Oral squamous cell carcinoma metabolite characteristic value screening method and system based on extreme random tree
Through the screening method of metabolites of oral squamous cell carcinoma metabolites based on extreme random trees, using plasma metabolomics data and machine learning algorithms, the early OSCC detection problem was solved, high accuracy and low cost diagnosis were achieved, and important clinical applications were used. value.
Patent Information
- Application Number
- CN202411021836.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-19
AI Technical Summary
The prior art is difficult to achieve early and low invasive oral squamous cell carcinoma (OSCC) detection, resulting in delayed diagnosis and missed treatment opportunities, and tissue biopsy may accelerate tumor progression and metastasis.
Using the oral squamous cell carcinoma metabolites characteristic value screening method based on extreme random trees, key metabolites characteristics are screened through plasma metabolomics data and machine learning algorithms, and an intelligent model for OSCC diagnosis is constructed to achieve early screening.
This method can accurately and reliably screen out key metabolites characteristics, improve the accuracy of early diagnosis of OSCC, reduce classification costs, and has good application prospects and practical significance.
Smart Images

Figure CN118983107A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an oral squamous cell carcinoma biomarker screening method and an intelligent diagnosis system based on an extreme random tree, and belongs to the technical field of biomedicine. Background Art
[0002] Oral squamous cell carcinoma (OSCC), as an important part of head and neck malignant tumors, not only poses a serious threat to the patient's life safety, but also greatly affects the patient's physiological functions such as chewing, speaking, and breathing. Even if cured, the loss of function will make it difficult for the patient to integrate into society, resulting in a serious social burden. Patients with early tumors usually rarely show obvious symptoms, resulting in delayed diagnosis and severe morbidity. Even with early diagnosis, recurrence is unpredictable. At present, tissue pathological biopsy is regarded as the "gold standard" for the diagnosis of OSCC, but because patients are often in the middle and late stages of the tumor when they are discovered, even if a biopsy is performed, the best time for tumor treatment may be missed. At the same time, tissue biopsy may accelerate the progression and metastasis of the tumor, thereby affecting the patient's prognosis. Therefore, how to achieve early and low-invasive OSCC detection is of great significance to improving the treatment effect and the quality of life of patients.
[0003] Metabolomics has been shown to be an important omics approach to understand the molecular pathways behind tumor phenotypes and identify new clinically useful markers, with the potential to serve as a diagnostic and prognostic tool for OSCC. The plasma metabolic profiles of OSCC patients show abnormal ketosis, lipogenesis, and energy metabolism. This abnormality is evident in the early stages, and as the disease progresses to the late stages, the abnormalities in these metabolic profiles become more pronounced. Assessing the plasma metabolic pathways of these population samples may lead to the discovery of new biomarkers and new strategies for the prevention and treatment of OSCC. Summary of the invention
[0004] The purpose of the present invention is to provide a biomarker screening method to obtain more reliable and accurate biomarkers for accurate diagnosis of early OSCC.
[0005] In order to achieve the above object, the technical solution of the present invention is to provide a method for screening characteristic values of oral squamous cell carcinoma metabolites based on an extreme random tree, characterized in that it comprises the following steps:
[0006] Step 1, obtaining combined data of n plasma metabolic markers to be screened from a certain number of OSCC patients and healthy people, where different plasma metabolic markers are defined as different features;
[0007] Step 2: preprocess the data set obtained in step 1, and divide the preprocessed data set into a training set and a test set;
[0008] Step 3, using the training set to build a model M for predicting the diagnosis result of OSCC, using the extreme random tree algorithm in the training process of the model M, constructing multiple decision trees by randomly selecting features and samples multiple times, and improving the classification performance by combining the results of these decision trees;
[0009] Step 4: Use the 5-fold cross validation method to train the model M constructed in step 3 J times, and sum the feature importance of each feature after J trainings to obtain the feature importance of n features. The following steps are used to calculate the feature importance of the lth feature: l , l=1,2,…,n:
[0010] Step 401: During each training, all decision trees are summarized, and the sum of the impurity reduction caused by the lth feature in all decision trees is calculated and then normalized to obtain the importance value, wherein in the construction process of each decision tree, different features are considered for each split, and when the lth feature is used for splitting, the impurity reduction caused by the current split is calculated;
[0011] Step 402: After completing J trainings, sum all importance values of the lth feature obtained in step 402 to obtain the feature importance I of the lth feature. l ;
[0012] Step 5: Sort the n features according to their importance to obtain the feature data set X sorted :
[0013] X sorted =X (1) ,X (2) ,…,X (n)}
[0014] In the formula, X (n) is the feature dataset X sorted The nth-ranked feature in ;
[0015] Step 6: Based on feature dataset X sorted Obtaining the accuracy list A includes the following steps:
[0016] Step 601: Iterate to create a new feature data set. The feature data set created in the kth iteration is represented by X () , k=0,1,…,n-1, then:
[0017] X () =X (1) ,X (2) ,…,X (n-k)}
[0018] Step 602: In the kth iteration, the feature data set X() is used as input to train the model M, and the model accuracy acc in the kth iteration is obtained. () :
[0019] acc () =Evaluate(M,X (k) ,y)
[0020] In the formula, Evaluate() is the model performance evaluation function, and y is the sample label;
[0021] Step 603: Update the accuracy list A, A = A∪{acc ()};
[0022] Step 604: After completing n rounds of iterations, obtain the final accuracy list A;
[0023] Step 7: Find the inflection point k of the value change from the accuracy list A * , inflection point k * is the feature dataset X sorted The dividing point between important features and unimportant features in the equation is expressed as
[0024] Preferably, the plasma metabolic markers to be screened in step 1 include:
[0025] 14 values: SM C24:1, C5, SM C16:0, PC ae C38:0, Ala, PC aa C34:4, PC aa C36:5, Total SM nonOH, PC aa C36:0, Met, PC aa C32:2, Gln, SM C18:0, Total SM;
[0026] 19 ratios: SM C24:1 / (Met / PC aa C40:3), Phe / PC aa C42:4, Val / C5, C3 / C4, (Ala / Gln) / (Tyr / Phe), Ala / PC aa C40:2, Met / PC aa C40:3, Ala / Gln, Met / PC aa C40:2, Phe / PC aa C40:3, Gln / Thr, Thr / Ser, C4 / C5, Glutaminolysis: (Ala+Asp+Glu) / Gln, Ala / lysoPC a C18:1, Met / lysoPC a C18:1, PUFA(PC) / MUFA(PC), PC_aa_C40:3 / PC_aa_C42:5, Pro / Orn.
[0027] Preferably, in step 3, the parameters of the model M after the model M is trained by out-of-bag accuracy tuning are constructed: 20 5-fold cross validations are performed, and a grid search or Bayesian optimization method is used to find the optimal model parameters according to the change in out-of-bag accuracy.
[0028] Preferably, before step 3 and before step 4, the method further includes the following steps: using a test set to evaluate the generalization ability of the model M constructed in step 3.
[0029] Preferably, the model M is evaluated using AUC, precision, recall and F1-score evaluation indicators.
[0030] Preferably, in step 401, the reduction in impurity is measured by Gini impurity or information gain.
[0031] Preferably, in step 7, the inflection point is found using a Kneed algorithm.
[0032] Another technical solution of the present invention is to provide an intelligent diagnosis system, characterized in that it includes the model M created in step 3 of the above-mentioned oral squamous cell carcinoma metabolite feature value screening method based on extreme random tree, and the important feature X obtained by the above-mentioned oral squamous cell carcinoma metabolite feature value screening method based on extreme random tree imp The real-time data is used as the input of the model M, and the model M predicts the diagnosis result of OSCC.
[0033] The technical solution disclosed in the present invention collects plasma metabolomics data sets of samples from OSCC patients and healthy controls, establishes an extreme random tree (ET) method to screen OSCC metabolomics ratios and metabolite-related diagnostic markers, and constructs a model for intelligent screening of OSCC. The present invention innovatively applies volatile metabolomics data, combined with advanced machine learning algorithms, to accurately identify plasma metabolite markers related to OSCC, and promotes the research progress of OSCC diagnosis in the field of precision medicine.
[0034] Compared with the prior art solutions, the present invention has the following beneficial effects:
[0035] (1) The present invention only selects a small number of key metabolite features to achieve accurate classification of different groups, with good reliability, high sensitivity and strong specificity;
[0036] (2) The method of the present invention can greatly reduce the classification cost;
[0037] (3) The important features screened by the method of the present invention and the prediction model established have strong predictive ability, adaptability and generalization ability;
[0038] (4) It can be used for early screening of OSCC in routine blood tests and physical examinations, and has good application prospects and practical significance. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 Flowchart for modeling and feature selection;
[0040] Figure 2 The ROC value of the EF model is shown;
[0041] Figure 3 Comparison bar chart of precision, recall and F1-score of EF model classification;
[0042] Figure 4 The changing trend of accuracy is shown. DETAILED DESCRIPTION
[0043] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.
[0044] One aspect of an embodiment of the present invention is to disclose a method for screening oral squamous cell carcinoma biomarkers based on an extreme random tree, comprising the following steps:
[0045] Step S1: Obtain plasma metabolite marker combination data of a certain number of OSCC patients and healthy people through public data sets, including experimental group and control group. The collected plasma metabolite marker combination data covers plasma samples of 72 primary OSCC patients who did not receive chemoradiotherapy and surgery and 61 healthy controls (see Table 1). Metabolomics analysis used p180Kit was used for mass spectrometry analysis, which can quantify 188 metabolites, including amino acids, biogenic amines, acylcarnitines, glycerophospholipids, and sphingomyelins. The analysis process used flow injection analysis tandem mass spectrometry (FIA-MS / MS) and liquid chromatography-mass spectrometry (LC-MS / MS), and was analyzed by SCIEX 4000 and Waters XEVO The analysis was performed by mass spectrometry, ensuring high accuracy and reliability of the data. In addition, metabolites that did not reach the detection limit were excluded from the data to improve the accuracy of the analysis.
[0046] In addition to quantifying 131 metabolites in plasma samples, several other variables were investigated, including metabolite ratios and sums of aromatic (AAA), branched (BCAA), essential (EAA), non-essential (NEAA) and selected amino acids, acylcarnitines, acylphosphatidylcholines (lysoPC), phosphatidylcholines (PC aa and PC ae), and sphingomyelin (SM), for a total of 104 sums and ratios. Therefore, the table contains sample name, label value (experimental group and control group) and features (metabolite concentration and metabolite ratio), with a total of 235 metabolites and their sums / ratios (Table 2 is a sample format).
[0047] The public dataset used in the embodiments of the present invention can be found at: https: / / doi.org / 10.1038 / s41598-023-34808-2.
[0048] Table 1 Collection information and missing data of OSCC patients
[0049]
[0050] Table 2 MS dataset of OSCC patients and controls (partial)
[0051]
[0052] In the embodiment of the present invention, the selected plasma metabolic marker combination data includes:
[0053] 14 values: SM C24:1, C5, SM C16:0, PC ae C38:0, Ala, PC aa C34:4, PC aa C36:5, Total SM nonOH, PC aa C36:0, Met, PC aa C32:2, Gln, SM C18:0, Total SM;
[0054] 19 ratios: SM C24:1 / (Met / PC aa C40:3), Phe / PC aa C42:4, Val / C5, C3 / C4, (Ala / Gln) / (Tyr / Phe), Ala / PC aa C40:2, Met / PC aa C40:3, Ala / Gln, Met / PC aa C40:2, Phe / PC aa C40:3, Gln / Thr, Thr / Ser, C4 / C5, Glutaminolysis: (Ala+Asp+Glu) / Gln, Ala / lysoPC a C18:1, Met / lysoPC a C18:1, PUFA(PC) / MUFA(PC), PC_aa_C40:3 / PC_aa_C42:5, Pro / Orn.
[0055] Step S2, preprocessing the OSCC data set obtained in step 1, including data cleaning, missing value filling and other processes. Data cleaning is to clean variables or special discrete values that are not related to the prediction, such as irrelevant socioeconomic characteristics or values greater than ±2 times the standard deviation of the mean. Since detection errors may lead to missing values, a suitable method is selected to fill or delete them according to the number of missing values. For example, when only one sample has missing values, the median is used to fill the missing values so that the data distribution is not greatly affected; when multiple samples have missing values, the feature is deleted. The processed OSCC data set is then divided into a training set and a test set, with a ratio of 80% and 20% respectively.
[0056] Step S3, use the training set to build an extreme random tree (Extra Trees) model, and tune the model parameters through the out-of-bag (OOB) accuracy without the need for a separate validation set. In an embodiment of the present invention, when tuning the model parameters, 20 5-fold cross validations are used, and a grid search or Bayesian optimization method is used to find the optimal model parameters according to the change in OOB accuracy. The adjusted model parameters are: the number of decision trees, the depth of the decision tree, and the maximum number of features used by a single decision tree. These three parameters are tuned in turn. Taking the grid search tuning process of the number of decision trees as an example, let the number of decision trees, use each value in the training set to train the model in turn, and obtain the model performance through the 5-fold cross validation method to obtain the optimal number of decision trees.
[0057] During the construction of the extreme random tree model, the input is the training set (i.e., the preprocessed metabolite values and ratio data), and the output is the diagnosis result of OSCC (positive or negative). The extreme random tree algorithm is used in the model training process. Its basic principle is to construct multiple decision trees by randomly selecting features and samples multiple times, and improve the classification performance by combining the results of these decision trees. In this embodiment, the construction process of each decision tree includes the following steps:
[0058] 1. Randomly extract samples from the training set to construct a subsample set. When training the Extra Trees model, the input parameters include the pre-processed metabolite values and ratio data, and the output parameters are the OSCC diagnosis results (positive or negative).
[0059] 2. For each node, randomly select some features to find the best split point.
[0060] 3. Repeat the process described in step 2 until the stopping condition (such as the maximum depth of the decision tree or the minimum number of samples) is met.
[0061] Step S4, model evaluation: Evaluate the generalization ability of the model on the test set, and use evaluation indicators such as AUC (area under the curve), precision, recall and F1-score to measure the model. The ROC curve area (AUC) of this model reaches 0.88, see Figure 2 ; For precision, recall and F1-score, see Figure 3 to Figure 4 .
[0062]
[0063] Among them, TP is the number of true positive examples, FP represents the number of false positive examples, FN represents the number of false negative examples, FP is the number of true negative examples, Precision is the precision rate (0.80), Recall is the recall rate (0.83), F1 is the F1-score (0.81), and Accuracy is the accuracy rate (0.82).
[0064] Step S5: For the model after parameter tuning obtained in step S4, the 5-fold cross validation method is used for training 100 times, and the feature importance of each metabolite marker in the plasma metabolite marker combination data is summed up for 100 times, and the features are sorted from large to small.
[0065] In the Extra Trees model, the calculation of feature importance is usually based on the contribution of each feature to the model prediction accuracy. Specifically, for the k-th metabolite marker in the j-th training, its feature importance can be calculated by the following steps:
[0066] 1. Reduction of impurity: During the construction of each decision tree, different features are considered for each split. When the k-th metabolite marker is used for splitting, the reduction in impurity caused by this split is calculated. The reduction in impurity is usually measured by Gini impurity or information gain (entropy).
[0067] 2. Cumulative importance: During a training process, all decision trees are summarized and the sum of the impurity reduction caused by the k-th metabolite marker in all decision trees is calculated.
[0068] 3. Standardization: In order to make the feature importance values comparable, the cumulative importance is standardized. This is usually done by dividing the importance of each feature by the sum of all feature importances.
[0069] 4. Sum and sort: After completing 100 trainings, the feature importance of each metabolite marker will be summed and sorted from large to small to obtain the final feature importance ranking.
[0070] Then, in the embodiment of the present invention, for the k-th metabolic marker in the j-th training, its characteristic importance can be calculated by the following formula:
[0071]
[0072] Wherein, {impurity reduction}{jk} of all decision trees is the impurity reduction when the k-th metabolite marker is used to split the node in the j-th training.
[0073] Assume that the data set consisting of all n metabolic markers (i.e., all n features) is represented by X = {x1, x2, …, x n}, the model after parameter tuning obtained in step S4 is M, and the set of feature importance sum values I is expressed as I = {I1, I2, ..., I n}, then the set of values of the sum of the importance of the sorted features is I sorted It is expressed as:
[0074] I sorted =I (1) ,I (2) ,…,I (n)},I (1) ≥I (2) ≥…≥I (n)
[0075] In the formula, I (n) The sum of the feature importances of the nth feature after sorting.
[0076] According to the collection I sorted The sorted feature dataset X obtained sorted It is expressed as:
[0077] X sorted =X (1) ,X (2) ,…,X (n)}
[0078] Where, X (n) is the nth feature after sorting.
[0079] Step S6: Based on the feature data set X sorted , iteratively create a new feature dataset X () , expressed as:
[0080] X () =X (1) ,X (2) ,…,X (n-k)}
[0081] In the formula, the value of k ranges from 0 to n-1;
[0082] Using feature dataset X () Train the model M obtained in step S4 after parameter adjustment and evaluate the model accuracy acc () , as shown below:
[0083] acc () =Evaluate(M,X (k) ,y)
[0084] Here, the Evaluate() function usually refers to evaluating the performance of the model M on a given dataset, and y usually represents the actual output or target variable. The Evaluate() function calculates the performance of the model M using a specific input X. (k) The corresponding output is then compared with the actual output y to determine the accuracy or other performance indicator of the model M.
[0085] Initialize a list A = [] for storing accuracy, and then use the following formula to update list A:
[0086] A=A∪{acc ()}
[0087] Repeat the above steps to complete all feature data sets X () Iterative creation and corresponding model accuracy acc () After calculation, the final list A is obtained.
[0088] Step S7: Diagnostic capability verification: Use the Kneed algorithm to find the inflection point k of the value change from the list A obtained in the previous step. * , inflection point k * This is the dividing point between important features and unimportant features (see Table 3 for important features):
[0089] k * =Kneed(A)
[0090] The important features finally obtained are expressed as
[0091] Table 3 Feature selection results of MS dataset for OSCC patients and controls
[0092]
[0093]
[0094] Another aspect of an embodiment of the present invention is to provide an intelligent diagnostic system, including the model M after parameter adjustment obtained in the above step S4, after the real-time data of the corresponding metabolic markers obtained by the above method is input into the model M, the model M predicts the diagnostic result of OSCC.
Claims
1. A method for screening characteristic values of oral squamous cell carcinoma metabolites based on extreme random trees, characterized in that: The following steps are involved: Step 1, obtaining combined data of n plasma metabolic markers to be screened from a certain number of OSCC patients and healthy people, where different plasma metabolic markers are defined as different features; Step 2: preprocess the data set obtained in step 1, and divide the preprocessed data set into a training set and a test set; Step 3, using the training set to build a model M for predicting the diagnosis result of OSCC, using the extreme random tree algorithm in the training process of the model M, constructing multiple decision trees by randomly selecting features and samples multiple times, and improving the classification performance by combining the results of these decision trees; Step 4: Use the 5-fold cross validation method to train the model M constructed in step 3 J times, and sum the feature importance of each feature after J trainings to obtain the feature importance of n features. The following steps are used to calculate the feature importance of the lth feature: l , l=1,2,…,n: Step 401: During each training, all decision trees are summarized, and the sum of the impurity reduction caused by the lth feature in all decision trees is calculated and then normalized to obtain the importance value, wherein in the construction process of each decision tree, different features are considered for each split, and when the lth feature is used for splitting, the impurity reduction caused by the current split is calculated; Step 402: After completing J trainings, sum all importance values of the lth feature obtained in step 402 to obtain the feature importance I of the lth feature. l ; Step 5: Sort the n features according to their importance to obtain the feature data set X sorted : X sorted =X (1) ,X (2) ,…,X (n) } Where, X (n) is the feature dataset X sorted The nth-ranked feature in ; Step 6: Based on feature dataset X sorted Obtaining the accuracy list A includes the following steps: Step 601: Iterate to create a new feature data set. The feature data set created in the kth iteration is represented by X () , k=0,1,…,n-1, then: X () =X (1) ,X (2) ,…,X (n-k) } Step 602: In the kth iteration, the feature data set X () As input, train the model M and obtain the model accuracy acc at the kth iteration () : acc () =Evaluate(M,X (k) ,y) In the formula, Evaluate() is the model performance evaluation function, and y is the sample label; Step 603: Update the accuracy list A, A = A∪{acc () }; Step 604: After completing n rounds of iterations, obtain the final accuracy list A; Step 7: Find the inflection point k of the value change from the accuracy list A * , inflection point k * is the feature dataset X sorted The dividing point between important features and unimportant features in the equation is expressed as 2. The method for screening characteristic values of oral squamous cell carcinoma metabolites based on extreme random trees according to claim 1, characterized in that: The plasma metabolic markers to be screened in step 1 include: 14 values: SM C24:1, C5, SM C16:0, PC ae C38:0, Ala, PC aa C34:4, PC aa C36:5, Total SM nonOH, PC aa C36:0, Met, PC aa C32:2, Gln, SM C18:0, Total SM; 19 ratios: SM C24:1 / (Met / PC aa C40:3), Phe / PC aa C42:4, Val / C5, C3 / C4, (Ala / Gln) / (Tyr / Phe), Ala / PC aa C40:2, Met / PC aa C40:3, Ala / Gln, Met / PC aa C40:2, Phe / PCaa C40:3, Gln / Thr, Thr / Ser, C4 / C5, Glutaminolysis: (Ala+Asp+Glu) / Gln, Ala / lysoPCaC18:1, Met / lysoPCa C18:1, PUFA(PC) / MUFA(PC), PC_aa_C40:3 / PC_aa_C42:5, Pro / Orn.
3. The method for screening characteristic values of oral squamous cell carcinoma metabolites based on extreme random trees according to claim 1, characterized in that: In step 3, the parameters of the model M after training by out-of-bag accuracy tuning are constructed: 20 5-fold cross validations are performed, and a grid search or Bayesian optimization method is used to find the optimal model parameters according to the change in out-of-bag accuracy.
4. The method for screening characteristic values of oral squamous cell carcinoma metabolites based on extreme random trees according to claim 1, characterized in that: Before step 4 and before step 3, the following step is also included: using a test set to evaluate the generalization ability of the model M constructed in step 3.
5. The method for screening characteristic values of oral squamous cell carcinoma metabolites based on extreme random trees according to claim 4, characterized in that: The model M is evaluated using AUC, precision, recall and F1-score evaluation indicators.
6. The method for screening characteristic values of oral squamous cell carcinoma metabolites based on extreme random trees according to claim 1, characterized in that: In step 401, the reduction in impurity is measured by Gini impurity or information gain.
7. The method for screening characteristic values of oral squamous cell carcinoma metabolites based on extreme random trees according to claim 1, characterized in that: In step 7, the inflection point is found using the Kneed algorithm.
8. An intelligent diagnostic system, characterized in that: The model M created in step 3 of the method for screening characteristic values of oral squamous cell carcinoma metabolites based on extreme random trees according to claim 1 is used to obtain the important feature X obtained by the method for screening characteristic values of oral squamous cell carcinoma metabolites based on extreme random trees according to claim 1. imp The real-time data is used as the input of the model M, and the model M predicts the diagnosis result of OSCC.
Citation Information
Cited By
Landslide susceptibility evaluation method based on Bayesian optimization adaptive enhancement integrated extreme random tree
CN119691363A
