A disease prediction method and system based on dynamic weighted fusion ensemble learning

Abstract

The application provides a disease prediction method and system based on dynamic weighted fusion ensemble learning. Specifically, the method constructs a dynamic weighted fusion model, which internally builds multiple basic classifiers composed of different feature selection algorithms and machine learning algorithms. By training and screening the model, and introducing a Softmax function with a temperature coefficient, the weight of each preferred classifier is measured, so that dynamic weighted fusion is performed according to the weight and the prediction value of each preferred classifier to obtain a DWF model corresponding to the disease. In the prediction task of platinum drug resistance of ovarian cancer, the performance of the DWF model is significantly better than that of the existing LR benchmark model, and the sensitivity and accuracy of the prediction are not affected by the treatment plan, and the DWF model has application prospects in early diagnosis and prediction of diseases.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence, machine learning, and clinical data analysis, specifically to a predictive method and system for disease risk classification based on routine clinical and laboratory data using a dynamic weighted fusion ensemble learning strategy, which is particularly suitable for the early prediction of platinum resistance in ovarian cancer. Background Technology

[0002] In clinical medicine, accurately predicting a patient's response to chemotherapy is crucial for developing personalized treatment plans. Taking ovarian cancer as an example, platinum-based resistance is a major bottleneck affecting patient prognosis. Current predictive methods mainly rely on retrospective platinum-free intervals (PFI), which suffer from significant time lag. While genomics- or radiomics-based predictive models have shown potential, their high cost and complex testing procedures limit their adoption in primary healthcare institutions. Routine laboratory tests (such as complete blood counts and biochemistry) offer advantages in terms of low cost and easy access, but the predictive specificity and sensitivity of single indicators are often insufficient.

[0003] With the rapid accumulation of medical big data and the continuous improvement of computing power, machine learning (ML) has become the core driving force for the development of precision diagnosis and treatment. However, existing machine learning methods face two major challenges when processing such medical data. First, class imbalance: in clinical cohorts, drug-resistant (positive) samples are usually far fewer than sensitive (negative) samples, causing models to tend to predict the majority class, thus missing high-risk patients. Second, the limitations of static ensembles: traditional ensemble learning (such as voting or simple averaging) or early multi-criteria fusion methods usually assign fixed weights to different models or weight them based on simple performance metrics. This approach cannot dynamically adapt to changes in data at different trade-offs in cross-validation and often smooths out the contribution of the best-performing model, introducing noise from mediocre models.

[0004] Therefore, there is an urgent need in this field for a disease prediction method and system that can effectively handle class-imbalanced data and employ an ensemble learning approach that can dynamically amplify the influence of high-performance models, in order to uncover the deep predictive value of routine laboratory data. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting diseases based on dynamic weighted fusion ensemble learning.

[0006] A first aspect of the present invention provides a method for constructing a weighted dynamic fusion prediction model for disease subtyping or diagnosis, the method comprising the steps of: (S1) Provide a raw multidimensional feature dataset; the raw multidimensional feature dataset includes a clinical baseline dataset and a laboratory test dataset; (S2) Preprocess the original multidimensional feature dataset to obtain a preprocessed multidimensional feature dataset; (S3) Construct a basic classifier library, which includes multiple basic classifiers consisting of M feature selection algorithms and N machine learning algorithms; where M and N are both integers > 1; (S4) Train each basic classifier in the basic classifier library using the preprocessed multidimensional feature dataset to obtain the prediction probability of each basic classifier; during the training process, evaluate each basic classifier according to the geometric mean and select a predetermined number of basic classifiers as preferred classifiers. (S5) Evaluate the overall score of the preferred classifier using its AUC and geometric mean; (S6) Evaluate the weight of each preferred classifier based on the comprehensive score of each preferred classifier and the Softmax function with a predetermined temperature coefficient; (S7) The weights and prediction probabilities of each preferred classifier are summed to obtain the final prediction probability, thereby constructing a weighted dynamic fusion prediction model for the disease.

[0007] In another preferred embodiment, step (S7) further includes: determining a classification threshold that maximizes the geometric mean based on the ROC curve of the weighted dynamic fusion prediction model.

[0008] In another preferred embodiment, the disease includes cancer.

[0009] In another preferred embodiment, the cancer includes ovarian cancer.

[0010] In another preferred embodiment, the original multidimensional feature dataset is derived from a disease patient group and an optional healthy group.

[0011] In another preferred embodiment, the patient population includes both platinum-resistant and platinum-sensitive ovarian cancer patients.

[0012] In another preferred embodiment, the original multidimensional feature dataset is selected from the group consisting of: imbalanced positive and negative sample data, or balanced positive and negative sample data.

[0013] In another preferred embodiment, the original multidimensional feature dataset is an imbalanced dataset of positive and negative samples.

[0014] In another preferred embodiment, the clinical baseline data includes age and body mass index (BMI).

[0015] In another preferred embodiment, the laboratory test dataset includes complete blood count and systemic immune / inflammatory markers, liver and kidney function and biochemical metabolic profiles, tumor marker combination, and coagulation and fibrinolysis functional profiles.

[0016] In another preferred embodiment, the blood routine and systemic immune / inflammatory indicators include: white blood cell count (WBC), lymphocyte percentage (LYM), intermediate cell percentage (MID), neutrophil percentage (GRAN), eosinophil percentage (EOS), basophil percentage (BASO), lymphocyte count (LYM#), intermediate cell count (MID#), neutrophil count (GRAN#), eosinophil count (EOS#), basophil count (BASO#), red blood cell count (RBC), hemoglobin (HGB), hematocrit (HCT), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), coefficient of variation of red blood cell distribution width (RDWCV), standard deviation of red blood cell distribution width (RDWSD), platelet count (PLT), mean platelet volume (MPV), platelet hematocrit (PCT), platelet distribution width (PDW), and neutrophil-to-lymphocyte ratio (NL).

[0017] In another preferred embodiment, the liver and kidney function and biochemical metabolic profile include: total bilirubin (TBIL), direct bilirubin (DBIL), indirect bilirubin (IBIL), alkaline phosphatase (ALP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), aspartate aminotransferase / alanine aminotransferase ratio (ASTM), lactate dehydrogenase (LDH), gamma-glutamyl transferase (GGT), total protein (TP), albumin (ALB), globulin (GELO), albumin / globulin ratio (A / G), non-esterified fatty acids (NEFA), prealbumin (PA), glutamate dehydrogenase (GLDH), cystatin C (CYSC), glycated albumin (GA), urea (UREA), creatinine (CRE), uric acid (UA), glucose (GLU), calcium, phosphorus, magnesium, potassium, sodium, chloride, and total carbon dioxide (TCO2).

[0018] In another preferred embodiment, the tumor marker combination includes a combination of gynecological tumor markers.

[0019] In another preferred embodiment, the combination of gynecological tumor markers includes: squamous cell carcinoma antigen (SCCA), carbohydrate antigen 19-9 (CA19-9), carbohydrate antigen 125 (CA125), alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), neuron-specific enolase (NSE), human chorionic gonadotropin β subunit (β-HCG), and human epididymal protein 4 (HE4).

[0020] In another preferred embodiment, the coagulation and fibrinolysis functional profile includes: prothrombin time (PT), fibrinogen (FIB), activated partial thromboplastin time (APTT), thrombin time (TT), international normalized ratio (INR), D-dimer (DDI), and fibrin degradation products (FDP).

[0021] In another preferred embodiment, in step (S2), the preprocessing includes methods selected from the group consisting of: handling outliers, filling missing data, correcting data distribution, standardizing data, and / or handling imbalanced positive and negative samples.

[0022] In another preferred embodiment, outliers are processed by setting a threshold based on a clinical reference range and truncating outliers to that threshold.

[0023] In another preferred embodiment, the method for filling missing data includes the multivariate chain equation interpolation (MICE) method.

[0024] In another preferred embodiment, the method for correcting the data distribution includes the Box-Cox transform.

[0025] In another preferred example, the method for standardizing the data includes min-max normalization.

[0026] In another preferred embodiment, the Borderline-SMOTE technique is used to handle the imbalanced positive and negative sample data.

[0027] In another preferred embodiment, the base classifier library includes M×N base classifiers.

[0028] In another preferred embodiment, there are 14 feature selection algorithms.

[0029] In another preferred embodiment, the feature selection algorithm includes: F-score (Analysis of Variance), T-score (Analysis of Variance), Two-Input Symmetric Correlation (DISR), Fisher score, Interaction Upper Cap (ICAP), Joint Mutual Information (JMI), Laplacian score, Logistic Loss L2,1 Norm Minimization (LL l21), Least Squares Loss L2,1 Norm Minimization (LS l21), Multi-Cluster Feature Selection (MCFS), Non-Negative Discriminant Feature Selection (NFDS), Relief-F algorithm (reliefF), Traceratio, and Unsupervised Discriminant Feature Selection (UDFS).

[0030] In another preferred embodiment, there are 12 machine learning algorithms.

[0031] In another preferred embodiment, the machine learning algorithm includes: Adaptive Boosting (AdaBoost), Balanced Random Forest (Balanced RF), Category Feature Gradient Boosting (CatBoost), Decision Tree (DT), ExtraTrees, Gradient Boosting, K-Nearest Neighbors (KNN), Lightweight Gradient Boosting (LGBM), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost).

[0032] In another preferred embodiment, in step (S4), the predetermined number of base classifiers are determined by a grid search method.

[0033] In another preferred embodiment, in step (S4), the preferred classifiers are 10.

[0034] In another preferred embodiment, the preferred classifier is: (1) A classifier combining the T-value selection method of analysis of variance and logistic regression; (2) A classifier combining the T-value selection method of analysis of variance and support vector machine; (3) A classifier combining the F-value selection method of analysis of variance and logistic regression; (4) A classifier combining trace ratio criterion and logistic regression; (5) A classifier combining Fisher's scoring method and logistic regression; (6) A classifier combining Fisher scoring and support vector machine; (7) A classifier combining unsupervised discriminative feature selection and logistic regression; (8) A classifier combining Laplace score and logistic regression; (9) A classifier combining the Relief-F algorithm and a balanced random forest; and (10) A classifier combining multi-cluster feature selection method and random forest.

[0035] In another preferred embodiment, in step (S5), the comprehensive score is calculated as follows: , in, For the first i The AUC value of the optimal classifier For the first i The geometric mean of the preferred classifiers and Let be the weighting coefficient, satisfying .

[0036] In another preferred embodiment, the , .

[0037] In another preferred embodiment, in step (S6), the weight is calculated as follows: , in, For the first i The combined score of the selected classifiers, where K is the number of selected classifiers and T is the temperature coefficient, satisfying 0 < K. T < 1.

[0038] In another preferred embodiment, T satisfies 0.05≤T≤0.15, 0.07≤T≤0.13, or 0.09≤T≤0.11, and preferably, T = 0.1.

[0039] In another preferred embodiment, the final prediction probability is calculated as follows: , in, The weights of the i-th preferred classifier are... Let be the predicted probability of the i-th preferred classifier.

[0040] In another preferred embodiment, the final prediction probability is calculated as follows: , Wherein, p1 is the predicted probability of the classifier combining (1) the ANOVA T-value selection method and logistic regression; p2 is the predicted probability of the classifier combining (2) the ANOVA T-value selection method and support vector machine; p3 is the predicted probability of the classifier combining (3) the ANOVA F-value selection method and logistic regression; p4 is the predicted probability of the classifier combining (4) the trace ratio criterion method and logistic regression; p5 is the predicted probability of the classifier combining (5) the Fisher score method and logistic regression; p6 is the classifier combining (6) the Fisher score method and support vector machine; p7 is the predicted probability of the classifier combining (7) the unsupervised discriminant feature selection method and logistic regression; p8 is the predicted probability of the classifier combining (8) the Laplace score method and logistic regression; p9 is the predicted probability of the classifier combining (9) the Relief-F algorithm and balanced random forest; and p10 is the predicted probability of the classifier combining (10) the multi-cluster feature selection method and random forest.

[0041] In another preferred embodiment, the weighted dynamic fusion prediction model is applicable to ovarian cancer patients with a platinum-free interval (PFI) of < 6 months or ≥ 12 months.

[0042] A second aspect of the present invention provides a disease typing / prediction system, the prediction system comprising: (M1) Input module, the input module is configured to input data, the data including the original multidimensional feature dataset of the test object, the original multidimensional feature dataset including clinical baseline dataset and laboratory test dataset; (M2) A typing / prediction module, comprising: (m2.1) A prediction unit configured as a prediction model for the disease, wherein the prediction model predicts the test object based on the original multidimensional feature dataset to obtain a final prediction probability; the prediction model is constructed by the method described in the first aspect of the present invention. (m2.2) A discrimination unit, configured to compare the final predicted probability of the object to be tested with a predetermined classification threshold to obtain a prediction result; (M3) Output module, which is configured to output the prediction results of the prediction module.

[0043] In another preferred embodiment, the typing / prediction system is used to predict whether the subject is a platinum-resistant ovarian cancer patient.

[0044] In another preferred embodiment, the typing / prediction system is used to type / predict ovarian cancer patients with a platinum-free interval (PFI) of < 6 months or ≥ 12 months.

[0045] In another preferred embodiment, the subjects to be tested include patients suspected of having platinum-resistant ovarian cancer.

[0046] In another preferred embodiment, the subject of the test is an ovarian cancer patient receiving different treatment regimens.

[0047] In another preferred embodiment, the ovarian cancer patients receiving different treatment regimens include patients who have undergone initial cytoreductive surgery and patients who have received neoadjuvant chemotherapy.

[0048] In another preferred embodiment, in the identification unit, if the final predicted probability is greater than the classification threshold, it indicates that the subject to be tested is a platinum-resistant ovarian cancer patient; otherwise, the subject to be tested is a platinum-sensitive ovarian cancer patient.

[0049] It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be described in detail here. Attached Figure Description

[0050] Figure 1 The flowchart of the construction and verification of the dynamic weighted fusion model proposed in this invention is shown.

[0051] Figure 2 This graph shows a performance comparison between the DWF model and the top ten base classifiers in this embodiment of the invention. The center point represents the mean, and the horizontal error bars represent the 95% confidence interval.

[0052] Figure 3 The confusion matrices of the DWF model and the LR baseline model are shown. (A) Confusion matrix of the DWF model constructed in this study; (B) Confusion matrix of the LR baseline model.

[0053] Figure 4 This diagram shows a comparison of the prediction performance of each base classifier in an example of the present invention. Each square in the grid corresponds to a unique combination of models, and its performance is evaluated using AUC, presented in both numerical form and color gradient.

[0054] Figure 5 This chart summarizes the Shapley Additive exPlanations (SHAP) analysis of the top 20 most important features in this embodiment of the invention. Each point represents a sample. The horizontal axis position represents the SHAP value; positive values ​​shift the prediction towards drug resistance, while negative values ​​shift it towards drug sensitivity. The color represents the feature value, transitioning from low (blue) to high (red).

[0055] Figure 6The KM survival curves of the PFI in the entire cohort of ovarian cancer patients predicted by the DWF model are shown.

[0056] Figure 7 The KM survival curves of PFI in ovarian cancer patients in the PDS subgroup predicted by the DWF model are shown.

[0057] Figure 8 The KM survival curves of PFI in ovarian cancer patients in the NAC subgroup predicted by the DWF model are shown. Detailed Implementation

[0058] Through extensive and in-depth research, the inventors have developed for the first time a Dynamically Weighted Fusion (DWF) model for disease prediction. This model employs the Borderline-SMOTE method during data preprocessing, effectively addressing the common problem of positive-negative sample imbalance in medical scenarios. During model training, the model incorporates multiple basic classifiers composed of different feature selection algorithms and machine learning algorithms. Training data is used to train and filter each basic classifier to obtain an optimal classifier. The comprehensive score of each optimal classifier is evaluated, and a Softmax function with a temperature coefficient is used to measure the weight of each optimal classifier. Dynamically weighted fusion is then performed based on the weights and the predicted values ​​of each optimal classifier to obtain the DWF model for the corresponding disease. This DWF model outputs the final risk probability obtained from the dynamically weighted fusion. This final risk probability is compared with the optimal classification threshold to predict the subject's disease. Based on this, the present invention is completed.

[0059] It should be understood that the specific methods and experimental conditions of the invention described below in varying degrees of detail are intended to provide a substantive understanding of the invention. Definitions of certain terms used in this specification are provided below. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0060] the term As used herein, the terms “containing” or “including (comprise)” can be open-ended, semi-closed, or closed-ended. In other words, the terms also include “consistently made of” or “made of”.

[0061] As used herein, the term “and / or” refers to and covers any and all possible combinations of one or more of the related listed items.

[0062] As used in this article, the term "significant" means that, in hypothesis testing, the observed effect (such as the difference between the experimental and control groups) is unlikely to be caused solely by random error. A hypothesis test includes: the null hypothesis (H0), which assumes that the observed effect does not exist (such as there is no difference between the experimental and control groups); p The value, i.e., the probability of observing the current or more extreme effect when H0 is true; and the significance threshold ( α The significance of a hypothesis test is usually determined by a significance threshold, which is generally set at 0.05. p If the value is ≤ α, then H0 is rejected, meaning the observed effect exists, and the result is called "significant".

[0063] As used in this article, the term "imbalanced positive and negative samples" refers to a significant difference in the number of samples between different classes (usually binary) in the training data for machine learning classification tasks, where the number of samples in one class (negative samples or majority class) is much larger than that in the other class (positive samples or minority class). This imbalance causes the trained model to be biased in predicting the majority class, while its ability to identify the minority class is poor, thus affecting the overall performance and generalization ability of the model.

[0064] As used in this article, the terms "platinum-free interval" and "PFI" are used interchangeably. They are important clinical indicators in oncology for assessing the efficacy and resistance of platinum-based chemotherapy, specifically referring to the time from the end of the last platinum-based chemotherapy session to disease recurrence or progression. A PFI ≥ 12 months is generally classified as platinum-sensitive, suggesting that re-administration of platinum-based regimens may still be effective. A PFI < 6 months is classified as platinum-resistant, indicating a significantly reduced objective response rate with re-administration of platinum-based regimens; clinical practice often involves switching to non-platinum-based drugs or targeted therapy. Generally speaking, the longer the PFI, the greater the likelihood that the patient will benefit from subsequent platinum-based retreatment.

[0065] The weighted dynamic fusion prediction model of the present invention This invention provides a method for constructing a weighted dynamic fusion prediction model for predicting diseases. The prediction model constructed using this method is particularly suitable for original training data with imbalanced positive and negative samples. Specifically, the method includes the following steps: (S1) Provide an original multidimensional feature dataset; the original multidimensional feature dataset comes from a disease patient group, and the original multidimensional feature dataset includes a clinical baseline dataset and a laboratory test dataset; (S2) Preprocess the original multidimensional feature dataset to obtain a preprocessed multidimensional feature dataset; (S3) Construct a basic classifier library, which includes multiple basic classifiers consisting of M feature selection algorithms and N machine learning algorithms; where M and N are both integers > 1; (S4) Train each basic classifier in the basic classifier library using the preprocessed multidimensional feature dataset to obtain the prediction probability of each basic classifier; during the training process, evaluate each basic classifier according to the geometric mean and select a predetermined number of basic classifiers as preferred classifiers. (S5) Evaluate the overall score of the preferred classifier using its AUC and geometric mean; (S6) Evaluate the weight of each preferred classifier based on the comprehensive score of each preferred classifier and the Softmax function with a predetermined temperature coefficient; (S7) The weights and prediction probabilities of each preferred classifier are summed to obtain the final prediction probability, thereby constructing a weighted dynamic fusion prediction model for the disease.

[0066] Preferably, the disease includes cancer. Preferably, the cancer includes ovarian cancer. Those skilled in the art will understand that the technical solution according to the present invention can be used to construct models for predicting other diseases (such as other cancers / tumors).

[0067] Preferably, the original multidimensional feature dataset comes from a disease patient population and an optional healthy population. Preferably, the disease patient population includes platinum-resistant and sensitive ovarian cancer patients. It is understood that a model trained using data from the platinum-resistant and sensitive ovarian cancer patients can be used to classify the drug sensitivity type of test subjects who have or are suspected of having ovarian cancer. It is further understood that a model trained using data from the disease patient population and the healthy population can be used for disease diagnosis.

[0068] Preferably, the original multidimensional feature dataset is selected from the group consisting of: imbalanced positive and negative sample data, or balanced positive and negative sample data. Preferably, the original multidimensional feature dataset is imbalanced positive and negative sample data.

[0069] Preferably, in step (S2), the preprocessing includes methods selected from the group consisting of: handling outliers, imputing missing data, correcting data distribution, standardizing data, and / or handling imbalanced positive and negative sample data. Preferably, the method for imputing missing data includes Multivariate Chain Equation Imputation (MICE). Preferably, the method for correcting data distribution includes Box-Cox transformation. Preferably, the method for standardizing data includes min-max normalization. Preferably, the Borderline-SMOTE technique is used to handle imbalanced positive and negative sample data. The experiments in Example 8 demonstrate that the Borderline-SMOTE method in the preprocessing step is particularly advantageous for training with imbalanced positive and negative sample training data; when training with balanced positive and negative sample training data, the method of this invention can also obtain a high-performance prediction model.

[0070] Preferably, the basic classifier library includes M×N basic classifiers. Preferably, the feature selection algorithm includes, but is not limited to: F-score (Analysis of Variance), T-score (Analysis of Variance), Two-Input Symmetric Correlation (DISR), Fisher score, Interaction Upper Cap (ICAP), Joint Mutual Information (JMI), Laplacian score, Logistic Loss L2,1 Norm Minimization (LL l21), Least Squares Loss L2,1 Norm Minimization (LS l21), Multi-Cluster Feature Selection (MCFS), Non-Negative Discriminant Feature Selection (NFDS), Relief-F algorithm, Trace ratio, and Unsupervised Discriminant Feature Selection (UDFS). Preferably, the machine learning algorithms include, but are not limited to: Adaptive Boosting (AdaBoost), Balanced Random Forest (Balanced RF), Category Feature Gradient Boosting (CatBoost), Decision Tree (DT), ExtraTrees, Gradient Boosting, K-Nearest Neighbors (KNN), Lightweight Gradient Boosting (LGBM), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost).

[0071] Preferably, in step (S5), the comprehensive score is calculated as follows: , Let AUC be the value of the i-th preferred classifier. Let be the geometric mean of the i-th preferred classifier. and Let be the weighting coefficient, satisfying .

[0072] Preferably, in step (S6), the weights are calculated as follows: ,in, For the first i The combined score of the selected classifiers, where K is the number of selected classifiers and T is the temperature coefficient, satisfying 0 < K. T < 1. The experiment in Example 9 demonstrates that the model containing the "softmax function based on temperature coefficient" outperforms the model without the softmax function and the simple average model in both accuracy and specificity.

[0073] Preferably, in the model trained using data from platinum-resistant and susceptible ovarian cancer patients, the preferred classifiers are: (1) a classifier combining ANOVA T-score selection and logistic regression; (2) a classifier combining ANOVA T-score selection and support vector machine; (3) a classifier combining ANOVA F-score selection and logistic regression; (4) a classifier combining trace ratio criterion and logistic regression; (5) a classifier combining Fisher score and logistic regression; (6) a classifier combining Fisher score and support vector machine; (7) a classifier combining unsupervised discriminative feature selection and logistic regression; (8) a classifier combining Laplace score and logistic regression; (9) a classifier combining Relief-F algorithm and balanced random forest; and (10) a classifier combining multi-cluster feature selection and random forest.

[0074] Preferably, the final prediction probability of the model is calculated as follows: , Wherein, p1 is the predicted probability of the classifier combining (1) the ANOVA T-value selection method and logistic regression; p2 is the predicted probability of the classifier combining (2) the ANOVA T-value selection method and support vector machine; p3 is the predicted probability of the classifier combining (3) the ANOVA F-value selection method and logistic regression; p4 is the predicted probability of the classifier combining (4) the trace ratio criterion method and logistic regression; p5 is the predicted probability of the classifier combining (5) the Fisher score method and logistic regression; p6 is the classifier combining (6) the Fisher score method and support vector machine; p7 is the predicted probability of the classifier combining (7) the unsupervised discriminant feature selection method and logistic regression; p8 is the predicted probability of the classifier combining (8) the Laplace score method and logistic regression; p9 is the predicted probability of the classifier combining (9) the Relief-F algorithm and balanced random forest; and p10 is the predicted probability of the classifier combining (10) the multi-cluster feature selection method and random forest.

[0075] The system of the present invention This invention provides a disease typing / prediction system. The system can be embodied as an electronic device, including but not limited to: smartphones, tablets, personal computers, servers, terminals, or other intelligent terminals with data processing capabilities. The system can also be embodied as a computer-readable storage medium, such as a disk, optical disk, solid-state drive, read-only memory, or flash memory, on which a computer program is stored. When the program is executed by one or more processors, it can implement the data processing flow. Furthermore, the system can also be embodied as a computer program product containing computer instructions that, when executed by a computer, cause the computer to perform all or part of the steps of the data processing flow.

[0076] The data processing flow can be loaded, deployed, or run on any of the aforementioned electronic devices. Through the collaboration of hardware resources and the software logic defined in the flow, the system can complete specific data acquisition, transmission, calculation, analysis, storage, or presentation tasks.

[0077] The main advantages of this invention include: (1) This invention integrates the advantages of multiple strong classifiers through dynamic screening and a weighting strategy with a temperature coefficient Softmax function, and constructs a dynamic weighted fusion model, which significantly improves AUC and accuracy, and is superior to a single machine learning model.

[0078] (2) The model of the present invention adopts the Borderline-SMOTE algorithm in the data preprocessing step, which can be applied to the prediction of imbalanced data queues and effectively improves the ability to identify minority classes.

[0079] (3) The model of the present invention only requires input of routine blood test data, without the need for expensive gene sequencing or complex pathological image analysis. It is easy to embed into existing hospital information systems, has controllable costs, and has extremely high clinical promotion value.

[0080] The present invention will be further illustrated below with reference to specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. It is also understood that the purpose of describing the present invention in conjunction with the embodiments is to cover other options or modifications that may be derived based on the claims of the present invention. To provide a deep understanding of the invention, many specific details will be included in the following description. The invention may also be practiced without using these details. Furthermore, to avoid confusion or obscuring the focus of the invention, some specific details will be omitted in the description.

[0081] Example 1: Construction of a Dynamic Weighted Fusion Model This embodiment relates to a construction process and method for a dynamic weighted fusion model. Figure 1 ).

[0082] (S1) Data Acquisition: Collect clinical baseline characteristics (such as age, BMI, etc.) and routine laboratory test data (such as blood routine, biochemical indicators, tumor markers, etc.) of the target patient group to form the original multidimensional feature dataset.

[0083] (S2) Data Preprocessing: Given that medical data often suffers from high noise, dimensionality heterogeneity, and class imbalance, a standardized preprocessing pipeline needs to be established. This includes: (1) Use the truncation method with clinical reference range to handle outliers; (2) For missing data, it is preferable to use the multivariate imputation by chained equations (MICE) algorithm, which utilizes the correlation between features, rather than simple mean imputation, in order to preserve data distribution information.

[0084] (3) Perform Box-Cox transformation on the non-normal distribution characteristics to make them approximately conform to a Gaussian distribution; (4) Feature normalization: Min-Max normalization is used to scale all features to the [0, 1] interval; (5) To address the common problem of imbalance between positive and negative samples in medical scenarios, the Borderline-Synthetic Minority Oversampling Technique (Borderline-SMOTE) is used during the training phase to generate synthetic samples near the decision boundary, thereby balancing the ratio of positive to negative samples and preventing the model from being biased towards the majority class.

[0085] (S3) Constructing a Basic Classifier Library: To capture complex nonlinear patterns in the data, this invention constructs a large-scale heterogeneous model library. Specifically, M feature selection algorithms with different principles are selected to filter feature subsets from the full set of features; each feature subset is then input into N different machine learning classification algorithms for training. This results in a total of M×N differentiated basic classifiers. These classifiers predict the same sample from different feature subspaces and decision logics, ensuring the diversity of the ensemble.

[0086] (S4) Dynamic Evaluation and Selection: In each fold of cross-validation, the performance of all base classifiers on the validation set is dynamically evaluated. G-means is used as the core evaluation metric to balance the model's ability to identify the majority and minority classes. Based on the evaluation results, the top K (Top-K) base classifiers are dynamically selected as a candidate set, and poorly performing models are eliminated. The predicted probability of each candidate base classifier is then obtained. .

[0087] (S5) Calculate the combined score: For each base classifier obtained from the screening, calculate the combined score (CS) of the K base classifiers based on the Area Under the Curve (AUC) and the Geometric Mean (G-mean). The formula for calculating CS is: , in, For the first The overall score of the basic classifiers and These are the weighting coefficients, and α + β = 1. Preferably, α =0.6, β =0.4.

[0088] (S6) Temperature Coefficient-Based Weighted Fusion: To further enhance the contribution of the high-performance model, this invention introduces a weighted fusion method based on the temperature coefficient. T The Softmax function is used to calculate the weights of candidate classifiers. First, the CS (Systemic Comparison) of each model is calculated based on AUC and G-mean. Then, the CS distribution is sharpened using a temperature coefficient T, giving the best-performing model significantly higher weights. The formulas for calculating the weights of each base classifier are as follows: , in, For the first The final weights of the base classifiers K The number of base classifiers selected. T It is a temperature coefficient, and 0 < T < 1. Preferably, the temperature coefficient T Setting it to 0.1, by setting a low temperature coefficient, makes the weight distribution more acute, thereby significantly amplifying the influence of the high-performance base classifier, while retaining the weak contributions of other models.

[0089] Fusion prediction probability, i.e., final risk probability .

[0090] (S7) Threshold Optimization and Result Output: The predicted probabilities of the K candidate classifiers are weighted and summed using the calculated weights to obtain the final fusion probability. Then, the optimal classification threshold that maximizes G-mean is searched on the ROC curve. Based on this threshold, the final binary classification prediction result, i.e., the fused prediction probability, is output. > If the event occurs at a certain time, it is considered a positive event; otherwise, it is considered a negative event.

[0091] Example 2: Dynamic weighted fusion model of platinum-based drug resistance in ovarian cancer This embodiment uses a cohort of ovarian cancer patients with platinum-based drug resistance and sensitivity as an example to illustrate the construction process of the dynamic weighted fusion model of the present invention.

[0092] (1) Research cohort: Inclusion criteria: All of the following conditions must be met for enrollment: ① confirmed by histopathological examination as primary ovarian cancer; ② received a standard first-line combination chemotherapy regimen based on platinum drugs; ③ have complete and standardized baseline routine laboratory test data before initial treatment upon admission.

[0093] Exclusion criteria: Patients meeting any of the following criteria were excluded: ① Patients with a history of other malignant tumors within 6 months prior to enrollment, or patients who were pregnant; ② Patients with incomplete follow-up data, incomplete records of core clinical outcomes (relapse time, survival status, etc.), or who could not complete the platinum resistance assessment; ③ Patients with a PFI of 6-12 months.

[0094] Baseline data were collected from 322 ovarian cancer patients, including 78 platinum-resistant (positive sample) and 244 platinum-sensitive (negative sample).

[0095] (2) Study cohort grouping: This study uses PFI as the core grouping basis. PFI is defined as the time interval from the end of the last platinum-based chemotherapy to the recurrence of ovarian cancer. The final included study cohorts were divided into the following two groups: Platinum-resistant group (PFI < 6 months): This endpoint is a recognized clinical surrogate indicator for primary platinum resistance and is for high-risk individuals with no sustained response to first-line platinum-based chemotherapy and rapid disease progression; Platinum-sensitive group (PFI ≥ 12 months): This is for individuals with a sustained and significant clinical response to first-line platinum-based chemotherapy.

[0096] In the traditional classification system for ovarian cancer clinical diagnosis and treatment, patients with a platinum-sensitive index (PFI) of 6–12 months are usually defined as a partial platinum-sensitive population. To improve the predictive efficacy of the model, this study actively excluded this intermediate phenotype population during the data preprocessing stage of machine learning model construction.

[0097] (3) Data collection: Clinical baseline data and laboratory test results were collected from the study cohort.

[0098] Clinical baseline data: age and BMI were used as the two characteristics. Data completeness for both characteristics was 100%.

[0099] Laboratory test indicators include four dimensions: complete blood count and systemic immune / inflammatory markers, liver and kidney function and biochemical metabolic profile, gynecological tumor marker combination, and coagulation and fibrinolysis function profile. The data completeness of each indicator in all four dimensions is above 90%. ① Complete blood count and systemic immune / inflammatory markers: white blood cell count (WBC), lymphocyte percentage (LYM), intermediate cell percentage (MID), neutrophil percentage (GRAN), eosinophil percentage (EOS), basophil percentage (BASO), lymphocyte count (LYM#), intermediate cell count (MID#), neutrophil count (GRAN#), eosinophil count (EOS#), basophil count (BASO#), red blood cell count (RBC), hemoglobin (HGB), hematocrit (HCT), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), coefficient of variation of red blood cell distribution width (RDWCV), standard deviation of red blood cell distribution width (RDWSD), platelet count (PLT), mean platelet volume (MPV), platelet hematocrit (PCT), platelet distribution width (PDW), neutrophil to lymphocyte ratio (NL).

[0100] ② Liver and kidney function and biochemical metabolic profile: Total bilirubin (TBIL), direct bilirubin (DBIL), indirect bilirubin (IBIL), alkaline phosphatase (ALP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), aspartate aminotransferase / alanine aminotransferase ratio (ASTM), lactate dehydrogenase (LDH), gamma-glutamyl transferase (GGT), total protein (TP), albumin (ALB), globulin (GELO), albumin / globulin ratio (A / G), non-esterified fatty acids (NEFA), prealbumin (PA), glutamate dehydrogenase (GLDH), cystatin C (CYSC), glycated albumin (GA), urea (UREA), creatinine (CRE), uric acid (UA), glucose (GLU), calcium (CA), phosphorus (P), magnesium (MG), potassium (K), sodium (NA), chloride (CL), and total carbon dioxide (TCO2).

[0101] ③ Gynecologic tumor marker combination: squamous cell carcinoma antigen (SCCA), carbohydrate antigen 19-9 (CA19-9), carbohydrate antigen 125 (CA125), alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), neuron-specific enolase (NSE), human chorionic gonadotropin β subunit (β-HCG), and human epididymal protein 4 (HE4).

[0102] ④ Coagulation and fibrinolysis functional spectrum: prothrombin time (PT), fibrinogen (FIB), activated partial thromboplastin time (APTT), thrombin time (TT), international normalized ratio (INR), D-dimer (DDI), fibrin degradation products (FDP).

[0103] (4) Data preprocessing: A standardized and normalized data preprocessing process was designed and implemented for the 70 clinical and laboratory features included.

[0104] ① The MICE algorithm was used for refined imputation of missing values. Specifically, this study employed an iterative chain imputator with a maximum of 30 iterations. The core mechanism of the MICE algorithm is as follows: for each variable with missing values, it is used as the dependent variable, and a regression model is constructed using all other variables with complete observations in the dataset as predictors; this chain iterative process is executed cyclically among all variables with missing values ​​until the imputation results converge.

[0105] ② Clinically driven outlier truncation: This study adopted a truncation strategy based on clinical reference range: extreme test values ​​that exceed the upper limit of the normal physiological reference range were uniformly truncated to the upper limit threshold of the clinical reference range of the indicator.

[0106] ③ Data standardization: Considering the significant differences in the dimensions and numerical scales of indicators from different laboratories, this study uses Min-Max Normalization to process all continuous features, mapping all continuous variables uniformly to the dimensionless interval [0, 1].

[0107] ④ Data distribution correction: This study introduces Box-Cox transformation for data distribution correction. It determines the optimal continuous transformation parameters through maximum likelihood estimation and performs a smooth transformation on non-normally distributed data to make it as close as possible to the normal (Gaussian) distribution.

[0108] ⑤ Imbalanced Sample Handling: This study introduces the Borderline-SMOTE algorithm for data equalization. The Borderline-SMOTE algorithm precisely focuses on minority class samples near the classification decision boundary, generating synthetic samples through K-nearest neighbor interpolation on these most easily misclassified boundary samples. This study strictly limits the execution scope of the Borderline-SMOTE algorithm to the training iterations within the cross-validation loop. That is, in each round of cross-validation iteration, the validation set always retains the original real-world class distribution characteristics and does not participate in the generation of synthetic samples.

[0109] (6) Construction of basic classifier library: In order to fully capture the multi-dimensional feature patterns of clinical data, from simple linear correlation to complex nonlinear biological interactions, this study constructed a basic classifier library with high diversity and heterogeneity.

[0110] Specifically, this study selected 14 feature selection algorithms, including: F-score (Analysis of Variance), T-score (Analysis of Variance), Two-Input Symmetric Correlation (DISR), Fisher's Score, Interaction Upper Cap (ICAP), Joint Mutual Information (JMI), Laplacian Score, Logistic Loss L2,1 Norm Minimization (LL l21), Least Squares Loss L2,1 Norm Minimization (LS l21), Multi-cluster Feature Selection (MCFS), Non-negative Discriminant Feature Selection (NFDS), Relief-F Algorithm (reliefF), Trace Ratio, and Unsupervised Discriminant Feature Selection (UDFS); and 12 machine learning algorithms: Adaptive Boosting Algorithm (AdaBoost), Balanced Random Forest (Balanced RF), Category Feature Gradient Boosting Algorithm (CatBoost), Decision Tree (DT), Extra Trees, and Gradient Boosting Machine (Gradient). The algorithms used include Boosting, K-Nearest Neighbors (KNN), Lightweight Gradient Boosting Machine (LGBM), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting Machine (XGBoost). These 12 machine learning algorithms cover four modeling paradigms: linear models, distance-based models, tree-based models, and the Boosting family of algorithms. Ultimately, 168 base classifiers were obtained.

[0111] (7) Screening of base classifiers and calculation of ensemble score (CS): G-mean was used as the core evaluation index for the initial screening of base classifiers in the outer validation set. In each fold cross-validation, the top 10 base classifiers with the highest G-mean were selected as the final ensemble members for that fold, and the predicted probability of each base classifier was obtained. Preferably, the number of preferred classifiers is set to 10. This value is the optimal hyperparameter determined through hyperparameter tuning (grid search method) after comprehensively considering the diversity of model integration and computational complexity. Finally, it is determined using the formula... Calculate the CS scores for these 10 base classifiers.

[0112] (8) Dynamic weighted fusion: incorporating temperature coefficients T The Softmax function is introduced into the weight allocation mechanism of clinical prediction models. The final weights of the base classifiers The calculation formula is: .

[0113] After parameter optimization, this study sets the temperature coefficient to a low value (T = 0.1). A low temperature coefficient can effectively sharpen the probability distribution of weight allocation, forming a non-absolute polarized weighting mode: it gives the best-performing base model the core decision-making power, while avoiding information loss caused by hard truncation. By retaining the supplementary contribution of the suboptimal model, it maintains the diversity and fault tolerance of the integrated system.

[0114] Ultimately, the fusion prediction probability of the DWF model The predicted probabilities of the above 10 basic classifiers The calculation formula is obtained through dynamic weighted summation: .

[0115] (9) Optimization calibration of classification threshold: By traversing all possible intercepts of the receiver operating characteristic curve (ROC) on the validation set, determine the threshold that allows for optimal classification. G-mean Maximize the optimal classification threshold When the patient's comprehensive risk score > At that time, it was predicted to be drug resistant (positive class).

[0116] To provide the best embodiment of the present invention and enable the model to be directly applied to actual clinical deployment, the present invention further utilizes the full cohort test sample set (containing all platinum-resistant and platinum-sensitive ovarian cancer patient samples) for final feature extraction and model training.

[0117] By applying the dynamic evaluation and screening method described in this invention to the entire queue of data, the top 10 preferred classifiers in terms of CS ranking were finally determined. Subsequently, according to the invention, the classifiers with temperature coefficients (…) T The deterministic weights of the 10 preferred classifiers in the final DWF model were calculated using the Softmax function (=0.1).

[0118] Specifically, the configuration parameters of the optimal embodiment of the DWF model are shown in Table 1. In this optimal embodiment, the final prediction probability of the DWF model is... It is obtained by multiplying the probability values ​​predicted by the above 10 basic classifiers for their respective selected feature subsets by the corresponding final fusion weights in Table 1, and then summing them up by weight.

[0119] Table 1. Base classifier configuration and weight allocation for the best implementation of the DWF prediction model Example 3: Performance Evaluation of the DWF Model This embodiment involves evaluating and comparing the performance of a platinum-resistant DWF prediction model for ovarian cancer, the top 10 baseline classifiers, and the clinically applicable benchmark (LR) model using ROC curve area under the curve (AUC), accuracy, sensitivity, and specificity metrics.

[0120] The results are as follows Figure 2 As shown, in the task of predicting platinum resistance in ovarian cancer, the DWF model constructed in this invention achieved an excellent performance of AUC 0.760 (95% CI: 0.683-0.837), which is significantly better than all single basic classifiers. While maintaining competitive sensitivity (0.736), it also achieved high specificity (0.780), which has good clinical significance and avoids misjudgment leading to unnecessary overtreatment.

[0121] Furthermore, the DWF model constructed in this invention was compared with the LR model, which is currently the most widely used model in ovarian cancer drug resistance research. This LR model uses... Figure 2 The optimal single-classifier feature engineering strategy (based on the top 10 core features selected by T score) has been proven in China. Performance evaluation results show that, limited by the inherent linear assumption, the predictive performance of the LR baseline model has significant limitations, with an AUC of only 0.660 (95% CI: 0.594~0.722) and an accuracy reduced to 0.602 (95% CI: 0.547~0.655).

[0122] Furthermore, by merging the confusion matrix of the 5-fold cross-validation results, the differences in the distribution of misclassified samples between the DWF model and the LR baseline model were systematically analyzed. Figure 2 ).

[0123] like Figure 3 As shown in Figure A, the DWF model exhibits highly consistent diagnostic efficacy across different risk populations: among 231 platinum-sensitive patients, the model correctly identified 180 cases, with a true negative rate (specificity) of 0.780; among 91 platinum-resistant patients, the model accurately predicted 67 cases, with a true positive rate (sensitivity) of 0.736. In comparison, Figure 3 B showed that the LR model had significant decision bias, with a specificity of only 0.589; 95 platinum-sensitive patients were incorrectly identified as drug-resistant by the model (false positives), accounting for 41.1%, while the number of false positives in the DWF model dropped to 51, accounting for only 22.1%.

[0124] These results demonstrate that traditional LR models suffer from a high risk of misdiagnosis due to their difficulty in fitting the decision boundaries of high-dimensional and complex data. In contrast, the DWF model, through its ensemble architecture, achieves refined optimization of the decision boundaries, significantly improving model specificity from 0.589 to 0.780. This performance improvement not only effectively avoids over-intervention or downgrading of treatment regimens in platinum-sensitive populations but also minimizes false positives and false negatives while enhancing the overall diagnostic efficacy of the model. From a clinical net benefit perspective, the high specificity and low misdiagnosis rate achieved by the DWF model through algorithmic optimization provide a superior benefit-risk ratio for its widespread application in real-world clinical scenarios.

[0125] Example 4: Performance Evaluation of the Basic Classifier This study comprehensively validates 168 independent base classifiers constructed by cross-combining 14 feature selection strategies with 12 mainstream machine learning algorithms within an internal cross-validation framework; based on the AUC of each model on the validation set, a global heatmap of the predictive performance of the base classifiers is plotted. Figure 4 ).

[0126] Depend on Figure 4 As shown in the matrix heatmap, there are significant differences in the predictive performance of different algorithm combinations in the task of identifying platinum-based drug resistance in ovarian cancer. From the overall distribution, the AUC values ​​of most single-model combinations are concentrated in the inefficient range of 0.500–0.600. After screening and comparison, among 168 base classifiers, the algorithm combination of T score and LR has the best predictive performance, but its peak AUC is only 0.656, confirming the limitations of a single algorithm.

[0127] The low-value regions in the heatmap (white blocks with AUC < 0.500) further reveal the inherent limitations of a single machine learning modeling paradigm when processing high-dimensional and complex clinical data. In the binary classification model evaluation system, a validation set AUC < 0.500 indicates that the classifier's discriminative power is lower than that of random guessing. This phenomenon suggests that the combination of some feature selection methods with the classifier fails to effectively capture the true biological differences between platinum resistance and susceptibility phenotypes. Instead, it may overfit to local noise or spurious correlations in the data, ultimately leading to systematic prediction biases in the model's decision boundary.

[0128] Example 5: Interpretability Analysis of the DWF Model To intuitively quantify the marginal contribution of each clinical biochemical indicator to the model's final prediction results, this study extracted the top 20 core features in terms of their contribution to the model's predictions and plotted a SHAP global summary diagram. Figure 5 ).

[0129] The distribution range of SHAP values ​​shows that the SHAP values ​​of all variables are concentrated in a narrow range of -0.06 to 0.06, suggesting that when the DWF model in this study determines the risk of platinum resistance in ovarian cancer, it does not rely on a single dominant variable or decisive feature, but rather accurately captures the weak signals of multiple conventional biomarkers, integrates the complex synergistic effects between various indicators, and finally constructs the model decision boundary.

[0130] SHAP analysis clarified the association between each indicator and the risk of platinum-based drug resistance, which highly aligns with the pathophysiological evolution of the tumor microenvironment. Taking the inflammation-related indicators (NL, GRAN, PLT) and coagulation indicators (FIB) that ranked highest in model contribution as examples, their high-value samples (red data points) were mainly concentrated on the right side of the vertical axis (SHAP>0), suggesting that increased systemic inflammatory burden and hypercoagulable state are core positive risk factors for platinum-based drug resistance. Conversely, low-value samples of LYM (blue data points) were largely clustered on the right side of the vertical axis, suggesting that immunosuppression (i.e., exhaustion of anti-tumor effector cell function) also significantly increases the risk of platinum-based drug resistance in patients.

[0131] Example 6: Risk Stratification of the Entire Cohort Based on DWF Model Scoring To further evaluate the application value of the DWF model in clinical prognostic assessment and survival prediction, this embodiment stratifies the entire cohort of patients based on the prediction results of the model and conducts survival analysis on the PFI of the stratified population.

[0132] Based on the discriminant results of the DWF model, the entire cohort of 322 patients was divided into a predicted platinum sensitivity group (n=204) and a predicted platinum resistance group (n=118). KM survival curves ( Figure 6 The results showed that there was a highly statistically significant difference in the PFI survival trajectories between the two groups of patients divided by the model (P<0.0001).

[0133] Specifically, for patients identified by the model as predicting drug resistance, the survival curve showed a sharp downward trend in the early follow-up period. The median progression-free survival (PFI) in this group was only 5.2 months, and 92 patients experienced endpoint events during the follow-up period, with an event incidence rate of 78.0%, indicating an extremely high risk of early relapse and disease progression in this population. In contrast, for patients identified by the model as being sensitive to prediction, the survival curve declined more gradually, demonstrating a significant long-term survival benefit, with the median PFI significantly prolonged to 36.0 months.

[0134] Notably, the median PFI predicted by the model for both the resistant and sensitive groups (5.2 months vs. 36.0 months) closely matches the median PFI observed during actual follow-up (3.0 months vs. 38.5 months) in both trend and magnitude. This survival analysis confirms the good clinical prognostic value of the DWF model, achieving accurate and robust prognostic risk stratification across the entire cohort. This stratification system is expected to assist clinicians in accurately identifying patients at high risk of resistance in the early stages of treatment, providing crucial decision-making support for the development of individualized maintenance therapy and follow-up strategies.

[0135] Example 7: Subgroup predictive efficacy and risk stratification validation under different treatment modalities To further verify the robustness and universality of the DWF model's predictive efficacy in different clinical scenarios, this embodiment divides the entire cohort into a primary debulking surgery (PDS) subgroup and a neoadjuvant chemotherapy (NAC) subgroup based on the patients' initial treatment regimen. The diagnostic predictive efficacy and longitudinal prognostic risk stratification ability of the model in the two subgroups are then explored.

[0136] The results are shown in Table 2. The DWF model demonstrated stable and excellent discrimination ability in both subgroups. Its AUC in the PDS subgroup and NAC subgroup were 0.755 (95% CI: 0.641~0.870) and 0.761 (95% CI: 0.659~0.864), respectively, indicating that the overall predictive and discriminative power of the model was not affected by the difference in treatment regimens.

[0137] Further analysis of the diagnostic indicators revealed differentiated diagnostic advantages across different subgroups: in the PDS population, the model exhibited higher overall accuracy (0.795) and specificity (0.829); while in the NAC subgroup, the model demonstrated superior ability to identify high-risk patients with drug resistance, achieving a sensitivity of 0.831. Notably, regardless of the treatment regimen subgroup, the model maintained high negative predictive value (NPV), at 0.879 in the PDS group and 0.891 in the NAC group, indicating good reliability in accurately identifying platinum-sensitive true negative patients and effectively reducing overtreatment due to misdiagnosis. Furthermore, the wide 95% CI range for some performance indicators is mainly related to the reduced sample size after subgroup division, but it did not affect the robustness of the model's overall predictive performance.

[0138] Table 2. Predictive efficacy of the DWF model under different initial treatment regimens. Note: Data are expressed as mean (95% CI).

[0139] In the PDS subgroup (n=209), the DWF model divided patients into a prediction-sensitive group (n=145) and a prediction-resistant group (n=64). KM survival analysis results ( Figure 7 The results showed a highly significant difference in prognostic risk index (PFI) survival between the two patient groups (P < 0.0001). In the predictive sensitive group, the survival curve remained consistently high throughout the follow-up period, demonstrating excellent long-term survival benefit, with a median PFI of NR. In contrast, among the 64 patients in the predicted resistant group, 47 experienced endpoint events, indicating rapid disease progression, with a median PFI of only 5.1 months. This significant difference in median PFI between the two groups confirms that the DWF model possesses excellent capabilities in identifying high-risk populations and stratifying prognostic risk in patients receiving standardized PDS treatment.

[0140] In the NAC subgroup (n=113), the DWF model also demonstrated stable and excellent risk stratification performance. Figure 8 The model divided NAC patients into a predictive sensitive group (n=59) and a predictive resistant group (n=54). Survival analysis showed that even in the NAC population with a poor overall prognostic baseline, the PFI predicted by the model in the sensitive group was significantly better than that in the resistant group. The median PFI in the two groups was 17.5 months and 5.5 months, respectively, and the difference was statistically significant (P=0.00074).

[0141] The combined analysis of the PDS and NAC subgroups clearly demonstrates that the prognostic stratification capability of the DWF model is independent of specific initial treatment regimens and can provide reliable individualized prognostic references for different types of ovarian cancer patients in complex real-world clinical decision-making scenarios.

[0142] Example 8: Robustness verification of the DWF model under different data balancing strategies As described in Example 3 above, the optimal embodiment of this invention using Borderline-SMOTE to process imbalanced positive and negative sample data achieves excellent clinical predictive efficacy with an AUC of 0.760, and a high balance between sensitivity and specificity. To further verify the robustness of the DWF integrated architecture described in this invention in the face of extreme data loss, and to prove that its core predictive efficacy does not solely rely on artificially synthesized data generated by oversampling techniques, this embodiment supplements an alternative embodiment based on Random Under-Sampling (RUS).

[0143] In this embodiment, the RUS strategy is used to process the original training set, that is, approximately 68% of the majority class (platinum-sensitive) samples are randomly removed, forcing a strict 1:1 ratio between the minority class (drug-resistant) and the majority class (sensitive). Subsequently, the same feature selection, basic classifier construction, and dynamic weighted fusion process as in Embodiments 1 and 2 are used to construct the DWF model of the alternative embodiment. The prediction performance of this alternative embodiment under 5-fold cross-validation is compared with the best embodiment in Embodiment 3, as shown in Table 3. Table 3. Performance comparison of DWF prediction models under different data balancing strategies As shown in Table 3, in the alternative embodiment employing the RUS strategy, although the model is forced to discard a large amount of original negative data and faces severe baseline information loss, the DWF model proposed in this invention still maintains an AUC value as high as 0.759, which is substantially on par with the best embodiment employing the oversampling strategy. This strongly demonstrates that the dynamic feature selection and temperature-coefficient-based Softmax weighted fusion mechanism adopted in this invention possesses extremely strong data adaptive mining capabilities. The discriminative ranking capability of the DWF architecture does not depend on a specific data balancing algorithm, proving the universality of this framework in medical small sample data processing.

[0144] While the RUS strategy improves the model's sensitivity to minority class samples, its lack of sufficient learning of the majority class distribution characteristics leads to a significant decrease in specificity, resulting in a high false positive rate in real-world applications. Therefore, this invention recommends the Borderline-SMOTE-based strategy described in Example 3 as the preferred clinical implementation plan to prevent the risk of overtreatment. In specific scenarios where extreme medical data is lacking, or where there is a clinical need to screen for high-risk drug-resistant patients, the DWF system of this invention can also seamlessly integrate with conventional downsampling strategies and maintain stable overall predictive performance.

[0145] Example 9: Ablation Experiment Verification of Core Weighted Fusion Mechanism To verify the irreplaceability of the "softmax function based on temperature coefficient" proposed in the dynamic weighted fusion module of this invention, and its outstanding and substantial technical effect on optimizing the clinical decision boundary, this embodiment conducted a rigorous ablation experiment on the weighting mechanism.

[0146] This experiment uses the same Borderline-SMOTE preprocessed data and 10 dynamically selected preferred base classifiers, only changing the weight allocation strategy during the final model fusion. Three comparison groups were set up: 1. Simple average group: assign equal weights to the 10 preferred classifiers.

[0147] 2. Standard Softmax Group (Ablation Group): The CS of the optimized classifier is directly substituted into the standard Softmax function for weight allocation, without introducing a temperature coefficient (equivalent to...). T =1.0).

[0148] 3. This invention group (with temperature coefficient Softmax): Adopting the scheme described in this invention, a low temperature coefficient ( T =0.1) After non-linear sharpening of the comprehensive score, weights are then allocated using the Softmax function. The performance comparison results of the ablation experiments are shown in Table 4: Table 4. Comparative Evaluation Table of Ablation Efficacy of Weighted Fusion Mechanism As can be seen from Table 4, when using the conventional simple average or conventional Softmax method in this field... T When using a strategy with a ratio of 1.0, although the model can maintain the basic AUC, its accuracy and specificity collapse significantly. This is because the weight distribution of the conventional fusion mechanism is too smooth, causing mediocre base classifiers to receive excessively high decision weights. These mediocre classifiers tend to produce decision bias, outputting a large number of false positive predictions (i.e., over-predicting drug resistance), resulting in low clinical specificity and a high risk of overtreatment.

[0149] When the low temperature coefficient unique to this invention is introduced ( T After setting the value to 0.1, the model's specificity jumped to 0.780, and the overall accuracy also jumped to 0.767. Mathematically, T =0.1 achieves exponential sharpening of the performance distribution, accurately suppresses noise interference from inferior base classifiers, and gives high-discrimination models absolute decision-making dominance.

[0150] The weighted strategy of this invention achieves a significant increase in specificity by fine-tuning sensitivity (maintaining it within a reasonable range of 0.736). In real-world clinical decision-making regarding platinum-resistant ovarian cancer, high specificity translates to an extremely low false diagnosis rate, minimizing the risk of sensitive patients being wrongly deprived of first-line standard chemotherapy. The temperature coefficient Softmax mechanism of this invention is not a simple replacement of conventional methods in the field, but rather a key architecture for solving the technical challenge of high false positive rates in medical ensemble learning, possessing significant inventiveness and clinical applicability.

[0151] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing teachings of this invention, those skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method for constructing a weighted dynamic fusion prediction model for disease subtyping or diagnosis, characterized in that, The method includes the following steps: (S1) Provide a raw multidimensional feature dataset; the raw multidimensional feature dataset includes a clinical baseline dataset and a laboratory test dataset; (S2) Preprocess the original multidimensional feature dataset to obtain a preprocessed multidimensional feature dataset; (S3) Construct a basic classifier library, which includes multiple basic classifiers consisting of M feature selection algorithms and N machine learning algorithms; where M and N are both integers > 1; (S4) Train each basic classifier in the basic classifier library using the preprocessed multidimensional feature dataset to obtain the prediction probability of each basic classifier; during the training process, evaluate each basic classifier according to the geometric mean and select a predetermined number of basic classifiers as preferred classifiers. (S5) Evaluate the overall score of the preferred classifier using its AUC and geometric mean; (S6) Evaluate the weight of each preferred classifier based on the comprehensive score of each preferred classifier and the Softmax function with a predetermined temperature coefficient; (S7) The weights and prediction probabilities of each preferred classifier are summed to obtain the final prediction probability, thereby constructing a weighted dynamic fusion prediction model for the disease.

2. The method as described in claim 1, characterized in that, The original multidimensional feature dataset is an imbalanced dataset of positive and negative samples.

3. The method as described in claim 1, characterized in that, The laboratory test dataset includes complete blood count and systemic immune / inflammatory markers, liver and kidney function and biochemical metabolic profiles, tumor marker combinations, and coagulation and fibrinolysis function profiles.

4. The method as described in claim 1, characterized in that, In step (S2), the preprocessing includes methods selected from the group below to finally obtain a preprocessed multidimensional feature dataset: handling outliers, filling missing data, correcting data distribution, standardizing data, and / or handling imbalanced positive and negative samples; We use boundary synthesis minority class oversampling techniques to handle imbalanced positive and negative sample data.

5. The method as described in claim 1, characterized in that, The feature selection algorithms include: ANOVA F-value selection method, ANOVA T-value selection method, two-input symmetric correlation method, Fisher score method, interaction upper bound method, joint mutual information method, Laplace score method, logistic loss l_2,1 norm minimization method, least squares loss l_2,1 norm minimization method, multi-cluster feature selection method, non-negative discriminative feature selection method, Relief-F algorithm, trace ratio criterion method, and unsupervised discriminative feature selection method.

6. The method as described in claim 1, characterized in that, The machine learning algorithms include: adaptive boosting algorithm, balanced random forest, category feature gradient boosting algorithm, decision tree, extreme random tree, gradient boosting machine, K-nearest neighbor algorithm, lightweight gradient boosting machine, logistic regression, random forest, support vector machine, and extreme gradient boosting machine.

7. The method as described in claim 1, characterized in that, In step (S5), the comprehensive score is calculated as follows: ， in, For the first i The AUC value of the optimal classifier For the first i The geometric mean of the preferred classifiers and Let be the weighting coefficient, satisfying ; The , .

8. The method as described in claim 1, characterized in that, In step (S6), the weights are calculated as follows: ， in, For the first i The combined score of the selected classifiers, where K is the number of selected classifiers and T is the temperature coefficient, satisfying 0 < K / T. T < 1.

9. The method as described in claim 1, characterized in that, The T satisfies 0.05≤T≤0.15, 0.07≤T≤0.13, or 0.09≤T≤0.11, preferably 0.09≤T≤0.11, and more preferably T=0.

1.

10. A disease typing / prediction system, characterized in that, The prediction system includes: (M1) Input module, the input module is configured to input data, the data including the original multidimensional feature dataset of the test object, the original multidimensional feature dataset including clinical baseline dataset and laboratory test dataset; (M2) A typing / prediction module, comprising: (m2.1) A prediction unit configured as a prediction model for the disease, wherein the prediction model predicts the test object based on the original multidimensional feature dataset to obtain a final prediction probability; the prediction model is constructed by the method described in claim 1. (m2.2) A discrimination unit, configured to compare the final predicted probability of the object to be tested with a predetermined classification threshold to obtain a prediction result; (M3) Output module, which is configured to output the prediction results of the prediction module.

Images