Multi-modal data integration and machine learning model for predicting curative effect and prognosis of PARP inhibitor by ovarian cancer patient

Abstract

The invention provides a multi-modal data integration and machine learning model for predicting the curative effect and prognosis of an ovarian cancer patient on a PARP inhibitor. The model comprises a data collection and processing module, a feature screening module, a model construction and prediction module and a model interpretation and application module. Data collection relates to clinical, pathological, biochemical indexes and the like, and feature screening is performed through correlation analysis, Cox regression and other methods. The model construction adopts random forest, logistic regression and other algorithms, and meta-learning is carried out through XGBoost or LightGBM. The performance of the model is evaluated through indexes such as AUC and accuracy, and the model is explained by using an SHAP value. According to the invention, the curative effect and prognosis of the PARP inhibitor by the ovarian cancer patient can be accurately predicted, and a scientific basis is provided for clinical treatment.

Description

Technical Field

[0001] The present invention relates to the field of utilizing multimodal data and machine learning technology, and more specifically, to a multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors. Background Art

[0002] In the medical field, ovarian cancer is one of the common malignant tumors in the female reproductive system, and its treatment and prognosis prediction have always been the focus of clinical research. Traditional evaluation of ovarian cancer treatment effects mainly relies on clinical and pathological characteristics, but these methods often lack accuracy and personalization. With the development of bioinformatics and machine learning technologies, more and more studies have begun to explore how to use multimodal data (including clinical, pathological, biochemical indicators, etc.) to improve the accuracy of treatment effect prediction. The integration and analysis of these data can provide more comprehensive disease characteristics, which can help to develop more personalized treatment plans.

[0003] However, existing technologies face problems such as difficulty in data integration, inaccurate feature selection, and limited model generalization when processing these multimodal data. In particular, in predicting the response and prognosis of ovarian cancer patients to PARP inhibitors, there is a lack of a model that comprehensively considers multiple data types and can provide high-accuracy predictions. In addition, existing models also have deficiencies in interpretability and clinical practicality, which limits their application in actual clinical decision-making.

[0004] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: the data integration and feature screening methods are not sophisticated enough, resulting in low model training efficiency and low prediction accuracy; the model construction lacks an in-depth understanding of the characteristics of different data modalities, resulting in limited generalization capabilities of the model in practical applications; the model is not interpretable enough, making it difficult to evaluate the contribution of each feature to the prediction results, limiting the application of the model in clinical decision-making; and the model evaluation method is not comprehensive enough to fully verify the clinical practicality and accuracy of the model. Summary of the invention

[0005] The present invention provides a multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors, comprising:

[0006] Data collection and processing module: collect clinical characteristics, pathological characteristics, biochemical indicators, treatment regimens, BRCA mutation / HRD status, and recurrence and survival status data of ovarian cancer patients, delete features with a missing rate of more than 50%, and perform Z-score standardization on KNN and SVM models;

[0007] Feature screening module: Calculate the correlation between features and remove features with a correlation greater than 0.8; use the Cox proportional hazard regression model for univariate analysis and retain features with a P value less than 0.05; remove features with a variance inflation factor greater than 10; use the elastic net method to select stable features;

[0008] Model building and prediction module: machine learning algorithms are used to build prediction models, and stacking models are used to summarize data from various modalities, including input layer, basic model layer, fusion layer and meta-model layer. In the basic model layer, random forest is used for genomic feature model, logistic regression is used for clinical feature model, and support vector machine is used for biochemical feature model. The fusion layer summarizes the output of the basic model. XGBoost or LightGBM is selected as the meta-learner in the meta-model layer, and the cross-entropy loss function is used. The basic model is trained multiple times, and the meta-model is trained using cross-validation or bootstrap method. The grid search method is used for parameter tuning, and the five-fold cross-validation method is used to evaluate the model performance. AUC, accuracy, F1 score, sensitivity and specificity indicators are selected;

[0009] Model interpretation and application module: Use SHAP values ​​to interpret the model, evaluate the contribution of each feature to model prediction, compare the consistency of predicted probability with actual results through calibration curve analysis, and use decision curve analysis to evaluate the clinical practicality of the model.

[0010] Furthermore, the clinical characteristics include age, gender, and weight.

[0011] Furthermore, the pathological characteristics include pathological stage, pathological classification, and metastasis type.

[0012] Furthermore, the biochemical indicators include routine blood tests and routine urine tests.

[0013] Furthermore, the treatment regimen includes a chemotherapy regimen and a PARP inhibitor type regimen.

[0014] Furthermore, the machine learning algorithms include Logistic regression, K-nearest neighbor algorithm, random forest, LightGBM, XGBoost, support vector machine, and naive Bayes.

[0015] Furthermore, in the basic model layer, each modality independently establishes a model and selects the optimal model.

[0016] Furthermore, the fusion method of the fusion layer includes simple splicing, averaging and weighted averaging.

[0017] Furthermore, total bile acid (TBA) and CA-199 were independent risk factors in patients with primary and recurrent ovarian cancer, respectively, identified by multivariate Cox regression analysis.

[0018] Furthermore, the performance of the model in patients with primary and recurrent ovarian cancer was evaluated by AUC, accuracy, and F1 score metrics.

[0019] The above-mentioned embodiments of the present invention have at least the following beneficial effects: the multimodal data integration and machine learning model can improve the prediction accuracy of the efficacy of PARP inhibitors and the prognosis of patients by comprehensively analyzing the clinical characteristics, pathological characteristics, biochemical indicators, treatment regimens, BRCA mutations / HRD status and other data of ovarian cancer patients. The feature screening and stability selection methods adopted by the model can eliminate irrelevant or unstable features and retain key information that substantially contributes to the prediction results, thereby enhancing the prediction ability and generalization of the model.

[0020] In addition, the model uses advanced machine learning algorithms such as random forests, logistic regression, support vector machines, etc., combined with the multi-level structure of the stacked model, to achieve in-depth mining and effective fusion of different modal data. This multi-level model construction method can not only improve the accuracy of prediction, but also clarify the contribution of each feature to the model prediction through explanatory tools such as SHAP values, thereby enhancing the interpretability of the model. At the same time, the clinical practicality of the model is verified through calibration curve and decision curve analysis, which can ensure the consistency between the model prediction results and the actual clinical results. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which:

[0022] Figure 1 A schematic diagram of the structure of a multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients to PARP inhibitors provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0023] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0024] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0025] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0026] Reference below Figure 1 , Figure 1 A schematic diagram of the structure of a multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients to PARP inhibitors provided by an embodiment of the present invention. Figure 1 As shown, a multimodal data integration and machine learning model 100 for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors includes:

[0027] Data collection and processing module 101: collect clinical characteristics, pathological characteristics, biochemical indicators, treatment regimens, BRCA mutation / HRD status, and recurrence and survival status data of ovarian cancer patients, delete features with a missing rate of more than 50%, and perform Z-score standardization on KNN and SVM models;

[0028] Feature screening module 102: calculate the correlation between features, remove features with a correlation greater than 0.8; use the Cox proportional hazard regression model to perform univariate analysis, retain features with a P value less than 0.05; remove features with a variance inflation factor greater than 10; use the elastic net method to select stable features;

[0029] Model building and prediction module 103: Use machine learning algorithms to build prediction models, and use stacking models to summarize data from various modalities, including input layer, basic model layer, fusion layer and meta-model layer. In the basic model layer, the genomic feature model uses random forest, the clinical feature model uses logistic regression, and the biochemical feature model uses support vector machine. The fusion layer summarizes the output of the basic model. The meta-model layer selects XGBoost or LightGBM as the meta-learner and uses the cross-entropy loss function. The basic model is trained multiple times, and the meta-model is trained using cross-validation or bootstrap. The grid search method is used for parameter tuning. The five-fold cross-validation method is used to evaluate the model performance, and the AUC, accuracy, F1 score, sensitivity and specificity indicators are selected;

[0030] Model interpretation and application module 104: Use SHAP values ​​to interpret the model, evaluate the contribution of each feature to the model prediction, compare the consistency of the predicted probability with the actual results through calibration curve analysis, and use decision curve analysis to evaluate the clinical practicality of the model.

[0031] It should be noted that the model includes a data collection and processing module, which is responsible for collecting clinical characteristics, pathological characteristics, biochemical indicators, treatment plans, BRCA mutation / HRD status, and recurrence and survival status data of ovarian cancer patients. Here, clinical characteristics refer to basic information such as age, gender, and weight of patients; pathological characteristics include pathological stage, pathological classification, metastasis type, etc.; and biochemical indicators cover laboratory test results such as blood routine and urine routine. After data collection, features with a missing rate of more than 50% are deleted to ensure the integrity and reliability of the data, and the KNN and SVM models are Z-score standardized to eliminate the influence between data of different dimensions.

[0032] Specifically, data collection can be achieved through channels such as hospital information systems, laboratory information systems, and electronic health records. When processing data, the parameters of Z-score standardization are set to a mean of 0 and a standard deviation of 1 to ensure that different features are comparable on the same scale. For missing data processing, if the missing value of a feature exceeds 50%, the feature will be deleted to avoid introducing too much uncertainty. In addition, for KNN and SVM models, the specific parameters of Z-score standardization are set to the mean and standard deviation of each feature, which can be calculated from the data set.

[0033] Preferably, in addition to Z-score standardization, other data preprocessing methods such as minimum-maximum standardization or decimal scaling standardization can also be considered to adapt to the distribution characteristics of different features. In terms of feature deletion, in addition to the criterion of missing rate exceeding 50%, feature importance evaluation methods such as information gain or mutual information can be introduced to further optimize the feature selection process. In addition, for the Z-score standardization of the model, more robust statistical methods such as median and interquartile range (IQR) can be considered to reduce the impact of outliers.

[0034] In some embodiments, the clinical characteristics include age, sex, and weight.

[0035] It should be noted that the model pays special attention to clinical characteristics, including three key parameters: age, gender, and weight. Here, clinical characteristics refer to basic information obtained directly from patients, which is essential for assessing the patient's health status and choosing treatment options. Age refers to the patient's life span, gender involves biological differences, and weight is related to the patient's nutritional status and drug dosage calculation.

[0036] Specifically, the collection of clinical characteristics can be completed through the initial assessment of the patient upon admission. Age can be confirmed through the patient's identification document, gender is determined based on the patient's physiological characteristics and / or self-report, and weight is measured on the hospital's scale when the patient is admitted. The collection of these data needs to comply with the hospital's privacy protection policy and data management specifications. In terms of parameter settings, age is usually expressed in years, gender is expressed in binary (male / female), and weight is recorded in kilograms.

[0037] Preferably, in addition to basic clinical features, other clinical parameters related to ovarian cancer treatment can also be considered, such as the patient's height, BMI (body mass index), past medical history, and family medical history. This additional information can provide more comprehensive patient information and help the model more accurately predict treatment effects and prognosis. During the data collection process, an electronic health record system can be used to automate and standardize data entry, reduce human errors, and ensure data consistency and integrity.

[0038] Furthermore, for weight measurement, one can consider using smart scales, which can automatically transmit data to the electronic health record system, further improving the efficiency and accuracy of data collection.

[0039] In some embodiments, the pathological features include pathological stage, pathological classification, and metastasis type.

[0040] It should be noted that the model covers pathological characteristics, including three key medical parameters: pathological stage, pathological classification, and metastasis type. Here, pathological characteristics refer to information about the nature and extent of the disease obtained through medical examinations, which is crucial for disease diagnosis and treatment planning. Pathological stage refers to the degree of development of the tumor, pathological classification is the classification of tumors based on histological characteristics, and metastasis type describes the way and location of tumor cell spread.

[0041] Specifically, the determination of pathological features relies on microscopic examination and analysis of tumor samples by pathologists. The pathological stage is usually determined according to the tumor-node-metastasis (TNM) staging system, which takes into account the size of the tumor (T), lymph node involvement (N), and the presence of distant metastasis (M). Pathological classification is based on the classification criteria of the World Health Organization (WHO), which classifies tumors according to different histological types and degrees of differentiation. The type of metastasis is determined by imaging examinations (such as CT, MRI) and biomarker testing, which describes the spread of tumor cells from the primary site to other parts of the body.

[0042] Preferably, in addition to the above-mentioned pathological characteristics, pathological parameters including tumor grade, molecular marker expression and other pathological parameters related to the prognosis of ovarian cancer can also be considered. These parameters can provide a more detailed description of tumor characteristics, which helps the model to more accurately predict the treatment effect. In the evaluation of pathological stages, more advanced imaging techniques, such as positron emission tomography (PET-CT), can be used to improve the accuracy of staging. For pathological classification, immunohistochemistry can be used to evaluate the specific protein expression of tumor cells, which helps to further refine the classification of tumors. In determining the type of metastasis, liquid biopsy technology can be combined to monitor tiny tumor residues and early metastases by detecting circulating tumor DNA (ctDNA) in the blood. The integration of these technologies can provide more comprehensive pathological information and enhance the predictive ability of the model.

[0043] In some embodiments, the biochemical indicators include routine blood tests and routine urine tests.

[0044] It should be noted that the model pays special attention to biochemical indicators, including two key laboratory test items: blood routine and urine routine. Here, biochemical indicators refer to the chemical composition data obtained by analyzing blood and urine samples, which can reflect the physiological and pathological status of patients. Blood routine refers to the detection of blood cell components, including red blood cells, white blood cells and platelets; urine routine involves various components in urine, such as protein, glucose, red blood cells, etc.

[0045] Specifically, the collection of biochemical indicators is completed through standard laboratory testing procedures. Routine blood tests usually include parameters such as red blood cell count, white blood cell count, hemoglobin level and platelet count, which can be measured by automated blood analyzers. Routine urine tests include the detection of components such as pH, specific gravity, protein, glucose, ketone bodies, red blood cells and white blood cells in urine, which are usually completed by urine analyzers. These test results provide clinicians with important information about the patient's health status, such as whether there is infection, inflammation, abnormal kidney function and other problems.

[0046] Preferably, in addition to routine blood and urine tests, other relevant biochemical indicators can also be considered, such as liver function tests (including ALT, AST, etc.), renal function tests (such as BUN, Cr, etc.), electrolyte levels (such as Na+, K+, etc.), and blood lipid analysis (such as total cholesterol, triglycerides, etc.). These additional biochemical indicators can provide more comprehensive metabolic and organ function information, which helps to more accurately assess the patient's health status and treatment effect. In terms of detection methods, more advanced technologies such as mass spectrometry or flow cytometry can be used to improve the sensitivity and accuracy of detection.

[0047] Furthermore, in order to improve detection efficiency and reduce sample processing time, it is possible to consider implementing automated sample processing and analysis processes, as well as using artificial intelligence-assisted image recognition technology to analyze complex samples such as urine sediment. The integration of these technologies can further improve the quality and efficiency of biochemical index detection.

[0048] In some embodiments, the treatment regimen comprises a chemotherapy regimen, a PARP inhibitor type regimen.

[0049] It should be noted that the model covers the treatment regimen, which includes two key treatment parameters: chemotherapy regimen and PARP inhibitor type regimen. Here, the treatment regimen refers to a specific treatment plan formulated for the patient's condition, which is crucial for disease control and prognosis improvement. The chemotherapy regimen refers to a plan for treatment with chemical drugs, while the PARP inhibitor type regimen refers to a plan for treatment with a specific type of PARP inhibitor.

[0050] Specifically, the treatment plan is determined based on the patient's condition and treatment response. The chemotherapy plan may include parameters such as the type of drug, dosage, dosing frequency, and treatment cycle. These parameters need to be adjusted according to the patient's weight, age, physiological conditions such as renal and liver function, and the pharmacokinetic properties of the drug. The PARP inhibitor type plan involves selecting specific PARP inhibitor drugs, such as olaparib, niraparib, etc., and determining the dosage and duration of treatment. The formulation of these treatment plans requires comprehensive consideration of the patient's pathological characteristics, biochemical indicators, and genetic information.

[0051] Preferably, in addition to chemotherapy regimens and PARP inhibitor type regimens, other adjuvant treatments such as targeted therapy, immunotherapy, etc. can also be considered. These treatments can be used in combination with chemotherapy or PARP inhibitor treatment to enhance the therapeutic effect. In chemotherapy regimens, drug combinations and dosages can be personalized based on the patient's gene expression profile and drug sensitivity test results. For PARP inhibitor type regimens, the most appropriate inhibitor can be selected based on the patient's BRCA mutation status and HRD status.

[0052] Furthermore, drug blood concentration monitoring and pharmacodynamic models can be used to optimize the dosing regimen to ensure that the drug is within a safe and effective concentration range. These refined operating steps and alternatives can further improve the personalization and effectiveness of treatment plans.

[0053] In some embodiments, the machine learning algorithm includes Logistic regression, K-nearest neighbor algorithm, random forest, LightGBM, XGBoost, support vector machine, and naive Bayes.

[0054] It should be noted that the model uses a variety of machine learning algorithms, including Logistic Regression, K-Nearest Neighbors, Random Forest, LightGBM, XGBoost, Support Vector Machine, and Naive Bayes. Here, machine learning algorithms refer to a series of statistical and computational techniques used to train computer systems to learn from data and make predictions or decisions. Logistic regression is a linear model for binary classification problems; K-Nearest Neighbors is an instance-based learning method; Random Forest is an ensemble learning method that improves prediction accuracy by building multiple decision trees; LightGBM and XGBoost are efficient gradient boosting frameworks; Support Vector Machine is a maximum margin classifier; Naive Bayes is a probability-based classification method.

[0055] Specifically, the application of these machine learning algorithms in the model involves specific parameter settings and conceptual understanding. For example, the parameters of Logistic regression include regularization strength and number of iterations; the parameter of the K-nearest neighbor algorithm is the number of neighbors K; the parameters of random forest include the number of trees and the depth of the trees; the parameters of LightGBM and XGBoost involve learning rate, number of trees and maximum depth, etc.; the parameters of support vector machines include kernel function type and regularization parameter; the parameters of naive Bayes involve probability estimation methods. The settings of these parameters need to be adjusted according to the characteristics of the data and the performance of the model to achieve the best prediction effect.

[0056] Preferably, in addition to the algorithms mentioned above, other machine learning algorithms, such as deep learning models, neural networks, etc., can also be considered to further improve the prediction ability of the model. In terms of parameter setting, automated parameter tuning techniques, such as grid search or random search, can be used to find the optimal parameter combination.

[0057] Furthermore, cross-validation can be used to evaluate the stability and generalization ability of the model. In terms of model integration, the stacking method can be considered to combine the prediction results of different algorithms to improve the overall prediction performance. These detailed operation steps and alternatives can further improve the accuracy and robustness of the model.

[0058] In some embodiments, in the basic model layer, each modality independently establishes a model and selects the optimal model.

[0059] It should be noted that in the basic model layer, the model independently builds a model for each modality and selects the optimal model. Here, modality refers to different sources or types of data, such as genomic features, clinical features, and biochemical features. Independent model building means designing and training a specialized machine learning model for the data characteristics of each modality to capture the unique information and patterns of the modality.

[0060] Specifically, for genomic feature models, the random forest algorithm can be used, which is capable of processing high-dimensional data and identifying the relationship between gene expression patterns and disease outcomes. For clinical feature models, logistic regression may be more suitable because it can evaluate the impact of multiple clinical parameters on disease outcomes and provide a risk ratio for each parameter. For biochemical feature models, support vector machine (SVM) may be a good choice because it performs well in dealing with classification problems, especially in distinguishing the impact of different biochemical index levels on disease prognosis. The parameter settings of these models need to be adjusted according to the characteristics of the specific data, such as the number and depth of trees in random forests, the regularization strength in logistic regression, and the kernel function and penalty parameters in SVM.

[0061] Preferably, in addition to the models and algorithms mentioned above, other algorithms can be considered to build the basic model layer, such as deep learning models to process complex genomic data, or gradient boosting machines (GBM) to process clinical and biochemical data. In the model selection process, grid search and cross-validation can be used to systematically evaluate the performance of different models and parameter settings to select the optimal model.

[0062] Furthermore, model fusion techniques, such as stacking or voting mechanisms, can be considered to combine the prediction results of different modality models to improve the accuracy and robustness of the overall model. These refined operation steps and alternatives can further enhance the model's predictive ability and clinical application value.

[0063] In some embodiments, the fusion method of the fusion layer includes simple concatenation, averaging, and weighted averaging.

[0064] It should be noted that the model uses different methods to integrate the outputs of the base models in the fusion layer. Here, the fusion layer refers to the process of combining the results of different base models into a comprehensive prediction. Simple concatenation is to directly connect the outputs of different models; averaging is to calculate the average of all model outputs; weighted averaging is to give different weights to each model according to its performance or importance, and then calculate the weighted average.

[0065] Specifically, the operation of the fusion layer involves how to effectively combine the prediction results of each basic model. For example, simple concatenation can directly connect the output vectors of random forest, logistic regression, and SVM to form a longer vector. The averaging method calculates the average of these outputs, which is suitable for situations where the output scales of all models are the same. Weighted averaging requires pre-determining the weights of each model, which can be assigned based on the model's accuracy, AUC value, or other performance indicators. Parameter setting includes determining the weight values, which are usually determined through parameter optimization techniques such as cross-validation and grid search.

[0066] Preferably, in addition to simple concatenation, averaging, and weighted averaging, other fusion techniques can be considered, such as stacking or model ensembling. In stacking, a new machine learning model can be used to learn how to best combine the predictions of the base models. Model ensembling can use a voting mechanism, such as using majority voting in classification problems to decide the final prediction result.

[0067] Furthermore, it is also possible to consider using deep learning methods to automatically learn how to extract the most useful information from the output of the basic model. These refined operation steps and alternatives can further improve the model's predictive performance and ability to adapt to different data sets.

[0068] In some embodiments, total bile acid (TBA) and CA-199 are independent risk factors in patients with de novo and recurrent ovarian cancer, respectively, identified by multivariate Cox regression analysis.

[0069] It should be noted that the model identified total bile acid (TBA) and CA-199 as independent risk factors in patients with primary and recurrent ovarian cancer, which was achieved through multivariate Cox regression analysis. Here, independent risk factors refer to factors that can still affect the prognosis of patients after considering the influence of other variables. Multivariate Cox regression analysis is a statistical method used to evaluate the impact of multiple variables on survival time.

[0070] Specifically, multivariate Cox regression analysis involves incorporating multiple candidate risk factors into the model simultaneously to assess their joint impact on survival time. In this case, total bile acid (TBA) and CA-199, as specific biochemical indicators, were selected as variables for analysis. TBA is an indicator of liver function, while CA-199 is a tumor marker that is often used to monitor the progression of certain cancers. Specific parameter settings include determining the cutoff value of candidate variables, encoding categorical variables, and setting the baseline risk function of the model. The setting of these parameters needs to be determined based on the data distribution and previous studies.

[0071] Preferably, in addition to TBA and CA-199, other known risk factors for ovarian cancer can also be considered, such as patient age, pathological stage or treatment response, etc. When building the model, a stepwise regression method can be used to select the most important variables, or a regularization method such as Lasso can be used to reduce the complexity of the model and prevent overfitting.

[0072] Furthermore, the model can be cross-validated to evaluate its predictive ability on independent datasets. These refined operational steps and alternatives can further improve the accuracy of the model and the reliability of clinical application.

[0073] In some embodiments, the performance of the model in patients with primary and recurrent ovarian cancer is evaluated by AUC, accuracy, and F1 score metrics.

[0074] It should be noted that the performance of the model in patients with primary and recurrent ovarian cancer was evaluated by AUC, accuracy, and F1 score indicators. Here, AUC refers to the area under the receiver operating characteristic curve, which measures the ability of the classification model to distinguish between two categories; accuracy refers to the proportion of correct predictions by the model; F1 score is the harmonic mean of precision and recall, which is used to measure the overall performance of the model.

[0075] Specifically, when evaluating model performance, the AUC value is calculated by comparing the probability predicted by the model with the actual events that occurred. The higher the value, the stronger the model's ability to distinguish. Accuracy refers to the proportion of samples predicted correctly by the model to the total number of samples, and the calculation formula is (true positive examples + true negative examples) / total number of samples. The F1 score is an indicator that strikes a balance between precision and recall, and the calculation formula is 2*(precision*recall) / (precision+recall). The specific parameter settings of these indicators include determining the threshold to divide the predicted results and the actual results, and selecting an appropriate evaluation data set.

[0076] Preferably, in addition to AUC, accuracy, and F1 score, other evaluation metrics such as sensitivity, specificity, and recall can also be considered. Sensitivity refers to the ability of the model to correctly identify positive examples, specificity refers to the ability of the model to correctly identify negative examples, and recall refers to the proportion of positive examples identified by the model to all actual positive examples.

[0077] Furthermore, in terms of operation steps, a cross-validation method can be used to evaluate the stability and generalization ability of the model, such as using K-fold cross-validation. In addition, a grid search can be performed on the model to find the optimal parameter setting, or a bootstrap method can be used to evaluate the uncertainty of the model.

[0078] The above-mentioned embodiments of the present invention have the following beneficial effects: the multimodal data integration and machine learning model described in the present invention can provide accurate predictions for the efficacy and prognosis of ovarian cancer patients on PARP inhibitors. By comprehensively considering data from multiple dimensions such as clinical, pathological, biochemical, and genetic, the model can capture more comprehensive patient characteristics, thereby improving the accuracy and reliability of the prediction. This comprehensive analysis method helps doctors develop more personalized treatment plans, optimize treatment decisions, and improve the quality of life and prognosis of patients.

[0079] In addition, the model can improve the stability and generalization ability of the model by adopting advanced machine learning techniques such as XGBoost or LightGBM as meta-learners, and using cross-validation and bootstrap methods for model training and verification. The performance of the model is strictly evaluated through indicators such as AUC, accuracy, F1 score, sensitivity and specificity to ensure the effectiveness of the model in practical applications. At the same time, by interpreting the model through methods such as SHAP value, the contribution of each feature to the prediction results can be clearly identified and evaluated, the transparency and credibility of the model can be enhanced, and a strong scientific basis for clinical decision-making can be provided.

[0080] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, a server, a mobile phone, and a tablet.

[0081] The above descriptions are only some preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention to form a technical solution.

Claims

1. A multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors, characterized in that: include: Data collection and processing module: collect clinical characteristics, pathological characteristics, biochemical indicators, treatment regimens, BRCA mutation / HRD status, and recurrence and survival status data of ovarian cancer patients, delete features with a missing rate of more than 50%, and perform Z-score standardization on KNN and SVM models; Feature screening module: calculates the correlation between features and removes features with a correlation greater than 0.8; Univariate analysis was performed using the Cox proportional hazards regression model, and features with a P value less than 0.05 were retained; Remove features with variance inflation factor greater than 10; Use the elastic net method to select features with stability; Model building and prediction module: The prediction model is built using machine learning algorithms, and the stacking model is used to summarize the data of each modality, including input layer, basic model layer, fusion layer and meta-model layer. In the basic model layer, the genomic feature model uses random forest, the clinical feature model uses logistic regression, and the biochemical feature model uses support vector machine. The fusion layer summarizes the output of the basic model. The meta-model layer selects XGBoost or LightGBM as the meta-learner and uses the cross-entropy loss function. The basic model is trained multiple times, and the meta-model is trained using cross-validation or bootstrap. The grid search method is used for parameter tuning. The five-fold cross-validation method is used to evaluate the model performance, and the AUC, accuracy, F1 score, sensitivity and specificity indicators are selected; Model interpretation and application module: Use SHAP values ​​to interpret the model, evaluate the contribution of each feature to model prediction, compare the consistency of predicted probability with actual results through calibration curve analysis, and use decision curve analysis to evaluate the clinical practicality of the model.

2. The multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors according to claim 1, characterized in that: The clinical characteristics include age, gender, and weight.

3. The multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors according to claim 1, characterized in that: The pathological characteristics include pathological stage, pathological classification, and metastasis type.

4. The multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors according to claim 1, characterized in that: The biochemical indicators include routine blood tests and routine urine tests.

5. The multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors according to claim 1, characterized in that: The treatment regimen includes a chemotherapy regimen and a PARP inhibitor type regimen.

6. The multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors according to claim 1, characterized in that: The machine learning algorithms include Logistic regression, K-nearest neighbor algorithm, random forest, LightGBM, XGBoost, support vector machine, and naive Bayes.

7. The multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors according to claim 1, characterized in that: In the basic model layer, each modality independently establishes a model and selects the optimal model.

8. The multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors according to claim 1, characterized in that: The fusion methods of the fusion layer include simple concatenation, averaging and weighted averaging.

9. The multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors according to claim 1, characterized in that: Total bile acid (TBA) and CA-199 were independent risk factors in patients with de novo and recurrent ovarian cancer, respectively, identified by multivariate Cox regression analysis.

10. The multimodal data integration and machine learning model for predicting the efficacy and prognosis of ovarian cancer patients on PARP inhibitors according to claim 1, characterized in that: The performance of the model in patients with primary and recurrent ovarian cancer was evaluated by AUC, accuracy, and F1 score metrics.

Images