical diseases based on multi-dimensional data, comprising the following steps: Step S1, Data Collection and Structuring: Collect and structure patient data from multiple dimensions through questionnaires to build a cervical disease database; Step S2, First-stage model construction: For the general screening population, using the questionnaire survey data collected in step S1, a first prediction model is constructed to predict the risk of HPV infection. Step S3, Second-stage model construction: For the HPV-positive or other suspicious individuals obtained by the first prediction model, integrate the high-dimensional clinical data collected in step S1 and use an ensemble learning strategy to construct a second prediction model for HSIL lesion risk assessment. Step S4, Model-connected early warning output: Connect the first prediction model in step S2 and the second prediction model in step S3 to form a dynamic two-stage data early warning chain from HPV infection risk prediction to HSIL lesion risk assessment, and output the risk level based on the model. Step S5, Generate Decision Support: Based on the risk level obtained in Step S4, automatically generate a visual risk assessment report and personalized clinical decision recommendations. The decision recommendations include recommending appropriate screening intervals for low-risk individuals and recommending further examinations for high-risk individuals.

[0006] The technical problem to be solved by the present invention can also be further achieved through the following technical solution: In step S1, the patient’s diagnosis and treatment data throughout the entire disease cycle, from community vaccination, screening, treatment to follow-up, are collected. The data dimensions include demographic characteristics, behavioral history characteristics, obstetric history characteristics, and laboratory test result characteristics.

[0007] The technical problem to be solved by the present invention can also be further achieved through the following technical solutions, wherein the method of constructing the first prediction model in step S2 includes the following steps: Step S21: Preprocess the questionnaire data. The preprocessing process includes label coding and one-hot coding. Step S22: The LightGBM algorithm is used to train and fit the preprocessed dataset to obtain an HPV infection risk prediction model. The output of the model is set as the positive probability of an individual being infected with HPV. Step S23: Perform multi-center validation on the first prediction model. The validation performance indicators include: area under the receiver operating characteristic curve not less than 0.92, sensitivity not less than 90%, and specificity not less than 96%.

[0008] The technical problem to be solved by the present invention can also be further achieved by the following technical solution: In step S22, the preset threshold division method output by the HPV infection risk prediction model is to set the preset threshold to 50% and divide the risk according to the preset threshold. When the probability is less than 50%, it is considered low risk and when the probability is not less than 50%, it is considered high risk.

[0009] The technical problem to be solved by the present invention can also be further achieved through the following technical solutions: the high-dimensional clinical data in step S3 includes HPV genotyping results, TCT cytology results, demographic characteristics, behavioral history characteristics, and obstetric history characteristics.

[0010] The technical problem to be solved by the present invention can also be further achieved through the following technical solution: the method of constructing the second prediction model in step S3 includes the following steps. Step S31: Using the Stacking fusion framework and LightGBM as the base model, combined with logistic regression and / or neural network models as meta-models for fusion training, a HSIL lesion risk prediction model is obtained. Step S32: Set the output of the second prediction model to the individual's risk percentage of HSIL, and clinically validate the second prediction model. The validation performance indicators include AUC not less than 0.83, sensitivity not less than 85%, and specificity not less than 82%.

[0011] This invention also discloses a two-stage dynamic risk prediction system for cervical diseases based on multi-dimensional data, comprising the following functional modules: a data acquisition module for collecting and standardizing multi-dimensional patient data through electronic questionnaires and hospital information system interfaces; a data processing and storage module for cleaning, encoding, and structuring the collected data and storing it in a cervical disease database; a model calculation engine module with a built-in first and second prediction model for receiving processed data and performing risk prediction calculations; a risk assessment and output interface for generating phased risk levels and forming easily understandable graphic reports based on the model calculation results; and a clinical decision support interface for pushing risk assessment results and suggestions to doctors' workstations or user terminal apps to assist in developing subsequent screening and / or treatment plans.

[0012] This invention also discloses a method for deploying the two-stage dynamic risk prediction model on a regional network and providing personalized services via mobile devices, comprising the following steps: Step A1, Regionalized Model Deployment and Service Packaging: Standardize and package the fully trained first and second prediction models, and deploy them in the risk assessment engine of a regional medical cloud platform or central server; Step A2, Mobile Application Interaction and Data Collection: Provide users with a mobile terminal APP, whose interface includes structured input forms corresponding to multi-dimensional parameters. The form design adopts guided questionnaires, drop-down selections, standardized options, etc.; Step A3, Secure Data Transmission and Remote Assessment: After the user submits input data through the APP, the APP client sends the anonymized data packets to the risk assessment engine API in the regional network through an encrypted transmission protocol. After receiving the data, the engine automatically calls the corresponding prediction model for real-time calculation; Step A4, Personalized Result Feedback and Health Management: The risk assessment engine sends the calculated risk level and corresponding clinical significance interpretation to the APP. The APP presents the assessment results to the user in a visual and easily understandable form for non-medical professionals. Based on the user's risk level, the APP intelligently matches and pushes personalized cervical health science popularization guidance and action suggestions from the preset knowledge base; Step A5, Regional Data Aggregation and Model Iteration: The regional platform aggregates and analyzes the anonymized assessment requests and result data to monitor the regional cervical disease risk distribution and regularly retrains and updates the deployed prediction model.

[0013] The technical problem to be solved by the present invention can also be further achieved through the following technical solution: the process of automatically calling the corresponding prediction model for real-time calculation in step A3 includes calling the first prediction model based on the questionnaire survey data to determine the HPV infection risk level; if the user inputs laboratory indicators such as HPV / TCT, or is determined to be high risk by the first prediction model, then automatically and / or after user confirmation, further calling the second prediction model to calculate the risk of HSIL lesions.

[0014] Compared with the prior art, the present invention has the following advantages: (1) This invention can integrate low-cost, easily accessible multi-dimensional data and use machine learning algorithms to construct a scientific risk stratification model to achieve a two-stage, continuous, dynamic early warning of "preliminary prediction of HPV infection risk" and "advanced assessment of HSIL lesion risk". The first prediction model serves as a low-cost, non-invasive alternative to large-scale screening in economically underdeveloped areas, while the second prediction model can significantly reduce unnecessary colposcopy referrals (verified to reduce by about 38%). Through a dynamic and continuous auxiliary decision-making process, a continuous assessment chain is constructed. Through visualized results and clear referral recommendations, it provides doctors with a powerful decision support tool and improves the systematicness and timeliness of screening. (2) This invention innovatively integrates traditional laboratory data (HPV, TCT) with low-cost, easily accessible questionnaire surveys and routine test data (such as blood routine tests), achieving high predictive performance. In particular, it has discovered and verified the important value of non-traditional indicators such as "education level", "number of pregnancies" and "age of first sexual intercourse" in data model prediction. (3) This invention extends professional medical risk prediction capabilities to medical institutions and individual users at all levels in a low-cost and high-efficiency manner through the architecture of “cloud model + terminal APP”. Users can obtain personalized risk assessment and scientific guidance based on authoritative models anytime and anywhere without special equipment or complicated operations, which greatly improves the accessibility and inclusiveness of high-quality screening resources. It is particularly conducive to the promotion and application in grassroots and resource-limited areas, and to the implementation of disease prevention and control at the forefront. Attached Figure Description

[0015] Figure 1 This is a flowchart of the two-stage dynamic risk prediction method for cervical diseases provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the ROC curve of the first stage (HPV infection risk prediction model) provided in the embodiment of the present invention, with an AUC value of 0.9266; Figure 3 This is a schematic diagram of the ROC curve of the second stage (HSIL lesion risk prediction model) provided in the embodiment of the present invention, with an AUC value of 0.8375; Figure 4 This is a schematic diagram showing the importance ranking of the main predictive factors for HSIL lesions provided in the embodiments of the present invention; Figure 5 This is a structural block diagram of the two-stage dynamic risk prediction system for cervical diseases provided in this embodiment of the invention; Figure 6 This is a schematic diagram of the user interface of a mobile application (APP) provided in an embodiment of the present invention. Detailed Implementation

[0016] Firstly, this invention provides a two-stage dynamic risk prediction method for cervical diseases, such as... Figure 1-4 As shown, it specifically includes the following steps: S1: Data Acquisition and Structuring: Construct a cervical disease database, collect and structure patient data across multiple dimensions. These data dimensions include, but are not limited to: demographic characteristics (e.g., age, education level), behavioral history (e.g., smoking history, age at first sexual intercourse), obstetric history (e.g., number of pregnancies), laboratory test results (e.g., HPV genotyping, TCT results), and blood routine-derived indicators (e.g., neutrophil percentage). Preferably, a total of 23 parameters are collected.

[0017] S2: Phase One Model Building and HPV Infection Risk Prediction: First, for the general screening population, the first prediction model is constructed using the questionnaire survey data collected in step S1 (such as 23 indicators including personal behavior, health status, and demographic characteristics).

[0018] The questionnaire data is then preprocessed, including label coding and one-hot coding. The LightGBM algorithm is used to train and fit the preprocessed dataset to obtain an HPV infection risk prediction model. The model outputs the positive probability of an individual being infected with HPV, and is divided into low risk (probability < 50%) and high risk (probability ≥ 50%) based on a preset threshold (e.g., 50%).

[0019] Finally, the first prediction model was validated in multiple centers, and its performance indicators were: AUC (area under receiver operating characteristic) not less than 0.92, sensitivity not less than 90%, and specificity not less than 96%.

[0020] S3: Phase Two Model Construction and HSIL Lesion Risk Assessment First, for individuals who are HPV positive or have other suspicious indications, higher-dimensional clinical data collected in step S1 can be integrated, including at least HPV genotyping results, TCT cytology results, and key questionnaire data.

[0021] Then, an ensemble learning strategy is used to construct a second prediction model. Preferably, a Stacking fusion framework is used, with LightGBM as the base model and logistic regression and / or neural network models as meta-models for fusion training, to obtain the HSIL lesion risk prediction model.

[0022] Finally, the second predictive model outputs the individual's percentage risk of developing HSIL. Clinically validated, its performance metrics are: AUC not less than 0.83, sensitivity not less than 85%, and specificity not less than 82%.

[0023] S4: Two-stage dynamic early warning and decision support: First, the first and second prediction models are connected in series to form a dynamic early warning chain of "HPV infection risk prediction → HSIL lesion risk assessment".

[0024] Then, based on the risk level output by the model, a visualized risk assessment report and personalized clinical decision recommendations are generated. These recommendations include, but are not limited to: recommending appropriate screening intervals for low-risk individuals and suggesting further diagnostic procedures such as colposcopy for high-risk individuals, thereby optimizing referral pathways and reducing unnecessary examinations.

[0025] Example 1: The two-stage dynamic risk prediction method for cervical disease of the present invention can be implemented in the following manner: 1. In relevant cooperating hospitals, 23 structured data items of female patients are prospectively collected through electronic questionnaires and hospital information systems, including age, education level, smoking history, age at first sexual intercourse, number of pregnancies, HPV genotyping (types 16 / 18 and the other 12 types), TCT results (ASC-US, LSIL, HSIL, etc.), and blood routine indicators. All data are stored in a centralized database after desensitization processing.

[0026] 2. First-Stage Model Application: For the general population participating in screening for the first time, the system invokes the first-stage HPV infection risk prediction model. The model calculates the HPV positivity probability value for the patient based on the 23 input questionnaire indicators. For example, an output probability of 65% indicates a "high risk of HPV infection." The system also provides visualization charts (such as...). Figure 2 The ROC curve (shown in the figure) demonstrates the overall discriminative ability of the model.

[0027] 3. Second-stage model application: For patients identified as high-risk in the first stage, or those with positive HPV test results, the system automatically recommends proceeding to the second-stage assessment. At this stage, the model integrates the patient's specific HPV genotype, detailed TCT results, and key questionnaire data (such as number of pregnancies), using a Stacking fusion model to calculate and output the percentage risk of developing HSIL. For example, the output risk value might be 72%. Figure 4 This demonstrates the significant contribution of various features (such as HPV16 / 18 positivity, number of pregnancies, etc.) to the prediction results in this stage of the model.

[0028] 4. Early Warning and Decision Support: The system summarizes the results from both stages and generates a comprehensive report. The report clearly states: "This patient has a high risk of HPV infection, and the risk of progression to HSIL is approximately 72%. Colposcopy is recommended as a priority examination." For low-risk patients, the report may state: "Current risk is low; routine follow-up is recommended in 12 months." This report is pushed to doctors' workstations in various hospitals via API and synchronized to the patient's mobile app.

[0029] Secondly, this invention provides a two-stage dynamic risk prediction system for cervical diseases that implements the above-mentioned method, comprising: a data acquisition module: used to collect and standardize multi-dimensional patient data through intelligent pre-registration consultation, electronic inpatient case registration, and AI-driven follow-up, as well as through hospital information system interfaces. A data processing and storage module: used to clean, encode, and structure the collected data, and store it in a cervical disease database. A model calculation engine: built with the first prediction model and the second prediction model, used to receive the processed data and perform risk prediction calculations. A risk assessment and output module: used to generate phased risk levels (such as high / low HPV infection risk, HSIL risk percentage) based on the model calculation results, and form an easy-to-understand graphic report. A clinical decision support interface: used to push the risk assessment results and suggestions to the doctor's workstation or user terminal APP to assist in the formulation of subsequent screening or treatment plans.

[0030] Example 2: The system implementing the above method includes: a backend service deployed on a cloud server or local server, comprising a data acquisition and processing module, and a model computation engine (loaded with pre-trained LightGBM and Stacking model files); a doctor-side web workstation or hospital HIS integration interface, serving as a clinical decision support interface; and a patient-side mobile app, serving as a terminal for data acquisition and result query. All modules communicate via an encrypted data transmission protocol to ensure patient privacy and data security. Figure 5 As shown.

[0031] Thirdly, the present invention also provides a method for deploying the two-stage dynamic risk prediction model on a regional network and providing personalized services through a mobile terminal application (APP), comprising the following steps: T1: Regionalized model deployment and service encapsulation: T11: Standardize and encapsulate the fully trained first prediction model (HPV infection risk prediction model) and second prediction model (HSIL lesion risk prediction model), and deploy them in the risk assessment engine of a regional medical cloud platform or central server.

[0032] T12: Create a unified application programming interface (API) for the risk assessment engine, which can be called by authorized client applications within the network.

[0033] T2: Mobile Application (APP) Interaction and Data Collection: T21: Provide users with a mobile terminal APP, whose interface includes structured input forms corresponding to the 23 multi-dimensional parameters. The form design adopts guided questionnaires, drop-down selections, standardized options, etc., to ensure that users can conveniently and accurately input personal information, behavioral history, and known simple test results (such as HPV genotyping and TCT result classification known to the user).

[0034] T22: The APP provides data verification and prompt functions to perform preliminary verification of the format and logic of the input data.

[0035] T3: Secure Transmission and Remote Assessment T31: After the user submits input data through the APP, the APP client sends the anonymized data packet to the risk assessment engine API in the regional network through an encrypted transmission protocol (such as HTTPS).

[0036] T32: After receiving the data, the risk assessment engine automatically calls the corresponding prediction model for real-time calculation. First, based on the questionnaire data, the first prediction model is called to determine the HPV infection risk level; if the user inputs laboratory indicators such as HPV / TCT, or the first model determines it to be high risk, then automatically or after user confirmation, the second prediction model is further called to calculate the HSIL lesion risk.

[0037] T4: Personalized Outcome Feedback and Health Management T41: The risk assessment engine will send the calculated risk level (such as "low risk of HPV infection" or "HSIL risk percentage: XX%) and its corresponding clinical significance interpretation to the APP.

[0038] T42: The app presents assessment results to users in a visual and easily understandable format for non-medical professionals (such as colored risk bars, emojis, and clear text conclusions).

[0039] T43: Crucially, based on the user's specific risk level, the app intelligently matches and pushes personalized cervical health education guidance and action suggestions from a pre-set knowledge base. These suggestions include, but are not limited to: for low-risk users: suggestions on routine screening intervals and guidance on healthy lifestyles; For high-risk users: It is strongly recommended to go to the hospital for further diagnostic examinations (such as those clearly marked "colposcopy recommended"), and provide consultation guidelines, precautions, and popular science articles or videos on related diseases.

[0040] Based on the risk stage (HPV infection stage or lesion stage), provide differentiated follow-up monitoring frequency recommendations.

[0041] T5: Regional Data Aggregation and Model Iteration: T51: On the premise of strictly following the privacy protection and desensitization norms, the regional platform can conduct aggregated analysis on anonymized assessment requests and result data for monitoring the risk distribution of cervical diseases in the region.

[0042] T52: Based on the continuously accumulating new data, the deployed prediction model is retrained and version-updated regularly, and dynamically upgraded through the network to ensure the timeliness and prediction accuracy of the model.

[0043] Example 3: For example, in a primary health service center, a doctor guides a screening female to use the dedicated APP provided by this center. Under the guidance of the APP, the female completed the filling of a 23-item questionnaire on personal circumstances and behavioral history in about less than 2 minutes, and entered the information of "HPV type 52 positive" and "TCT result is ASC-US" obtained from recent tests in the community hospital, as Figure 6 shown.

[0044] After submission, the data is encrypted and uploaded to the cloud server of the provincial cervical disease prevention and control platform. The risk assessment engine in the server successively calls the first and second prediction models for calculation. After a few seconds, the APP receives and displays the result: "Your HPV infection risk is [High risk]. According to your HPV typing and cytology results, the risk of progression to high-grade lesions (HSIL) is approximately 65% (belonging to medium-high risk). It is strongly recommended that you: go to the gynecology department of a regular hospital for a colposcopy examination within 3 months to confirm the diagnosis. At the same time, please view the 'What to Do After HPV Positive' science popularization guide prepared for you." Through the result and science popularization guidance, the female clearly understood her own risks and subsequent steps, and then made an appointment for an outpatient visit at a superior hospital through the navigation function in the APP. The platform background anonymously records this assessment for model optimization. The specific parameter combinations (which may include 23 items), coding methods, specific hyperparameter configurations of the fusion of LightGBM and Stacking, risk threshold settings, etc. used in the model have all been verified as relatively optimal examples through a large number of experiments, but they are not the only limitations of the present invention. Based on the same inventive concept, the adjustments and optimizations made to the data dimensions and algorithm parameters should all fall within the protection scope of the present invention.

[0045] A multi-dimensional risk prediction model constructs a two-stage dynamic early warning system, achieving accurate prediction of cervical cancer risk by integrating multiple key data parameters. This system incorporates 23 parameters, including obstetric history, smoking history, sexual history, HPV genotyping, and TCT results. The LightGBM (Light Gradient Boosting Machine) algorithm is used to analyze and process this data, training a scientifically sound risk stratification model that can construct a two-stage dynamic early warning chain based on the individual characteristics of different patients: "HPV infection risk prediction → HSIL lesion risk assessment".

[0046] In terms of clinical translation, this two-stage dynamic early warning system has played a crucial role. For the general population, based on resources from a cervical disease database, a machine learning-based HPV infection risk prediction model was developed, aiming to provide a low-cost, non-invasive alternative for cervical cancer screening. Traditional HPV testing is limited in its widespread application due to its relatively high cost and low coverage. This model collects 23 indicators, including personal behavior, health status, and demographic characteristics, through a questionnaire survey. The LightGBM algorithm is used to label and encode these data, followed by one-hot encoding. The dataset is then fitted to train a scientifically sound risk stratification model, outputting the HPV positive probability (low: <50%; high: ≥50%). Multicenter validation shows that the model's overall AUC is 0.9266, sensitivity is 90.24%, and specificity is 96.31%. The study found that education level (used for 14,025 node splits), number of pregnancies (used for 12,820 node splits), and age at first sexual intercourse (used for 9,044 node splits) were the main predictors of HSIL HPV positivity, and the model was better at predicting HPV infection of types 16 and 18 (AUC=0.9368).

[0047] Compared to traditional testing, this model offers cost-free assessment and provides users with easily understandable results through visual analysis. Since the model cannot replace laboratory testing for definitive diagnosis, data needs to be updated regularly to adapt to epidemiological changes. A HSIL lesion prediction model has been developed for individuals with suspected lesions. Addressing the issues of high false-negative rates, excessive referrals, and resource waste associated with traditional HPV+TCT screening for HSIL, a critical stage in cervical cancer, this invention provides an HSIL lesion risk prediction model. An AI-powered intelligent prediction model has emerged, innovatively integrating HPV genotyping, TCT cytology, and questionnaire data. It employs a LightGBM algorithm architecture combining Stacking with logistic regression and neural networks to output the HSIL risk percentage. Validated by a study of 2396 cases, the model achieved an AUC of 0.8375, with sensitivity and specificity of 85.1% and 82.3%, respectively. Incorporating high-weight features such as "HPV16 / 18 positive," "positive for the remaining 12 HPV types," "TCT results," and "number of pregnancies" can reduce unnecessary colposcopy examinations by 38%. Despite its reliance on standardized data and the need for regular updates, this model makes significant contributions to the accurate screening and management of precancerous cervical lesions. It provides physicians with valuable referral recommendations, improving diagnostic accuracy and treatment timeliness. By using algorithms to drive early intervention in clinical care, it identifies high-risk individuals at the stage of HPV infection or low-grade lesions, intelligently recommends screening frequencies and intervention points, and shifts the focus of prevention and control from "post-treatment" to "pre-lesion."

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A two-stage dynamic risk prediction method for cervical diseases based on multi-dimensional data, characterized in that: Includes the following steps, Step S1, Data Collection and Structuring: Collect and structure patient data from multiple dimensions through questionnaires to build a cervical disease database; Step S2, First-stage model construction: For the general screening population, using the questionnaire survey data collected in step S1, a first prediction model is constructed to predict the risk of HPV infection. Step S3, Second-stage model construction: For the HPV-positive or other suspicious individuals obtained by the first prediction model, integrate the high-dimensional clinical data collected in step S1 and use an ensemble learning strategy to construct a second prediction model for HSIL lesion risk assessment. Step S4, Model-connected early warning output: Connect the first prediction model in step S2 and the second prediction model in step S3 to form a dynamic two-stage data early warning chain from HPV infection risk prediction to HSIL lesion risk assessment, and output the risk level based on the model. Step S5, Generate Decision Support: Based on the risk level obtained in Step S4, automatically generate a visual risk assessment report and personalized clinical decision recommendations. The decision recommendations include recommending appropriate screening intervals for low-risk individuals and recommending further examinations for high-risk individuals.

2. The method for predicting the two-stage dynamic risk of cervical diseases based on multi-dimensional data according to claim 1, characterized in that: Step S1 includes collecting the patient's diagnosis and treatment data throughout the entire disease cycle, from community vaccination, screening, treatment to follow-up. The data dimensions include demographic characteristics, behavioral history characteristics, obstetric history characteristics, and laboratory test result characteristics.

3. The method for predicting the two-stage dynamic risk of cervical diseases based on multi-dimensional data according to claim 2, characterized in that: The method for constructing the first prediction model in step S2 includes the following steps: Step S21: Preprocess the questionnaire data. The preprocessing process includes label coding and one-hot coding. Step S22: The LightGBM algorithm is used to train and fit the preprocessed dataset to obtain an HPV infection risk prediction model. The output of the model is set as the positive probability of an individual being infected with HPV. Step S23: Perform multi-center validation on the first prediction model. The validation performance indicators include: area under the receiver operating characteristic curve not less than 0.92, sensitivity not less than 90%, and specificity not less than 96%.

4. The method for predicting the two-stage dynamic risk of cervical diseases based on multi-dimensional data according to claim 3, characterized in that: In step S22, the preset threshold division method output by the HPV infection risk prediction model is to set the preset threshold to 50% and divide the risk according to the preset threshold. When the probability is less than 50%, it is considered low risk and when the probability is not less than 50%, it is considered high risk.

5. The method for predicting the two-stage dynamic risk of cervical diseases based on multi-dimensional data according to claim 1, characterized in that: The high-dimensional clinical data in step S3 includes HPV genotyping results, TCT cytology results, demographic characteristics, behavioral history characteristics, and obstetric history characteristics.

6. The method for predicting the two-stage dynamic risk of cervical diseases based on multi-dimensional data according to claim 1, characterized in that: The method for constructing the second prediction model in step S3 includes the following steps. Step S31: Using the Stacking fusion framework and LightGBM as the base model, combined with logistic regression and / or neural network models as meta-models for fusion training, a HSIL lesion risk prediction model is obtained. Step S32: Set the output of the second prediction model to the individual's risk percentage of HSIL, and clinically validate the second prediction model. The validation performance indicators include AUC not less than 0.83, sensitivity not less than 85%, and specificity not less than 82%.

7. A two-stage dynamic risk prediction system for cervical diseases based on multi-dimensional data, characterized in that: Includes the following functional modules, Data acquisition module: used to collect and standardize multi-dimensional patient data through smart consultation before registration, electronic inpatient case registration and follow-up driven by both human and AI, and hospital information system interface. Data processing and storage module: used to clean, encode and structure the collected data and store it in the cervical disease database; Model calculation engine module: It has a built-in first prediction model and a second prediction model, which are used to receive processed data and perform risk prediction calculations; Risk assessment and output interface: used to generate phased risk levels and form easy-to-understand graphic reports based on the model calculation results; Clinical decision support interface: used to push risk assessment results and recommendations to doctors' workstations or user terminal apps to assist in the development of subsequent screening and / or treatment plans.

8. A method for deploying the two-stage dynamic risk prediction model on a regional network and providing personalized services via mobile terminals, characterized in that: Includes the following steps, Step A1, Regionalized Model Deployment and Service Packaging: Standardize and package the trained first and second prediction models, and deploy them in the risk assessment engine of the regional medical cloud platform or central server. Step A2, Mobile Application Interaction and Data Collection: Provide users with a mobile terminal APP, whose interface includes structured input forms corresponding to multi-dimensional parameters. The form design adopts guided questionnaires, drop-down selections, standardized options, and other methods. Step A3, Secure Data Transmission and Remote Assessment: After the user submits input data through the APP, the APP client sends the anonymized data packet to the risk assessment engine API in the regional network through an encrypted transmission protocol. After receiving the data, the risk assessment engine automatically calls the corresponding prediction model to perform real-time calculation. Step A4, Personalized Result Feedback and Health Management: The risk assessment engine sends the calculated risk level and corresponding clinical significance interpretation to the APP. The APP presents the assessment results to the user in a visual and easily understandable form for non-medical professionals. Based on the user's risk level, the APP intelligently matches and pushes personalized cervical health science popularization guidance and action suggestions from the preset knowledge base. Step A5, Regional Data Aggregation and Model Iteration: The regional platform aggregates and analyzes anonymized assessment requests and results data to monitor the regional cervical disease risk distribution and regularly retrains and updates the deployed prediction models.

9. A two-stage dynamic risk prediction system for cervical diseases based on multi-dimensional data as described in claim 8, characterized in that: The process of automatically calling the corresponding prediction model for real-time calculation in step A3 includes calling the first prediction model based on the questionnaire survey data to determine the HPV infection risk level; if the user inputs HPV / TCT laboratory indicators or is determined to be high risk by the first prediction model, then automatically and / or after user confirmation, further calling the second prediction model to calculate the risk of HSIL lesions.

Images