Machine learning algorithm-based ovarian cancer diagnosis model and construction method and application thereof

Abstract

The invention discloses an ovarian cancer diagnosis model based on a machine learning algorithm and a construction method and application thereof, and the construction method comprises the following steps: obtaining external vesicle proteomics data of a clinical ovarian cancer patient and a cell line, and carrying out the normalization preprocessing of the data through Python medium-high throughput analysis; differential protein screening is carried out by utilizing a bioinformatics method, feature selection is carried out in combination with a machine learning algorithm, and nine kinds of external vesicle protein biomarker diagnosis models are constructed; evaluating model performance by using a working characteristic curve of a subject, calculating an area under the curve as a discrimination performance index, optimizing a diagnosis threshold value, and obtaining an optimal model through a ridge regression method; and testing an external test set by using the optimal model, and evaluating the performance of the model by using the working characteristic curve of the subject. According to the diagnostic model disclosed by the invention, the analysis efficiency of exosome data can be remarkably improved, the generalization ability of the model is improved, and the false positive rate is effectively reduced, so that the clinical applicability of the diagnostic model is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an ovarian cancer diagnosis model based on a machine learning algorithm, a construction method thereof, and an application thereof. Background Art

[0002] Ovarian cancer is one of the most lethal malignancies of the female reproductive system. Its high mortality rate is mainly attributed to the hidden early symptoms, which result in most patients being diagnosed in the late stage. Although imaging screening (such as transvaginal ultrasound) and serological tests (such as CA125) have made some progress in the early diagnosis of ovarian cancer in recent years, these methods still have major limitations such as insufficient sensitivity and low specificity. In addition, some detection methods have not been widely used in clinical practice due to their high invasiveness and high false positive rate. Therefore, there is an urgent need to develop an efficient, low-invasive and highly accurate early diagnosis strategy for ovarian cancer to enhance screening results and improve patient prognosis.

[0003] In recent years, exosomes have become widely present in body fluids and can carry tumor-derived pathological information and biomolecules, making them an ideal tool for early cancer detection. Studies have shown that exosomes have shown significant potential in the early diagnosis of various cancers, including lung cancer, hepatocellular carcinoma, and pancreatic cancer. However, most current exosome-based cancer diagnosis studies focus only on a single or a few biomarkers and lack systematic, multi-factor integrated analysis, resulting in low model generalization capabilities. In addition, some machine learning models are prone to overfitting due to imbalanced training data and imperfect feature screening methods, limiting their clinical application value.

[0004] Therefore, there is an urgent need to build a high-precision ovarian cancer diagnostic model based on machine learning algorithms that can effectively integrate multidimensional biomarker information from exosomes, optimize feature selection and model training strategies, and achieve more efficient, less invasive and highly accurate prediction capabilities to make up for the shortcomings of traditional biomarkers, provide more accurate and efficient technical support for the early diagnosis of ovarian cancer, improve screening accuracy, reduce misdiagnosis rate, thereby detecting the disease earlier and improving patient prognosis. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides an ovarian cancer diagnostic model based on a machine learning algorithm, as well as its construction method and application. Through data-driven feature selection and optimization algorithms, it achieves efficient integration of multiple biomarkers derived from exosomes, thereby improving the accuracy and stability of diagnosis.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] A method for constructing an ovarian cancer diagnostic model based on a machine learning algorithm comprises the following steps:

[0008] (1) Sample set acquisition: extracellular vesicle proteomics data from clinical ovarian cancer patients and cell lines were obtained, and the data were normalized and preprocessed using high-throughput analysis in Python;

[0009] (2) Construction of random forest model: Using bioinformatics methods to screen differentially expressed proteins and combining them with machine learning algorithms for feature selection, a diagnostic model for nine extracellular vesicle protein biomarkers, including HEL2, RAB11B, HSPA4, ACTG1, CD147, GLUT3, RHOA, CD46, and RAN, was constructed;

[0010] (3) Obtaining the optimal model: The receiver operating characteristic curve was used to evaluate the model performance, the area under the curve was calculated as the discriminant performance indicator, and its diagnostic threshold was optimized to obtain the optimal model through the ridge regression method;

[0011] (4) Model validation: Use the optimal model to test the external test set, and use the receiver operating characteristic curve to evaluate the performance of the model to achieve model validation.

[0012] In the above scheme, in step (2), the feature selection process uses ridge regression to screen variables, and the regularization parameter λ is optimized through 5-fold cross validation.

[0013] In the above scheme, in step (1), the sample includes a serum sample, a plasma sample, a cell fluid sample, a urine sample or a saliva sample.

[0014] An application of an ovarian cancer diagnostic model based on a machine learning algorithm, wherein the ovarian cancer diagnostic model utilizes a combination of exosomal protein biomarkers for early prediction and diagnosis of ovarian cancer, and can be combined with the clinical biomarker CA125 to construct a combined diagnostic model.

[0015] In the above scheme, the ovarian cancer model can be used for clinical serum sample testing, and exosome proteomics analysis can be used to distinguish ovarian cancer patients from healthy individuals.

[0016] A system for constructing an ovarian cancer diagnostic model based on a machine learning algorithm, which is applied to any of the construction methods described above, comprising:

[0017] Data acquisition module, used for data collection and obtaining sample data sets;

[0018] A data processing module is used to extract valid samples from the sample data set that can be used to build an evaluation model;

[0019] A model building module is used to randomly divide the incomplete data set of valid samples into a training set and a validation set, and use a random forest method to fit the training set, and record the optimal model parameters according to the out-of-bag error;

[0020] The threshold calculation module is used to calculate the model classification threshold using the validation set according to the ROC curve.

[0021] An ovarian cancer diagnosis system based on a machine learning algorithm, comprising:

[0022] It is the pre-input module of the evaluation model, used to input the data to be diagnosed;

[0023] An ovarian cancer diagnostic model constructed by the method according to any one of claims 1 to 3, used to evaluate the data to be evaluated;

[0024] Display module, used to display analysis and diagnosis results.

[0025] A computer program product stored on a computer-readable medium includes a computer-readable program, which, when executed on an electronic device, provides a user input interface for applying the ovarian cancer diagnosis system based on a machine learning algorithm as claimed in claim 7.

[0026] A computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to apply the ovarian cancer diagnosis system based on a machine learning algorithm as claimed in claim 7.

[0027] Through the above technical solution, the present invention provides an ovarian cancer diagnosis model based on a machine learning algorithm, and its construction method and application have the following beneficial effects:

[0028] The present invention provides a machine learning-based biomarker screening and diagnostic modeling method that can significantly improve the analysis efficiency of exosome data, enhance the generalization ability of the model, and effectively reduce the false positive rate, thereby enhancing its clinical applicability. Compared with the existing technology, the present invention has achieved the following outstanding improvements:

[0029] 1. The present invention identifies and constructs the P9 model, which is composed of nine biomarkers: HEL2, RAB11B, HSPA4, ACTG1, CD147, GLUT3, RHOA, CD46, and RAN. The P9 model demonstrates excellent accuracy in distinguishing early-stage ovarian cancer from benign gynecological diseases, demonstrating its excellent diagnostic efficacy and significant clinical application value. The P9 model is combined with the commonly used clinical indicator CA125 to construct the CAP9 model, further enhancing its diagnostic capabilities.

[0030] 2. This paper adopts a data-driven machine learning strategy and combines ridge regression to optimize feature selection, making the model more accurate in processing high-dimensional and complex exosome proteomics data, avoiding the subjective bias of traditional methods;

[0031] 3. The model of the present invention has undergone rigorous cross-validation and parameter optimization to ensure its stability on different data sets, and reduces the risk of overfitting through regularization, thereby improving the reliability of clinical testing;

[0032] 4. This invention breaks through the limitations of single biomarker detection, integrates multi-marker information, and constructs a more diagnostically effective prediction model, significantly improving the sensitivity and specificity of early screening for ovarian cancer and providing strong support for clinical precision diagnosis.

[0033] Through the above innovative design, the present invention not only optimizes the feature extraction and modeling strategies, but also makes up for the traditional method's reliance on a single marker. It also realizes multidimensional data fusion and precise classification through machine learning, providing an intelligent and efficient diagnostic method for ovarian cancer screening and diagnosis based on serum exosome biomarkers. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments or the prior art.

[0035] Figure 1 A heat map showing differentially expressed proteins in ovarian cancer patients, ovarian cancer cell lines, and healthy women.

[0036] Figure 2 The principal component analysis of serum exosome proteomes of ovarian cancer patients and healthy controls is shown in Figure 2.

[0037] Figure 3 The principal component analysis of the exosome proteome of ovarian cancer cell lines and normal ovarian epithelium;

[0038] Figure 4 Volcano map of differentially expressed proteins between ovarian cancer patients and healthy women;

[0039] Figure 5 This is a volcano plot of differentially expressed proteins between ovarian cancer cell lines and normal control cells;

[0040] Figure 6 A Venn diagram showing the overlap of exosome proteins in patient serum and cell line exosomes.

[0041] Figure 7 KEGG enrichment analysis of 75 differentially expressed proteins in serum and cell line exosomes;

[0042] Figure 8 A heat map showing the top 20 highly expressed proteins and KEGG enriched proteins in cell lines and patients;

[0043] Figure 9 for Figure 8GO functional enrichment analysis of differentially expressed proteins;

[0044] Figure 10 ROC curves for predicting ovarian cancer using a single protein marker in the P9 model and the entire P9 model in the training set;

[0045] Figure 11 The prediction performance of the P9 model for different stages of ovarian cancer in the training set and validation set;

[0046] Figure 12 The sensitivity and specificity of the P9 and CAP9 models for overall, early-stage, and advanced ovarian cancer patients in the training and validation sets are shown in Table 1.

[0047] Figure 13 The ROC curve for CA125 in predicting ovarian cancer in the training set;

[0048] Figure 14 The prediction results of CA125 and P9 models for ovarian cancer in the training set respectively;

[0049] Figure 15 ROC curve of CAP9 model for predicting ovarian cancer in the training set;

[0050] Figure 16 The following are the prediction results of ovarian cancer by the CAP9 model in the training set and validation set. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0052] The present invention provides a method for constructing an ovarian cancer diagnostic model based on a machine learning algorithm, comprising the following steps:

[0053] (1) Sample set acquisition: Extracellular vesicle proteomics data of clinical ovarian cancer patients and cell lines were obtained, and the data were normalized and preprocessed using high-throughput analysis (HTSeq) in Python to eliminate experimental batch effects.

[0054] (2) Construction of random forest model: Using bioinformatics methods to screen differentially expressed proteins and combining them with machine learning algorithms for feature selection, a diagnostic model for nine extracellular vesicle protein biomarkers, including HEL2, RAB11B, HSPA4, ACTG1, CD147, GLUT3, RHOA, CD46, and RAN, was constructed;

[0055] The specific construction method is as follows:

[0056] 1. Screening of differentially expressed proteins

[0057] The distribution of differentially expressed proteins was visualized using a volcano plot. For cell samples, the screening criteria were |fold difference| > 3 and p value < 0.05. For serum samples, the screening criteria were |fold difference| > 1.5 and p value < 0.05. The distribution of differentially expressed proteins was visualized using a volcano plot.

[0058] 2. Bioinformatics analysis

[0059] KEGG pathway enrichment analysis: The R package "clusterProfiler" was used to analyze the key signaling pathways of differentially expressed proteins and explore potential pathogenic mechanisms.

[0060] GO functional enrichment analysis: The R package “clusterProfiler” was used to analyze the enrichment of differentially expressed proteins in terms of cellular components, biological processes, and molecular functions to reveal the functional characteristics of proteins.

[0061] (3) Obtaining the optimal model: The receiver operating characteristic curve was used to evaluate the model performance, the area under the curve was calculated as the discriminant performance indicator, and its diagnostic threshold was optimized to obtain the optimal model through the ridge regression method;

[0062] Ridge regression was used to address the high-dimensional nature of proteomics data to reduce multicollinearity and improve model stability. Five-fold cross-validation was performed using the R package "glmnet" to optimize the model's regularization parameters to improve generalization.

[0063] (4) Model validation: Use the optimal model to test the external test set, and use the receiver operating characteristic curve to evaluate the performance of the model to achieve model validation.

[0064] The R package “pROC” was used to calculate the receiver operating characteristic (ROC) curve to evaluate the diagnostic accuracy of the model, and the area under the curve was calculated as the discriminant performance indicator.

[0065] Example

[0066] 1. Research subjects and sample inclusion: From March 2021 to February 2024, the present invention included a total of 200 patients with ovarian cancer, all of whom underwent surgery and were pathologically confirmed, including patients in different stages I-IV, covering different histological subtypes such as serous carcinoma, mucinous carcinoma, clear cell carcinoma, and endometrial cancer. There were 136 healthy female controls, excluding various malignant tumors, inflammation and ovarian-related diseases. The training set included 125 ovarian cancer patients and 84 healthy women, and the validation set included 63 ovarian cancer patients and 42 healthy women. 10 mL of peripheral blood was collected from all subjects, and the serum was separated and exosomes were enriched and isolated for exosome proteome mass spectrometry detection. The cell lines used were two ovarian cancer cell lines, Skov3 and A2780, and the normal ovarian epithelial cell line T1074 was used as a control. Exosomes in the isolated cell culture medium were enriched and isolated for proteome mass spectrometry detection.

[0067] 2. Mass spectrometry identification: Liquid chromatography tandem mass spectrometry was used for proteomic analysis. The raw data were first normalized using HTSeq to eliminate the experimental batch effect. Exosome proteomics identified 1088 serum exosome proteins and 3629 cell line exosome proteins ( Figure 1 ). The exosome protein profiles of ovarian cancer and healthy subjects can be distinguished by two-dimensional principal component analysis ( Figure 2 ), tumor cell lines and normal cell lines can also be distinguished ( Figure 3 ), suggesting that our strategy of using exosome proteome to search for biomarkers is correct. Subsequently, the R package "DESeq" was used to screen differentially expressed proteins: for cell samples, |fold difference|>3, p value <0.05 were used as screening criteria, and for serum samples, |fold difference|>1.5, p value <0.05 were used as screening criteria. 332 differentially expressed proteins in serum exosomes were obtained ( Figure 4 ), 1321 differential proteins in cell line exosomes ( Figure 5 There are 75 overlapping proteins between patient-derived and cell-line-derived differentially expressed proteins ( Figure 6 ). These differentially expressed proteins were subjected to KEGG enrichment analysis and found to be mainly enriched in platelet activation, intercellular junctions, etc. Then, 72 proteins were screened by taking the top 20 differentially expressed proteins from each source and the important pathway proteins enriched in KEGG. Figure 7 ). Then, GO functional enrichment analysis was performed on these 72 key differential proteins, and it was found that they were mainly concentrated in platelet activation, intercellular connection, vesicle cavity, etc. ( Figure 8Based on commercially available antibodies, 12 proteins were ultimately selected for in-depth testing. Comparison of the detection signal results led to the selection of nine proteins for constructing the P9 model, including HEL2, RAB11B, HSPA4, ACTG1, CD147, GLUT3, RHOA, CD46, and RAN.

[0068] Next, we will describe the process of building an ovarian cancer diagnostic model based on ridge regression. We used ridge regression to construct the P9 model for ovarian cancer diagnosis. The key advantage of ridge regression is its ability to effectively address multicollinearity between biomarkers, improving model stability and predictive accuracy. During model training, we utilized cross-validation for hyperparameter optimization and combined ROC curve analysis to evaluate the model's diagnostic performance, ensuring its generalization across both training and validation sets. To ensure model reproducibility, the entire modeling process was implemented in R.

[0069] Using the above method, the final results are as follows: In the training set, the area under the curve of a single protein marker can be as high as 0.950, and the area under the curve of the P9 model of all 9 markers is further improved to 0.991 ( Figure 10 ), with a sensitivity of 94.4% and a specificity of 98.8%, and showed good performance in all clinical stages ( Figure 11 , Table 1), specifically, the sensitivity of early (stage I-II) and late (stage III-IV) stages reached 88.9% and 96.5%, respectively ( Figure 12 , Table 2). In another independent validation set, the sensitivity and specificity of the P9 model were 91.7% and 100%, respectively ( Figure 12 , Table 2). In comparison, the area under the curve of CA125, a marker currently used in clinical practice, is 0.910, with a sensitivity and specificity of 85.6% and 78.6%, respectively ( Figure 13 ) and resulted in 18 false positives, while P9 had only 1 false positive ( Figure 14 ). Combining CA125 as a marker, we optimized and constructed the CAP9 model combining P9 with CA125, which has a higher diagnostic efficacy and achieved 100% specificity in both the training set and the validation set ( Figure 12 , Figure 16 , Table 3, Table 4), and the sensitivity of phase II and IV in the validation set was further improved ( Figure 15 , Figure 16 , Table 4). In summary, the ovarian cancer diagnostic model constructed based on machine learning in the present invention can effectively reduce the false positive rate and improve the early detection capability. In particular, the model constructed by combining it with the existing clinical marker CA125 has more significant clinical application potential.

[0070] Table 1 Sensitivity and specificity of the P9 model for predicting ovarian cancer at different stages

[0071] All periods Phase I Phase II Stage III Stage IV Training set sensitivity (%) 94.4 92.3 90.0 95.6 97.1 Training set specificity (%) 98.8 98.8 98.8 98.8 98.8 Validation set sensitivity (%) 93.7 91.7 91.7 95.0 94.7 Validation set specificity (%) 100.0 100.0 100.0 100.0 100.0

[0072] Table 2 Sensitivity and specificity of the P9 model for the prediction of overall, early and late ovarian cancer

[0073] All periods Early Late stage Training set sensitivity (%) 94.0 88.9 96.5 Training set specificity (%) 98.8 98.8 98.8 Validation set sensitivity (%) 91.7 89.5 93.1 Validation set specificity (%) 100.0 100.0 100.0

[0074] Table 3 Sensitivity and specificity of the CAP9 model for the prediction of overall, early and late ovarian cancer

[0075] All periods Phase I and II Stages III and IV Training set sensitivity (%) 94.0 88.9 96.5 Training set specificity (%) 100.0 100.0 100.0 Validation set sensitivity (%) 93.8 89.5 96.6 Validation set specificity (%) 100.0 100.0 100.0

[0076] Table 4 Sensitivity and specificity of the CAP9 model for predicting ovarian cancer at different stages

[0077] All periods Phase I Phase II Stage III Stage IV Training set sensitivity (%) 94.4 92.3 90.0 95.6 97.1 Training set specificity (%) 100.0 100.0 100.0 100.0 100.0 Validation set sensitivity (%) 95.2 91.7 91.7 95.0 100.0 Validation set specificity (%) 100.0 100.0 100.0 100.0 100.0

[0078] An application of an ovarian cancer diagnostic model based on a machine learning algorithm. The ovarian cancer diagnostic model uses a combination of exosomal protein biomarkers for early prediction and diagnosis of ovarian cancer, and can be combined with the clinical biomarker CA125 to construct a combined diagnostic model.

[0079] This ovarian cancer model can be used for clinical serum sample testing and to differentiate ovarian cancer patients from healthy individuals through exosome proteomic analysis.

[0080] A system for constructing an ovarian cancer diagnostic model based on a machine learning algorithm, which is applied to any of the above construction methods, comprising:

[0081] Data acquisition module, used for data collection and obtaining sample data sets;

[0082] A data processing module is used to extract valid samples from the sample data set that can be used to build an evaluation model;

[0083] The model building module is used to randomly split the incomplete dataset of valid samples into training and validation sets, fit the training set using the random forest method, and record the optimal model parameters based on the out-of-bag error;

[0084] The threshold calculation module is used to calculate the model classification threshold using the validation set according to the ROC curve.

[0085] An ovarian cancer diagnosis system based on a machine learning algorithm, comprising:

[0086] It is the pre-input module of the evaluation model, used to input the data to be diagnosed;

[0087] An ovarian cancer diagnostic model constructed by the method according to any one of claims 1 to 3, used to evaluate the data to be evaluated;

[0088] Display module, used to display analysis and diagnosis results.

[0089] A computer program product stored on a computer-readable medium comprises a computer-readable program, which, when executed on an electronic device, provides a user input interface for applying the ovarian cancer diagnosis system based on a machine learning algorithm as claimed in claim 7.

[0090] A computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to apply the ovarian cancer diagnosis system based on a machine learning algorithm as claimed in claim 7.

[0091] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing an ovarian cancer diagnostic model based on a machine learning algorithm, characterized in that: The following steps are involved: (1) Sample set acquisition: extracellular vesicle proteomics data from clinical ovarian cancer patients and cell lines were obtained, and the data were normalized and preprocessed using high-throughput analysis in Python; (2) Construction of random forest model: Using bioinformatics methods to screen differentially expressed proteins and combining them with machine learning algorithms for feature selection, a diagnostic model for nine extracellular vesicle protein biomarkers, including HEL2, RAB11B, HSPA4, ACTG1, CD147, GLUT3, RHOA, CD46, and RAN, was constructed; (3) Obtaining the optimal model: The receiver operating characteristic curve was used to evaluate the model performance, the area under the curve was calculated as the discriminant performance indicator, and its diagnostic threshold was optimized to obtain the optimal model through the ridge regression method; (4) Model validation: Use the optimal model to test the external test set, and use the receiver operating characteristic curve to evaluate the performance of the model to achieve model validation.

2. The method for constructing an ovarian cancer diagnostic model based on a machine learning algorithm according to claim 1, characterized in that: In step (2), the feature selection process uses ridge regression to screen variables, and the regularization parameter λ is optimized through 5-fold cross validation.

3. The method for constructing an ovarian cancer diagnostic model based on a machine learning algorithm according to claim 1, characterized in that: In step (1), the sample includes a serum sample, a plasma sample, a cell fluid sample, a urine sample or a saliva sample.

4. An application of an ovarian cancer diagnosis model based on a machine learning algorithm, characterized in that: The ovarian cancer diagnostic model utilizes a combination of exosomal protein biomarkers for early prediction and diagnosis of ovarian cancer, and can be combined with the clinical biomarker CA125 to construct a combined diagnostic model.

5. The use according to claim 4, characterized in that This ovarian cancer model can be used for clinical serum sample testing and to differentiate ovarian cancer patients from healthy individuals through exosome proteomic analysis.

6. A system for constructing an ovarian cancer diagnostic model based on a machine learning algorithm, applied to the construction method according to any one of claims 1 to 3, comprising: Data acquisition module, used for data collection and obtaining sample data sets; A data processing module is used to extract valid samples from the sample data set that can be used to build an evaluation model; A model building module is used to randomly divide the incomplete data set of valid samples into a training set and a validation set, and use a random forest method to fit the training set, and record the optimal model parameters according to the out-of-bag error; The threshold calculation module is used to calculate the model classification threshold using the validation set according to the ROC curve.

7. An ovarian cancer diagnosis system based on machine learning algorithm, characterized in that: include: It is the pre-input module of the evaluation model, used to input the data to be diagnosed; An ovarian cancer diagnostic model constructed by the method according to any one of claims 1 to 3, used to evaluate the data to be evaluated; Display module, used to display analysis and diagnosis results.

8. A computer program product stored on a computer-readable medium, comprising a computer-readable program, which, when executed on an electronic device, provides a user input interface for applying the ovarian cancer diagnosis system based on a machine learning algorithm as claimed in claim 7.

9. A computer-readable storage medium storing instructions, which, when executed on a computer, causes the computer to apply the ovarian cancer diagnosis system based on the machine learning algorithm according to claim 7.

Images