Machine learning-based intermediate trophoblastic tumor prediction model and application

Abstract

The application relates to a machine learning-based intermediate trophoblastic tumor prediction model and application, which has the functions of online input, automatic calculation, real-time output of survival probability and dynamic curve, can provide quantitative basis for doctors, assist in formulating treatment strategies, and makes up for the deficiencies of the prior art in disease specificity, interpretability and clinical practicability. The model comprises the following modules: a data input module, wherein a user inputs basic parameters of a patient through a webpage interface; an index calculation module, which calculates inflammatory biomarkers; a risk prediction module, which adopts a pre-trained RSF model to perform prediction after standardizing the input data, and performs high / low risk stratification based on a risk threshold determined in a training set; a model explanation module, which displays a summary plot of model variables through SHAP value visualization technology; and a result display and interaction module, which is based on a webpage front end built on a Shiny platform, and generates a prediction probability value and an individual dynamic survival curve.

Description

Technical Field

[0001] This invention relates to the technical field of medical artificial intelligence and tumor clinical decision support, and in particular to a machine learning-based intermediate trophoblastic tumor prediction model, and the application of this machine learning-based intermediate trophoblastic tumor prediction model in survival prediction and risk stratification of patients with intermediate trophoblastic tumors. Background Technology

[0002] Intermediate trophoblastic tumors (ITTs) are a rare group of malignant tumors originating from placental trophoblasts, including placental site trophoblastic tumors (PSTTs) and epithelioid trophoblastic tumors (ETTs). These diseases are characterized by low hCG secretion, insensitivity to chemotherapy, and significant heterogeneity in biological behavior. Traditional FIGO scoring systems and risk stratification methods are primarily applied to high-β-hCG tumors such as choriocarcinoma, and are inaccurate for prognostic assessment of ITT patients, easily leading to underestimation of risk or inadequate treatment. Currently, clinical practice mainly relies on single factors such as FIGO staging and gestational interval to assess risk, lacking standardized, multi-parameter integrated quantitative risk assessment tools.

[0003] Existing literature on personalized prognostic assessment of tumors largely focuses on common cancer types, such as breast cancer and lung cancer. A few researchers have attempted to apply machine learning methods to the construction of survival models for all trophoblastic tumors, for example, using algorithms such as Cox regression and random forests to process data from public databases like SEER for risk prediction. However, at present, there is no integrated, dedicated risk prediction device for intermediate-sized trophoblastic tumors.

[0004] Most existing techniques for prognostic assessment of trophoblastic tumors are based on general statistical models, such as Cox regression or clinically empirical scoring systems. These methods have the following major drawbacks in patients with intermediate trophoblastic tumors (ITT):

[0005] 1. Lack of disease-specific models: Existing models are mostly based on data from high-hCG-secreting choriocarcinoma, which are not applicable to PSTT / ETT patients with low hCG, insidious disease course, and high heterogeneity. This results in insufficient accuracy in predicting ITT risk and may delay the identification and treatment of high-risk patients.

[0006] 2. Limited variable selection dimensions and failure to integrate immune inflammatory indicators: Traditional models rely on a small number of clinical parameters and do not introduce blood inflammatory indicators that have been proven to have prognostic value, such as the neutrophil / lymphocyte ratio (NLR) and the systemic inflammatory response index (SIRI), which limits the integrity and recognition ability of the model.

[0007] 3. Limited prediction methods and inability to quantify individual risk: Most existing methods only provide a rough classification of high / low risk, lack precise survival probability prediction, and cannot dynamically generate individual survival curves that change over time, making it difficult to provide specific quantitative basis for individualized treatment.

[0008] 4. The models are not interpretable and are difficult for clinicians to understand and accept: Even when using machine learning methods to build models, existing studies often do not provide a transparent explanation mechanism. Clinicians find it difficult to understand the basis of the "black box" predictions, which affects their adoption and trust in actual clinical decision-making.

[0009] 5. Lack of practical clinical tools and difficulty in translation: Most models remain in the research stage and have not been developed into online tools with input interfaces, automatic calculation and visualization output functions, which makes it impossible for them to be widely used in clinical practice. Summary of the Invention

[0010] To overcome the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide an intermediate trophoblastic tumor prediction model based on machine learning, which has functions such as online input, automatic calculation, and real-time output of survival probability and dynamic curves. It can provide doctors with quantitative evidence, assist in the formulation of treatment strategies, and make up for the deficiencies of existing technologies in terms of disease specificity, interpretability and clinical applicability.

[0011] The technical solution of this invention is: this machine learning-based intermediate trophoblastic tumor prediction model, which includes:

[0012] The data input module allows users to input the patient's basic parameters through a web interface. These parameters include the International Federation of Gynecologic and Obstetric Societies FIGO tumor stage, previous pregnancy interval IFAP, Ki-67 index, and cell count, which consists of peripheral blood neutrophils, lymphocytes, or monocytes.

[0013] The indicator calculation module calculates inflammatory biomarkers, including the neutrophil-to-lymphocyte ratio (NLR) and the systemic inflammatory response index (SIRI).

[0014] The risk prediction module uses a pre-trained RSF model to standardize the input data and make predictions, outputting the probability of patients progressing to PFS within 1 to 5 years, and performing high / low risk stratification based on the risk threshold determined in the training set.

[0015] Model Interpretation Module: Visualizes a summary graph of model variables using SHAP value visualization technology;

[0016] The results display and interaction module is a web front-end built on the Shiny platform, which generates predicted probability values ​​and individual dynamic survival curves.

[0017] This invention is specifically designed for intermediate trophoblastic tumors (ITT). It combines FIGO staging, Ki-67 index, gestational interval IFAP, neutrophil-lymphocyte ratio (NLR), and the systemic immune inflammation index (SIRI) as key indicators to construct a disease-specific random survival forest (RSF) prediction model, significantly improving the accuracy of individualized progression risk prediction. The device integrates the interpretable algorithm SHAP, intuitively presenting the contribution of each variable to risk prediction, solving the problem of existing "black box" models being clinically uninterpretable. Real-time input, automatic calculation, and dynamic survival curve output are achieved through a web platform, enhancing physicians' ability to identify risks in ITT patients and improving the efficiency of treatment decisions.

[0018] It also provides an application of a machine learning-based intermediate trophoblastic tumor prediction model, which is used in the assessment of progression risk and prediction of individualized survival probability for patients with intermediate trophoblastic tumors. Attached Figure Description

[0019] Figure 1 The diagram shown is a schematic representation of a machine learning-based intermediate trophoblastic tumor prediction model according to the present invention.

[0020] Figure 2 The internal validation results of the Random Survival Forest (RSF) model in this invention are presented. Figure A shows the Out-of-Bag (OOB) survival probability prediction curve for each patient, with red representing the overall median survival trend; Figure B shows the OOBBrier score changing over time, reflecting the prediction error for patients in different risk strata, with the red line representing the overall level; Figure C shows the OOB Continuous Rank Probability Score (CRPS) curve, used to evaluate the model's prediction distribution accuracy in each risk stratum; Figure D shows the model's OOB (Out-of-Bag) prediction results for patient mortality risk at different time points, with blue dots representing patients who actually died and black dots representing those who did not experience the event. The axis represents the model's predicted mortality risk score, and the horizontal axis represents the follow-up time. The figures show that deceased patients have a higher predicted risk in the early stages, with the blue dots concentrated at the top, indicating good time sensitivity and discriminative ability; Figure E shows the relationship between the number of trees in the model and the OOB error rate, showing that the predictive performance tends to stabilize when the number of trees exceeds approximately 100. These results indicate that the model has good survival prediction ability and generalization performance.

[0021] Figure 3This demonstrates the operational relationships of the online prediction platform in this invention. Users sequentially input basic patient parameters on the left-hand interface, including FIGO stage, previous pregnancy interval (IFAP), Ki-67 index, and peripheral blood neutrophil, lymphocyte, and monocyte counts. The system then automatically calculates the inflammatory biomarkers NLR and SIRI, and feeds the standardized input values ​​into a pre-trained Random Survival Forest (RSF) model for prediction. At a specified time point (e.g., 12 months), the model outputs the individual progression probability at that time point (0.404 in the figure), and simultaneously plots the corresponding individual dynamic survival curve in the right-hand chart, achieving a visual display of survival probability and interactive risk assessment, facilitating physicians to develop individualized treatment strategies. Detailed Implementation

[0022] like Figure 1 As shown, this machine learning-based intermediate trophoblastic tumor prediction model includes:

[0023] The data input module allows users to input basic patient parameters via a web interface. These parameters include FIGO tumor stage, previous pregnancy interval (IFAP), Ki-67 index, and cell count (peripheral blood neutrophils, lymphocytes, or monocytes). The indicator calculation module calculates inflammatory biomarkers, including the neutrophil-to-lymphocyte ratio (NLR) and the systemic inflammatory response index (SIRI). (Specific calculation method: NLR = neutrophil count × 10^6) 9 / L)÷Lymphocyte count (×10) 9 / L); SIRI = neutrophil count (×10 9 / L)×Monocyte count(×10) 9 / L)÷Lymphocyte count (×10) 9 / L); After the user inputs the cell count parameters, this system will automatically complete the above ratio calculation and use the result as one of the input variables for the subsequent risk prediction module to participate in the quantitative assessment of individual survival probability.

[0024] The risk prediction module uses a pre-trained RSF (Random Survival Forest) model to standardize the input data and make predictions, outputting the probability of patients progressing to PFS within 1 to 5 years, and performing high / low risk stratification based on the risk threshold determined in the training set.

[0025] Model Interpretation Module: Visualizes a summary graph of model variables using SHAP value visualization technology;

[0026] The results display and interaction module is a web front-end built on the Shiny platform, which generates predicted probability values ​​and individual dynamic survival curves.

[0027] This invention is specifically designed for intermediate trophoblastic tumors (ITT). It combines FIGO staging, Ki-67 index, gestational interval IFAP, neutrophil-lymphocyte ratio (NLR), and the systemic immune inflammation index (SIRI) as key indicators to construct a disease-specific random survival forest (RSF) prediction model, significantly improving the accuracy of individualized progression risk prediction. The device integrates the interpretable algorithm SHAP, intuitively presenting the contribution of each variable to risk prediction, solving the problem of existing "black box" models being clinically uninterpretable. Real-time input, automatic calculation, and dynamic survival curve output are achieved through a web platform, enhancing physicians' ability to identify risks in ITT patients and improving the efficiency of treatment decisions.

[0028] Preferably, in the indicator calculation module, NLR = neutrophil count ÷ lymphocyte count; SIRI = neutrophil count × monocyte count ÷ lymphocyte count; the calculation results of NLR and SIRI are used as one of the input variables of the risk prediction module to participate in the quantitative assessment of individual survival probability.

[0029] Preferably, in the risk prediction module, multiple decision trees are constructed. During the training process, each decision tree uses only a portion of the samples and features, and the nodes are divided using the log-rank splitting criterion. For each patient, the survival probability estimate is obtained by aggregating the leaf node distributions in each decision tree.

[0030] Preferably, in the risk prediction module, constructing a single decision tree includes: firstly, randomly selecting approximately 63% of the samples from the training set using the Bootstrap method to form the training set of the current tree; the unselected portion is used for out-of-bag (OOB) error estimation; subsequently, at each split node, a subset is randomly selected from all input features to participate in the split criterion evaluation to avoid feature selection bias; each node uses the log-rank statistic as the splitting criterion, calculates the survival difference for each candidate feature and its different splitting thresholds, selects the combination with the largest log-rank value as the optimal splitting method for that node, and recursively divides the data into left and right child nodes until the minimum number of samples is reached or further splitting is not possible; finally, the input data of each patient is placed into each tree, falling into the corresponding leaf node, aggregating the survival information of similar samples in the leaf node, generating an individual survival probability estimate using the Kaplan-Meier method, and averaging the prediction results of all trees to obtain the final comprehensive survival curve.

[0031] Preferably, when randomly selecting a subset, if there are 5 variables, then 2 of them are randomly selected.

[0032] Preferably, in the risk prediction module, some samples and features are obtained by using bootstrap resampling and random feature selection.

[0033] Preferably, in the model interpretation module, the SHAP method is used to deduce the direction and weight of the influence of each input variable on the final risk value.

[0034] Preferably, in the model interpretation module, the SHAP method is based on the Shapley value principle in game theory, assuming that each feature is a participant, and attributing the marginal contribution of the model output by traversing different feature combinations.

[0035] Preferably, in the model interpretation module, SHAP first constructs a local linear approximation function for the model. While keeping other features constant, it observes the change in the risk prediction value of each variable before and after its inclusion in the model. This change is the SHAP value of that variable; a positive or negative sign indicates the variable's promoting or protective effect on the predicted risk, and the magnitude indicates the strength of its influence. The model interpretation results are output as a variable summary plot, a force plot for a single patient, or a dependency plot for a single patient. This achieves transparency in individual prediction results and assists doctors in understanding the model's rationale.

[0036] It also provides an application of a machine learning-based intermediate trophoblastic tumor prediction model, which is used in the assessment of progression risk and prediction of individualized survival probability for patients with intermediate trophoblastic tumors.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A machine learning-based intermediate trophoblastic tumor prediction model, characterized by: It includes: The data input module allows users to input the patient's basic parameters through a web interface. These parameters include the International Federation of Gynecologic and Obstetric Societies FIGO tumor stage, previous pregnancy interval IFAP, Ki-67 index, and cell count, which consists of peripheral blood neutrophils, lymphocytes, or monocytes. The indicator calculation module calculates inflammatory biomarkers, including the neutrophil-to-lymphocyte ratio (NLR) and the systemic inflammatory response index (SIRI). The risk prediction module uses a pre-trained RSF model to standardize the input data and make predictions, outputting the probability of patients progressing to PFS within 1 to 5 years, and performing high / low risk stratification based on the risk threshold determined in the training set. Model Interpretation Module: Visualizes a summary graph of model variables using SHAP value visualization technology; The results display and interaction module is a web front-end built on the Shiny platform, which generates predicted probability values ​​and individual dynamic survival curves.

2. The machine learning-based intermediate trophoblastic tumor prediction model according to claim 1, characterized in that: In the indicator calculation module, NLR = neutrophil count ÷ lymphocyte count; SIRI = neutrophil count × monocyte count ÷ lymphocyte count; the calculation results of NLR and SIRI are used as one of the input variables of the risk prediction module to participate in the quantitative assessment of individual survival probability.

3. The machine learning-based intermediate trophoblastic tumor prediction model according to claim 2, characterized in that: In the risk prediction module, multiple decision trees are constructed. During the training process, each decision tree uses only a portion of the samples and features, and the nodes are divided using the log-rank splitting criterion. For each patient, the survival probability estimate is obtained by aggregating the leaf node distributions in each decision tree.

4. The machine learning-based intermediate trophoblastic tumor prediction model according to claim 3, characterized in that: In the risk prediction module, constructing a single decision tree includes: first, randomly selecting approximately 63% of the samples from the training set using the Bootstrap method to form the training set for the current tree; the unselected portion is used for out-of-bag (OOB) error estimation; then, at each split node, a subset is randomly selected from all input features to participate in the split criterion evaluation to avoid feature selection bias; each node uses the log-rank statistic as the splitting criterion, calculating the survival difference for each candidate feature and its different splitting thresholds, selecting the combination with the largest log-rank value as the optimal splitting method for that node, thereby recursively dividing the data into left and right child nodes until the minimum number of samples is reached or further splitting is not possible; finally, the input data of each patient is fed into each tree, falling into the corresponding leaf node, aggregating the survival information of similar samples in the leaf node, generating an individual survival probability estimate using the Kaplan-Meier method, and averaging the prediction results of all trees to obtain the final comprehensive survival curve.

5. The machine learning-based intermediate trophoblastic tumor prediction model according to claim 4, characterized in that: When randomly selecting a subset, if there are 5 variables, then 2 of them are randomly selected.

6. The machine learning-based intermediate trophoblastic tumor prediction model according to claim 5, characterized in that: In the risk prediction module, some samples and features are obtained using bootstrap resampling and random feature selection.

7. The machine learning-based intermediate trophoblastic tumor prediction model according to claim 6, characterized in that: In the model interpretation module, the SHAP method is used to deduce the direction and weight of the influence of each input variable on the final risk value.

8. The machine learning-based intermediate trophoblastic tumor prediction model according to claim 7, characterized in that: In the model interpretation module, the SHAP method is based on the Shapley value principle in game theory. It assumes that each feature is a participant and attributes the marginal contribution of the model output by traversing different feature combinations.

9. The machine learning-based intermediate trophoblastic tumor prediction model according to claim 8, characterized in that: In the model interpretation module, SHAP first constructs a local linear approximation function of the model. While keeping other features unchanged, it observes the magnitude of the change in the risk prediction value of each variable before and after it is added to the model. This magnitude is the SHAP value of the variable. Positive and negative values ​​indicate the promoting or protective effect of the variable on the predicted risk, and the magnitude indicates the strength of its influence. The output of the model interpretation results is a variable summary plot, a force plot of a single patient, or a dependency plot of a single patient.

10. The application of the machine learning-based intermediate trophoblastic tumor prediction model according to claim 1, characterized in that: Its application is in the assessment of progression risk and prediction of individualized survival probability in patients with intermediate trophoblastic tumors.

Images